Soft computing based compressive sensing techniques in signal processing: A comprehensive review

2.1 Compressed Sensing techniques

CS is a widely used image reconstruction technique which can reconstruct an original image from its reduced sample itself. The computational time taken by CS algorithms for image reconstruction is usually higher than the state-of-art reconstruction techniques. Sang-Hoong Jung et al. [56] introduced a greedy algorithm to overcome this complexity by using an iteration approach. The greedy algorithm used is Orthogonal Matching Pursuit (OMP), and it provides a high dimensional image quality in a short period of time. The high rate data applications consist of ambiguity such as heavy noise, interference, outliers, and channel fading present in the information to be broadcasted. In these types of applications, long term wideband sensing is quite costly to be applied. Jie Zhao et al. [54] has proposed a method of scheduling sequential CS sensing to solve the above problem. This approach is a combination of CS and sequential periodic detection techniques, which give the following outcomes, such as improved sensor quality, reduced CS recovery overhead, and appropriate wideband sensing. To provide optimal performance, CS techniques incorporate random structures, and they are widely applied to restore the missing signal samples. Christina Knill et al. [55] used CS post-processing for enhancing the Multi-Input Multi-Output (MIMO) approach, which leads to an effective regain of the full processing of the state-of-art Orthogonal Frequency-Division Multiplexing (OFDM). This method makes benefits for the information present in the CS, which leads to effective reconstruction, accelerated processing, and low computational complexity.

The Agriculture, Forestry, and Urban Planning field make use of high-resolution remote sensing images. These images exhibit the texture and features of the images in a more clear representation. The processing of these images includes various levels of complexities, such as texture Similarity, Massive Storage Space, and Information Leakage. A method called a finite-state chaotic CS cloud remote sensing image registration [62] is used to solve these complexities. The problem of texture similarity is overcome by using improved Scale-Invariant Feature Transform(SIFT) for both local and global information. CS is also applied to Hyperspectral Image(HSI) Reconstruction. The main problem to be solved is to identify the key characteristics present in the HSI image. Li Wang et al. [63] proposed a CS reconstruction algorithm for HIS images using spectral unmixing characteristics. The HIS image is sampled both spatially and spectrally. The reconstruction of the HIS image is obtained iteratively by solving a joint optimization problem which consists of the endmember and abundance matrix.

Zhitao et al. [1] have presented a method for medical signal compression in which Electro Cardiogram (ECG) information is compressed using “Set Partitioning In Hierarchical Trees (SPIHT)” method. The signals were selected from the “open source database, and the performance analyzed that their proposed codec was fundamentally more productive in compression and calculation than existing techniques. It is currently outstanding that one can recreate sparse or compressible signals precisely from a very set number of estimations, conceivably contaminated with noise. Emmanuel et al. [2] have proposed a procedure known as "compressed sensing" or "compressive sampling" depends on properties of the sensing grid, for example, the confined isometry property. In this note, we build up new outcomes about the exactness of the recreation from under sampled estimations, which enhance before appraisals and have the benefit of being progressively exquisite. Fred Chen et al. [3] have proposed a signal “Agnostic CS” obtaining framework is exhibited that tends to send telemetry data under the exchange capacity limitations of remote sensors. Luisa F.Polania et al. [4], has displayed by utilizing wavelet transform and iterative threshold technique; at that point, CS is accomplished to build the information-packed. In the wake of performing CS, “Bayesian CS (BCS)” is utilized to reproduce the first data. Standard Wavelet Dictionaries [17] is used to compress and decompress ECG signals which provide low-computational complexity for encryption and decryption. One-bit quantization CS [27] technique preserves the single sign information resulting in less storage cost and hardware complexity where the sparse signals can be reconstructed with high probability.

2.2 Medical Signal Compression methods

CS is another methodology for the obtaining and recuperation of a thin signal that empowers examining rates essentially underneath the traditional Nyquist rate. Anna M.R. Dixon et al. [5] have proposed a CS-based methodology for signal compression. ECG signal, for the most part, show excess between contiguous heartbeats because of its quasi-periodic structure. They demonstrated that their repetition inferred a huge division of regular help among successive heartbeats. In empowering nonstop remote cardiovascular observing in Wireless body sensor systems (WBSN), they can accomplish improved personalization and nature of consideration, expanded capacity of counteractive action and early conclusion, and enhanced patient autonomy, portability, and wellbeing. Amongst them, power productivity can be improved through implanted ECG compression, to diminish broadcast appointment above energy-hungry remote connections. Hossein Mamaghanian et al. [6], have evaluated the capability of the developing CS signal acquisition worldview for low-multifaceted nature energy effective ECG compression on the cutting edge Shimmer WBSN bit. Curiously, their outcomes demonstrate that CS speaks to an aggressive choice to cutting edge “Digital Wavelet Transform (DWT)” - based ECG compression arrangements. All the more explicitly, while expectedly displaying second rate compression execution than its DWT-based partner for a given recreated signal quality, its significantly lower multifaceted nature and CPU processing time empowers it to ultimately outperform their technique as far as by and large energy productivity.

CS is a quickly rising signal handling method that empowers the exact catch and reproduction of sparse signals from just a small amount of Nyquist rate tests, altogether decreasing the information rate and system power utilization. ZHANG Hong-xin1 et al. [7] suggested an in-depth relative inquiry into the current best estimates of class CS recovery. Reliability, accuracy, resistance to noise, time of calculation, and are used as primary measures. Besides, ECG signals are studied to investigate the execution of real-world bio-signals. Fred Chen et al. [8] showed an evaluated and authorized the capability of the developing CS worldview for ongoing power-efficient Electro Cardiogram (ECG) compression on resource-constrained sensors. In their research, they are applying sparsity models to exploit basic data in recovery algorithms. More precisely, returning to known sparse reconstruction algorithms, they identified creative scheme-based adjustments for the vigorous recovery of reduced signals such as ECG.

Anna M. R. Dixon et al. [9] present the utilization of CS calculations for information compression in remote sensors to address the energy and telemetry data transfer capacity in regular to remote sensor hubs. Consequences of the examination demonstrate that an advanced usage is fundamentally high power products for the remote sensor space wherever signal needs maximum gain and minimum to high goals. WBANs comprise of little insightful biomedical remote sensors appended on or embedded in the body to gather crucial biomedical information from the human anatomy giving continuous physical fitness monitoring systems [10]. Be that as it may, the utilization of regular ECG framework is confined by the patient’s versatility, transmission limit, and physical size. Along these lines, Monica Fira et al. [11] have improved remote ECG frameworks. Because of these, CS methodology as another sampling approach and the coordinated effort of sensing matrix selection algorithm dependent on unique thresholding approach were utilized to give a hearty low- robust low-complexity detection calculation in portals and passageways with high likelihood and enough precision. Jeevan K et al. [12], have proposed to utilize the block sparse Bayesian learning system to compress/ remake non-inadequate crude FECG chronicles. Especially, every segment of the network can contain just two nonzero passages. This demonstrates the system, contrasted with different calculations, for example, current CS calculations and wavelet calculations can enormously decrease code execution in CPU in the data compression stage.

2.3 Data Acquisition Techniques

In this correspondence, Anurag Singh and S. Dandapat et al. [13], have proposed and talk about similarly a few procedures for ECG signal compression propelled from the basics of CS theory, concentrating on securing systems, projection lattices and remaking word references and on the impacts of the preprocessing included. The primary methodology for ECG signal compression depends on the immediate CS securing of the signal with no preprocessing of the waveforms before taking the projections, neither for the development of the word references. This "certified" CS we will call patient-specific classical compressed sensing (PSCCS) since the word reference is worked from patient starting accounts. The second methodology executes a particular preprocessing stage intended to upgrade sparsity and improve recoverability, because of dividing the signal into single heart thumps (otherwise called cardiovascular examples) - signified further as cardiac patterns compressed sensing – (CPCS) since for this situation the gained signal and the word reference atoms are preprocessed portioned heart pulsates without or with focusing of the R wave.

The signal recovery algorithm [14] depends on limiting the pseudo-norm of the second-order difference, called the pseudo-norm, of the signal. The enhancement included is done utilizing a sequential conjugate-gradient algorithm. The lexicon learning calculation utilizes an iterative system wherein a signal reproduction, and a word reference update steps are rehashed until an assembly standard is fulfilled. The signal reproduction step is actualized by utilizing the proposed signal recovery algorithm and the word reference update step is executed by utilizing the Linear Least-Squares strategy. Broad recreation results exhibit their calculation defers improved reproduction execution for transiently associated ECG signals with respect to the cutting edge - regularized least squares and Bayesian learning-based calculations.

In this work, Anurag Singh et al. [15], proposed a Distributed Compressive Sensing (DCS) is to develop the fundamental connection structure between various channels of Multi-channel ECG signals. The joint remaking capacity of DCS lessens the number of compacted estimations required for precise reproduction without influencing the contortion level. Luisa F. Polanıa et al. [16], has inspected the fusion of CS with an effective lossy compression technique on ECG signals. The utilization of word reference figuring out how to naturally make the lexicon is depicted. Two strategies forword reference creation were presented: Patient Agnostic and specific. A comprehensive analysis of both methodologies is portrayed. Considering mobile ECG checking as an application, every system is broke down for a wide scope of CR.

Ongoing outcomes in telecardiology demonstrate that CS is a promising instrument to bring down energy utilization in WBAN for ECG monitoring. In this paper, Dana Al Akil1 et al. [18], have proposed to abuse the structure of the wavelet portrayal of the ECG signal to support the exhibition of CS-based strategies for compression and reproduction of ECG signals. Benyuan Liu et al. [19], have exhibited the structure of a low-power and territory proficient equipment motor for multi-channel compression of electrocardiogram (ECG) signals. CS is especially appropriate for low-control executions since it can drastically lessen circuit intricacy for compression tasks. Another element of the proposed design is that it is reasonable for multi-channel frameworks. The structure is actualized in a specific group of Field-Programmable Gate Arrays (FPGA) which are appropriate for low-control applications. Their estimation results demonstrate a compelling decrease in intensity utilization between 20 to 40 percent at various working frequencies utilizing a power gating system. The power utilization of a 4-channel framework and the 8-channel framework is expanded uniquely by 7.1% and 11.2% individually, contrasted with the single-channel framework. These frameworks cannot be structured together because there are a few limitations to be followed. Energy Utilization, Information Compression, And Gadget Cost are the significant requirements considered, which helps in developing an information compression system.

Andrianiaina Ravelomanantsoa et al. [20], have proposed to utilize a created CS calculation which can recoup such non-inadequate physiological signal. In this paper, Yishan Wang et al. [21] have recommended that the measurement of signal reconstruction quality, for example, PRD, isn’t sufficient for the compression method inspired by CS theory, particularly when neglecting the sampling rate of the crude signal. Among the current uses of WBSNs, a Wearable Health Monitoring System (WHMS) is the most significant. In common WHMS, scaled-down remote biosensors joined to or embedded in the human body, gather bio-signal, to give constant and consistent wellbeing checking. Yuvraj V. Parkale et al. [22], have exhibited a CS-based way to deal with compress and recuperate the detected physiological data from the remote biosensors. The CS encoding procedure has a minimum execution difficulty and is appropriate for use in energy-compelled frameworks, for example, WHMS.

This plan results in a decrease of capacity necessity and low power utilization of framework contrasted with the Nyquist sampling theory, where the examining recurrence must be at any rate twofold the most extreme recurrence present in the data signal for the accurate recreation of the data. This paper, Nidhi R, Bhadravati et al. [23], have introduced a top to bottom investigation on late patterns in CS concentrated on ECG compression. Giulia Da Poian et al. [24] have exhibited a wearable, and wireless ECG framework is right off the bat structured with Bluetooth Low Energy (BLE). It can recognize a 3-lead ECG signal, and it is remote. Besides, the advanced CS is actualized to build the energy effectiveness of a remote ECG sensor. The distinctive sparsifying premise, different compression proportions, and a few remaking calculations are recreated and reviewed. At long last, the recreation is done by the Android Application (App) on advanced mobile phones to show the signal progressively [25].

2.4 Analog to Digital Conversion

The analog-to-digital conversion (ADC) phase is one of the fundamental bottlenecks of rapid media communications frameworks. This part shows a study of various plausible simple to advanced transformation systems that are appropriate to conquer these challenges and to get the Software-Defined Radio (SDR) worldview, where most functionalities, rather than being performed in the simple area (i.e., channels and blenders), are performed in the computerized space. In SDR, the analog-to-digital transformation is executed following the reception apparatus, and the radio frequency (RF) signal is straightforwardly changed over to advance with no past blending stage. Since it is not possible to approach this thought with the conventional ADC from current business gadgets, this section depicts a few systems that might be utilized. Even though the proposed frameworks have progressively prohibitive determinations, these arrangements lessen the last multifaceted nature, as will be itemized in this section. Three diverse promising procedures, such as sub-sampling, interleaving, and CS, helps to perform this function efficiently.

3.1 Convex Relaxation

With the advancement of quick strategies for Linear Programming in the eighties, the possibility of convex relaxation turned out to be encouraging. These class calculations take care of a curved improvement issue through linear programming to get recreation [41]. The quantity of estimations required for accurate recreation is little. However, techniques are computationally unpredictable. “Basis Pursuit”(BP) [44, 45], “Basis Pursuit De-Noising (BPDN),” “Least Absolute Shrinkage and Selection Operator (LASSO)” and “Least Angle Regression (LARS)” are a few instances of such calculations. BP is a standard for breaking down a signal into an "ideal" superposition of word reference components, where ideal methods having the littlest l1norm of co-efficients among every such disintegration. BP, in exceptionally over complete lexicons, prompts enormous scale streamlining issues [53].

3.2 Non-Convex Minimization Algorithms

Numerous viable issues of significance are non- convex, and most non- convex issues are hard (if certainly feasible) to explain precisely in a sensible time. In substitute minimization systems [38], the optimization is done with certain factors that are held fixed in recurrent style and linearization strategies, in which the destinations and limitations are changed (or approximated by a convex work) [39]. Different procedures incorporate inquiry calculations (for example, hereditary calculations), which depend on basic arrangement update principles to advance.

3.3 Greedy Iterative Algorithm

Because of the quick reconstruction and low complexity of numerical structure, a group of iterative ravenous calculations has been generally utilized in compressive sensing as of late. These class calculations take care of the remaking issue by finding the appropriate response, well ordered, in an iterative style. The quick and exact reproduction calculations have been the focal point of the investigation of CS, they will be the key advancements for the utilization of CS. At present, the most important greedy algorithms include matching pursuit and gradient pursuit [43]. The thought is to choose sections of θ voraciously. At every emphasis, the segment of θ that associates most with is chosen. “Matching Pursuit” (MP) [58, 59], “Orthogonal Matching Pursuit” (OMP) [48, 49, 50] and “Compressive Sampling Matching Pursuit” (CoSaMP)” [51] are the normally utilized ravenous iterative calculations because of their low usage cost and fast of reconstruction [43].

3.4 Combinatorial / Sublinear Algorithms

This class of calculations recovers sparse signals through gathering testing. They are incredibly quick and efficient, when contrasted with convex relaxation or greedy calculations yet require specific design in the estimations, Φshould be sparse. Agent calculations are Fourier Sampling Algorithm; Chaining Pursuit proper is an iterative algorithm, Heavy Hitters on Steroids (HHS) [42].

3.5 Iterative Thresholding Algorithms

These algorithms are faster than other types of algorithms. Here, right estimations are recuperated by delicate or hard thresholding, beginning from uproarious estimations given the signal is inadequate [40]. Its capacity relies on number cycles and issues setup within reach.

These calculations can provide hypothetical certification with its execution, which can be appeared in the specific one. The fundamental thought of the thresholding method is to pursue a decent possibility for the gauge of help set, which fits the estimation. “Message Passing (MP)” calculations are a significant modification of iterative thresholding calculations in which fundamental factors are related to coordinated diagram edges. “Expander Matching Pursuits,” “Sparse Matching Pursuits,” and “Sequential Sparse Matching Pursuits” areas of late proposed calculations in this space accomplish close straight recuperation time.

3.6 Adaptive Filtering

Adaptive filtering techniques are outstanding, while CS is a mainstream point as of late, so it is astonishing that no writing utilizes an adaptive filtering structure in CS remaking issue. The reason may be that the point of CS is to remake a sparse signal while the answers for general adaptive filtering calculations are not sparse. A few LMS varieties, with some sparse limitations, included their cost capacities, exist in inadequate framework distinguishing proof. Subsequently, these techniques can be connected to take care of the CS issue. We propose another method for adaptive proof of sparse frameworks dependent on the CS theory. We control the transmitted pilot (input data) and the received data signal with the end goal that loads of adaptive filtering approach the compressed form of the inadequate framework rather than the first framework. To this end, we utilize irregular channel structure at the transmitter to shape the estimation lattice as per the CS system. The ordinary recuperation calculations can remake the first sparse framework. Thus, the denoising property of CS can be conveyed in the proposed strategy at the recuperation arrange.

3.7 Distributed Sensing and Processing

In the DCS issue, we are normally keen on improving the recuperation exactness of xl, or the necessities on the estimation gadget, contrasted with the solo-sensor case by accepting that the information among sensor hubs in a system have corresponded. The thought is then that by gathering or sharing some data among the hubs and misusing this relationship, and we can accomplish better execution. The signal connection is generally displayed contrastingly, relying upon the application.

4.1 Evaluation Measures of 2000-2018 Review paper

The performance of various acquisition and reconstruction methods is analyzed here in Figure 1 to Figure 5. From the results (2000-2018), we can conclude that the SPSA with noise is quite difficult to minimize because of the nonlinear nature of the recovered noise. In graph 1, graph 2, graph 3, and graph 4 we have analyzed the performance measures such as compression ratio, compressed signal reconstruction, compression sensing and compression measures based on the algorithms such as Simultaneous Perturbation Stochastic Approximation (SPSA), Health Technology Assessment(HTA), Analysis by Synthesis ECG compressor (ASEC) and SPIHT to be reviewed and the measures to be analyzed in the graphical representation.

Figure 1

Graphical representation of 2000-2010 review papers

Figure 2

Graphical representation of 2011-2013 review results

Figure 3

Graphical representation of 2014-2016 review results

Figure 4

Graphical representation of 2017-2018 review papers

Figure 5

Comparison of NMSE

An exhibition examination of the normal standardized mean square error (NMSE) between calculations from every classification is investigated underneath. From the Convex Relaxation classification, FISTA calculations are actualized. The BCS was executed, speaking to the Non-convex optimization classification. Lastly, from the Greedy calculations, the MP and OMP were executed. The near examination is given in Figure 5.

It very well may be seen that the exhibitions of the considerable number of calculations increment when the quantity of estimations M increments. In any case, it tends to be seen that a low Mesteem (M < N) enables the calculations to recover the inadequate signal bringing about low NMSE values. Among the calculations investigated, the BCS presents the best execution.

4.2 Analysis of Various Reconstruction methods in CS

The significant element of CS is that it requires proficient reconstruction calculations. The reproduction of compressed inspected signals includes an arrangement of an underdetermined system of direct conditions and, like this, has limitlessly numerous arrangements. The signal reconstruction procedure is basically to pick the best gauge of the first signal from all the potential arrangements acquired from the above converse condition. This might be accomplished by the raised improvement calculation. Different reconstruction calculations utilized for the reproduction of compressed samples signal might be delegated (Table 1),

• Convex Relaxation

• Non-Convex minimization

• Greedy Iterative

• Combinatorial

• Iterative Thresholding

4.3 Complexity analysis of algorithms

The complexity of reconstruction methods are presented in Table 2,

4.4 Performance Evaluation Metrics for CS

The CS system’s efficiency is measured by the following performance evaluation metrics [58, 59, 60, 61] illustrated, as shown below.

• Compression Ratio: It calculates the degree to which the algorithm used removes the unwanted(redundant) data. The Compression ratio obtains the information as shown in equation (5)

Where ODB_O is the number of bits taken to represent the original signal, and CDB_C is the number of bits taken to represent the compressed signal. If the compression ratio’s output value is high, then less memory space is required to store the data. The output value also consists of additional information related to the original signal, which helps in retrieving it later for processing.

• Mean Square Error(MSE): MSE is the average square difference between the original and reconstructed signal. The output of MSE is always a non-negative integer and neither zero. If the model’s MSE output value is close to zero, then the model offers significant performance.

• Percentage Root Mean Square Difference (PRMSD): PRMSD measures the acceptable level of fidelity and degree of distortion undergone by the algorithms used for compression and decompression. Visual Inspection is required to determine the adequacy of the reconstructed signal. Here the reconstructed signal’s distorted level is compared with the original signal with the following equation (6)

Where A_r is the reconstructed signal, and A_o is the original signal.

• Normalized Percentage Root Mean Square Difference (NPRMSD): It is the normalized value of PRMSD, which does not consider the mean value of the signal(A). The information of the signal lies in the variance. It can offer accurate error estimates when compared to PRMSD. The reconstructed signal fidelity issue cannot be solved from the mean value of the data obtained.

• Quality Score(QS): The quality score rates the overall performance of the compression technique used. If a higher value is obtained for QS, then the performance is said to have a potential effect in the signal compression. It is stated as the ratio between the Compression Ratio(CR) and PRMSD.

• Root Mean Square Error(RMSE): It calculates the quantity of error present in the reconstructed signal when compared to the original signal. The RMS method preserves the original signal quality, which serves as an added advantage. This method is more effective when compared to PRMSD.

• • Normalized Root Mean Square Error(NRMSE): The NRMSE of RMSE is similar to the PRMSD value obtained, except the multiplication done by a hundred.

• Covariance: The correlation between the original signal and the reconstructed signal is measured using the covariance function.

Where E is the expected value for the original and reconstructed signal obtained.

• Signal to Noise Ratio(SNR): The difference between the original and reconstructed signal is taken as the noise. The SNR compares the reconstructed signal with its background noise. The measurement used to express SNR is a logarithmic decibel scale. When compared with PRMSD and NPRMSD, the output value of SNR can predict the accuracy of the system more clearly.

• Peak Signal to Noise Ratio(PSNR): PSNR can be represented as the ratio between the maximum intensity of the signal and the background noise. This is used to estimate the quality of the reconstructed signal. A higher PSNR value indicates a high-quality signal.

In equation (14), the value MA used is constant and denotes the maximum intensity of the signal.