Underwater image compression using energy based adaptive block compressive sensing for IoUT applications

Abstract

Internet of Underwater Things (IoUT) consists of a large number of interconnected resource-constrained underwater devices that are capable of monitoring vast unexplored water bodies. Specifically, these devices are equipped with cameras to capture the underwater scenes and communicate them with each other and also with the cloud. However the data generated is very high which limits the performance of the IoUT devices in terms of computational capabilities and battery lifetime. Block Compressed Sensing technique which performs block by block fixed sampling can be utilized to achieve data compression however it ends up in image distortions after reconstruction. To unravel this issue, Adaptive Block Compressive Sensing technique is used. In this paper, Energy based Adaptive Block Compressive Sensing (EABCS) with Orthogonal Matching Pursuit reconstruction algorithm is proposed to improve the sampling performance and visual quality of the reconstructed image. Sparse binary random matrix is used as measurement matrix as it is highly sparse. With this energy based adaptive strategy, higher measurements are assigned to blocks with higher energy and vice versa. The proposed EABCS technique has achieved better compression with approximately 25–30% of measurements/samples with an increase in Peak signal to noise ratio of about 3–5 dB and structural similarity Index of around 0.1–0.3 with respect to other adaptive strategies. Percentage of space saving is also about 60–70%.

1 Introduction

IoUT [7] comprises of large number of underwater devices or things that are wirelessly connected together for interacting with each other without human intervention. The devices are particularly used for underwater surveillance applications to prevent theft of precious items and treasures, underwater exploration for natural resources discovery, Tsunami and earthquake prevention. The things in IoUT usually refers to processors, sensors, actuators, HD cameras and communication devices to capture, process and send the data they acquire from physical objects/devices. To minimize the Production cost of IoUT devices, they are equipped with low power embedded computational devices [26]. Table 1 gives the memory statistics of low power embedded devices in IoUT.

Table 1 The memory statistics of low power embedded devices in IoUT [26]

It is inferred from the Table 1, that those devices do not provide much storage or memory capabilities. These memory constraint makes it very important to use technologies that can effectively manage the IoUT system.

The overall architecture of IoUT based surveillance system is shown in Fig. 1.

Fig. 1

IoUT based surveillance system

In the surveillance system shown, Nodes (UWSN-Underwater Wireless Sensor Network) with cameras and IAUV’s (Imaging Autonomous Underwater Vehicles) captures the image for underwater monitoring. The captured image is sent to onshore station via surface station or buoys, which is then transported to cloud for storage and future reference. OMP can be used for image reconstruction. However, those devices cannot address large amount of data generated because of the computational complexities and energy constraints [5, 11]. There is a pressing need for preprocessing the image and compressing them using compressed sensing to cut down the data, power, energy and storage space [6].

1.1 Related works

Compressed sensing (CS) [3, 8, 10] reconstructs the signal from fewer samples in contradiction to those obtained using nyquist sampling. But traditional CS suffers from poor reconstructed signal quality [14] and are not suitable for large scale applications, high resolution and high density images.

To overcome the drawbacks of CS, BCS [14, 15] was proposed. This technique has small sensing matrix unlike traditional CS. Also the whole image is divided into individual blocks and processed by sensing matrix simultaneously. Thus it has low computational complexity and reconstruction quality is improved. In our previous work, we have applied BCS for medical image compression [24]. However, number of samples selected from each and every block is fixed in BCS without taking into account the block structure. To overcome this problem several adaptive strategies have been developed.

Saliency information of the image based on human vision system is used as adaptive strategy in [1, 22, 35, 39]. Another saliency method [12] for wildlife monitoring is based on image total variation. All these techniques involves usage of costly sensors to obtain the saliency map and hence not suited for IoUT. Masking effect of the human vision system is employed in [31] and weight matrix is designed in [34, 42]. In Adaptive sampling [16], a measurement matrix is designed using energy values. All these schemes have the drawback of designing masking matrix separately to extract information leading to computation and storage burden. Entropy, standard deviation and wavelet coefficients are used to adaptively extract measurements in [19, 33, 37] respectively. The image reconstruction quality is good however, they are not implemented for colour images. In [38], sparsity of the image block is exploited for adaptively allotting measurements. As it uses fixed threshold for samples allocation, it does not achieve convincing allocation of adaptive samples. Gradient based BCS on the basis of sparsity of the block is proposed in [41]. It has blocking artifacts in the reconstructed images. Measurement allocation based on block’s sparse degree is utilized in [20] for a green IoT. It does not offer convincing compression leading to poor rate-distortion performance. Multi-shape block split strategy [40] based ABCS increases burden of the processors in real time applications, as block size is not equal and dividing the image into blocks itself is a tedious process.

Adaptive sampling scheme using texture information is proposed in [28]. The image is reconstructed at the encoder side itself during the sampling process for achieving adaptive sampling rate allocation resulting in high computational complexity. Information about the edges of the image are used to adaptively extract measurements in [4]. These techniques revealed the block’s structural complexity well but, they ignored the correlation between the blocks resulting in poor visual quality. An image ABCS algorithm that is based on distribution of error between the blocks was proposed in [9, 21]. It can reveal the variation of texture and edge details between the blocks but there is no significant improvement in the visual quality.

Wildlife monitoring system with wireless multimedia sensor networks by using image extraction method is used in [23]. It can adaptively extract the target area by using characteristics of image pixels. It can reduce the energy loss of nodes but accurate reconstruction is not possible. Deep learning model using convolutional neural network for underwater image compression and machine learning based image compression is used in [18, 25] respectively. They require larger data set for training making the process slow and computationally costly. A self-adaptive learning dictionary based algorithm is used in [36]. It is compact and converges fastly but, it uses an iteration algorithm for optimization. The aforementioned algorithms had their own disadvantages like increased computational complexity, poor reconstruction quality, power and energy constraints and hence not suited for IoUT.

1.2 EABCS contributions

In order to overcome the issues of employing existing ABCS algorithms for IoUT application, we propose EABCS algorithm using energy distribution of the blocks to improve visual quality and reduce computational complexity [13]. Both fixed and adaptive measurements are used for better reconstruction of blocks with less energy. Blocks with less energy contains less low frequency components and in turn less information content. In contrast blocks with high energy have more information content. The algorithm aims in allotting more measurements for blocks with high energy than low energy ones. The proposed algorithm EABCS is tested for underwater image compression application. The advantages of the proposed methods include

• 1. Selection of a minimum number of samples from the entire image still maintaining good reconstruction quality.

• 2. Easy to implement in hardware platform as the proposed EABCS method uses simple mathematical operations rather than complicated matrix calculations. Therefore the proposed method is compatible with any existing resource-constrained CS systems.

• 3. Most of the blocks are reconstructed with almost less or no distortions.

• 4. The proposed EABCS can easily capture the important low-frequency components from the image much better than other parameters like standard deviation, entropy, saliency, error, contrast and sparsity having simple calculation procedures.

• 5. The algorithm maintains minimal run time as complexity is less.

• 6. SS of proposed algorithm is much high than other methods.

The remainder of this article is structured as follows. Section 2 presents the overview of BCS. Section 3 discusses about the proposed energy based Adaptive BCS technique. Section 4 provides the experiment results. Section 5 gives the conclusion of the paper and discusses about the future scope.

2 Overview of block compressive sensing (BCS)

In CS [3, 8, 10] the entire image is processed as a whole by using a large measurement operator which reduces the reconstruction performance. To overcome this, BCS is used where the whole image is split into smaller blocks and measurement operator is multiplied with each and every block separately and simultaneously. Hence BCS is considered to be simpler and efficient. Some of the other advantages are

• 1. Measurement matrix size is small. Therefore geometry of all regions of the image can be captured effectively.

• 2. Sampling time is less as blocks are processed simultaneously. Like CS, BCS demands sparsity and compressibility of the signal in particular domain.

Fig. 2

Overall architecture of EABCS

2.1 BCS sampling scheme

The blockwise sampling scheme comprises of the following steps

• 1. Let $N=r\times c$ represent the total pixel values in an image with ‘r’ rows and ‘c’ columns.

• 2. The image is divided into blocks. Let $N1=B^2$ represent the total pixel values in a block of size ‘$B \times B$’.

• 3. The blocks are sampled simultaneously and independently using the measurement matrix $\phi _b$.

• 4. The sampled vector can be defined by

  $$\begin{aligned} y_i=\phi _b x_i \end{aligned}$$

  (1)

  where $i=1,2,\ldots (N/N1)$.

• 5. Small block size makes the computation complex but the storage memory required is less where as large block size reduces reconstruction performance but reduces complexity.

In BCS sampling rate is fixed where as in ABCS it is adaptive. This adaptive concept helps in selecting sufficient number of measurements from each block thereby reducing redundancy.

3 Energy based adaptive block compressive sensing (EABCS)

Low frequency DCT coefficients attract the human eyes prominently than those with high frequencies. The low frequency coefficients possess high energy and contribute more for the reconstructed image quality. Hence, the proposed EABCS is based on energy value calculation of each block. If the energy content of a block is high, then more information is to be extracted from that particular block. More measurements are allotted to that block making it possible to allot different number of measurements to different blocks based on the energy content.

The overall framework of the EABCS scheme is shown in Fig. 2. The image is divided into blocks and varying number of measurements is chosen from each block. EABCS is performed on the measurements and the image is reconstructed using OMP algorithm.

3.1 EABCS sampling scheme

Energy is “measure” of information and is defined as rate of change of magnitude of the pixel values over local areas. If energy value of a block is high, then more low frequency components are present in that particular block. Higher sampling rate is required to reconstruct that particular block. Similarly low sampling rate is sufficient for a block with low energy value. Thus compressibility of any block with varying information content can be estimated by using the energy value. The various steps involved in the proposed algorithm is shown below

• 1. Consider an colour image with R, G, B planes. Each plane is divided into ‘n’ blocks (i = 1, 2, 3, ...n) of size $B \times B$ each. The image blocks are then converted to column vector $x_i$ and the following calculations are performed on each plane. Given the value of sampling rate ‘SR’, the sampling frequency (SF) can be calculated by

  $$\begin{aligned} \mathrm{SF}=\mathrm{SR}\times B \times B \times n \end{aligned}$$

  (2)

  where ‘n’ represents total number of blocks. To select measurements lesser than the total pixel values, we fix the upper bound (UB) by using the formula

  $$\begin{aligned} \mathrm{UB}=0.4\times B^2 \end{aligned}$$

  (3)

• 2. To ensure better reconstruction quality of the blocks with low energy, we assign fixed sampling rate (FSR) to each block by using the formula

  $$\begin{aligned} \mathrm{FSR}=W\times \mathrm{SR} \end{aligned}$$

  (4)

  where ‘W’ is fixed sampling allocation parameter

• 3. Fixed number of measurements $(\mathrm{FM}_i)$ to each block (i = 1, 2 ...n) are chosen by using

  $$\begin{aligned} \mathrm{FM}_i=\frac{\mathrm{FSR}\times \mathrm{SF}}{n} \end{aligned}$$

  (5)

  FSR represents the rate at which we try to obtain the fixed samples from the whole block set and SF represents the possible number of samples that can be obtained from any image block. Fixed measurements are calculated by using these values.

• 4. Energy $(E(x_i))$ is calculated as the sum of square of mean ($\mu$) and variance ($\sigma$) of DC coefficients

  $$\begin{aligned} E(x_i)=\sum _{i=1}^{n}|\mu _i|^2+\sigma _i \end{aligned}$$

  (6)

  Mean and variance of the blocks are found out by using the equation,

  $$\begin{aligned} \mu _i=\frac{\sum _{r=1}^{B} \sum _{c=1}^{B} I_\mathrm{dc}(r,c)}{B^2} \end{aligned}$$

  (7)
  $$\begin{aligned} \sigma _i^2=\frac{1}{B^2} \sum _{r=1}^{B} \sum _{c=1}^{B} (I_\mathrm{dc}(r,c)-\mu _i)^2 \end{aligned}$$

  (8)

  where $I_\mathrm{dc}(i,j)$ refers to DC coefficient in (i, j) position and $i=1,2,3\ldots n$

• 5. Percentage ($P_i$) energy contribution of each block is given by,

  $$\begin{aligned} P_i=\frac{E(x_i)}{\sum _{i=1}^{n}E(X_i)} \end{aligned}$$

  (9)

• 6. Adaptive measurements ($\mathrm{AM}_i$) for each block is determined by,

  $$\begin{aligned} \mathrm{AM}_i=P_i(\mathrm{SF}-n\mathrm{FM}_i) \end{aligned}$$

  (10)

• 7. The sum of excess measurements are calculated by using

  $$\begin{aligned} E=E+(\mathrm{AM}_i-\mathrm{UB}) \end{aligned}$$

  (11)

  Maximum limit of samples (UB) that a block can have is fixed, otherwise the purpose of using CS will not be met. For this case after adaptive measurement assignment to each block, the measurement count is checked with UB. If the allotted count exceeds UB, the excess value is calculated and the average of the excess value is added to all the blocks. This process is repeated till all blocks have measurements within the limit. This is largest percentage allocation principle [42].

• 8. The total measurements that can be allotted to each block are determined by

  $$\begin{aligned} \mathrm{TM}_i=\mathrm{FM}_i+\mathrm{AM}_i \end{aligned}$$

  (12)

• 9. The EABCS column vectors are then obtained by,

  $$\begin{aligned} y_i=\sum _{i=1}^{n}\phi _{ai}x_i \end{aligned}$$

  (13)

  where $\phi _{ai}$ is sparse binary random matrix [2, 29] of size $\mathrm{TM}_i \times B^2$. The size of $y_i$ is $\mathrm{TM}_i\times 1$.

The principle behind the proposed method is usage of second order moments (mean square plus variance) which is essentially the energy of the frequency components to project the subspace quadratically approach the direction towards the signal components having more energy. Therefore more information about the frequency components with high energy can be easily captured and reconstruction can be performed more accurately. It is the process of amplifying coefficients with high energy value and shrinking smaller ones.

Fig. 3

Flowchart for measurement allocation based on the energy content

3.2 OMP recovery

The OMP algorithm [27, 30] can be stated as follows

• 1. A measurement matrix $\phi$, A measurement vector y, and Sparsity level k are the inputs to the algorithm

• 2. Set the residue value $r_0=y$,the sparse set $\wedge _0=\big \{ \big \}$ and iteration number $t=1$

• 3. The index can be found from the following equation

  $$\begin{aligned} \lambda _t: \lambda _t=\mathrm{arg} \max _{i=1,2,\ldots n} \big | \big \langle r_{t-1},\varphi _i \big \rangle \big | \end{aligned}$$

  (14)

  This equation solves the optimization problem.

• 4. Update the index set by using

  $$\begin{aligned} \wedge _t=\wedge _{t-1} \cup \big \{ \lambda _t \big \} \end{aligned}$$

  (15)

  The matrix of chosen atoms is updated by using

  $$\begin{aligned} \phi _t= \big [ \phi _{t-1} \varphi _{\lambda _t} \big ] \end{aligned}$$

  (16)

• 5. The least square problem is solved to get the estimate

  $$\begin{aligned} s_t=\mathrm{arg} \min _s\Big \Vert |\phi _t s - y\Big \Vert _2 \end{aligned}$$

  (17)

• 6. New approximation of data and new residue are calculated by,

  $$\begin{aligned} a_t=\phi _t s_t \end{aligned}$$

  (18)
  $$\begin{aligned} r_t=y-a_t \end{aligned}$$

  (19)

• 7. Increment the value of t and iterate till z.

Table 2 PSNR (dB) and SSIM comparison of various standard test images for block size ($8\times 8$)

Table 3 Comparison between PSNR, number of measurements and block size for EABCS

This algorithm initially selects one index which has dominant value, adds the selected index to the sparse set. It then groups the column vectors corresponding to these sparse indices, extracts the signal components and then updates the residue vector. The value of estimate in the component $\lambda _t$ is equal to ith component of $x_t$.The flowchart of the EABCS algorithm is shown in Fig. 3.

4 Experimental results and discussion

The performance of the EABCS algorithm is compared with other ABCS techniques [19, 21, 37]. Adaptive compressive sensing algorithms are used because of its simplicity and suitability for resource constrained underwater IoUT applications. To evaluate the quality of the reconstructed image, standard colour test images like Lena ($512 \times 512$), Barbara ($720 \times 576$), Goldhill ($360 \times 288$) and Baboon ($512\times 512$) are used. The experiment is conducted using MATLAB (R2019a) and DCT is used as sparse transform. PSNR and SSIM [17] are averaged over five independent trials as the measurement matrix-sparse binary random matrix is purely random. The PSNR and SSIM are calculated by using the formulas given below,

$$\begin{aligned} \mathrm{PSNR}=20\log _{10}\frac{\mathrm{MAX}_I^2}{\mathrm{MSE}} \end{aligned}$$

(20)

where MSE is mean square error which is given by the formula,

$$\begin{aligned} \mathrm{MSE}=\frac{1}{\mathrm{MN}}\sum _{m=1}^M\sum _{n=1}^N(O(m,n)-R(m,n))^2 \end{aligned}$$

(21)

where O(m, n) is the pixel value of original image and R(m, n) is the pixel value of the recovered image.

$$\begin{aligned} \mathrm{SSIM}=\frac{(2 \mu _x \mu _y + S_1)(2\sigma _{xy}+S_2)}{(\mu _x^2+\mu _y^2+S_1)(\sigma _x^2+\sigma _y^2+S_2)} \end{aligned}$$

(22)

where the stabilization factors $S_1$=$(k_1L)^2$, $S_2=(k_2L)^2$, $k_1=0.01,k_2=0.03$ and L is the dynamic range if pixel values [32]. Table 2 shows the PSNR comparison of the various standard test images for various ABCS algorithms.

The results show that the proposed EABCS achieves highest PSNR and SSIM values for all standard images irrespective of the sampling rates in most of the cases. The proposed EABCS has achieved approximately 1–5 dB increase in PSNR values with respect to Error-ABCS and STD-ABCS and 0.5–2 dB increase with respect to ENT-ABCS techniques.There is an increase in SSIM value of around 0.1–0.2.

4.1 Proposed algorithm’s performance comparison for underwater images

Important parameter for image reconstruction is the block size ‘$B\times B$’. If ‘$B \times B$’ is too small the computation of energy values of the block becomes more complex. This is because smaller the block size, more will be the number of blocks. Also if ‘$B\times B$’ is too large then the energy values may not correctly reflect the characteristic of the block. Sampling rate allocation parameter ‘W’ affects the reconstructed image quality. These two are the main parameters. Table 3 presents the comparison between the PSNR (dB), number of measurements and block size for a sea algae ($320\times 240$). It is inferred that high PSNR value is achieved with less number of measurements only for the block size of $8 \times 8$. Hence $8\times 8$ is used as block size for the simulation of the proposed EABCS in this paper.

Subjective comparison of sea algae $(320\times 240)$ for various block sizes is shown in Fig. 4. Figure 4a is the original image and (b), (c) and (d) are reconstructed images for block sizes of $8\times 8, 16\times 16, 32\times 32$ respectively.

Fig. 4

Visual quality comparison a original image b, c and d reconstructed Sea algae $(320\times 240)$ image by EABCS for block of $8\times 8, 16 \times 16$ and $32\times 32$ respectively

Table 4 lists the comparison between the PSNR (dB) and samples allocation parameter ‘W’, for SR $=$ 0.5 and block size of $8\times 8$. As ‘W’ increases PSNR decreases verifying the fact that adaptive measurement allocation (small W) improves the reconstructed image quality.

Figure 5 shows the graphical representation of the relationship between PSNR and ‘W’. ‘W’ value of 0.5 is fixed in the simulation of the proposed EABCS as very low ‘W’ value leads to complete adaptiveness and very high ‘W’ value corresponds to conventional CS.

Table 5 compares the sampling time of various algorithms for sea algae image of block size $(8\times 8)$

It is evident from the table that the run time taken for sampling in proposed technique (EABCS) is smallest for a block size of $8\times 8$. This makes the proposed EABCS more suitable for resource constrained IoUT applications. To check the consistency of the proposed EABCS, it is applied for few other underwater images and results are shown below. Figure 6 presents the visual quality comparison of underwater fish image, reconstructed by proposed method, standard BCS algorithm [14] and adaptive algorithm used in [19, 21, 35, 37] at sampling rate of 0.1, 0.3 and 0.5.

From the Fig. 6 it is inferred that the proposed algorithm reconstructs all blocks of the image with visually pleasing quality whereas, other algorithms gives poor reconstruction quality of certain blocks.The proposed EABCS was tested for a dataset containing more than 30 images and the visual quality of few underwater images and standard test images reconstructed using EABCS at SR $=$ 0.5 is shown in Figs. 7 and 8 respectively.

The proposed EABCS reconstructs standard test images with high quality which can be observed from Fig. 8. To justify the better performance of the proposed algorithm at even low SR of about 0.1, few other underwater images are simulated and the results are shown in the Fig. 9.

From the Fig. 9 it is visualized that the images are reconstructed with good quality at SR $=$ 0.1. Choosing low sampling rate reduces the number of samples selected. so, there is little distortion in reconstructed image. Local magnification of few underwater images after reconstruction by EABCS at SR $=$ 0.5 is shown in Fig. 10.

Table 4 Comparison between PSNR and sampling rate allocation parameter for EABCS and adaptive algorithms

Fig. 5

The relationship between PSNR (dB) and parameter ‘W’

Table 5 Comparison between runtime of adaptive algorithms and Proposed EABCS

Fig. 6

Visual quality comparison a original image. b, c and d recovered images by adaptive algorithm at SR $=$ 0.1, 0.3, 0.5. From left, column 1: BCS, column 2: S-EABCS, column 3: STD-ABCS; column 4: ENT-ABCS; column 5: proposed EABCS

Fig. 7

Image courtesy: unsplash website

Visual quality of underwater images reconstructed using proposed EABCS at SR $=$ 0.5. Column 1 and 3: original images, column 2 and 4: reconstructed images

Fig. 8

Reconstructed Standard test images by EABCS at SR $=$ 0.5 a lena b baboon, c barbara, d goldhill

Fig. 9

Visual quality comparison. a, c and e original images; b, d and f images reconstructed by using proposed EABCS at SR $=$ 0.1

Fig. 10

Local magnification of underwater images reconstructed by proposed EABCS at SR $=$ 0.5

From Fig. 10, It is inferred that magnification of reconstructed images at SR $=$ 0.5, gives minute detail about the object. Objective evaluation for Sea horse $(320\times 200)$, Sea fishes $(640\times 400)$, Sea weed $(336\times 448)$, Lion fish $(496\times 368)$ and Turtle $(496\times 368)$ is shown in the Tables 6, 7 and 8. It compares PSNR, SSIM, Number of measurements and percentage measurements used for EABCS and other adaptive schemes used in standard BCS algorithm [14] and adaptive algorithm used in [19, 21, 35, 37] . It can be concluded that the proposed EABCS works well for various images with good PSNR, SSIM and reduced number of measurements irrespective of the size, colour or texture. $\Delta$ represents the difference value between proposed EABCS and other adaptive techniques used for comparison.

From Table 6 it is concluded that the proposed EABCS achieves highest average PSNR values at any sampling rate. Only for sea weed image at a sampling rate of 0.3, ENT-ABCS outperforms EABCS. Proposed EABCS has achieved increase in PSNR of 1–2 dB, 1.5–3.5 dB, 2.5–4 dB, 3–5 dB and 4–6 dB with respect to ENT-ABCS, STD-ABCS, Error-ABCS, S-ABCS and BCS respectively. By using proposed EABCS there is a significant increase in the PSNR value as the sampling rate increases verifying the fact that increasing the number of measurements increases the reconstruction quality. ENT-ABCS has PSNR and SSIM values closer to the proposed EABCS technique but, the number of measurements used is high which can be verified from the Table 8.

From Table 7, it is inferred that the proposed EABCS has higher SSIM values in most of the cases. There is some SSIM degradation when reconstructing seahorse and turtle at SR $=$ 0.3 and lion fish at SR $=$ 0.1. Proposed EABCS has improved the quality of the reconstructed image which is proved from both subjective and objective point of view. The number of measurements chosen plays an important role in enhancing the quality of the reconstructed image. Choosing less number of measurements with high PSNR and SSIM is the highlight of the EABCS technique. The % Measurements (% M’s) used is given by,

$$\begin{aligned} \%M's=\frac{\mathrm{Measurements}\,(M's)}{{\hbox {Total-measurements}}} \end{aligned}$$

(23)

Other adaptive schemes have good PSNR and SSIM in few cases. But the number of measurements chosen is always greater than proposed EABCS scheme. The Proposed EABCS is the most suitable technique in reconstructing the image with lesser measurements and good quality. The graphical representation of results obtained is shown in Fig. 11.

Table 6 PSNR(dB) comparison of various underwater images

Table 7 SSIM comparison of various underwater images

Table 8 Measurement comparison of various underwater images

Fig. 11

Graphical representation. Sampling rate vs. a PSNR b SSIM c %Measurements

From Fig. 11 it can be concluded that the proposed EABCS outperforms all other ABCS schemes.

4.2 Proposed EABCS suitability for IoUT

Minimal sample selection From Table 8, we can see EABCS can reconstruct images with 25–30% of samples which is very less than other schemes. Due to lesser number of sample selection the reconstructed image size is much less than the original image size which can be observed from Table 9. Minimal sample selection helps in improving memory and storage issues in IoUT.

Simple mathematical operations Complicated operations increases the computational time thereby wasting power and energy of low power embedded devices. Proposed EABCS uses only simple operations rather than using complicated convolution or matrix vector calculations.

Good visual quality with less distortions Visual quality is not directly related with memory or power issues. But, achieving good reconstruction quality would be an additional benefit for obtaining certain minute details for analysis or security purposes.This is the highlight of proposed EABCS.

Minimal run time From Tables 5 and 9 we can conclude that the proposed EABCS executes faster due to its less complex nature. This helps improving the battery life of low power embedded devices.

Better space saving Space saving is calculated from the size of the original and compressed file. Table 9 shows the compressed file size and space saving of proposed EABCS compared with all other schemes and it can be concluded that proposed EABCS has SS of about 60–70% which is much higher than other methods.

Table 9 Comparison of various algorithms

5 Conclusion and future work

In this research paper, an Adaptive BCS technique-EABCS which utilizes the energy content of blocks to achieve the goal of adaptively allocating measurements is proposed for IoUT devices with limited resources. The visual quality of the reconstructed image was significantly improved by the proposed EABCS scheme compared with the other adaptive BCS schemes. Due to the ease in energy value calculation and remarkable quality improvement, the proposed technique is very useful for practical IoUT applications. The advantage of the proposed method is that it eliminates block measurement redundancy, achieves good subjective and objective results irrespective of the colour, size, texture, smoothness and works well for even low sampling rate. The future work is to implement EABCS scheme for underwater surveillance videos.