Rate adaptive compressed sampling based on region division for wireless sensor networks Authors: Wei Wang, Xiaoping Jin, Daying Quan, Mingmin Zhu, Xiaofeng Wang, Ming Zheng, Jingjian Li, Jianhua Chen Published: 2024 ABSTRACT Image acquisition and transmission in wireless sensor networks (WSN) are core issues for some resource-deficient multimedia sensing applications. Reducing sampling rates and data transmission lowers sensor node costs and energy, addressing communication bottlenecks. Block compressed sampling (BCS) can meet the above requirements. For BCS, the sparsity or smoothness of the block signal is a crucial parameter, which determines the setting range of the sampling rate. For the sampling side of the sensor node, we cannot directly obtain the complete digital signal. Therefore, it becomes difficult to perform adaptive rate compressed sampling. In this paper, a novel adaptive sampling rate allocation scheme based on region division is proposed. First, we use a simple auxiliary vector to determine the complex and smooth regions of the current image. For the smooth region, we use a mean vector to divide each block into a residual block and a mean value block. Then the proposed prior probability sparsity estimation model is used to estimate the sparsity order of each residual block, while each mean value block requires only one measurement to restore losslessly. For the complex region, we first set a higher baseline sampling rate for it, and then adaptively allocate the remaining supplementary sampling rate based on the statistical characteristics of each block itself. Experiment results show that the proposed scheme can allocate an appropriate sampling rate to each block, reduce the total sampling rate, and significantly improve the signal reconstruction quality simultaneously. Introduction Wireless Sensor Networks (WSNs) consist of low-cost and low-power sensor nodes that can wirelessly transmit data. These devices are capable of sensing, processing, and communicating simultaneously, enabling WSNs to be widely applied in various fields for monitoring and identifying critical changes in environmental or physical states1–4. Wireless image sensors networks can provide rich visual information for many potential applications, such as wireless surveillance systems5,6and submarine underwater wireless sensor systems7. These visual sensors are capable of capturing and transmitting key visual information about the surrounding environment, which has driven the development of various emerging applications. These include visual monitoring of wildlife, battlefield surveillance, particularly suitable for hard-to-reach or hostile areas, detecting snipers, intruders and tracking their movements, observing traffic flow, and applications in mobile multimedia, etc. The image collection and transmission process is one of the key steps in wireless image sensor networks. Since the energy and bandwidth resources of each node in a wireless sensor network are limited, it is not suitable to sample or transmit a large amount of data at full rate without compression. By decreasing the sampling rate and minimizing the data volume during the communication transmission process, the hardware expenses and power usage of each sensor node can be significantly lessened, thus addressing the communication bottleneck. Some research works have highlighted that the transmission of uncompressed images via current Wireless Sensor Networks (WSNs) will considerably drain the energy of each node8–10. Hence, implementing a compression process prior to data transmission becomes crucial. An image compression scheme based on WSNs uses the neighborhood correlation sequence algorithm to successfully achieve efficient compression effects11. However, this method increases the cycle time of the processor due to the increased complexity of the algorithm, which in turn increases the power consumption of the sensor and may shorten the service life of the battery. Therefore, reducing energy consumption is particularly critical when collecting and compressing images in WSNs. Chen et al.12 proposed an innovative image compression approach for WSNs, utilizing scrambling block compressed sampling (CS). This method significantly surpasses traditional compression techniques in. Fig. 1The reconstruction results under different sampling rates. (a) is the original image, (b) is the reconstructed image when the sampling rate is too low, (c) is the reconstructed image when the sufficient sampling rate is used, and (d) is the reconstructed image when more than the sufficient sampling rate is used. The reconstruction results under different sampling rates. (a) is the original image, (b) is the reconstructed image when the sampling rate is too low, (c) is the reconstructed image when the sufficient sampling rate is used, and (d) is the reconstructed image when more than the sufficient sampling rate is used. terms of rate-distortion performance, enhancing the peak signal-to-noise ratio (PSNR) by 2 to 3 dB. Nevertheless, the algorithm demands considerable computational resources, attributable to the intricate processes involved in encoding and quantization. In resource-deficient wireless sensor networks, designing efficient image compression and transmission schemes faces several challenges. First, traditional image compression algorithms need to find a balance between computational overhead and data compression effect. Secondly, traditional image compression requires that the sensor node must completely obtain the pixel domain information of the original image before performing compression processing. Unlike traditional compression methods, CS13,14can achieve high compression rates without having to completely collect pixel data from the original image, and it has a high probability of accurate image reconstruction. It achieves the integration of sampling and compression, avoiding full-rate data collection and encoding compression on the sampling side of each sensor node, thereby significantly reducing energy consumption. In BCS imaging systems, traditional sampling strategies assign the same sampling rate to each image block. However, in reality, each block of an image exhibits different statistical characteristics, that is, their sparsity level is different. Generally, it is difficult for image blocks with different structures to share the same sampling rate and measurement matrix. Using a uniform sampling rate for all blocks will lead to larger reconstruction errors for non-sparse blocks, while sparse blocks will be oversampled, resulting in unreasonable allocation of sampling resources. For example, as shown in Fig. 1, in the case of uniform sampling, when the sampling rate is very low, the quality of the reconstructed image is poor, especially in the complex region. When the sampling rate is sufficient or much higher than the sufficient rate, the reconstructed results look almost the same. This means that it is necessary to assign appropriate sampling rates to different areas of the image based on the statistical characteristics of the image signal itself. In order to be able to construct a suitable adaptive sampling scheme on the sampling side, Yu et al.15 proposed a scheme that uses complementary sensors to obtain low-resolution images, and uses pulsed cosine transform to extract prominent features in the image. This method achieves adaptive block compressed sampling (ABCS) by allocating more measurements to salient areas and less measurements to non-salient areas. However, this scheme requires obtaining low-resolution images before adaptive sampling, which increases hardware cost and energy consumption. Warnell et al.16introduced an adaptive CS technique that utilizes a cross-validation matrix to estimate the reconstruction error. This method can effectively estimate the sparsity of the signal by constructing a probability distribution model of the video signal. In17, researchers proposed a sampling allocation method to select more participants in areas with uneven data distribution, thereby improving reconstruction accuracy. Experiment results show that the proposed method exhibits a more reasonable allocation of sampling resources. In18, researchers proposed an adaptive reweighted sampling adaptive block compressed sensing (ARS-ABCS) method in the fast Haar wavelet transform (FHWT) domain. This method allocates more samples to areas with low reconstruction quality or severe artifacts through adaptive selection and reweighting process, thereby improving the overall image reconstruction quality. Experiment results show that ARS-ABCS improves PSNR by 1.5 to 5dB compared with other non-weighted adaptive schemes. In order to improve the performance of CS image coding, Chen et al.19proposed a CS multi-layer residual coding scheme to correct this problem to a certain extent. This scheme divides CS measurements into multiple levels and utilizes all previous level measurements to predict measurement results at a specific level. By converting CS measurements into multi-layer residual coefficients, these coefficients are easier to compress. By calculating the residuals between the quantized true CS measurements and the corresponding inferred measurements, and using Huffman coding to associate each residual quantization index with a binary code, the redundancy between CS measurements is effectively reduced. In addition, the prediction and quantization processes are designed to be layer independent, which can save a lot of encoding time. Experiment results show that this scheme can significantly outperform JPEG2000 on some test images and approach or reach the performance level of HEVC-Intra. In20, a scheme combining deep learning regularization term and proximal operator is proposed to solve the CS reconstruction problem. A residual regression network measures the distance between damaged and clean images to determine its subspace. A dilated residual channel attention network processes the proximal operator to map distorted images to the clean set. Adaptive proximal selection strategy is embedded in CS reconstruction loop, and self-integration strategy enhances restoration performance. In21, an adaptive CS method using fuzzy rules is proposed. This method automates sampling rate selection by leveraging image saliency and edge features. A fuzzy logic system (FLS) is used to determine the adaptability of the block sampling ratio based on edge and saliency information. In22, researchers proposed a novel adaptive sampling rate image compressive sensing neural network called RACSNet. This scheme utilizes the reconstruction distortion in the measured domain to guide the sampling rate allocation of different image patches without accessing the original image. RACSNet achieves adaptive sampling rate allocation and significant performance at arbitrary sampling rates by progressively training reusable sampling matrices, pyramidal initial reconstruction subnetworks, hierarchical deep reconstruction subnetworks, and improved loss functions and reconstruction distortion maps. Image compressed sensing effect that is better than existing methods. In23, researchers proposed a local random sampling method. For each set of local image measurements, this method first randomly selects an image pixel and then measures its nearby pixels based on the probability of being close to the primary pixel. In24, researchers proposed a post-transform frequency saliency CS method, which calculates the frequency saliency information of the image in the post-wavelet domain through the transformed post-wavelet coefficients. Specifically, the wavelet coefficients are treated as pixels of a block megapixel sensor. Experimental results show that this method outperforms traditional saliency-based methods in terms of peak signal-to-noise ratio, average structural similarity index, and visual quality. A multi-scale progressive CS method for high-resolution imaging is proposed25. The image resolution is gradually improved through progressive sampling, and low-resolution information is used to improve efficiency. The multi-scale feature accelerates deep learning (DL) reconstruction and solves the limitations of high-resolution image training. A compressed sensing image encryption scheme based on optimized orthogonal measurement matrix was proposed26, which achieves high-security image encryption and high-quality decryption reconstruction by combining discrete wavelet transform (DWT) and orthogonal matching pursuit (OMP) algorithm with chaos theory. Liu et al.27proposed a measurement allocation strategy that enhances the efficiency of CS sampling by prioritizing important blocks with more measurements while assigning fewer measurements to less crucial blocks, guided by the user’s reconstruction quality requirements for each block. In28, the energy content of the image blocks has been used as the allocation factor. The higher the energy value, the greater the number of measurements assigned to that block, and vice versa. In29, researchers introduced an innovative adaptive compressive sensing method that uses fuzzy rules. This approach automatically determines the sampling rate by considering the saliency and edge characteristics of the image. The fuzzy logic system adjusts the sampling rate adaptively, relying on these two key features: edge information and saliency. In22, researchers proposed a novel scheme to allocate sampling rates based on content characteristics without requiring access to the original image. This is achieved through a dual-branch network structure, where a lightweight base sampling branch provides feature information to guide the adaptive sampling rates. Testa and Magli30proposed an ABCS imaging scheme. This scheme is able to estimate the complexity of individual blocks directly from measurement data. Specifically, they adaptively determined the optimal number of measurements required for each block using previously calculated correlation coefficients from the autoregressive model. In a previous study31, we proposed an ABCS scheme based on empirical mode decomposition technique. This scheme can guide the process of adaptive sampling based on the energy distribution map of high-frequency components extracted from the energy matrix of measurement results through data-driven Zigzag EMD. However, even though the scheme used an improved one-dimensional EMD technique to reduce computational complexity, it still increased the computational burden on the sampling side. For BCS systems, the rational allocation of sampling rates is crucial as it impacts the sampling efficiency of each sensor node and the accuracy of the image restoration. Considering that natural images are usually not sparse and exhibit considerable spatial redundancy, this redundancy can be exploited as prior knowledge in the sampling stage to increase the sparsity of the image signal to be sampled. Based on this analysis, aiming to enhance the sparsity of the signal to be sampled, diminish the total sampling rate, and more effectively tackle the challenge of sampling rate allocation, we propose a novel ABCS scheme based on region division. First, we classify the image to be sampled into complex and smooth regions in the measurement domain using an auxiliary vector. For the smooth region, we use a mean vector to divide each block into a sparse residual block and a mean block. Then, the proposed prior probability sparsity estimation model is utilized to estimate the sparsity order of each residual block, while the mean block, containing equal elements, can be succinctly represented by a single value. For the complex region, in order to reconstruct the original signal with high accuracy, we set a higher baseline sampling rate for them, followed by adaptive supplementary sampling rate allocation based on the statistical characteristics of the block itself. The proposed scheme can split the signal into complex and smooth regions and perform appropriate adaptive rate sampling processes for each region. In addition, the proposed scheme can further sparse the signal to be sampled based on the statistical characteristics of the image signal, significantly improving the efficiency of the sampling process, thereby significantly improving the signal reconstruction accuracy at the same or lower sampling rate. Block CS overview Suppose the sample signal of the current scene is \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\varvec{v}$$\end{document}. The signal \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\varvec{v}$$\end{document} consists of N non-overlapping n-dimensional vectorized block signals \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\varvec{v}}_{i}$$\end{document} (i = 1, 2, …, N). Then, the measurement vector of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\varvec{v}}_{i}$$\end{document} can be obtained after \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\varvec{v}}_{i}$$\end{document} is sampled by the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:m\times\:n$$\end{document} measurement matrix \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\varvec{\varPhi\:}}_{B}(m{F}_{mean}$$\end{document}, it means that the current i-th block contains higher spatial frequency, it will be categorized into the complex region; otherwise, it will be classified into the smooth region. Given the lower spatial frequency of the smooth region signal, pixel values in this areas exhibit smoother transitions, that is, the pixel values within each smooth region block are convergent and the fluctuation range is extremely limited. Consequently, the smooth region contains a surplus of redundant information. It becomes imperative to enhance sparsity in the smooth region by further eliminating redundant data. This optimization aims to curtail sampling costs and bolster sampling efficiency. Assume an all-one vector \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\varvec{b}=[\text{1,1},\dots\:,1]$$\end{document} with length n, and use \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\varvec{b}$$\end{document} to sample the block signal \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\varvec{v}}_{i,smooth}$$\end{document} in the smooth region. Then we can calculate the mean value of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\varvec{v}}_{i,smooth}$$\end{document} as:5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:E\left({\varvec{v}}_{i,smooth}\right)\:=\frac{\varvec{b}{\varvec{v}}_{i,smooth}}{n}.$$\end{document} Further, we can calculate the mean vector of the signal \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\varvec{v}}_{i,smooth}$$\end{document}:6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\varvec{E}}_{i,smooth}^{*}\:=E\left({\varvec{v}}_{i,smooth}\right){\varvec{b}}^{T}.$$\end{document} Based on the above analysis, in order to further remove the redundant components of each block in the smooth region, we have,7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\varvec{v}}_{i,smooth,res}\:=\:{\varvec{v}}_{i,smooth}-\:{\varvec{E}}_{i,smooth}^{*}.$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\varvec{v}}_{i,smooth,res}$$\end{document} represents the residual signal, theoretically expected to be sparser than \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\varvec{v}}_{i,smooth}$$\end{document}. In other words, a greater number of elements in the vector tend towards zero or values close to zero, with only a few elements having large magnitudes. Then, we have the variance of the residual signal \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\varvec{v}}_{i,smooth,res}$$\end{document} as:8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\text{D}\left({\varvec{v}}_{i,smooth,res}\right)=\frac{{\left|\right|{\varvec{v}}_{i,smooth,res}\left|\right|}_{2}^{2}}{n}.$$\end{document} However, it is infeasible to directly calculate the signal variance at the sampling side of a sensor node according to (8). Considering that the sampling process of CS is essentially a restricted dimensionality reduction projection process of the signal, according to the Johnson-Lindenstrauss Lemma37,38, theoretically we can estimate the energy of the original signal based on the compressed measurement values. Suppose the initial sampling matrix \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\varvec{\varPhi\:}}_{init}\in\:{\varvec{R}}^{\gamma\:\times\:n}$$\end{document}, n is the length of the block signal, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\gamma\:\ge\:C{\alpha\:}^{-2}\text{l}\text{o}\text{g}(1/2\beta\:)$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\alpha\:$$\end{document} is the accuracy parameter, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\beta\:$$\end{document}is the confidence parameter, 𝐶 is a constant value, for Gaussian and Bernoulli random matrices the constant 𝐶 can be set as 839. Since \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:(1-\alpha\:)>0$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\parallel\varvec{v}_{i}\parallel>0$$\end{document}, then combined with the above analysis we have:9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{(1-\alpha\:)}^{2}\le\:\frac{{\parallel{\varvec{y}}_{i,init}\parallel}_{2}^{2}}{{\parallel{\varvec{v}}_{i}\parallel}_{2}^{2}}\le\:{(1+\alpha\:)}^{2}.$$\end{document} is satisfied with a probability exceeding \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:(1-\beta\:)$$\end{document}, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\varvec{y}}_{i,init}={\varvec{\varPhi\:}}_{init}{\varvec{v}}_{i}$$\end{document} is the initial measurements. At the sampling side of the actual WSN imaging system, we can approximate it as10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\parallel{\varvec{y}}_{i,init}\parallel}_{2}^{2}\cong\:{\parallel{\varvec{v}}_{i}\parallel}_{2}^{2}.$$\end{document} Combining (10) and considering (8) and (7), we can further derive the following approximation:11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:D\left({\varvec{v}}_{i,smooth,res}\right)\:\cong\:\frac{{\parallel{\varvec{\varPhi\:}}_{init}{\varvec{v}}_{i,smooth,res}\parallel}_{2}^{2}}{n}=\frac{{\parallel{\varvec{\varPhi\:}}_{init}{\varvec{v}}_{i,smooth}-{\varvec{\varPhi\:}}_{init}{\varvec{E}}_{i,smooth}^{*}\parallel}_{2}^{2}}{n}.$$\end{document} (11) can be executed on the sampling side, which means that according to (11), the variance of each residual block signal in the smooth region can be estimated in the compressed domain. Combining the above analysis and the characteristics of the smooth region, we can assume that the pixel value in the i-th residual block approximately obeys a Gaussian distribution: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:N(0,\:{\sigma\:}_{i,smooth,res}^{2})$$\end{document}, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\sigma\:}_{i,smooth,res}^{2}$$\end{document} represents the variance value of the i-th residual block in the current smooth region. In order to estimate the number of large values (i.e. the sparsity order) within each sparse block in the smooth region, we need to set a threshold \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\tau\:}_{L}$$\end{document}. The pixel value exceeding \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\tau\:}_{L}$$\end{document} is considered large value, otherwise the value is considered to be 0 or close to 0 and is not counted. According to the above analysis, the probability that the pixel value in the i-th residual block in the smooth region is greater than \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\tau\:}_{L}$$\end{document} is:12\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{P}_{i,smooth,L}=2{\int\:}_{{\tau\:}_{L}}^{\infty\:}N(0,\:{\sigma\:}_{i,smooth,res}^{2})dx.$$\end{document} Then, the sparsity order of the i-th residual block can be calculated as:13\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\varGamma\:}_{i,smooth,res}=n\cdot\:{P}_{i,smooth,L}.$$\end{document} By traversing each residual block in the smooth region, the sparsity order of each block can be finally estimated. According to the estimated sparsity order of each block, we can further calculate the sampling number required for each residual block40:14\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{M}_{i,smooth,res}\ge\:2{\varGamma\:}_{i,smooth,res}\text{log}(n/{M}_{i,smooth,res}).$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{M}_{i,smooth,res}$$\end{document} is the sampling number required for the i-th residual block (generally what we discuss on the sampling side is the minimum sampling rate that meets the reconstruction requirements). If the estimated number \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{M}_{i,smooth,res}$$\end{document} is greater than the initial measurement number, a corresponding number of supplementary sampling will be taken. If the estimated number \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{M}_{i,smooth,res}$$\end{document} is less than the initial measurement number, no additional sampling is needed, but the excess initial measurements should be removed. Each block in the smooth region has a mean vector where every element is identical, allowing it to be fully described with a single value. On the reconstruction side, each mean image block (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\varvec{B}}_{1},{\varvec{B}}_{2,\dots\:,}{\varvec{B}}_{N}$$\end{document}) in the smooth region can be restored using the received mean values. These mean blocks are then spliced together to form the mean image \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\varvec{I}}_{smooth}^{mean}$$\end{document} for the smooth region based on the category index of the corresponding smooth region. Since the blocks in the image are categorized into complex and smooth regions, recording each block’s area category only requires one bit. This minimal bit requirement has a negligible impact on transmission bandwidth. Block statistical characteristics estimation and sampling rate adaptive allocation for complex region For the complex region, the spatial frequency of each block signal is relatively high, and the signal volatility is strong. In order to ensure high-accuracy recovery of the complex image, we first set a baseline sampling rate (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{R}_{base}$$\end{document}) for each complex block. Moreover, in order to reduce the blocking effect in the complex region, we also appropriately increase the size of the sampling block in the complex region to reduce the spectrum leakage in the process of reconstruction, so the dimension of the block signal in the complex region is correspondingly increased to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{n}_{C}$$\end{document}, and the initial measurement matrix is adjusted to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\varvec{\varPhi\:}}_{init,C}$$\end{document} accordingly. Next, we need to estimate the statistical characteristics of each block in the complex region. Variance is a good measure of how much a signal deviates, indicating the signal’s level of fluctuation. By estimating the variance for each block in the complex region, we can then allocate an appropriate specific sampling rate for each block. Use an all-one vector \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\varvec{b}=[\text{1,1},\dots\:,1]$$\end{document} of length \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{n}_{C}$$\end{document} to sample the block signal \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\varvec{v}}_{i,comp}$$\end{document} in the complex region, and then we can calculate the mean value of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\varvec{v}}_{i,comp}$$\end{document} as :15\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:E\left({\varvec{v}}_{i,comp}\right)\:=\frac{\varvec{b}{\varvec{v}}_{i,comp}}{{n}_{C}}.$$\end{document} Further, we can calculate the mean vector of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\varvec{v}}_{i,comp}$$\end{document}:16\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\varvec{E}}_{i,comp}^{*}\:=E\left({\varvec{v}}_{i,comp}\right){\varvec{b}}^{T}.$$\end{document} Then the variance of the signal \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\varvec{v}}_{i,comp}$$\end{document} can be recorded as:17\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:D\left({\varvec{v}}_{i,comp}\right)=\frac{{\left|\right|{\varvec{v}}_{i,comp}-{\varvec{E}}_{i,comp}^{*}\left|\right|}_{2}^{2}}{{n}_{C}}.$$\end{document} According to the previous analysis and (10), we have18\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\left|\right|{\varvec{v}}_{i,comp}-{\varvec{E}}_{i,comp}^{*}\left|\right|}_{2}^{2}\cong\:{\left|\right|{\varvec{\varPhi\:}}_{init,C}({\varvec{v}}_{i,comp}-{\varvec{E}}_{i,comp}^{*})\left|\right|}_{2}^{2}.$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\varvec{\varPhi\:}}_{init,C}\in\:{\varvec{R}}^{{\gamma\:}_{C}\times\:{n}_{C}}$$\end{document}. Further, we have19\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:D\left({\varvec{v}}_{i,comp}\right)\cong\:\frac{{\left|\right|{\varvec{\varPhi\:}}_{init,C}({\varvec{v}}_{i,comp}-{\varvec{E}}_{i,comp}^{*})\left|\right|}_{2}^{2}}{{n}_{C}}.$$\end{document} Thus, we can estimate the statistical characteristics of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\varvec{v}}_{i,comp}$$\end{document} in the compressed domain. Based on the estimated statistical characteristics, we can allocate an appropriate sampling rate to each block in the complex region:20\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{M}_{i,comp}={R}_{base}\cdot\:{n}_{C}+(1-{R}_{base})\cdot\:{n}_{C}\cdot\:{D\left({\varvec{v}}_{i,comp}\right)}^{*}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{M}_{i,comp}$$\end{document} is the sampling number of each block in the complex region,\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\:{R}_{base}=\:{{\gamma\:}_{C}/n}_{C}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{D\left({\varvec{v}}_{i,comp}\right)}^{*}=(D\left({\varvec{v}}_{i,comp}\right)-{D\left({\varvec{v}}_{i,comp}\right)}_{min})/({D\left({\varvec{v}}_{i,comp}\right)}_{max}-{D\left({\varvec{v}}_{i,comp}\right)}_{min})$$\end{document} is the normalized variance value, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{D\left({\varvec{v}}_{i,comp}\right)}_{min}$$\end{document} is the minimum value of the estimated variance, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{D\left({\varvec{v}}_{i,comp}\right)}_{max}$$\end{document} is the maximum value of the estimated variance. According to the above proposed scheme, the sampling rates of each block in the smooth region and the complex region are calculated respectively, and then the sampling matrices \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\varvec{\varPhi\:}}_{i,s}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\varvec{\varPhi\:}}_{i,c}$$\end{document} of the secondary sampling can be generated according to their respective sampling rates. Adaptive rate compressed sampling can be performed respectively according to the secondary sampling matrices. The sampled measurement results are transmitted to the reconstruction side, and then the original signal can be restored by using a reconstruction algorithm. To demonstrate the process of sampling rate estimation for the proposed ABCS scheme, we outline the corresponding pseudo-code in Algorithm 1. Fig. aAlgorithm 1 Algorithm 1 Simulation experiments and analysis In this section, we test the proposed ABCS scheme on standard test images commonly used in the field of adaptive rate compressed sampling (http://www.imageprocessingplace.com/root_files_V3/image_databases.htm, https://www.quinticsports.com/sample-videos/. visited on December 14, 2023). To facilitate objective comparison, we uniformly set the size of the test image to 512 × 512. The detailed experiment parameters setting are elaborated in Section IV.A. In Section IV.B, we provide a comprehensive comparison of the sampling performance among various ABCS schemes: ABCS-MC41, ABCS-SD15,ABCS-Entropy42, IEMD31,Zigzag EMD31, and analyze their respective reconstruction results. Experiment parameters setting For smooth region, the image block size is 8 × 8, while for complex region with more high-frequency components and main information, in order to avoid obvious blocking effect, the block size needs to be adjusted appropriately. However, increasing the block size leads to a higher computational load during sampling. To manage this load effectively, we set a block size of 16 × 16 for the complex region, where the block signal length \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{n}_{C}=16\:\times\:\:16$$\end{document}, which is common in image coding. In order to ensure an objective comparison, we adjusted the sampling rate of the compared schemes to be consistent with the proposed scheme. Since a high initial sampling number will increase the sampling burden of the sensor node, and this value should remain small relative to the total length of the signal, we set the value of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\gamma\:=n\cdot\:\sigma\:$$\end{document},\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\:{{\gamma\:}_{C}=n}_{C}\cdot\:\sigma\:$$\end{document}, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\sigma\:=0.3$$\end{document}. The threshold parameter \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\tau\:}_{L}$$\end{document} is set to 2. In CS systems, sparsity refers to the number of large values or the number of non-zero values in a signal. These large values usually contain the main information of the signal, while small values have lower information content. In this paper, we consider pixel values above the threshold \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\tau\:}_{L}$$\end{document} as larger values, and those below this threshold as smaller values. Although this threshold is set based on experience, it is also a commonly used threshold when processing grayscale images. Image reconstruction accuracy is assessed using the peak signal-to-noise ratio (PSNR) and the structural similarity index (SSIM). The simulation is conducted in MATLAB, using CPU 12th Gen Intel(R) Core(TM) i7-12700 H 2.30 GHz, RAM 16.0 GB. Simulation results of adaptive rate sampling In this section, we compare the performance of various state-of-the-art adaptive compressed sampling schemes, where in order to fairly compare the reconstruction accuracy, the reconstruction process of all schemes adopts the algorithm proposed in31. ABCS-MC starts by using preliminary measurement results to calculate the saliency value for each block within the compressed domain. Following this, it adapts the compressed sampling process based on these saliency values. ABCS-SD is an adaptive rate sampling method that uses an auxiliary imaging device. This scheme allocates block sampling rates based on the saliency values of a low-resolution image produced in advance by the auxiliary device. ABCS-Entropy is a scheme that allocates sampling rates based on the perceived entropy of each block signal in a reduced-dimensional space. This scheme is actually a scheme that allocates sampling rates based on signal complexity. ABCS-IEMD is an adaptive rate estimation scheme based on image empirical mode decomposition, which requires two-dimensional empirical mode decomposition in the compressed domain at the sampling side. ABCS-Zigzag EMD is an enhanced adaptive sampling method derived from empirical mode decomposition, capable of more precise rate allocation through one-dimensional empirical mode decomposition. It features lower complexity compared to ABCS-IEMD and delivers improved accuracy in image reconstruction. From the comparison data in Table 1, we can see that the PSNR of the images generated by the proposed adaptive scheme is significantly higher than that of all other schemes. This also further verifies that compared with other schemes, the proposed scheme can perform higher-accuracy sparsity estimation based on a small number of measurement results at the sampling side and perform higher-performance adaptive rate sampling. For example, for the test image Crowd, when the sampling rates of other schemes are set to the same as the sampling rate of the proposed scheme, the PSNR of the proposed adaptive scheme is higher than that of other comparative schemes. Specifically, it exceeds ABCS-MC by 2.48 dB, ABCS-SD by 2.45 dB, ABCS-Entropy by 6.88 dB, ABCS-IEMD by 0.65 dB, and ABCS-Zigzag EMD by 0.97 dB. Table 1Comparison of PSNR Across various ABCS schemes.V. ImagesAdaptive Sampling SchemesABCS-MCABCS-SDABCS-EntropyABCS-IEMDABCS-Zigzag EMDProposedRate/PSNR(dB)Rate/PSNR(dB)Rate/PSNR(dB)Rate/PSNR(dB)Rate/PSNR(dB)Rate/PSNR(dB)Fishing Boat0.42/33.350.42/33.270.42/31.870.42/32.870.42/32.800.42/33.59Clown0.38/34.240.38/34.510.38/31.220.38/34.390.38/34.450.38/35.12Crowd0.48/33.470.48/33.500.48/29.070.48/35.300.48/34.980.48/35.95Horse Jump(frame104)0.40/40.170.40/40.590.40/37.300.40/41.000.40/40.950.40/41.30100 m Start(frame001)0.45/37.230.45/35.740.45/33.270.45/36.840.45/36.260.45/37.44Bowler(frame342)0.37/35.360.37/35.410.37/33.590.37/35.450.37/35.640.37/35.76London 2012 Landing(frame120)0.40/33.280.40/33.290.40/30.970.40/34.070.40/34.080.40/34.69London 2012 Take Off (frame169)0.37/32.190.37/33.030.37/31.220.37/34.100.37/33.940.37/35.01 Table 1Comparison of PSNR Across various ABCS schemes.V. ImagesAdaptive Sampling SchemesABCS-MCABCS-SDABCS-EntropyABCS-IEMDABCS-Zigzag EMDProposedRate/PSNR(dB)Rate/PSNR(dB)Rate/PSNR(dB)Rate/PSNR(dB)Rate/PSNR(dB)Rate/PSNR(dB)Fishing Boat0.42/33.350.42/33.270.42/31.870.42/32.870.42/32.800.42/33.59Clown0.38/34.240.38/34.510.38/31.220.38/34.390.38/34.450.38/35.12Crowd0.48/33.470.48/33.500.48/29.070.48/35.300.48/34.980.48/35.95Horse Jump(frame104)0.40/40.170.40/40.590.40/37.300.40/41.000.40/40.950.40/41.30100 m Start(frame001)0.45/37.230.45/35.740.45/33.270.45/36.840.45/36.260.45/37.44Bowler(frame342)0.37/35.360.37/35.410.37/33.590.37/35.450.37/35.640.37/35.76London 2012 Landing(frame120)0.40/33.280.40/33.290.40/30.970.40/34.070.40/34.080.40/34.69London 2012 Take Off (frame169)0.37/32.190.37/33.030.37/31.220.37/34.100.37/33.940.37/35.01 Comparison of PSNR Across various ABCS schemes. In the comparison of SSIM in Table 2, we can also see that the proposed scheme is generally better than other adaptive schemes. For example, for the test image Bowler, the SSIM value of the proposed scheme is significantly higher than the other schemes. The above experiment results also further demonstrate the rationality and effectiveness of the proposed sampling rate allocation algorithm. Table 2Comparison of SSIM Across various ABCS schemes.VI. ImagesAdaptive Sampling SchemesABCS-MCABCS-SDABCS-EntropyABCS-IEMDABCS-Zigzag EMDProposedRate/SSIMRate/ SSIMRate/SSIMRate/ SSIMRate/ SSIMRate/ SSIMFishing Boat0.42/0.97210.42/0.96510.42/0.94100.42/0.97580.42/0.97490.42/0.9790Clown0.38/0.96560.38/0.96930.38/0.92950.38/0.97930.38/0.97930.38/0.9773Crowd0.48/0.98950.48/0.99020.48/0.96130.48/0.99510.48/0.99450.48/0.9952Horse Jump(frame104)0.40/0.99190.40/0.99350.40/0.97720.40/0.99570.40/0.99580.40/0.9948100 m Start(frame001)0.45/0.99390.45/0.99480.45/0.98060.45/0.99680.45/0.99630.45/0.9979Bowler(frame342)0.37/0.93710.37/0.94270.37/0.92260.37/0.95530.37/0.95720.37/0.9646London 2012 Landing(frame120)0.40/0.96330.40/0.96400.40/0.91490.40/0.98080.40/0.98060.40/0.9815London 2012 Take Off (frame169)0.37/0.94310.37/0.95520.37/0.90750.37/0.97640.37/0.97510.37/0.9779 Table 2Comparison of SSIM Across various ABCS schemes.VI. ImagesAdaptive Sampling SchemesABCS-MCABCS-SDABCS-EntropyABCS-IEMDABCS-Zigzag EMDProposedRate/SSIMRate/ SSIMRate/SSIMRate/ SSIMRate/ SSIMRate/ SSIMFishing Boat0.42/0.97210.42/0.96510.42/0.94100.42/0.97580.42/0.97490.42/0.9790Clown0.38/0.96560.38/0.96930.38/0.92950.38/0.97930.38/0.97930.38/0.9773Crowd0.48/0.98950.48/0.99020.48/0.96130.48/0.99510.48/0.99450.48/0.9952Horse Jump(frame104)0.40/0.99190.40/0.99350.40/0.97720.40/0.99570.40/0.99580.40/0.9948100 m Start(frame001)0.45/0.99390.45/0.99480.45/0.98060.45/0.99680.45/0.99630.45/0.9979Bowler(frame342)0.37/0.93710.37/0.94270.37/0.92260.37/0.95530.37/0.95720.37/0.9646London 2012 Landing(frame120)0.40/0.96330.40/0.96400.40/0.91490.40/0.98080.40/0.98060.40/0.9815London 2012 Take Off (frame169)0.37/0.94310.37/0.95520.37/0.90750.37/0.97640.37/0.97510.37/0.9779 Comparison of SSIM Across various ABCS schemes. In Figs. 3 and 4, we further compare the reconstruction accuracy of various adaptive sampling schemes through curve graphs, including both PSNR and SSIM. The comparison results clearly show that the scheme proposed in this paper achieves significant gains in PSNR and SSIM across various standard test images. For example, for the image Clown at a sampling rate of 0.38, the reconstructed image using the proposed scheme has a PSNR of 35.12, which is higher by 0.88 dB, 0.61 dB, 3.9 dB, 0.73 dB, and 0.67 dB respectively, compared to ABCS-MC, ABCS-SD, ABCS-Entropy, ABCS-IEMD, and ABCS-Zigzag EMD. Fig. 3PSNR Comparison of Various ABCS Schemes on different test images. PSNR Comparison of Various ABCS Schemes on different test images. Fig. 4SSIM Comparison of Various ABCS Schemes on different test images. SSIM Comparison of Various ABCS Schemes on different test images. In Fig. 5, we compare the visual quality of images generated by various state-of-the-art schemes in detail. Corresponding to the objective data comparisons mentioned above, the images generated by the scheme proposed in this paper have the best visual effects and are closest to the original images. Whether in terms of detail information or smoother areas, the proposed scheme significantly enhances clarity and noticeably reduces block artifacts in the generated images. For example, in the zoomed details of the test image Fishing Boat, the letters generated by the proposed scheme are clearer and closer to the original image compared to other schemes. Similarly, for the test image Clown, images generated by the compared schemes exhibit more ghosting and block effects due to inappropriate sampling rate allocation. In contrast, the image generated by the proposed scheme have more completely reconstructed details with almost no block effects. These experiment results further demonstrate that the adaptive rate allocation scheme proposed in this paper can assign a more suitable sampling rate to each block signal, thereby significantly improving the sampling performance of each sensor node. Fig. 5Comparison of visual quality for various ABCS Schemes. From left to right, the images are arranged as follows: the original image, ABCS-MC, ABCS-SD, ABCS-Entropy, ABCS-IEMD, ABCS-Zigzag EMD, and the proposed adaptive sampling scheme. The rows are labeled as follows: (a) Fishing Boat, (b) Clown, (c) Crowd, (d) Horse Jump(frame104), (e) 100 m Start(frame001), (f) Bowler(frame342), (g) London 2012 Landing(frame120), (h) London 2012 Take Off (frame169). Comparison of visual quality for various ABCS Schemes. From left to right, the images are arranged as follows: the original image, ABCS-MC, ABCS-SD, ABCS-Entropy, ABCS-IEMD, ABCS-Zigzag EMD, and the proposed adaptive sampling scheme. The rows are labeled as follows: (a) Fishing Boat, (b) Clown, (c) Crowd, (d) Horse Jump(frame104), (e) 100 m Start(frame001), (f) Bowler(frame342), (g) London 2012 Landing(frame120), (h) London 2012 Take Off (frame169). Conclusion In this paper, we propose an ABCS scheme based on image region division to address the issue of sampling resource allocation at nodes in WSNs. The proposed scheme is capable of estimating the complexity of the signal to be collected based solely on the measurement results of the signal, without pixel knowledge of the original image signal, and then allocating a matching sampling rate for different signals. Extensive experiment results validate the effectiveness of the proposed scheme. Due to the effectiveness of the proposed adaptive rate estimation scheme, it can accurately estimate the required sampling rate for the signal to be sampled and achieve higher quality image signal reconstruction with the same total sampling resource consumption, effectively removing the commonly present blocking effects in block-based image compressed sampling. In future research, we will apply the proposed scheme to real-time multimedia systems in WSNs to improve the irrational allocation of sampling resources.