Image data compression methods, apparatus, electronic devices and storage media

Abstract

The image data compression method, apparatus, electronic device, and storage medium proposed in this application include: acquiring image data to be compressed and acquiring a measurement matrix; performing linear measurements on the image data to be compressed based on the measurement matrix to obtain initial global observation data and initial local observation data for multiple local regions; selecting sensing regions one by one from the multiple local regions, obtaining row-layer associated regions of the sensing regions based on row-layer direction, and obtaining column-block consensus regions of the sensing regions based on column-block direction; updating the sensing local observation data based on all row-layer local observation data and obtaining local estimation data, and performing consensus updates until all sensing local observation data updates are completed; and performing global refinement processing based on all local estimation data to obtain image compressed data corresponding to the image data to be compressed, ensuring both computational efficiency and compression accuracy during image data compression.

Description

Technical Field

[0001] This application relates to the field of image data processing technology, and in particular to image data compression methods, apparatus, electronic devices and storage media. Background Technology

[0002] With the widespread adoption of applications such as streaming media, social media, and online education, image and video data have become a dominant component of network traffic. This makes image compression technology a key pillar in balancing data volume and transmission smoothness. In recent years, compressed sensing (CS) theory has become a popular and widely applied technique in the field of image compression. This technique utilizes the sparsity of image signals in a transform domain (such as the wavelet transform domain) to obtain low-dimensional observation data by linearly measuring the image at a frequency much lower than the Nyquist sampling rate using a measurement matrix. Related techniques typically employ various reconstruction algorithms, such as convex optimization algorithms, soft / hard thresholding algorithms, or matching pursuit algorithms, to solve for sparse solutions from this observation data, and then recover the compressed image through inverse transformation.

[0003] However, existing image compression algorithms based on compressed sensing (CS) suffer from common drawbacks. First, they generally neglect the structural design of the measurement matrix, treating it as a single unit in product operations regardless of the matrix used. This approach fails to utilize the internal structural characteristics of the matrix, resulting in high computational costs when processing large-scale images. Second, when the number of measurements (i.e., the sampling rate) is increased to improve compression quality, the expanded problem size significantly increases the computational cost of existing algorithms. Therefore, current image data compression techniques cannot guarantee both computational efficiency and compression accuracy. Summary of the Invention

[0004] This application provides an image data compression method, apparatus, electronic device, and storage medium that can ensure both computational performance and compression accuracy when compressing image data.

[0005] To achieve the above objectives, a first aspect of this application proposes an image data compression method, the method comprising:

[0006] The image data to be compressed is obtained, as well as the measurement matrix, which is generated by scaling and stitching together multiple orthogonal matrices.

[0007] Linear measurements are performed on the image data to be compressed based on the measurement matrix to obtain initial global observation data. Based on the initial global observation data, initial local observation data of multiple local regions in the observation area are obtained, and each local region corresponds to each measurement submatrix in the measurement matrix.

[0008] A perception region is selected one by one from the multiple local regions. The row-level associated region corresponding to the perception region is obtained from the local regions based on the row-level direction, and the column-block consensus region corresponding to the perception region is obtained from the local regions based on the column-block direction.

[0009] The sensing local observation data is updated based on all row-level local observation data, and the local estimation data corresponding to the sensing region is obtained based on the updated sensing local observation data. Consensus updates are performed on all column block local observation data until all sensing local observation data updates are completed. The row-level local observation data is the initial local observation data of the row-level associated region, the sensing local observation data is the initial local observation data of the sensing region, and the column block local observation data is the initial local observation data of the column block consensus region.

[0010] Based on all the local estimation data, a global refinement process is performed to obtain the image compression data corresponding to the image data to be compressed.

[0011] In some embodiments, obtaining the measurement matrix includes:

[0012] Obtain multiple orthogonal matrices, and obtain a first scaling factor and a second scaling factor;

[0013] Based on the region where the sensing region is located, a first scaling matrix is ​​selected from a plurality of orthogonal matrices, and the other orthogonal matrices are used as a second scaling matrix;

[0014] Multiply all the first scaling matrices by the first scaling factor to obtain the first measurement submatrix, and multiply all the second scaling matrices by the second scaling factor to obtain the second measurement submatrix;

[0015] The measurement matrix is ​​obtained by concatenating all the first measurement sub-matrices and all the second measurement sub-matrices.

[0016] In some embodiments, the step of performing linear measurements on the image data to be compressed based on the measurement matrix to obtain initial global observation data includes:

[0017] The image data to be compressed is subjected to an orthogonal linear transformation to obtain a sparse representation vector;

[0018] The initial global observation data are obtained by multiplying the sparse representation vector based on the measurement matrix.

[0019] In some embodiments, selecting the sensing region from the plurality of local regions one by one includes:

[0020] Divide the observation area into squares to obtain at least one divided observation sub-region;

[0021] According to the order of dividing the observation sub-regions, in each of the divided observation sub-regions, the local area on the diagonal is selected one by one as the sensing area.

[0022] In some embodiments, updating the sensing local observation data based on all row-level local observation data includes:

[0023] Based on the accumulated value of all the row-level local observation data, the local accumulated value is obtained;

[0024] Based on the difference between the initial global observation data and the local accumulated value, updated local observation data is obtained, and the updated local observation data is used as the new local observation data for the sensing region.

[0025] In some embodiments, obtaining local estimation data corresponding to the sensing region based on the updated sensing local observation data, and performing consensus update on the local observation data of all columns, includes:

[0026] The initial local estimation data is obtained by multiplying the transpose of the first measurement submatrix corresponding to the sensing region in the measurement matrix with the updated sensing local observation data, and then dividing by the square of the first scaling factor.

[0027] Thresholding is performed on the initial local estimation data to obtain the local estimation data;

[0028] Multiply the first measurement sub-matrix corresponding to the sensing area with the local estimation data to obtain the first updated local observation data, and use the first updated local observation data as the new sensing local observation data;

[0029] The second measurement submatrix corresponding to each column block consensus region is multiplied with the local estimated data to obtain the second updated local observation data, and the second updated local observation data is used as the new column block local observation data.

[0030] In some embodiments, the step of performing global refinement based on all the local estimation data to obtain image compressed data corresponding to the image data to be compressed includes:

[0031] All the local estimation data are concatenated in column block order to obtain a sparse estimation vector;

[0032] The first index set is obtained by selecting multiple local estimation data with the largest absolute value from the sparse estimation vector;

[0033] Multiple measurement sub-matrices corresponding to the first indicator set are selected from the measurement matrix to obtain the first indicator measurement matrix;

[0034] Solving the least squares model of the first index yields a wide-domain refined value. The least squares model of the first index is obtained based on the product of the transpose of the measurement matrix of the first index and the initial global observation data.

[0035] The image compression data is obtained based on the wide-area refinement value.

[0036] In some embodiments, obtaining the image compressed data based on the wide-area refinement value includes:

[0037] Based on multiple wide-domain refined values, a wide-domain vector is obtained, and multiple wide-domain refined values ​​with the largest absolute values ​​are selected from the wide-domain vector to obtain a second index set;

[0038] Multiple measurement sub-matrices corresponding to the second indicator set are selected from the measurement matrix to obtain the second indicator measurement matrix;

[0039] Solving the least squares model of the second index yields a narrow-domain refined value. The least squares model of the second index is based on the product of the transpose of the measurement matrix of the second index and the initial global observation data.

[0040] Based on all the aforementioned narrow-domain refined values, a sparse image vector is obtained;

[0041] The sparse image vector is subjected to an inverse orthogonal transform to obtain the compressed image data.

[0042] To achieve the above objectives, a second aspect of this application provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the image data compression method as described in the first aspect.

[0043] To achieve the above objectives, a third aspect of the present application provides a storage medium, which is a computer-readable storage medium storing a computer program that, when executed by a processor, implements the image data compression method as described in the first aspect.

[0044] The image data compression method, apparatus, electronic device, and storage medium proposed in this application include: First, acquiring image data to be compressed and acquiring a measurement matrix, wherein the measurement matrix is ​​generated by scaling and stitching multiple orthogonal matrices; then, performing linear measurements on the image data to be compressed based on the measurement matrix to obtain initial global observation data, and based on the initial global observation data, obtaining initial local observation data for multiple local regions within the observation area, wherein each local region corresponds to each measurement submatrix in the measurement matrix; subsequently, selecting a sensing region from the multiple local regions one by one, obtaining the row-layer associated region corresponding to the sensing region from the local regions based on the row-layer direction, and obtaining the local regions based on the column-block direction. The first step is to obtain the column block consensus region corresponding to the perceived region within the domain. Next, the perceived local observation data is updated based on all row-level local observation data. Based on the updated perceived local observation data, local estimation data corresponding to the perceived region is obtained, and consensus updates are performed on all column block local observation data until all perceived local observation data updates are complete. The row-level local observation data serves as the initial local observation data for the row-level associated region, the perceived local observation data serves as the initial local observation data for the perceived region, and the column block local observation data serves as the initial local observation data for the column block consensus region. Finally, based on all local estimation data, global refinement processing is performed to obtain the image compression data corresponding to the image data to be compressed. This application's embodiments design the measurement matrix as a structured matrix composed of multiple small scaled orthogonal matrices, and decompose large-scale computation into local sparse sensing and information consensus. The method of this application only performs the core sparse sensing operation on the selected "sensing region" one by one to obtain local estimation data, and uses the update mechanism of the "column block consensus region" to efficiently integrate the estimation data. This completely replaces the high-cost method of participating in the product operation of the entire measurement matrix in the prior art, significantly reducing the computational complexity. More importantly, when the number of measurements is increased to improve the compression quality (e.g., by increasing the row layer of the matrix), the number of core sensing calculations in this application remains unchanged, avoiding the surge in computational cost caused by the expansion of the problem scale in the prior art. Finally, by performing global refinement processing on all local estimation data, the solution of this application obtains high-quality image compressed data while achieving high computational efficiency and high stability when compressing image data, effectively overcoming the defects of the prior art.

[0045] Other features and advantages of this application will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the application. The objectives and other advantages of this application may be realized and obtained by means of the structures particularly pointed out in the description, claims and drawings. Attached Figure Description

[0046] Figure 1 This is a flowchart of an image data compression method provided in an embodiment of this application.

[0047] Figure 2 yes Figure 1 The flowchart for step 101.

[0048] Figure 3 This is a schematic diagram of a plurality of first scaling matrices and a plurality of second scaling matrices provided in another embodiment of this application.

[0049] Figure 4 This is a schematic diagram of a measurement matrix splicing mode provided in another embodiment of this application.

[0050] Figure 5 This is a schematic diagram of an extended form of the measurement matrix after increasing the sampling amount, provided in another embodiment of this application.

[0051] Figure 6 yes Figure 1 The flowchart for step 102.

[0052] Figure 7 yes Figure 1 The flowchart for step 103.

[0053] Figure 8 This is a schematic diagram of selecting a sensing area provided in another embodiment of this application.

[0054] Figure 9 yes Figure 1 The flowchart for step 104.

[0055] Figure 10 yes Figure 1 Another flowchart for step 104.

[0056] Figure 11 yes Figure 1 The flowchart for step 105.

[0057] Figure 12 yes Figure 11 The flowchart for step 1105.

[0058] Figure 13 This is a flowchart illustrating an image data compression method provided in an embodiment of this application.

[0059] Figure 14 This is a schematic diagram of another plurality of first scaling matrices and a plurality of second scaling matrices provided in an embodiment of this application.

[0060] Figure 15 This is a schematic diagram of another method for selecting a sensing area according to an embodiment of this application.

[0061] Figure 16 This is a schematic diagram of the hardware structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0062] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0063] It should be noted that although functional modules are divided in the device schematic diagram and the logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than the module division in the device or the order in the flowchart.

[0064] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0065] With the widespread adoption of applications such as streaming media, social media, and online education, image and video data have become a dominant component of network traffic. This makes image compression technology a key pillar in balancing data volume and transmission smoothness. In recent years, compressed sensing (CS) theory has become a popular and widely applied technique in the field of image compression. This technique utilizes the sparsity of image signals in a transform domain (such as the wavelet transform domain) to obtain low-dimensional observation data by linearly measuring the image at a frequency much lower than the Nyquist sampling rate using a measurement matrix. Related techniques typically employ various reconstruction algorithms, such as convex optimization algorithms, soft / hard thresholding algorithms, or matching pursuit algorithms, to solve for sparse solutions from this observation data, and then recover the compressed image through inverse transformation.

[0066] However, existing image compression algorithms based on compressed sensing (CS) suffer from common drawbacks. First, they generally neglect the structural design of the measurement matrix, treating it as a single unit in product operations regardless of the matrix used. This approach fails to utilize the internal structural characteristics of the matrix, resulting in high computational costs when processing large-scale images. Second, when the number of measurements (i.e., the sampling rate) is increased to improve compression quality, the expanded problem size significantly increases the computational cost of existing algorithms. Therefore, current image data compression techniques cannot guarantee both computational efficiency and compression accuracy.

[0067] To ensure both computational efficiency and compression accuracy during image data compression, this application's embodiments design the measurement matrix as a structured matrix composed of multiple small scaled orthogonal matrices. Large-scale computation is decomposed into local sparse sensing and information consensus. The method of this application performs core sparse sensing operations only on selected "sensing regions" one by one to obtain local estimation data, and utilizes the update mechanism of "column block consensus regions" to efficiently integrate this estimation data. This completely replaces the high-cost method in existing technologies that involves the entire measurement matrix in product operations, significantly reducing computational complexity. More importantly, when the number of measurements is increased to improve compression quality (e.g., by increasing the row layers of the matrix), the number of core sensing calculations in this application remains unchanged, avoiding the surge in computational costs caused by the expansion of the problem scale in existing technologies. Finally, by performing global refinement processing on all local estimation data, the solution of this application achieves high computational efficiency and high stability while obtaining high-quality compressed image data, effectively overcoming the shortcomings of existing technologies.

[0068] First, let's explain the symbols: u ← z means assigning the value of z to u, where u and z are column vectors of the same dimension; Describes the transpose of matrix D. The matrix D represents the inverse of matrix D; W represents a certain orthogonal linear transformation. This represents the inverse transform of W. This means retaining the K components with the largest absolute values ​​of vector u, and setting the remaining components to 0. Let u be a vector of the same dimension and contain at most K non-zero elements; suppose a matrix D is divided into q rows and p columns, each sub-block being an m x m matrix, then let [i,j] denote the region corresponding to the i-th row and j-th column block, and let... This represents the submatrix corresponding to the region [i,j], where i=1, ...,q; j=1, ...,p; express A column vector u of dimension L consists of L segments It is pieced together column by column, with each segment All are m-dimensional column vectors.

[0069] Understandably, the sparsity of an image is primarily determined by the absolute values ​​of its coefficients on the image representation basis (e.g., wavelet basis). If the absolute values ​​of the vast majority of coefficients are small, and only a small number of coefficients have large absolute values, then the image can be called a sparse image. Compression speed refers to the running time required for a compression algorithm to complete image compression; the shorter the time, the faster the compression speed. Compression quality refers to the difference between the compressed image and the original image; the smaller the difference, the higher the compression quality.

[0070] The image data compression method, apparatus, electronic device, and storage medium in the embodiments of this application will be described in detail below. First, the image data compression method in the embodiments of this application will be described. (Refer to...) Figure 1 This is an optional flowchart of the image data compression method provided in the embodiments of this application. Figure 1 The method may include, but is not limited to, steps 101 to 105. It is also understood that this embodiment... Figure 1 The order of steps 101 to 105 is not specifically limited; the order of steps can be adjusted or certain steps can be added or removed according to actual needs. The image data compression method provided in this application can be applied to any control system (such as a server, computer, etc.) with computing resources connected to the main communication channel.

[0071] Step 101: Obtain the image data to be compressed and the measurement matrix. The measurement matrix is ​​generated by scaling and stitching multiple orthogonal matrices.

[0072] Step 101 will be described in detail below.

[0073] In some embodiments, image data to be compressed is first acquired, and a measurement matrix is ​​acquired, wherein the measurement matrix is ​​generated by scaling and stitching together multiple orthogonal matrices. Specifically, the "image data to be compressed" can refer to the original image data X, or more specifically, to the sparse representation vector x obtained after the image data X undergoes some orthogonal linear transformation (such as DCT or DWT). The "measurement matrix" A is a large matrix with a specific internal structure.

[0074] In this application, the measurement matrix is ​​required to have at least two rows, where each sub-block is an m-row, m-column scaled orthogonal matrix. Without loss of generality, the implementation process of the algorithm is described below using a 3-row, 9-column block (i.e., q=3, p=9) example (Note: In practical applications, the value of m can be given according to the user's needs, and then the number of columns p=n / m can be determined).

[0075] The following section will further describe how to generate this measurement matrix.

[0076] Reference Figure 2 To obtain the measurement matrix, the steps 201 to 204 are as follows.

[0077] Step 201: Obtain multiple orthogonal matrices, and obtain the first scaling factor and the second scaling factor.

[0078] Step 202: Based on the region where the sensing region is located, select the first scaling matrix from multiple orthogonal matrices, and use the other orthogonal matrices as the second scaling matrix.

[0079] Step 203: Multiply all first scaling matrices by the first scaling factor to obtain the first measurement submatrix, and multiply all second scaling matrices by the second scaling factor to obtain the second measurement submatrix.

[0080] Step 204: Concatenate all the first measurement sub-matrices and all the second measurement sub-matrices to obtain the measurement matrix.

[0081] Steps 201 to 204 are described in detail below.

[0082] In some embodiments, multiple orthogonal matrices are first obtained, along with a first scaling factor and a second scaling factor. This step is a preparatory stage for constructing the measurement matrix. First, "multiple orthogonal matrices" refers to a series (e.g., q*p matrices given a value of m) of small size and identical dimensions (e.g., m rows and m columns). B i,j , i=1,2,3 ; j=1, ...,9 These matrices all satisfy the mathematical property of "orthogonality" (i.e., their transpose is equal to their inverse). They can be selected from Discrete Cosine Transform (DCT) matrices, Discrete Wavelet Transform (DWT) matrices, or generated by Gram-Schmidt orthogonalization of random matrices (but the same matrix blocks cannot be reused). The above matrix blocks are concatenated into a d-row, n-column (d=3m, n=9m) matrix B (3 rows, 9 columns) as shown below.

[0083]

[0084] Additionally, two specific positive values ​​need to be obtained: the "first scaling factor" α and the "second scaling factor" β. These two factors are used for subsequent weighted scaling of the orthogonal matrix, and they typically need to satisfy specific constraints (e.g., α 2 + ( q -1) β 2 =1 Where q is the row level, for example: α=1 / 3, β=2 / 3).

[0085] Next, based on the region where the sensing region is located (the sensing region is selected by selecting sub-matrices located in different row layers and different column blocks, such as alternating selection of row layers 1, 2, and 3, and column blocks increasing sequentially from 1 to 9, as explained in subsequent steps), a first scaling matrix is ​​selected from multiple orthogonal matrices, and the other orthogonal matrices are used as the second scaling matrix. The "sensing region" refers to a specific set of regions pre-selected in the logical layout of the measurement matrix for performing the core sparse sensing operation (e.g., regions [1,1], [2,2], [3,3]... selected alternately according to different row layers and different column blocks). "Based on the region where the sensing region is located," the orthogonal matrices corresponding to the "sensing region" (e.g., ...) are... B 1,1 , B 2,2 , B 3,3 , B 1,4 , B 2,5 , B 3,6 , B 1,7 , B 2,8 , B 3,9 The selected matrices are categorized as the "first scaling matrix," while all other orthogonal matrices not selected as "sensory regions" are categorized as "other." It is then classified as a "second scaling matrix".

[0086] Reference Figure 3 This is a schematic diagram of multiple first scaling matrices and multiple second scaling matrices provided in an embodiment of this application. Figure 3 The image shows a set of (3*9=27) m-row, m-column "orthogonal matrices". They are arranged into 3 rows and 9 columns. In this diagram, the "sensing area" corresponds to the 9 orthogonal matrices highlighted in blue. B 1,1 , B 2,2 , B 3,3 , B 1,4 , B 2,5 , B 3,6 , B 1,7 , B 2,8 , B 3,9These are selected as the "first scaling matrix"; while the orthogonal matrices corresponding to all the unhighlighted white areas in the diagram (such as...) B 1,2 , B 1,3 , B 2,1 (etc.) are then classified as "second scaling matrix".

[0087] Next, all first scaling matrices are multiplied by the first scaling factor to obtain the first measurement submatrix, and all second scaling matrices are multiplied by the second scaling factor to obtain the second measurement submatrix. This step performs the actual "scaling" operation. The obtained "first scaling factor" α is multiplied by the selected "all first scaling matrices" (i.e., the matrices located in the sensing region), and the result of the operation ( A i,j= αB i,j This is the "first measurement sub-matrix". Similarly, multiplying the "second scaling factor" β by "all second scaling matrices" (i.e., matrices not located in the sensing region) yields the result ( A i,j= βB i,j This is the "second measurement sub-matrix". After this step, all orthogonal matrices... They were all converted into "scaled orthogonal matrices". .

[0088] Finally, the measurement matrix is ​​obtained by concatenating all the first and second measurement sub-matrices. This step generates both the "all first measurement sub-matrices" and the "all second measurement sub-matrices". They are arranged and pieced together in an orderly manner according to their original logical layout (e.g., q rows and p columns). The final combination forms a large-scale, structured matrix of d rows and n columns (d=m*q, n=m*p), which is the "measurement matrix" A shown below.

[0089]

[0090] Reference Figure 4 This is a schematic diagram of a measurement matrix splicing mode provided in an embodiment of this application. For example... Figure 4 The diagram illustrates one possible concatenation pattern for a "measurement matrix." This "measurement matrix" is "obtained" by concatenating all first and second measurement sub-matrices. In this schematic, the "measurement matrix" is constructed as a structure with 2 "row layers" and 12 "column blocks," containing a total of 24 measurement sub-matrices. (Where i=1,2; j=1,...,12). These They are all scaled orthogonal matrices with m rows and m columns, and they are specific instances of the generated "first measurement submatrix" and "second measurement submatrix".

[0091] Reference Figure 5 This is a schematic diagram illustrating an extended form of the measurement matrix after increasing the sampling amount, as provided in an embodiment of this application. For example... Figure 5 The diagram illustrates an extended form of the "measurement matrix" when increasing the image sampling rate. The figure shows a "measurement matrix" consisting of 3 "row layers" and 12 "column blocks." The portion labeled "original measurement matrix" (rows 1 and 2) corresponds to a basic measurement matrix (as shown in the attached diagram). Figure 4 As shown); while the third row of layers marked "Add Measurement Section" (as shown) A 3,j The measurement submatrix (shown in gray shading) is added to improve compression quality. This figure visually illustrates the flexibility of this technical solution, allowing for the addition of "row layers" to increase image sampling data. According to the technical solution, this addition does not significantly increase the core perception computation cost because the number of subproblems requiring a solution (i.e., the number of "perception regions") remains unchanged.

[0092] Through steps 201 to 204 above, a structured method for designing and generating a "measurement matrix" is provided. This method abandons the traditional approach of treating the measurement matrix as an undifferentiated whole and instead introduces the concept of a "perceptual region," dividing the matrix into two types of "orthogonal matrices" with different functions (perceptual and non-perceptual). Then, by using a "first scaling factor" and a "second scaling factor," these two types of matrices are scaled differently to generate a "first measurement sub-matrix" and a "second measurement sub-matrix." This strategy aims to effectively reduce the mutual coherence of the overall measurement matrix. The structured and differentiated design of the final spliced ​​measurement matrix gives it a natural ability to perceive local sparseness in images, making it possible to decompose the large-scale compression problem into low-cost computation of "local perception" and "column consensus."

[0093] Step 102: Perform linear measurements on the image data to be compressed based on the measurement matrix to obtain initial global observation data, and based on the initial global observation data, obtain initial local observation data for multiple local regions in the observation area, with each local region corresponding to each measurement submatrix in the measurement matrix.

[0094] Step 102 is described in detail below.

[0095] Next, linear measurements are performed on the image data X to be compressed based on the measurement matrix A to obtain initial global observation data. Based on this initial global observation data, initial local observation data for multiple local regions within the observation area are obtained, with each local region corresponding to a measurement submatrix in the measurement matrix. Here, "linear measurement" refers to multiplying the generated "measurement matrix" A with the "image data X to be compressed" (or the sparse vector x) to obtain low-dimensional "initial global observation data" y (i.e., y≈Ax). This y vector is "global." For subsequent iterative calculations, initial local observation data needs to be assigned to the "multiple local regions" [i,j] within the "observation area." Here, "local region" [i,j] is a logical concept, corresponding to each measurement submatrix in the "measurement matrix." A i,j One-to-one correspondence. Since A is a concatenation of 3-row matrix blocks, y can also be viewed as 3 m-dimensional column vectors. y 1, y 2, y 3 is concatenated column by column, that is y = ( y 1, y 2, y 3). These can be viewed as the observation data corresponding to the 1st, 2nd, and 3rd rows of matrix A, as shown below.

[0096]

[0097] For each local region [i,j] (i=1,2,3; j=1, ..., 9), define y i,j =A i,j x j and treat it as a local image. x j Scaling orthogonal matrices A i,j The following observational data. Therefore, y 1, y 2, y 3 represents the linear superposition of the observation data from the local images of rows 1, 2, and 3, respectively. y i = y i,1 + y i,2 + ... + y i,9 , i=1, 2, 3 .

[0098] Because at the start of the iteration, precise y i,jSince it is unknown, it cannot be solved by directly solving the local subproblem. y i,j =A i,j x j Obtain the x j The estimation of principal components (this process is called the "sparse sensing" problem of the local region). Therefore, this step will set an initial value "based on...initial global observation data" y, for example, the data of the i-th row layer in the y vector. y i The p "local regions" evenly distributed across this row layer are shown below.

[0099]

[0100] The following section will further describe how to perform linear measurements on the image data to be compressed based on the measurement matrix to obtain initial global observation data.

[0101] Reference Figure 6 The initial global observation data is obtained by performing linear measurements on the image data to be compressed based on the measurement matrix, including the following steps 601 to 602.

[0102] Step 601: Perform an orthogonal linear transformation on the image data to be compressed to obtain a sparse representation vector.

[0103] Step 602: Perform a product operation on the sparse representation vector based on the measurement matrix to obtain the initial global observation data.

[0104] Steps 601 to 602 are described in detail below.

[0105] In some embodiments, an orthogonal linear transformation W is performed on the image data X to be compressed to obtain a sparse representation vector x. This step preprocesses the original "image data to be compressed" (i.e., the high-dimensional image vector X) to utilize its inherent sparsity. Since image data is generally not sparse in the spatial domain but exhibits sparsity in other transform domains (such as the wavelet domain or frequency domain), this step employs an "orthogonal linear transformation" W (e.g., Discrete Cosine Transform DCT or Discrete Wavelet Transform DWT) to transform the image vector X (i.e., x = W(X)). The resulting "sparse representation vector" x has most coefficients approaching zero, with only a small number of significant non-zero coefficients. This sparsity forms the mathematical basis for subsequent efficient measurements using compressed sensing theory.

[0106] Then, the sparse representation vector is multiplied based on the measurement matrix to obtain the initial global observation data. This step is the core "linear measurement" process of compressed sensing. This step utilizes the structured "measurement matrix" A with d rows and n columns generated in the previous step (where the number of measurements d is much smaller than the dimension n of the "sparse representation vector" x). By performing the "product operation" (i.e., y ≈ Ax), the algorithm collects information from the high-dimensional "sparse representation vector" x at a rate much lower than the Nyquist sampling rate. The resulting d-dimensional vector y is the "initial global observation data". This low-dimensional y vector contains the main information needed to reconstruct the original image and is the input source for all subsequent local sensing and consensus iterative calculations.

[0107] Through steps 601 and 602 above, the original "image data to be compressed" is converted into a "sparse representation vector" through an "orthogonal linear transformation." This utilizes the sparsity of the signal, which is a prerequisite for applying compressed sensing theory. Then, "based on" this sparsity, a "product operation" is performed using a structured "measurement matrix," thereby efficiently obtaining low-dimensional "initial global observation data" y from the high-dimensional sparse vector x with far fewer measurements than the Nyquist sampling rate. This process achieves efficient "linear measurement" of image data, ensuring information acquisition while laying the data foundation for subsequent execution of low-cost, high-precision "local sensing collaborative algorithms."

[0108] Step 103: Select the perception region from multiple local regions one by one, obtain the row-level associated region corresponding to the perception region from the local regions based on the row-level direction, and obtain the column-block consensus region corresponding to the perception region from the local regions based on the column-block direction.

[0109] Step 103 will be described in detail below.

[0110] Furthermore, sensing regions are selected sequentially from multiple local regions. Based on the row-level direction, the row-level associated regions corresponding to the sensing regions are obtained from the local regions, and based on the column-block direction, the column-block consensus regions corresponding to the sensing regions are obtained from the local regions. This step defines the path of iterative computation and the data collaboration relationship. First, in this algorithm, core computation is not performed on all "local regions," but rather a portion of them are selected sequentially as "sensing regions." The selection rule for "sensing regions" is (e.g.) "alternating selection of different row levels and different column blocks" to ensure the uniformity and coverage of computation (e.g., selection in the order of [1,1]→[2,2]→[3,3]→[1,4]...). For the currently selected "sensing region" (assumed to be [i,j]), the "row-level associated region" defined "based on the row-level direction" refers to all other local regions located in the same "row level" i as this sensing region (e.g., [i,k], k≠j). The "column consensus region" defined by "column block direction" refers to all other local regions located in the same "column block" j as the perception region (e.g., [k,j], k≠i).

[0111] The following section will further describe how to select the sensing regions one by one.

[0112] Reference Figure 7 The process involves selecting a sensing region from multiple local regions one by one, including the following steps 701 to 702.

[0113] Step 701: Divide multiple local regions in the observation area into squares to obtain at least one sub-region of observation.

[0114] Step 702: According to the order of dividing the observation sub-regions, select the local area on the diagonal of each divided observation sub-region as the sensing area.

[0115] Steps 701 to 702 are described in detail below.

[0116] In some embodiments, when selecting the sensing region, multiple local regions within the observation region can be first divided into squares to obtain at least one sub-observation region. This step logically groups the macroscopic layout of the "observation region." The "observation region" refers to a complete logical grid (e.g., a q-row, p-column grid) composed of all "local regions" [i,j]. This step "divides" it into squares; for example, if the "observation region" is a 3-row, 9-column grid, it can be logically "divided" into three consecutive 3*3 sub-observation regions. This "division" itself is not a physical segmentation, but a logical grouping method for subsequent orderly and uniform selection.

[0117] Then, according to the order in which the observation sub-regions are divided, local regions on the diagonal are selected one by one as the sensing regions within each divided observation sub-region. This step is based on logical grouping to perform the specific "sensing region" selection operation. The algorithm will perform the selection "in each... sub-region" according to the "order in which the observation sub-regions are divided" (e.g., starting from the first 3*3 sub-region, then the second, and so on). The selection rule is "selecting local regions on the diagonal one by one": for example, in the first 3*3 "divided observation sub-region", local regions [1,1], [2,2], [3,3] on its main diagonal are selected; then, in the second 3*3 "divided observation sub-region" (corresponding to columns 4 to 6 of the original grid), local regions [1,4], [2,5], [3,6] on its main diagonal are selected. All "local regions" selected in this way are determined as "sensing regions", and they will perform the core sparse sensing operation in subsequent steps.

[0118] Reference Figure 8 This is a schematic diagram illustrating the selection of a sensing region according to an embodiment of this application. Figure 8 As shown, the "observation area" consists of "multiple local regions" (in this example, 3 rows and 9 columns). This "observation area" is "divided into squares," logically resulting in three 3*3 "divided observation sub-regions" (i.e., columns 1-3, 4-6, and 7-9). Subsequently, according to the order of dividing the observation sub-regions, "in each divided observation sub-region," "local regions on the diagonal are selected one by one as sensing regions." As shown in the blue highlighted area in the figure, the selected "sensing regions" are [1,1], [2,2], [3,3], [1,4], [2,5], [3,6], [1,7], [2,8], [3,9]. The red arrows in the figure further indicate the execution order of "selecting one by one," and these selected "sensing regions" will perform the core "sparse sensing" operation in subsequent steps.

[0119] Understandably, if the sensing area cannot be precisely divided into multiple rectangular regions, a few columns of 0 elements can be added at the end to make it precisely divided into multiple rectangular regions, thereby enabling the selection of the sensing area.

[0120] Through steps 701 and 702 above, a systematic, low-cost, and highly efficient strategy for "selecting the sensing region" is provided. The combination of "square division" and "selecting diagonals one by one" defines a specific selection rule, which essentially ensures that the selected "sensing region" is evenly distributed across "different row layers" and "different column blocks" (i.e., alternating selection mode) within the macroscopic grid of the "observation region." This strategy brings dual advantages: First, it achieves "uniform and fair" coverage of the entire "observation region," avoiding computational bias; second, it "greatly reduces the number of sensing nodes," that is, it selects only p (number of column blocks) "local regions" as the "sensing region" to perform core operations, without needing to calculate all p*q regions, realizing the "low-cost computation" design concept of this application, and significantly reducing the computational complexity of the algorithm while ensuring sensing coverage.

[0121] Step 104: Update the sensing local observation data based on all row-level local observation data, obtain the local estimation data corresponding to the sensing region based on the updated sensing local observation data, and perform consensus update on all column block local observation data until all sensing local observation data updates are completed. The row-level local observation data is the initial local observation data of the row-level associated region, the sensing local observation data is the initial local observation data of the sensing region, and the column block local observation data is the initial local observation data of the column block consensus region.

[0122] Step 104 is described in detail below.

[0123] Furthermore, according to the selected sensing regions in order, for each sensing region, the sensing local observation data is updated based on all row-level local observation data. Based on the updated sensing local observation data, the corresponding local estimation data for the sensing region is obtained, and consensus updates are performed on all column-block local observation data until all sensing local observation data updates are completed. This step is the core iterative computation loop. For the selected current "sensing region" [i,j], firstly, its "sensing local observation data" is updated. y i,j This "update" is implemented "based on all row-level local observation data," meaning it is based on the global observation data of that row level. y i Subtract the sum of the current "local observation data" for all "row-level associated regions" from the middle. Then, "based on the updated sensory local observation data"... y i,j By solving a local, small-scale system of linear equations ( x j ←( A i,j ) Ty i,j / α 2 ), and perform threshold processing on its solution ( x j ← H K ( x j Thus, we "obtain" the "local estimation data" corresponding to the "perceived area". x j Finally, using this newly calculated... x j To perform the "consensus update": the algorithm "updates the local observation data of all column blocks", that is, recalculates the "local observation data" of all regions in column block j (including the "sensing region" [i,j] and all its "column block consensus regions" [k,j], k≠i). This "update-solve-consensus" process is executed sequentially for all "sensing regions" "until all sensing local observation data are updated".

[0124] The process will be described in more detail below.

[0125] Reference Figure 9 The sensing local observation data is updated based on all row-level local observation data, including the following steps 901 to 902.

[0126] Step 901: Obtain the local cumulative value based on the cumulative value of all row-level local observation data.

[0127] Step 902: Based on the difference between the initial global observation data and the local accumulated value, the updated local observation data is obtained, and the updated local observation data is used as the new sensing local observation data for the sensing area.

[0128] Steps 901 to 902 are described in detail below.

[0129] In some embodiments, a local cumulative value is obtained based on the cumulative value of all row-level local observation data. This step updates the "sensory local observation data" for a specific "sensory region" [i,j]. y i,j The preparatory calculations performed. Specifically, the algorithm accesses all other "row-level associated regions" [i,k] (where k≠j) located in the same "row layer" i as the current "sensory region" [i,j]; then, it stores the "row-level local observation data" of these "row-level associated regions". y i,k To perform a summation, we execute an "accumulation" operation. The result of this "accumulation" operation ( k≠j yi,k This is defined as a "local accumulation value".

[0130] Then, based on the difference between the initial global observation data and the local accumulated values, updated local observation data is obtained, and this updated local observation data is used as the new sensing local observation data for the sensing region. This step performs the actual update operation. It first obtains the component corresponding to the current row layer i in the "initial global observation data". y i Then, the algorithm calculates the observation data for that row layer. y i The obtained "local cumulative value" ( k≠j y i,k The "difference" between ) is used to obtain updated local observation data, such as y i - k≠j y i,k Finally, this value of "updating local observation data" is used as "new local observation data" for the current "sensory region" [i,j], i.e., an assignment operation is performed, such as y i,j ←y i - k≠j y i,k .

[0131] Through steps 901 and 902 above, an efficient estimation and "update" mechanism for "perceived local observation data" is realized. Since the "initial global observation data" of the row layer is a linear superposition of all "local observation data" of the row layer, by calculating the "local cumulative value" (i.e. the sum of contributions from all other regions except the current sensing region) and then performing the "difference" operation, an accurate estimate of the "perceived local observation data" corresponding to the current "sensing region" is cleverly extracted from the total data of the row layer. This "update" operation provides an accurate and decoupled input for subsequent steps, thereby decomposing a large-scale, coupled global solution problem into a series of small-scale, low-cost local solution problems, significantly improving the computational efficiency of the algorithm.

[0132] The following section will further describe how to perform consensus updates.

[0133] Reference Figure 10 Based on the updated local observation data of the sensing area, the local estimation data corresponding to the sensing area is obtained, and consensus update is performed on the local observation data of all columns and blocks, including the following steps 1001 to 1004.

[0134] Step 1001: Based on the transpose of the first measurement submatrix corresponding to the sensing area in the measurement matrix and the product of the updated sensing local observation data, and then divided by the square of the first scaling factor, the initial local estimation data is obtained.

[0135] Step 1002: Perform thresholding based on the initial local estimation data to obtain local estimation data.

[0136] Step 1003: Multiply the first measurement sub-matrix corresponding to the sensing area with the local estimation data to obtain the first updated local observation data, and use the first updated local observation data as the new sensing local observation data.

[0137] Step 1004: Multiply the second measurement sub-matrix corresponding to the consensus region of each column block by the local estimated data to obtain the second updated local observation data, and use the second updated local observation data as the new column block local observation data.

[0138] Steps 1001 to 1004 are described in detail below.

[0139] In some embodiments, initial local estimation data is obtained by multiplying the transpose of the first measurement submatrix corresponding to the sensing region in the measurement matrix with the updated sensing local observation data, and then dividing by the square of the first scaling factor. This step is the core computation for solving local sparse sensing. It utilizes the "updated sensing local observation data" (e.g., ...) from the previous steps. y i,j ), and the "first measurement submatrix" corresponding to the "sensing region" [i,j] in the "measurement matrix" (i.e. A i,j The calculation is as follows: First, calculate... A i,j The transpose of the matrix and y i,j The product of these factors is then divided by the square of the first scaling factor α, for example, for the sensing region [1,1]. x 1←( A 1,1 ) T y 1,1 / α 2 The mathematical essence of this operation is solving local equations, because... A i,j It is a scaled orthogonal matrix. The result of the operation is defined as the "initial local estimate data". x j It is a preliminary estimate of the local image fragment, which will be sparsified in the next step.

[0140] Then, thresholding is performed on the initial local estimation data to obtain the final local estimation data. This step involves processing the obtained "initial local estimation data". x j Sparsity reduction is performed. Since this technical solution is based on the sparsity assumption of compressed sensing, it is necessary to perform sparsification on the solution. x j Perform "threshold processing". This "threshold processing" is typically a hard threshold operation, its function being to preserve... x j The K components with the largest absolute values ​​in the vector are selected, while all other components are set to zero. The resulting sparse vector after this "thresholding" process represents the final "local estimate" of the "sensory region" in this iteration. x j ← H K ( x j ).

[0141] Next, the first measurement submatrix corresponding to the sensing region is multiplied by the local estimated data to obtain the first updated local observation data, which is then used as the new sensing local observation data. This step is the first part of the "column consensus update," that is, updating the "local observation data" of the "sensing region" [i,j] itself. The "first measurement submatrix" corresponding to the "sensing region" [i,j] is then... A i,j This is in contrast to the sparse "local estimation data" obtained. x j Perform product operations (i.e.) A i,j* x j The result of the operation is called the "first updated local observation data," which represents the data based on the current best local estimate. x j The data that should be observed. Finally, this "first updated local observation data" is used as the new local observation data for the sensing region [i,j] (i.e., y i,j ← A i,j* x j ), for use in subsequent iterative calculations.

[0142] Furthermore, the second measurement sub-matrix corresponding to each column block consensus region is multiplied by the local estimated data to obtain the second updated local observation data, which is then used as the new column block local observation data. This step is the core of the "column consensus update" and embodies the concept of "consensus." This step uses the same "local estimated data" obtained from the same process. xj This is shared among all "column block consensus regions" (i.e., [k,j], k≠i) located in the same column block j. The algorithm will sequentially assign the "second measurement submatrix" corresponding to each "column block consensus region" [k,j]. A k,j , and the shared "local estimation data" x j Perform product operations (i.e.) A k,j* x j This is used to obtain the "second updated local observation data". Finally, this "second updated local observation data" is used as the "new column block local observation data" (i.e., ...) for the corresponding "column block consensus region" [k,j]. y k,j ← A k,j x j , k=1, 2, 3 ).

[0143] In one example, each "sparse sensing" operation executes three instructions (initial update of local observation data; solving the local linear equation system; and thresholding the solution), and outputs the sparse solution corresponding to that module. We will introduce the specific implementation process of the "sparse sensing + column consensus" operation in the order of the sensing regions.

[0144] When performing the "sparse sensing" operation on the first sensing region [1,1], the three instructions executed are (i)-(iii):

[0145] (i) Initial Update as follows:

[0146]

[0147] (ii) Solve the system of equations , get

[0148]

[0149] (iii) Perform threshold processing:

[0150]

[0151] At this point, the "sparse sensing" operation on the sensing region [1,1] is complete.

[0152] (iv) Perform consensus update That is, the first column of matrix A is shared. Update the observation data for all local regions corresponding to the first column block:

[0153]

[0154] Entering the second sensing region [2,2], perform the same "sparse sensing + column consensus" operation as follows:

[0155] (i) Initial update as follows:

[0156]

[0157] (ii) Solve the system of equations ,get

[0158]

[0159] (iii) To Perform threshold processing:

[0160]

[0161] At this point, the "sparse sensing" operation on the sensing region [2,2] is complete.

[0162] (iv) Perform consensus update That is, the second column of matrix A is shared. Update the observation data for all local regions corresponding to the second column block:

[0163]

[0164] Proceed to the "sparse sensing + column consensus" operation on the sensing region [3,3], ..., until the last sensing region. In this example, the last sensing region is [3,9], and its corresponding "sparse sensing + column consensus" operation is as follows:

[0165] (i) Initial Update as follows:

[0166]

[0167] (ii) Solve the system of equations ,get

[0168]

[0169] (iii) Perform threshold processing:

[0170]

[0171] At this point, the "sparse perception" operation on the perception region [3,9] is complete.

[0172] (iv) Perform consensus update That is, the 9th column of matrix A is shared. Update the observation data for all local regions corresponding to the 9th column block:

[0173]

[0174] At this point, the "sparse sensing" operation and the corresponding "column consensus" operation for the selected sensing region have been completed.

[0175] Through steps 1001 to 1004 above, an efficient and low-cost "local perception collaboration" computation mechanism is realized. By using "sparse perception" computation, the large-scale global problem is decomposed into small-scale local solutions based on the "first measurement sub-matrix" (scaling orthogonal matrix) for the "perception region", resulting in extremely low computational cost. Then, an innovative information collaboration mechanism is established through "column consensus update": only the "perception region" is responsible for calculating the "local estimation data", while all other "column consensus regions" in the same column block share this estimation result and update their own "local observation data" based on their respective "second measurement sub-matrix". This "local sparse perception + column consensus" design avoids the repeated and costly solution of all local regions, greatly reduces the overall computational complexity of the algorithm, and ensures the global consistency integration of information in the column block direction.

[0176] Step 105: Based on all local estimation data, perform global refinement to obtain the image compression data corresponding to the image data to be compressed.

[0177] Step 105 is described in detail below.

[0178] Finally, based on all local estimation data, global refinement is performed to obtain the image compressed data corresponding to the image data to be compressed. After the iterative loop is completed, the algorithm has obtained the "local estimation data" of all column blocks. This step combines all these "local estimation data" (e.g., x i , i=1, ... ,9 ) are concatenated in column order to form a complete sparse estimation vector (e.g. x= ( x 1, ... , x9) Subsequently, to further improve accuracy, a "global refinement" process is performed on the vector x. This "refinement" process can include "wide thresholding" and "narrow thresholding": first, a wider set of indices (e.g., 2K maximum components) is selected and a least-squares problem is solved to obtain a "wide-domain refined value"; then, based on this result, a narrower set of indices (e.g., K maximum components) is selected and a final least-squares problem is solved to obtain a "narrow-domain refined value," thus obtaining the final sparse image vector x. This refined sparse vector x is the sparse representation of the "image compressed data," and can finally be restored to the final compressed image through inverse transformation, as described below.

[0179] Reference Figure 11 Based on all local estimation data, global refinement is performed to obtain the image compression data corresponding to the image data to be compressed, including the following steps 1101 to 1105.

[0180] Step 1101: Concatenate all local estimation data in column block order to obtain a sparse estimation vector.

[0181] Step 1102: Select the multiple local estimation data with the largest absolute value from the sparse estimation vector to obtain the first index set.

[0182] Step 1103: Select multiple measurement sub-matrices corresponding to the first indicator set from the measurement matrix to obtain the first indicator measurement matrix.

[0183] Step 1104: Solve the first index least squares model to obtain the wide-domain refined value. The first index least squares model is obtained by multiplying the transpose of the first index measurement matrix with the initial global observation data.

[0184] Steps 1101 to 1104 are described in detail below.

[0185] In some embodiments, all local estimation data are concatenated in column block order to obtain a sparse estimation vector. This step marks the beginning of the "global refinement process." It collects all local estimation data computed for all p column blocks in the previous steps (…). x 1, ... , x p Then, in column block order (i.e.) x 1 Follow x 2. And so on) These m-dimensional local vectors are "concatenated" in the column direction to form an n-dimensional (n=mp) complete "sparse estimation vector" representing the entire image, such as... x= ( x 1, ... , x9), this vector is the basis for subsequent refining operations.

[0186] Then, the first set of indices is obtained by selecting the multiple local estimates with the largest absolute values ​​from the sparse estimation vector. This step is to identify the most important components in the sparse estimation vector x. The algorithm examines all n elements in vector x and selects multiple (e.g., 2K) components with the largest absolute values, which contribute the most numerically to the image, to obtain the first set of indices. S ={ i 1 ,...,i 2K},in i 1 <...<i 2K All are positive integers no greater than n, which is the set of indices corresponding to the positions of these 2K "local estimated data" components in the n-dimensional vector x.

[0187] Next, multiple measurement sub-matrices corresponding to the first indicator set are selected from the measurement matrix to obtain the first indicator measurement matrix. This step constructs a relevant sub-matrix for the subsequent least squares solution. Using the obtained "first indicator set" S, the process returns to the original, complete d x n "measurement matrix" A. Then, from the n column vectors of the "measurement matrix" A, it selects those columns whose column indices are contained in all columns of the "first indicator set" S (i.e., the "corresponding multiple measurement sub-matrices," referring to the column vectors of A). These selected columns (e.g., 2K columns) are combined in their original order to form a new d x 2K sub-matrix, which is the "first indicator measurement matrix." A S .

[0188] Next, the least squares model of the first index is solved to obtain the wide-domain refined value. The least squares model of the first index is obtained by multiplying the transpose of the first index measurement matrix with the initial global observation data. This step is the core calculation for performing the "wide threshold operation". This step constructs a "least squares model of the first index" [( A S ) T A S ] -1 ( A S ) T y The model aims to find a 2K-dimensional solution vector u that lies within the "first index measurement matrix". A S Under the influence of this, it can best fit the "initial global observation data" y. This model is typically solved using the standard least squares method (…). u ← [(A S ) T A S ] -1 ( A S ) T y This is achieved through a process called ), where the resulting 2K-dimensional vector u is the "wide-domain refined value," which is a high-precision estimate of the original signal over the subspace defined by the "first index set" S.

[0189] By executing steps 1101 to 1104, the first key stage of "global refinement processing," namely "wide threshold operation," is achieved. After concatenating the complete "sparse estimation vector," the main energy support subspace of the signal is accurately identified by constructing the "first index set" and the "first index measurement matrix." Finally, by solving the "first index least squares model," the signal components in this subspace are globally corrected and refined using the "initial global observation data," thereby "obtaining" the "wide-domain refined value." Compared with the local estimation in the previous steps, this "global refinement" process significantly improves the estimation accuracy, providing a higher quality vector foundation for subsequent "narrow threshold operation" or direct output of compressed data.

[0190] Step 1105: Obtain image compressed data based on the wide-area refined values.

[0191] Step 1105 will be described in detail below.

[0192] In some embodiments, after obtaining the wide-range refined value u, let x It is an n-dimensional column vector that is equal to u on the index set S and 0 outside the index set, as shown below.

[0193]

[0194] To further improve the reliability of image data compression, a narrow threshold operation is performed as described below.

[0195] Reference Figure 12 Based on the wide-area refined value, image compressed data is obtained, including the following steps 1201 to 1205.

[0196] Step 1201: Based on multiple wide-area refined values, obtain a wide-area vector, and select the wide-area refined values ​​with the largest absolute values ​​from the wide-area vector to obtain the second index set.

[0197] Step 1202: Select multiple measurement sub-matrices corresponding to the second index set from the measurement matrix to obtain the second index measurement matrix.

[0198] Step 1203: Solve the least squares model of the second index to obtain the narrow-domain refined value. The least squares model of the second index is obtained by multiplying the transpose of the second index measurement matrix with the initial global observation data.

[0199] Step 1204: Based on all narrow-domain refined values, obtain the sparse image vector.

[0200] Step 1205: Perform an inverse orthogonal transform on the sparse image vector to obtain compressed image data.

[0201] Steps 1201 to 1205 are described in detail below.

[0202] In some embodiments, a wide-domain vector is obtained based on multiple wide-domain refined values, and the multiple wide-domain refined values ​​with the largest absolute values ​​are selected from the wide-domain vector to obtain a second index set. This step is the beginning of "narrow threshold" refinement. First, it constructs an n-dimensional "wide-domain vector" based on the previously obtained "multiple wide-domain refined values" u. x The vector x The value on the "first index set" S is equal to u, while the value outside the index set S is 0. Next, the algorithm applies this n-dimensional "wide-field vector"... x The analysis selects "multiple" (e.g., K, where K is usually less than 2K from the previous step) components with the largest absolute values; these components are considered the core building blocks of the signal. The indexes of the locations of these K components constitute the "second index set". Λ ={ r 1 ,...,r K},in r 1 <...<r K All are positive integers not greater than n.

[0203] Then, multiple measurement sub-matrices corresponding to the second indicator set are selected from the measurement matrix to obtain the second indicator measurement matrix. This step prepares the matrix for the final least squares solution. It utilizes the obtained, smaller, and more accurate "second indicator set". Λ ={ r 1 ,...,r K The algorithm then returns to the original, complete "measurement matrix" A. Next, from the n column vectors of A, it selects "multiple measurement submatrices" (i.e., columns of A) whose column indices are contained in the "second indicator set". Combining these selected K columns yields the "second indicator measurement matrix". A Λ The dimension of this matrix is ​​d*K.

[0204] Next, the second index least squares model is solved to obtain narrow-domain refined values. The second index least squares model is obtained by multiplying the transpose of the second index measurement matrix with the initial global observation data. This step performs the final coefficient refinement. It constructs a "second index least squares model"[( A Λ ) T A Λ ] -1 ( A Λ ) T y The model aims to utilize a "second indicator measurement matrix". A Λ And "initial global observation data" y. By "solving" the model (i.e. calculating v ← [( A Λ ) T A Λ ] -1 ( A Λ ) T y This yields a K-dimensional solution vector v, which is the "narrow-domain refined value". This vector v represents the most accurate amplitude estimate of the signal component in the K-dimensional subspace defined by the "second index set" Λ.

[0205] Furthermore, based on all the narrow-domain refined values, a sparse image vector is obtained. This step is the process of constructing the final sparse solution. Based on the obtained K-dimensional "narrow-domain refined value" vector v, the algorithm creates an n-dimensional zero vector, and then places all the narrow-domain refined values ​​from vector v one by one into this n-dimensional vector, their positions determined by the "second index set". Λ Uniquely specified. The final n-dimensional vector "obtained" is the "sparse image vector" as shown in the following formula.

[0206]

[0207] It is equal to v on the index set Λ and 0 outside the index set. It is a sparse vector containing at most K non-zero elements and is the optimal representation of the original image in the sparse domain. Vector x is the sparse image generated in this iteration, containing at most K non-zero elements.

[0208] Furthermore, to further improve image quality, based on the currently obtained approximate image x, the observation data of all local regions are updated, such as... Then, repeat the above calculation process. If the difference in x between two iterations is less than a pre-set precision value (e.g., 10), then... -4The algorithm terminates and outputs a compressed image through the following steps.

[0209] Finally, an inverse orthogonal transform is performed on the sparse image vector to obtain compressed image data. This step is the final step in recovering the image from the sparse solution. It takes the obtained "sparse image vector" x (which lies in the transform domain). The algorithm performs an inverse orthogonal transform on this vector x. W -1 (This transformation corresponds to the "orthogonal linear transformation" W mentioned above). The result of the inverse transformation (i.e. X=W -1 ( x )) is an n-dimensional vector X, which is no longer sparse but has been restored to the spatial domain of the original image. It is the final image compressed data that can be viewed or stored.

[0210] Through steps 1201 to 1205 above, the second and more refined "narrow thresholding" stage of the "global refinement processing" is achieved, and the final image data generation is completed. By constructing a "second index set" and a "second index measurement matrix," the solution space is narrowed from a "wide domain" to a more precise "narrow domain." Then, by solving the "second index least squares model," highly accurate "narrow domain refined values" are calculated in this smaller and more precise subspace. This two-stage refinement strategy from "wide" to "narrow" is more robust in determining the main support set of the signal and obtaining more accurate coefficients compared to single thresholding. This ensures that the resulting "sparse image vector" is the final sparse solution of the refinement process. Finally, the optimal sparse solution is transformed back into the final "image compressed data" through "orthogonal inverse transformation," thereby achieving high-precision image reconstruction while maintaining a high compression ratio.

[0211] Understandably, if the dimension n of a sparse vector x is not an integer multiple of 9, some zeros can be added to x, thus changing the dimension of x to a multiple of 9. Therefore, x can be divided into 9 segments: x = ( x 1 , x 2 , ..., x 9), of which each segment x j All are m-dimensional column vectors.

[0212] Reference Figure 13 This is a flowchart illustrating an image data compression method provided in an embodiment of this application. Figure 13 The diagram shows a complete flowchart of an image data compression method according to this application. The process begins with a preparation phase, which includes "inputting integer values ​​n, m, q" and "generating pq distinct m*m orthogonal matrices". Subsequently, the process execution... Perform scaling to obtain the scaling matrix. The data are then spliced ​​together to form a measurement matrix A.

[0213] After obtaining the measurement matrix A, the process "obtain the observation vector y=Ax of image x" is executed. Following this, the procedure "input any initial vector" is executed. satisfy This corresponds to the initialization operation.

[0214] After entering the main iteration loop, the process first selects the [i, j]th region according to the 'execution region selection rule'. Then, it performs 'local region sparse sensing operation' on this 'perceptual region' to generate local image segments. x j "Estimation", and "Execute the corresponding fragment" x j The algorithm performs a 'column consensus operation' to obtain updated local observation data. At the end of this loop (i.e., traversing all receptive regions), the algorithm synthesizes an estimated image. It then performs a wide thresholding operation, followed by a narrow thresholding operation to obtain an updated image x.

[0215] After performing the narrow thresholding operation, the process reaches a decision point: "Does the termination criterion meet?". If "yes" (e.g., the difference in x between two iterations is less than a preset precision value), the algorithm executes "Stop" and outputs the final compressed image data. If "no", the process enters the update step: "Update the local observation vector based on the current estimated x." After the update is complete, the process returns and restarts the next round of "selection according to the 'execution region selection rule'", thus forming a complete iterative loop until the termination criterion is met.

[0216] Understandably, the calculation principle for the case of q=2 rows is exactly the same as that for the case of 3 rows, as shown in the example below.

[0217] Reference Figure 14 This is a schematic diagram of another plurality of first scaling matrices and a plurality of second scaling matrices provided in the embodiments of the application. For example... Figure 14 The diagram shows matrix B in the case of q=2 rows and p=9 columns. This diagram illustrates a set of m x m "orthogonal matrices". In this schematic, the sensing area corresponds to... Figure 14 The nine orthogonal matrices highlighted in blue are... B 1,1 , B 2,2 ..., these are selected as the "first scaling matrix"; while the orthogonal matrices corresponding to all the unhighlighted white areas in the figure (such as...) B 1,2 (etc.) are then classified as "second scaling matrix".

[0218] Reference Figure 15 This is a schematic diagram of another selected sensing area provided in the application embodiment. For example... Figure 15 The diagram illustrates the sequence of selecting the sensing region in a 2-row, 9-column block configuration. This diagram shows the sequence determined by the "measurement submatrix". A i,j The diagram shows the "observation area," with the blue highlighted area representing the selected "perception area." The red arrows in the diagram clearly indicate the "one-by-one" order in which the "local sparse perception" operation is performed.

[0219] Based on this, the implementation scheme for the q=2 row layer case is described again below (since steps 601, 1101-1105, and 1201-1205 are exactly the same as the 3 row layer case, the specific details are omitted):

[0220] The image X is sparsified to obtain its sparse representation vector x.

[0221] Generate orthogonal matrix blocks. Given a value of m, generate some m x m orthogonal matrix blocks, using... B i,j , i=1,2 ; j=1, ...,9 This means that the above matrix blocks are concatenated into a d-row, n-column (d=2m, n=9m) matrix B (2-row, 9-column blocks):

[0222]

[0223] A structured measurement matrix A is generated using orthogonal matrix blocks, where each sub-block is a scaled orthogonal matrix. The following conditions are met: α 2 + ( q -1) β 2 =1 Two positive numbers α and β (q=2 in this example). For example: α=3 / 5, β=4 / 5. Select submatrices located in different row levels and different column blocks (rows 1 and 2 are selected alternately, and column blocks increase sequentially from 1 to 9): B 1,1 , B 2,2 , B 1,3 , B 2,4 , B 1,5 , B 2,6 , B 1,7, B 2,8 , B 1,9 The areas they correspond to are called "sensory regions", such as Figure 14 As shown.

[0224] Then, based on matrix B and scaling factors α and β, a structured measurement matrix A is designed. The specific method is as follows: when [i,j] belongs to the sensing region, the sub-block is... B i,j Multiply by α to obtain the scaled orthogonal matrix A i,j ,Right now A i,j= αB i,j When [i,j] does not belong to the perception region, the sub-block is... B i,j Multiplying by β yields the scaled orthogonal matrix. A i,j ,Right now A i,j= βB i,j The above scaling orthogonal matrix A i,j , i=1,2 ; j=1, ...,9 Concatenate them into a d-row, n-column (d=2m, n=9m) structured measurement matrix A (2-row, 9-column block):

[0225]

[0226] The total observed data y of the image is obtained using a linear overlay technique. The observed data y is an approximation of Ax, and is a column vector with d=2m dimensions. Since A is a concatenation of two row-level modules, y can also be viewed as two m-dimensional column vectors. y 1, y 2 is concatenated column by column, that is y = ( y 1, y 2). These can be viewed as the observation data corresponding to the 1st and 2nd rows of matrix A, respectively, i.e.:

[0227]

[0228] For each local region [i,j] (i=1,2; j=1, ..., 9), define y i,j =A i,j x j and treat it as a local image. x j Scaling orthogonal matrices Ai,j Precise observational data is required. Therefore, y 1, y 2 represents the linear superposition of the observation data of the local images in the first and second rows, respectively. y i = y i,1 + y i,2 + ... + y i,9 , i=1, 2 .

[0229] Set initial observation data for a local area y i, j At the start of the iteration, any setting is allowed. y i,j ( i=1,2 ; j = 1,...,9 The initial value of ), for example:

[0230]

[0231] Sparse sensing computation and local observation data in local regions y i,j The process involves updates and iterations. Local observation data is acquired through sparse sensing of a few local regions, aided by column consensus technology. y i,j Gradual updates. For example... Figure 15 As shown, the order of the perceptual regions that perform the "sparse perception" operation is: [1,1]→[2,2]→[1,3]→[2,4]→[2,5]→[2,6]→[1,7]→[2,8]→[1,9].

[0232] Each "sparse sensing" operation executes three instructions (initial update of local observation data; solving the local linear equation system; and thresholding the solution), and outputs the sparse solution corresponding to that module. We will introduce the specific implementation process of the "sparse sensing + column consensus" operation according to the selection order of the sensing region.

[0233] When performing the "sparse sensing" operation on the first sensing region [1,1], the three instructions executed are (i)-(iii):

[0234] Initial update as follows:

[0235]

[0236] Solve the system of equations ,get:

[0237]

[0238] right Perform threshold processing:

[0239]

[0240] At this point, the "sparse sensing" operation on the sensing region [1,1] is complete.

[0241] Execute consensus update That is, the first column of matrix A is shared. Update the observation data for all local regions corresponding to the first column block:

[0242]

[0243] Entering the second sensing region [2,2], perform the same "sparse sensing + column consensus" operation as follows:

[0244] (i) Initial update as follows:

[0245]

[0246] (ii) Solve the system of equations ,get:

[0247]

[0248] (iii) To Perform threshold processing:

[0249]

[0250] At this point, the "sparse sensing" operation on the sensing region [2,2] is complete.

[0251] (iv) Perform consensus update That is, the second column of matrix A is shared. Update the observation data for all local regions corresponding to the second column block:

[0252]

[0253] Proceed to the "sparse sensing + column consensus" operation on the sensing region [1,3], ..., until the last sensing region. In this example, the last sensing region is [1,9], and its corresponding "sparse sensing + column consensus" operation is as follows:

[0254] Initial update as follows:

[0255]

[0256] Solve the system of equations ,get:

[0257]

[0258] right Perform threshold processing:

[0259]

[0260] At this point, the "sparse perception" operation on the perception region [1,9] is complete.

[0261] Execute consensus update That is, the 9th column of matrix A is shared. Update the observation data for all local regions corresponding to the 9th column block:

[0262]

[0263] At this point, the "sparse sensing" operation and the corresponding "column consensus" operation for the selected sensing region have been completed.

[0264] Then perform wide threshold calculation and narrow threshold calculation separately.

[0265] Then, the iteration continues. To further improve image quality, based on the currently obtained approximate image x, the observation data for all local regions are updated:

[0266]

[0267] Then, repeat the above calculation process. If the difference in x between two iterations is less than a pre-set precision value (e.g., 10), then the calculation is repeated. -4 If the algorithm terminates, the compressed image is output through the following steps.

[0268] Inverse transformation of sparse vector x W -1 The compressed image X can then be obtained, i.e. X=W -1 ( x ).

[0269] Furthermore, when increasing the number of measurements, taking a two-row layer as an example, assuming the addition of one measurement layer results in a total of three row layers (the scaling factor of each sub-block in measurement matrix A needs to be adjusted); the first two row layers are the original measurement layers, and the third row layer is the newly added measurement layer. The total number of "sparse sensing" operations remains the same as before the addition of measurements (both are 9 times), with only a small number of column consensus operations added, which have a relatively low computational cost.

[0270] The proposed solution designs the measurement matrix as a pattern of small-scale orthogonal matrix stitching, with each sub-matrix performing local sparse sensing only on the image, resulting in low computational complexity. Compared to existing technologies in processing large-scale images, the proposed solution significantly reduces computational costs. The proposed solution injects a locally scaled orthogonal structure into the measurement matrix, and the algorithm avoids using parameters, helping to ensure the stability of the compression process. Existing technologies are susceptible to numerical errors in large-scale images, facing difficulties in parameter selection, often leading to unstable numerical results. The proposed solution allows for increasing the image sampling data to improve image compression quality without significantly increasing computational costs. For example, as... Figure 5 As shown, the measurement matrix increases from 24 blocks (2 rows, 12 columns) to 36 blocks (3 rows, 12 columns), but the number of subproblems requiring a solution remains unchanged. In existing technologies, as the sampling rate increases, the computational cost rises significantly, while the processing efficiency decreases accordingly.

[0271] This application's solution employs a "local sparse sensing + column consensus → global information integration" technique to accurately capture the main feature data of an image, thereby improving the quality of image compression. Furthermore, the solution integrates local sensing results through consensus technology, ensuring global consistency. This avoids solving numerous sub-problems and prevents local errors from affecting the overall result. Existing technologies lack sufficient refinement when processing local details and directly embed the measurement matrix as a whole during computation, failing to consider its internal structural features. In conclusion, this application's solution surpasses existing technologies in many aspects, including computational efficiency, compression quality, flexibility, stability, and the sophistication of its algorithm design principles, representing a breakthrough in the field of image compression.

[0272] This application also provides an electronic device, including:

[0273] At least one memory;

[0274] At least one processor;

[0275] At least one program;

[0276] The program is stored in memory, and the processor executes at least one program to implement the image data compression method described above in this application. The electronic device can be any smart terminal, including mobile phones, tablets, personal digital assistants (PDAs), and in-vehicle computers.

[0277] Please see Figure 16 , Figure 16 The hardware structure of an electronic device according to another embodiment is illustrated. The electronic device includes:

[0278] The processor 1601 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this application.

[0279] The memory 1602 can be implemented in the form of ROM (Read Only Memory), static storage device, dynamic storage device, or RAM (Random Access Memory). The memory 1602 can store the operating system and other applications. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory 1602 and is called and executed by the processor 1601 using the image data compression method of the embodiments of this application.

[0280] The input / output interface 1603 is used to implement information input and output;

[0281] The communication interface 1604 is used to enable communication and interaction between this device and other devices. Communication can be achieved through wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.).

[0282] Bus 1605 transmits information between various components of the device (e.g., processor 1601, memory 1602, input / output interface 1603, and communication interface 1604);

[0283] The processor 1601, memory 1602, input / output interface 1603 and communication interface 1604 are connected to each other within the device via bus 1605.

[0284] This application embodiment also provides a storage medium, which is a computer-readable storage medium, storing a computer program that, when executed by a processor, implements the above-described image data compression method.

[0285] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs. Furthermore, memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, memory may optionally include memory remotely located relative to the processor, and these remote memories can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0286] The embodiments described in this application are for the purpose of more clearly illustrating the technical solutions of the embodiments of this application, and do not constitute a limitation on the technical solutions provided by the embodiments of this application. As those skilled in the art will know, with the evolution of technology and the emergence of new application scenarios, the technical solutions provided by the embodiments of this application are also applicable to similar technical problems.

[0287] Those skilled in the art will understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of this application, and may include more or fewer steps than shown, or combine certain steps, or different steps.

[0288] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0289] Those skilled in the art will understand that all or some of the steps in the methods disclosed above, as well as the functional modules / units in the systems and devices, can be implemented as software, firmware, hardware, or suitable combinations thereof.

[0290] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0291] It should be understood that in this application, "at least one (item)" means one or more, and "more than" means two or more. "And / or" is used to describe the relationship between related objects, indicating that three relationships can exist. For example, "A and / or B" can represent three cases: only A exists, only B exists, and both A and B exist simultaneously, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one (item) of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one (item) of a, b, or c can represent: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.

[0292] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of the units described above is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. The coupling or direct coupling or communication connection between the shown or discussed units may be through some interfaces, or indirect coupling or communication connection between the apparatus or units, and may be electrical, mechanical, or other forms.

[0293] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0294] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0295] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes multiple instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing programs, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0296] The preferred embodiments of the present application have been described above with reference to the accompanying drawings, but this does not limit the scope of the claims of the present application. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and substance of the embodiments of the present application shall be within the scope of the claims of the present application.

Claims

1. An image data compression method characterized by, The method comprises: acquiring image data to be compressed, and acquiring a measurement matrix generated by scaling and splicing based on a plurality of orthogonal matrices; performing linear measurement on the image data to be compressed based on the measurement matrix to obtain initial global observation data, and based on the initial global observation data, obtaining initial local observation data of a plurality of local regions in an observation region, each of the local regions corresponding to each measurement sub-matrix in the measurement matrix; selecting a perception region one by one from the plurality of local regions, obtaining a row layer associated region corresponding to the perception region from the local region in a row layer direction, and obtaining a column block consensus region corresponding to the perception region from the local region in a column block direction; updating the perception local observation data based on all row layer local observation data, and obtaining local estimation data corresponding to the perception region based on the updated perception local observation data, and performing consensus update on all column block local observation data until all perception local observation data updates are completed, the row layer local observation data being the initial local observation data of the row layer associated region, the perception local observation data being the initial local observation data of the perception region, and the column block local observation data being the initial local observation data of the column block consensus region; performing global refinement processing based on all the local estimation data to obtain image compression data corresponding to the image data to be compressed.

2. The image data compression method according to claim 1, characterized by, The acquisition of the measurement matrix comprises: acquiring a plurality of orthogonal matrices, and acquiring first and second scaling coefficients; selecting a first scaling matrix from the plurality of orthogonal matrices based on the region where the perception region is located, and taking other orthogonal matrices as second scaling matrices; multiplying all the first scaling matrices by the first scaling coefficient to obtain first measurement sub-matrices, and multiplying all the second scaling matrices by the second scaling coefficient to obtain second measurement sub-matrices; splicing based on all the first measurement sub-matrices and all the second measurement sub-matrices to obtain the measurement matrix.

3. The image data compression method of claim 1, wherein, The linear measurement on the image data to be compressed based on the measurement matrix to obtain initial global observation data comprises: performing orthogonal linear transformation on the image data to be compressed to obtain a sparse representation vector; performing product operation on the sparse representation vector based on the measurement matrix to obtain the initial global observation data.

4. The image data compression method of claim 1, wherein, The selection of the perception region one by one from the plurality of local regions comprises: performing square division on the plurality of local regions in the observation region to obtain at least one divided observation sub-region; according to the order of the divided observation sub-regions, selecting the local region on the diagonal line in each divided observation sub-region as the perception region one by one.

5. The image data compression method of claim 1, wherein, The updating of the perception local observation data based on all row layer local observation data comprises: obtaining a local accumulation value based on the accumulation value of all the row layer local observation data; Based on the difference between the initial global observation data and the local accumulated value, updated local observation data is obtained, and the updated local observation data is taken as new perception local observation data of the perception area.

6. The image data compression method of claim 2, wherein, The updated perception local observation data is used to obtain local estimation data corresponding to the perception area, and all column block local observation data is updated by consensus, including: Based on the product of the transpose of the first measurement sub-matrix corresponding to the perception area in the measurement matrix and the updated perception local observation data, divided by the square of the first scaling coefficient, initial local estimation data is obtained; The initial local estimation data is threshold processed to obtain the local estimation data; The first measurement sub-matrix corresponding to the perception area is multiplied by the local estimation data to obtain first updated local observation data, and the first updated local observation data is taken as new perception local observation data; The second measurement sub-matrix corresponding to the column block consensus area is multiplied by the local estimation data one by one to obtain second updated local observation data, and the second updated local observation data is taken as new column block local observation data.

7. The image data compression method of claim 1, wherein, Based on all the local estimation data, global refinement processing is performed to obtain image compression data corresponding to the image data to be compressed, including: All the local estimation data is spliced in column block order to obtain a sparse estimation vector; A first index set is obtained by selecting multiple local estimation data with the largest absolute value from the sparse estimation vector; A first index measurement matrix is obtained by selecting multiple measurement sub-matrices corresponding to the first index set from the measurement matrix; A wide-range refinement value is obtained by solving a first index least squares model based on the product of the transpose of the first index measurement matrix and the initial global observation data; The image compression data is obtained based on the wide-range refinement value.

8. The image data compression method of claim 7, wherein, The image compression data is obtained based on the wide-range refinement value, including: A wide-range vector is obtained based on multiple wide-range refinement values, and a second index set is obtained by selecting multiple wide-range refinement values with the largest absolute value from the wide-range vector; A second index measurement matrix is obtained by selecting multiple measurement sub-matrices corresponding to the second index set from the measurement matrix; A narrow-range refinement value is obtained by solving a second index least squares model based on the product of the transpose of the second index measurement matrix and the initial global observation data; A sparse image vector is obtained based on all the narrow-range refinement values; The sparse image vector is orthogonally inverse transformed to obtain the image compression data.

9. An electronic device, comprising: The computer program is executed by the processor to realize the image data compression method of any one of claims 1-8.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to realize the image data compression method of any one of claims 1-8.

Images