Method of compressing and decompressing multi-dimensional data

By using superpixel-based dictionary learning and sparse coding techniques, the problem of long computation time in multidimensional data compression is solved, achieving fast compression and high-quality reconstruction, and improving memory utilization efficiency and compression ratio.

CN119032567APending Publication Date: 2024-11-26HONG KONG APPLIED SCI & TECH RES INST
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202480001552.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Priority Date
2024-07-19
Filing Date
2024-07-26
Publication Date
2024-11-26

AI Technical Summary

Technical Problem

Existing technologies have long computation times in multidimensional data compression, making it difficult to achieve high-quality data reconstruction and high compression ratio memory savings.

Method used

A superpixel-based dictionary learning method is adopted. The dictionary is trained by generating superpixel clusters, and sparse coding is used to compress and decompress multidimensional data. The dictionary training process is optimized by superpixel cluster similarity and batch orthogonal matching pursuit algorithm.

Benefits of technology

It achieves rapid multidimensional data compression and high-quality data reconstruction, improving memory usage efficiency and compression rate.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119032567A_ABST
    Figure CN119032567A_ABST
Patent Text Reader

Abstract

The invention provides a method for compressing multi-dimensional data with a plurality of data points into a compressed file. The method comprises the following steps: generating a plurality of superpixels from the multi-dimensional data; clustering the plurality of superpixels into a plurality of superpixel clusters; respectively training a plurality of dictionaries by using the plurality of superpixel clusters; dividing the multi-dimensional data into a plurality of data sub-blocks; vectorizing the plurality of data sub-blocks to obtain a plurality of vectorized data sub-blocks; selecting a dictionary for each vectorized data sub-block; performing sparse coding on each vectorized data sub-block by using the selected dictionary to obtain compressed representation; and storing the resulting compressed representation and the index of the corresponding selected dictionary in a compressed file.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to multidimensional data compression, and in particular to multidimensional data compression based on dictionary learning of sparse representation. Background Art

[0002] In recent years, sensors such as cameras, LIDAR, and inertial measurement units (IMUs) have been widely used, generating a large amount of multidimensional data to describe complex phenomena or systems and capture the relationship and interaction between different dimensions. Since multidimensional data requires high precision for further inspection and measurement tasks, the amount of multidimensional data is very large. In addition, multidimensional data may need to be stored for a long time, occupying a large amount of memory.

[0003] Dictionary learning based on sparse representation has been applied to data compression in various applications such as image and video compression, medical imaging, and signal processing, where the goal is to efficiently represent signals or data using a sparse set of basic elements. Dictionary learning allows the dictionary to be adjusted to better fit the characteristics of the data being compressed. This adaptive property allows for more efficient representation of complex signals by learning a dictionary that captures the essential features of the data in a sparse manner. By promoting sparsity, sparse representation based dictionary learning compression techniques are very valuable as they help to store and transmit data more efficiently while retaining important information. However, most existing dictionary learning methods learn training samples in a serial manner, which requires long computational time. Summary of the invention

[0004] The following summary is illustrative only and is not limiting in any way. That is, the following summary is provided to introduce the concepts, highlights, benefits, and advantages of the novel and non-obvious technologies described herein. The following detailed description will further introduce selected embodiments. Therefore, the following summary is not intended to determine the essential features of the claimed subject matter, nor is it intended to determine the scope of the claimed subject matter.

[0005] The purpose of the present invention is to propose a solution to the above problems, to compress multidimensional data in different forms, to reconstruct the compressed data with high quality, to save memory usage with high compression ratio, and to achieve fast compression speed. Therefore, the present invention provides a multidimensional data compression mechanism based on superpixels, trained dictionaries and sparse coding; and a data reconstruction mechanism for reconstructing multidimensional data with high quality.

[0006] In one aspect of the present invention, a method for compressing multidimensional data having multiple data points into a compressed file is provided. The method comprises: generating multiple superpixels from the multidimensional data; clustering the multiple superpixels into multiple superpixel clusters; using the multiple superpixel clusters to respectively train multiple dictionaries; dividing the multidimensional data into multiple data sub-blocks; vectorizing the multiple data sub-blocks to obtain multiple vectorized data sub-blocks; selecting a dictionary for each vectorized data sub-block; sparsely encoding each vectorized data sub-block using the corresponding selected dictionary to obtain a compressed representation; storing the obtained compressed representation and the index of the corresponding selected dictionary in a compressed file. Each dictionary is selected for the corresponding vectorized data sub-block by the following method: calculating the similarity between the corresponding vectorized data sub-block and each superpixel cluster; finding the best matching superpixel cluster with the highest similarity to the vectorized data sub-block; selecting the dictionary trained with the best matching superpixel cluster as the encoding dictionary. Each dictionary is trained with the corresponding superpixel cluster in the following manner: dividing the corresponding superpixel cluster into a plurality of training data sub-blocks; vectorizing the plurality of training data sub-blocks to obtain a plurality of initial representations respectively; calculating a similarity matrix between each pair of training data sub-blocks; calculating an average similarity value for each training data sub-block; selecting a plurality of initial training data sub-blocks from the plurality of training data sub-blocks according to the calculated average similarity values; initializing the dictionary by setting the initial training data sub-blocks as initial atoms of the dictionary; training the dictionary by executing a batch orthogonal matching pursuit (batchOMP) algorithm constrained by a target sparsity and a target residual threshold to construct a sparse coding representation for each training data sub-block; and updating the dictionary atom by atom using the constructed sparse coding representation to obtain a trained dictionary.

[0007] In another aspect of the present invention, a method for decompressing multidimensional data from a compressed file obtained by the method of the first aspect is provided. The method comprises: decoding a plurality of compressed representations to obtain a plurality of decompressed data sub-blocks; and connecting the plurality of decompressed data sub-blocks to reconstruct the multidimensional data. Each decompressed data sub-block is obtained by: retrieving a corresponding dictionary according to a corresponding index; multiplying the compressed representation with the retrieved dictionary to reconstruct a vectorized data sub-block; and devectorizing the reconstructed vectorized data sub-block to obtain a decompressed data sub-block.

[0008] It is worth noting that although the description provided herein may be in the context of certain data processing techniques for high-quality compressed multidimensional data based on sparse representation, other data processing techniques using algorithms such as orthogonal matching pursuit, batch orthogonal matching pursuit, simple linear iterative clustering, Laplace sparse subspace clustering, etc., as well as the proposed concepts, schemes, and any variants / derivatives thereof, can be implemented in and by other types of wireless access technologies, networks, and network topologies. Therefore, the scope of the present disclosure is not limited to the examples described herein. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] The embodiments of the present invention are described in more detail below in conjunction with the accompanying drawings, in which:

[0010] Figure 1 An example scenario of a multidimensional data compression process according to an embodiment of the present invention is shown;

[0011] Figure 2 is a flow chart of a multidimensional data compression method according to an embodiment of the present invention;

[0012] Figure 3 is a flowchart of generating superpixels according to an embodiment of the present invention;

[0013] Figure 4 is a flow chart of generating superpixels according to an embodiment of the present invention;

[0014] Figure 5 is a flowchart of generating superpixels according to an embodiment of the present invention;

[0015] Figure 6 is a flow chart of superpixel clustering according to an embodiment of the present invention;

[0016] Figure 7 is a continuous flow chart of superpixel clustering according to an embodiment of the present invention;

[0017] Figure 8 The figure is a flow chart of a multi-dimensional data decompression method according to an embodiment of the present invention. DETAILED DESCRIPTION

[0018] In the following description, the method of compressing and decompressing multidimensional data, etc. will be described as a preferred example. It will be apparent to those skilled in the art that modifications, including additions and / or substitutions, can be made without departing from the scope and spirit of the present invention. Specific details may be omitted in order not to obscure the present invention; however, the disclosure of the present invention is intended to enable those skilled in the art to practice the teachings herein without undue experimentation.

[0019] refer to Figure 1 , which shows an example scenario of multidimensional data compression. Multidimensional data 100 includes a plurality of multidimensional data points (not shown). A method for compressing multidimensional data 100 includes: Figure 2 Steps S302-S306 and Figure 3 Specifically, the method for compressing multidimensional data 100 includes: generating multiple superpixels 110 from multidimensional data (eg, image) 100, clustering the multiple superpixels 110 into multiple superpixel clusters 111, and using the multiple superpixel clusters 111 to train multiple dictionaries 300 respectively.

[0020] The multidimensional data compression method 100 also includes: dividing the multidimensional data 100 into multiple data sub-blocks (e.g., image blocks) 130; vectorizing the multiple data sub-blocks 130 to obtain multiple vectorized data sub-blocks; selecting a dictionary for each vectorized data sub-block; sparsely encoding each vectorized data sub-block using the corresponding selected dictionary to obtain a compressed representation; and storing the obtained compressed representation and the index of the corresponding selected dictionary in a compressed file.

[0021] By calculating the similarity between the corresponding vectorized data sub-block and each superpixel cluster, the best matching superpixel cluster with the highest similarity to the vectorized data sub-block is found, and the dictionary trained with the best matching superpixel cluster is selected as the encoding dictionary, thereby selecting the encoding dictionary for the corresponding vectorized data sub-block.

[0022] Specifically, one of the simple linear iterative clustering (SLIC) algorithm, the Felzenszwalb superpixel segmentation, the simple linear iterative clustering over-segmentation (SLICO) algorithm, the TurboPixels algorithm, and the Quick-Shift algorithm can be used to generate the superpixel 110, but is not limited thereto. The SLIC algorithm is used here as an exemplary implementation of generating the superpixel 110. The SLIC algorithm is a segmentation algorithm that aims to segment data sub-blocks into compact regions that are similar in color and texture.

[0023] For details of steps S402-S410 of the SLIC algorithm, see Figure 4 When the SLIC algorithm is started, some seed points are selected as initial superpixel centers 111 . These seed points can be obtained according to a regular grid or according to the color gradient of the multidimensional data 100 .

[0024] Then, multiple overlapping superpixel regions centered on and corresponding to multiple superpixel centers are defined respectively. The size of each overlapping superpixel region is equal to twice the maximum interval S of the superpixel centers. The maximum interval S can be calculated as Wherein N represents the area of ​​the multi-dimensional data 100 , and K represents the number of super-pixel centers 111 .

[0025] In order to assign each multidimensional data point to a corresponding superpixel region and update the current position of each superpixel center 111, it is necessary to calculate the distance between the data point and each adjacent superpixel, where the adjacent superpixel can be defined as a superpixel within an overlapping superpixel region including the multidimensional data point. Since superpixel regions are overlapped and cut, each multidimensional data point can be located within multiple superpixel regions.

[0026] Based on the calculated distance, a multidimensional data point can be assigned to a superpixel region that contains the nearest neighboring superpixels with the shortest distance to the data point.

[0027] In one embodiment, the distance D between each superpixel center 111 can be To obtain.

[0028] in, Represents two colors in multidimensional data 100 (l i ,a i ,b i ) and (l j ,a j ,b j )’s color distance; Represents two data points (x i ,y i ,z i ) and (x j ,y j ,z j )’s spatial distance between two spatial positions; m represents a constant, which is usually used to balance the influence of color distance and spatial distance.

[0029] By adjusting the constant m, the formation process of the superpixel 110 can be fine-tuned according to the specific characteristics and requirements of the multi-dimensional data, thereby obtaining a better superpixel segmentation effect.

[0030] refer to Figure 5 , the following steps S412-S418, assign each multidimensional data point to the nearest superpixel center according to its color and spatial distance. More specifically, the distance between each multidimensional data point and each superpixel center is calculated. Then, each multidimensional data point is assigned to the nearest superpixel center. In other words, each multidimensional data point is classified as belonging to the superpixel region closest to the superpixel center 111.

[0031] The current position of each superpixel center 111 is updated to the center point position of all multidimensional data points assigned to the corresponding superpixel region. Then, the residual value between the updated position and the current position of each superpixel center 111 is calculated. The residual value is usually calculated using Euclidean distance or other distance metrics to measure the position change of the superpixel center 111 between two iterations. Specifically, for each superpixel center i, the residual value between two iterations can be calculated as:

[0032] in and They represent the position of superpixel center i at the t-th and t-1-th iterations respectively.

[0033] The threshold is used to determine whether to terminate the iteration process. Usually, the threshold is set to a very small value, such as 0.1 or 0.01, to ensure that the iteration stops when the position change of the superpixel center 111 is small enough. By comparing the calculated residual value and the threshold, it can be determined whether the algorithm has converged. If the residual value is less than the threshold, it can be concluded that the algorithm has converged and no further iteration is required.

[0034] The purpose of calculating the residual value in the algorithm is to monitor the convergence of the algorithm. By checking the movement of the superpixel center 111 during the iteration process, it can be determined whether the algorithm has converged, thereby avoiding unnecessary iterations and improving the efficiency of the algorithm.

[0035] The super pixels 110 can be clustered by executing any of the following clustering algorithms, such as Laplacian sparse subspace clustering (LSSC) algorithm, K-means clustering algorithm, spectral clustering algorithm, hierarchical clustering algorithm, density-based clustering algorithm and bottom-up hierarchical clustering algorithm, but not limited to this. This article uses the LSSC algorithm as an example implementation of super pixel 110 clustering. The LSSC algorithm is an efficient high-dimensional data subspace clustering method. It is designed to process data that is roughly located in a low-dimensional subspace in a high-dimensional ambient space. The LSSC algorithm uses the sparsity assumption of data points in the subspace and uses graph Laplacian to capture the intrinsic structure of the data.

[0036] See also Figure 6 The detailed steps S602-S610 of the superpixel clustering process and Figure 7 Detailed steps S702-712 of the superpixel clustering process in FIG. At the beginning of the LSSC algorithm, feature vectors of the superpixels 110 are extracted respectively. The feature vectors may include color histograms, texture features, gradient features, or any other suitable descriptors that capture the features of the superpixels 110. The feature vectors may be obtained by:

[0037] Among them, u i represents the feature vector of the i-th superpixel, Γ i represents the number of pixels in the i-th superpixel, f jrepresents the feature vector of the ,th pixel, l represents the brightness value in the color space, a represents the a-axis to represent the color components from red to green, b represents the b-axis to represent the color components from yellow to blue, I X Represents the first-order derivative of the intensity of each pixel along the x-axis, I Y Represents the first-order derivative of each pixel along the y-axis, I Z Represents the first-order derivative of each pixel along the z-axis, I XX Represents the second-order derivative of each pixel along the x-axis, I YY represents the second-order derivative of each pixel along the y-axis, O ZZ represents the second-order derivative of each pixel along the z-axis, β represents the scaling factor, β×x represents the pixel position information along the x-axis, β×y represents the pixel position information along the y-axis, and β×z represents the pixel position information along the z-axis.

[0038] Then, the covariance matrix of each superpixel 110 is obtained as follows:

[0039] Among them, M i (a, b) represents the covariance matrix elements of features a and b in superpixel i.

[0040] A similarity matrix is ​​calculated between each superpixel to quantify the pairwise similarity between elements in the dataset. The similarity matrix is ​​calculated as follows:

[0041] in, represents the similarity between two superpixels i1 and i2, r1 and r2 represent different indices or identifiers of superpixels, The feature vector M representing superpixels i1 and i2 i1 and M i2 The distance or dissimilarity between superpixels i1 and i2, e represents the base of the natural logarithm, which is approximately equal to 2.71828.

[0042] A diagonal matrix is ​​then calculated that summarizes the similarity of a particular superpixel r′1 with all other superpixels in the multidimensional dataset 100. The diagonal matrix is ​​calculated as follows:

[0043]

[0044] Construct a weighted undirected graph with R nodes representing R feature vectors. Edges are connected to the nodes. Each edge has a weight representing the similarity value between a pair of nodes connected by the edge. The Laplacian matrix (L matrix) of the weighted undirected graph can be obtained by subtracting the diagonal matrix from the similarity matrix. The L matrix encodes the local and global structure of the data. In spectral clustering, the L matrix is ​​used to capture the graph structure of the data, where each superpixel is regarded as a node in the graph and the similarity between superpixels determines the weight of the edge. The L matrix can be expressed as:

[0045] L=ES

[0046] Then, a minimization problem constrained by the Laplacian matrix is ​​defined. By solving the minimization problem for each superpixel, the sparse coefficients for sparse coding can be obtained. The sparse coefficients can be obtained by the following formula:

[0047]

[0048]

[0049]

[0050] Aims to find the minimum and u i The difference While maintaining sparsity. i is used to The linear transformation matrix projected to the low-dimensional space, u i is the target vector in low-dimensional space. express and other data points The L1 norm of a vector is the sum of the absolute values ​​of its components. In other words, the L1 norm of a vector is the sum of the absolute values ​​of its components. S(i, i′) represents the similarity between i and i′ in the similarity matrix. Therefore, by optimizing the above objective function, the sparse coefficients can be obtained. Data reconstruction error, sparsity and similarity penalty are considered at the same time to achieve effective data representation.

[0051] The sparse coefficients also need to be converted into matrices for calculation. The sparse coefficient matrix can be expressed as:

[0052]

[0053] The similarity matrix is ​​updated according to the sparse coefficients to obtain a symmetric matrix, which is used to adjust the original similarity matrix to better reflect the similarity between data points. Usually, the original similarity matrix may not satisfy the symmetry. Therefore, by adjusting the original similarity matrix, a new matrix with symmetry is obtained, that is, a symmetric matrix.

[0054] The last step of the LSSC algorithm is to partition the weighted undirected graph according to the updated symmetric matrix to obtain superpixel clusters 111.

[0055] Each dictionary 300 is trained using the corresponding superpixel cluster, and the specific steps are as follows: the corresponding superpixel cluster is divided into a plurality of training data sub-blocks, the plurality of training data sub-blocks are respectively vectorized to obtain a plurality of initial representations, a similarity matrix of each pair of training data sub-blocks is calculated, and an average similarity value of each training data sub-block is calculated; according to the calculated average similarity value, a plurality of initial training data sub-blocks are selected from the plurality of training data sub-blocks, the initial training data sub-blocks are set as initial atoms of the dictionary, and the dictionary is initialized; the dictionary is trained by executing a batch orthogonal matching pursuit (batch OMP) algorithm constrained by a target sparsity and a target residual threshold, and a sparse coding representation is constructed for each training data sub-block; and the dictionary is updated atom by atom using the constructed sparse coding representation to obtain a trained dictionary.

[0056] Specifically, the batch OMP algorithm is iteratively executed in the following manner to construct a sparse coded representation for each training data subblock: a) calculating the Gram matrix of the dictionary; b) calculating the correlation vector of the dictionary with respect to the data subblock, where each element of the correlation vector corresponds to the correlation between the dictionary atom and the data subblock; c) finding the best matching atom in the dictionary with the maximum correlation with the data subblock; d) performing Cholesky decomposition on the subset of the Gram matrix containing the index of the best matching atom to obtain the Cholesky decomposition result; e) solving the sparse coded representation using the Cholesky decomposition result; f) updating the correlation vector by subtracting the product of the sparse coded representation and the Gram matrix; g) multiplying the dictionary with the sparse coded representation to obtain the reconstructed signal; and h) updating the residual between the reconstructed signal and the training data subblock. If the target sparsity is not reached, or the updated residual is greater than the target residual threshold, the above steps (c) to (h) are repeated. If the target sparsity is reached or the updated residual is less than the target residual threshold, the sparse coded representation of the training 2D / 3D data subblock is output for further operation.

[0057] In an exemplary embodiment of batch OMP, the dictionary is denoted as Φ and the signal (or data sub-block) is denoted as Gram matrix G = Φ T Φ, initial correlation vector Signal The square norm∈ 0 , is constructed based on the upper limit of the required sparsity K and the residual norm (square) threshold ∈. Before iteration, the index vector I is initialized to an empty vector, the lower triangular matrix L obtained by Cholesky decomposition is initialized and filled with 1 (i.e., unit vector), and the initial sparse representation vector All zeros, support vector (or related vector) is also initialized. In each iteration n, k is set to When n is greater than 1, the submatrix containing the atomic index in G is decomposed by Cholesky, denoted as G I,I , and its result is used to update L. Afterwards, the sparse representation vector γ is solved using the Cholesky decomposition result at the stored atomic index, that is, I is updated by adding k. Sparse representation vector the element at the given index in I Set to Right now In addition, renew Right now The update residual is also calculated / updated in is the reconstructed signal, δ n and δ n-1 are equal to the nth iteration and the n-1th iteration respectively. when When the number of non-zero elements of is less than K or the updated residual is greater than the standard residual threshold, the above iterations are repeated.

[0058] The method of decompressing multidimensional data 100 from a compressed file includes: Figure 8 Specifically, multiple compressed representations in the compressed file are decoded to obtain multiple decompressed data sub-blocks, and then the multiple decompressed data sub-blocks are connected to reconstruct the multi-dimensional data.

[0059] Specifically, each decompressed data sub-block is obtained by: retrieving a corresponding dictionary according to a corresponding index; multiplying the compressed representation with the retrieved dictionary to reconstruct a vectorized data sub-block; and devectorizing the reconstructed vectorized data sub-block to obtain a decompressed data sub-block.

[0060] The embodiments disclosed herein can be implemented using computing devices, computer processors or electronic circuits, including but not limited to application specific integrated circuits (ASICs), field programmable gate arrays (FPGAs), microcontrollers and other programmable logic devices configured or programmed according to the teachings of the present disclosure. Computer instructions or software codes running in computing devices, computer processors or programmable logic devices can be easily prepared by those skilled in the art of software or electronics according to the teachings of the present disclosure.

[0061] All or part of the methods according to various embodiments may be performed in one or more computing devices, including server computers, personal computers, laptop computers, and mobile computing devices, such as smartphones and tablet computers.

[0062] Embodiments of the present invention may include computer storage media, transient and non-transitory memory devices, in which computer instructions or software codes are stored, which can be used to program or configure a computing device, a computer processor or an electronic circuit to perform any method process of the present invention. The storage medium, transient and non-transitory memory devices may include, but are not limited to, floppy disks, optical disks, Blu-ray disks, DVDs, CD-ROMs and magneto-optical disks, ROMs, RAMs, flash memory devices, or any type of medium or device suitable for storing instructions, codes and / or data.

[0063] Embodiments of the present invention may also be implemented in a distributed computing environment and / or a cloud computing environment, where all or part of the machine instructions are executed in a distributed manner by one or more processing devices interconnected by a communication network, such as an intranet, a wide area network (WAN), a local area network (LAN), the Internet, and other forms of data transmission media.

[0064] The above description of the present invention is provided for the purpose of illustration and description. It is not intended to be exhaustive or to limit the present invention to the precise form disclosed. For those skilled in the art, many modifications and variations are apparent.

[0065] The embodiment was chosen and described in order to best explain the principles of the invention and its practical application, thereby enabling others skilled in the art to understand the invention for various embodiments and with various modifications as are suited to the particular use contemplated.

Claims

1. A method for compressing multidimensional data having a plurality of data points into a compressed file, comprising: generating a plurality of superpixels from the multidimensional data; Clustering the plurality of superpixels into a plurality of superpixel clusters; Using the plurality of superpixel clusters to train a plurality of dictionaries respectively; dividing the multidimensional data into a plurality of data sub-blocks; Vectorizing the multiple data sub-blocks to obtain multiple vectorized data sub-blocks; Select a dictionary for each vectorized data sub-block; Sparsely encode each vectorized data sub-block using the corresponding selected dictionary to obtain a compressed representation; as well as storing the resulting compressed representation and the corresponding index of the selected dictionary in the compressed file; where each dictionary is selected for the corresponding sub-block of vectorized data in the following way: Calculating the similarity between the corresponding vectorized data sub-block and each superpixel cluster; Find the best matching superpixel cluster with the highest similarity to the vectorized data sub-block; Selecting the dictionary trained with the best matching superpixel cluster as the encoding dictionary; and wherein each dictionary is trained with the corresponding superpixel cluster in the following manner: Dividing the corresponding superpixel cluster into a plurality of training data sub-blocks; vectorizing the plurality of training data sub-blocks to obtain a plurality of initial representations respectively; Calculate the similarity matrix between each pair of training data sub-blocks; Calculate the average similarity value of each training data sub-block; selecting a plurality of initial training data sub-blocks from the plurality of training data sub-blocks according to the calculated average similarity value; Initializing the dictionary by setting the initial training data sub-block as an initial atom of the dictionary; Training the dictionary to construct a sparse coded representation for each training data sub-block by executing a batch orthogonal matching pursuit (batch OMP) algorithm with a target sparsity and a target residual threshold as constraints; and The dictionary is updated atom by atom using the constructed sparse coding representation to obtain a trained dictionary.

2. The method of claim 1, wherein executing the batch OMP algorithm constructs a sparse coded representation for each training data sub-block in the following manner: (a) calculating the Gram matrix of the dictionary; (b) calculating a correlation vector of the dictionary with respect to the data sub-block, wherein each element of the correlation vector corresponds to a correlation between an atom of the dictionary and the data sub-block; (c) finding the best matching atom in the dictionary that has the greatest correlation with the data sub-block; (d) performing Cholesky decomposition on a subset of the Gram matrix containing the index of the best matching atom to obtain a Cholesky decomposition result; (e) solving the sparse coding representation using the Cholesky decomposition result; (f) updating the correlation vector by subtracting the product of the sparse coded representation and the Gram matrix; (g) multiplying the dictionary by the sparse coding representation to obtain a reconstructed signal; (h) updating the residual between the reconstructed signal and the training data sub-block; (i) if the updated residual is greater than the target residual threshold, repeating steps (c) to (h); as well as (j) If the target sparsity is reached or the updated residual is less than the target residual threshold, outputting the sparse coded representation of the training data sub-block for further operation.

3. The method according to claim 1, wherein: The plurality of superpixels are generated by: initializing a plurality of superpixel centers at a maximum interval; Updating the positions of the multiple superpixel centers; and (a) defining a plurality of overlapping superpixel regions centered on the plurality of superpixel centers and corresponding to the plurality of superpixel centers respectively, wherein the size of each cluster region is equal to twice the maximum interval between the superpixel centers; (b) for each multidimensional data point: within each corresponding overlapping superpixel region, calculating a plurality of distances between a plurality of adjacent superpixel centers and the multidimensional data point; and assigning the multidimensional data point to a superpixel region containing a nearest adjacent superpixel having the shortest distance to the multidimensional data point; (c) for each superpixel center: updating the current position of the superpixel center to the center point position of all multidimensional data points assigned to the corresponding superpixel region; and (d) calculating the residual value between the updated position of each superpixel center and the current position; (e) If any residual value is greater than the threshold, repeat steps (b) and (d).

4. The method according to claim 3, wherein: The distances from the neighboring superpixel centers to the multidimensional data points are calculated based on the color distance and spatial distance from the neighboring superpixel centers to the image pixels.

5. The method according to claim 4, wherein the plurality of superpixels are clustered by executing a Laplace sparse subspace clustering algorithm; the Laplace sparse subspace clustering algorithm is executed in the following manner: Extracting a plurality of feature vectors of the plurality of superpixels respectively; Get the covariance matrix of each superpixel; Get the similarity value of each pair of superpixels; Constructing a weighted undirected graph, the weighted undirected graph having a plurality of nodes representing a plurality of feature vectors and edges connecting the nodes; wherein, The weight of each edge represents the similarity value between a pair of nodes connected by the edge; The weighted undirected graph is segmented to obtain a plurality of superpixel clusters.

6. The method according to claim 5, wherein the feature vector of each superpixel is the average of the feature vectors of all pixels in the superpixel.

7. The method according to claim 6, wherein the superpixel C i The covariance matrix M i It is given by the following formula: in, M i (a,b) represents the covariance matrix element of features a and b in superpixel i, f j and u i are the pixel feature vector and mean feature vector in superpixel j, respectively, and f j (a′) represents the a′th element of the eigenvector, u i (a″) represents the a″th element of the mean eigenvector, f j (b′) represents the b′th element of the eigenvector, u i (b″) represents the b″th element of the mean eigenvector. 8 . The method according to claim 7 , wherein the similarity value between superpixel i1 and superpixel i2 is calculated based on the distance between the feature vectors of the superpixels i1 and i2 .

9. The method according to claim 8, wherein: The weighted undirected graph is partitioned in the following way: Obtaining the Laplace matrix of the weighted undirected graph; defining a minimization problem subject to the Laplacian matrix constraints; Solving the minimization problem for each superpixel to obtain a sparse coefficient of the superpixel; Construct a sparse coefficient matrix for a superpixel of R sub-pixels; Updating a symmetric matrix using the sparse coefficient matrix; The weighted undirected graph is partitioned according to the updated symmetric matrix.

10. According to the method of claim 9, the minimization problem is defined by the sparse coefficients, a linear transformation matrix for projecting the sparse coefficients into a low-dimensional space, a target vector in the low-dimensional space, and an L1 norm of the sparse coefficients and the sparse coefficients of other data points.

11. A method for decompressing multidimensional data from a compressed file obtained by the method of claim 1, the method comprising: decoding the plurality of compressed representations to obtain a plurality of decompressed data sub-blocks; as well as concatenating a plurality of decompressed data sub-blocks to reconstruct multi-dimensional data; as well as Each decompressed data sub-block is obtained in the following way: Retrieve the corresponding dictionary according to the corresponding index; multiplying the compressed representation with the retrieved dictionary to reconstruct the vectorized data sub-block; and The reconstructed vectorized data sub-block is devectorized to obtain the decompressed data sub-block.

12. The method of claim 11, wherein executing the batch OMP algorithm constructs a sparse coded representation for each training data sub-block by: (a) Calculate the Gram matrix of the dictionary; (b) calculating a correlation vector of the dictionary with respect to the data sub-block, wherein each element of the correlation vector corresponds to a correlation between an atom of the dictionary and the data sub-block; (c) finding the best matching atom in the dictionary that has the greatest correlation with the data sub-block; (d) performing Cholesky decomposition on a subset of the Gram matrix containing the index of the best matching atom to obtain a Cholesky decomposition result; (e) solving the sparse coding representation using the Cholesky decomposition result; (f) updating the correlation vector by subtracting the product of the sparse coded representation and the Gram matrix; (g) multiplying the dictionary by the sparse coding representation to obtain a reconstructed signal; (h) updating the residual between the reconstructed signal and the training data sub-block; (i) if the updated residual is greater than the target residual threshold, repeat steps (c) to (h); and (j) If the target sparsity is reached or the updated residual is less than the target residual threshold, outputting the sparse coded representation of the training data sub-block for further operation.

13. The method according to claim 11, wherein the plurality of superpixels are generated by: initializing a plurality of superpixel centers at a maximum interval; Updating the positions of the multiple superpixel centers; and (a) defining a plurality of overlapping superpixel regions centered on the plurality of superpixel centers and corresponding to the plurality of superpixel centers respectively, wherein the size of each cluster region is equal to twice the maximum interval between the superpixel centers; (b) for each multidimensional data point: within each corresponding overlapping superpixel region, calculating a plurality of distances between a plurality of adjacent superpixel centers and the multidimensional data point; and assigning the multidimensional data point to a superpixel region containing a nearest adjacent superpixel having the shortest distance to the multidimensional data point; (c) for each superpixel center: updating the current position of the superpixel center to the center point position of all multidimensional data points assigned to the corresponding superpixel region; and (d) Calculate the residual value between the updated position and the current position of each superpixel center; (e) If any residual value is greater than the threshold, repeat steps (b) and (d).

14. The method according to claim 13, wherein: The distance from the center of an adjacent superpixel to a multidimensional data point is calculated based on the color distance and spatial distance from the center of the adjacent superpixel to an image pixel.

15. The method according to claim 14, wherein the plurality of superpixels are clustered by executing a Laplace sparse subspace clustering algorithm; the Laplace sparse subspace clustering algorithm is executed in the following manner: Extracting a plurality of feature vectors of the plurality of superpixels respectively; Get the covariance matrix of each superpixel; Get the similarity value of each pair of superpixels; Constructing a weighted undirected graph, the weighted undirected graph having a plurality of nodes representing the plurality of feature vectors and edges connecting the nodes; wherein, The weight of each edge represents the similarity value between a pair of nodes connected by the edge; The weighted undirected graph is segmented to obtain a plurality of superpixel clusters.

16. The method according to claim 15, wherein the feature vector of each superpixel is the average of the feature vectors of all pixels in the superpixel.

17. The method according to claim 16, wherein the superpixel C i The covariance matrix M i It is given by the following formula: in, M i (a,b) represents the covariance matrix element of features a and b in superpixel i, f j and u i are the pixel feature vector and mean feature vector in superpixel j, respectively, and f j (a′) represents the a′th element of the eigenvector, u i (a″) represents the a″th element of the mean eigenvector, f j (b') represents the b'th element of the eigenvector, u i (b″) represents the b″th element of the mean eigenvector. 18 . The method according to claim 17 , wherein the similarity value between superpixel i1 and superpixel i2 is calculated based on the distance between the feature vectors of the superpixels i1 and i2 .

19. The method according to claim 18, wherein: The weighted undirected graph is partitioned in the following way: Obtaining the Laplace matrix of the weighted undirected graph; defining a minimization problem subject to the Laplacian matrix constraints; Solving the minimization problem for each superpixel to obtain a sparse coefficient of the superpixel; Construct a sparse coefficient matrix for a superpixel of R sub-pixels; Updating a symmetric matrix using the sparse coefficient matrix; The weighted undirected graph is partitioned according to the updated symmetric matrix.

20. According to the method of claim 19, the minimization problem is defined by the sparse coefficients, a linear transformation matrix for projecting the sparse coefficients into a low-dimensional space, a target vector in the low-dimensional space, and an L1 norm of the sparse coefficients and the sparse coefficients of other data points.