Microscopic measurement data compression sensing method

Abstract

The application provides a microscopic measurement data compression sensing method. The method can be used in the technical field of digital signal processing. The method comprises the following steps: after multiple classification data are obtained by classifying microscopic measurement data according to a BP neural network, sparse transformation is performed on each classification data according to a sparse transformation base in a pre-set sparse transformation base set, and a target sparse matrix corresponding to each classification data is determined in the obtained multiple sparse matrices; target measurement vectors are obtained by performing compression processing on the target sparse matrix corresponding to each classification data; finally, the target measurement vectors corresponding to each classification data are reconstructed by a pre-set reconstruction algorithm to obtain target reconstruction data, the target reconstruction data are reorganized to obtain reconstructed microscopic measurement data, and high-quality processing of the microscopic measurement data is realized.

Description

Technical Field

[0001] This application relates to digital signal processing technology, and more particularly to a method for compressed sensing of microscopic measurement data. Background Technology

[0002] With the continuous development and advancement of information technology, big data has become a hot topic in today's society. Big data refers to the process of collecting, storing, managing, analyzing, and applying massive, high-speed, diverse, and heterogeneous data through a series of technical means. The emergence of big data technology has provided us with more opportunities and challenges.

[0003] In existing technologies, the sampling process for data acquisition generally follows the Nyquist sampling theorem. Then, according to a certain procedure, data is transmitted from the data source to the data terminal through one or more data links, realizing point-to-point information transmission and exchange. Microscopic measurement data can be used to inspect or repair chips or microelectromechanical systems (MEMS) devices, thus creating a greater demand for processing massive amounts of microscopic measurement data.

[0004] In summary, how to achieve high-quality processing of microscopic measurement data is a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0005] This application provides a compressed sensing method for microscopic measurement data to address the problem of how to achieve high-quality processing of microscopic measurement data.

[0006] In a first aspect, this application provides a method for compressed sensing of microscopic measurement data, comprising:

[0007] The microscopic measurement data were classified using a BP neural network, resulting in multiple categories of data, each with a different frequency.

[0008] Based on the sparse transformation basis in the pre-defined sparse transformation basis set, sparse transformation is performed on each category of data to obtain multiple sparse matrices corresponding to each category of data. The sparse transformation basis set includes multiple sparse transformation bases.

[0009] For each category of data, a target sparse matrix is ​​determined from multiple sparse matrices corresponding to the category data. The target sparse matrix is ​​the matrix with the largest sparsity among the multiple sparse matrices.

[0010] The target sparse matrix corresponding to each category of data is compressed to obtain the target measurement vector;

[0011] The target measurement vector corresponding to each category of data is reconstructed using a preset reconstruction algorithm to obtain the target reconstruction data;

[0012] The target reconstruction data corresponding to each category of data is reorganized to obtain reconstructed microscopic measurement data.

[0013] In conjunction with the first aspect, in some embodiments, the compression of the target sparse matrix corresponding to each classification data to obtain the target measurement vector includes:

[0014] Based on the target sparse matrix, constraint isometry, and pre-defined matching model corresponding to each category of data, the target measurement matrix corresponding to each category of data is determined.

[0015] For each category of data, the target measurement vector is obtained by performing an inner product calculation on the target sparse matrix and the target measurement matrix corresponding to the category data.

[0016] The matching model is a model that calculates the matching degree between each category of data and each corresponding measurement matrix, and the constraint equidistant condition is used to filter the target measurement matrix corresponding to each category of data.

[0017] In conjunction with the first aspect, in some embodiments, determining the target measurement matrix corresponding to each category of data based on the target sparse matrix, constraint isometry, and a pre-defined matching model includes:

[0018] The correlation coefficient between the target sparse matrix and the measurement matrix corresponding to each category of data is solved until it approaches 0, thus obtaining the first set of measurement matrices corresponding to each category of data. The first set of measurement matrices includes all measurement matrices obtained by solving the correlation coefficient approaching 0.

[0019] The first set of measurement matrices that satisfy the constraint equidistance condition in the first set of measurement matrices corresponding to each category of data is determined as the second set of measurement matrices corresponding to each category of data.

[0020] Using the matching model, the matching degree is calculated for each category data and each measurement matrix in the corresponding second measurement matrix set to obtain all matching degrees for each category data.

[0021] The measurement matrix corresponding to the largest matching degree among all matching degrees for each category of data is determined as the target measurement matrix.

[0022] In conjunction with the first aspect, in some embodiments, the method further includes:

[0023] The target measurement vector corresponding to each category data is encoded using the LZW string table compression algorithm to obtain a new target measurement vector after compression.

[0024] Accordingly, the step of reconstructing the target measurement vector corresponding to each category data using a preset reconstruction algorithm to obtain target reconstruction data includes:

[0025] The new target measurement vector corresponding to each category of data is reconstructed using the preset reconstruction algorithm to obtain the target reconstruction data.

[0026] In conjunction with the first aspect, in some embodiments, the step of reconstructing the target measurement vector corresponding to each classification data using a preset reconstruction algorithm to obtain target reconstruction data includes:

[0027] For each category of data, based on pre-set constraints, the target measurement vector corresponding to the category of data is iteratively calculated using a dictionary learning algorithm to obtain a dictionary matrix and a coefficient matrix;

[0028] The inner product of the dictionary matrix and the coefficient matrix corresponding to each category of data is calculated to obtain the target reconstructed data.

[0029] The reconstruction algorithm includes the dictionary learning algorithm.

[0030] In conjunction with the first aspect, in some embodiments, the method further includes:

[0031] The dictionary matrix and coefficient matrix corresponding to each category of data are optimized by an optimization model to obtain the optimized dictionary matrix and coefficient matrix. The optimization model is obtained by training a Bayesian sparse model based on a pre-acquired set of dictionary matrices and coefficient matrices, and is used to optimize the dictionary matrix and coefficient matrix.

[0032] Accordingly, the step of performing an inner product calculation on the dictionary matrix and the corresponding coefficient matrix for each category of data to obtain the target reconstructed data includes:

[0033] The inner product of the optimized dictionary matrix and the optimized coefficient matrix corresponding to each category of data is calculated to obtain the target reconstructed data.

[0034] In conjunction with the first aspect, in some embodiments, before performing sparse transformation on each category data according to the sparse transformation basis in the pre-set sparse transformation basis set to obtain multiple sparse matrices corresponding to each category data, the method further includes:

[0035] The multiple categories of data are labeled with their locations to obtain the location number corresponding to each category of data.

[0036] Accordingly, the reorganization of the target reconstruction data corresponding to each category of data to obtain reconstructed microscopic measurement data includes:

[0037] Based on the location number corresponding to each category of data, the target reconstruction data corresponding to each category of data is reorganized to obtain reconstructed microscopic measurement data.

[0038] In conjunction with the first aspect, in some embodiments, the set of sparse transform bases includes at least the following sparse transform bases:

[0039] The first sparse transform basis is obtained by combining the wavelet basis with the discrete cosine transform (DCT) basis;

[0040] The second sparse transform basis obtained by combining the wavelet basis and the Curvelet basis; and...

[0041] The third sparse transform basis is obtained by combining the wavelet basis and the Shearlet basis.

[0042] In conjunction with the first aspect, in some embodiments, determining the target sparse matrix from multiple sparse matrices corresponding to each category of data includes:

[0043] For each category of data, the sparsity of each sparse matrix corresponding to the category of data is calculated according to the sparsity calculation formula.

[0044] From the multiple sparse matrices corresponding to the classification data, the sparse matrix with the largest sparsity is determined as the target sparse matrix.

[0045] Secondly, this application provides a microscopic measurement data compression sensing device, comprising:

[0046] The classification model is used to classify microscopic measurement data based on a BP neural network, resulting in multiple classification data, each with a different frequency.

[0047] The sparse module is used to perform sparse transformation on each category of data according to the sparse transformation basis in the pre-defined sparse transformation basis set, so as to obtain multiple sparse matrices corresponding to each category of data. The sparse transformation basis set includes multiple sparse transformation bases.

[0048] The determination module is used to determine a target sparse matrix from multiple sparse matrices corresponding to each type of data, wherein the target sparse matrix is ​​the matrix with the largest sparsity among the multiple sparse matrices;

[0049] The compression module is used to compress the target sparse matrix corresponding to each category of data to obtain the target measurement vector;

[0050] The reconstruction module is used to reconstruct the target measurement vector corresponding to each category of data using a preset reconstruction algorithm to obtain the target reconstruction data.

[0051] The reconstruction module is used to reconstruct the target reconstruction data corresponding to each category of data to obtain reconstructed microscopic measurement data.

[0052] This application provides a compressed sensing method for microscopic measurement data. After classifying the microscopic measurement data into multiple categories using a BP neural network, a sparse transformation is performed on each category of data according to a pre-set sparse transformation basis set. A target sparse matrix corresponding to each category of data is determined from the resulting multiple sparse matrices. The target sparse matrix corresponding to each category of data is then compressed to obtain a target measurement vector. Finally, the target measurement vector corresponding to each category of data is reconstructed using a pre-set reconstruction algorithm to obtain target reconstructed data. This reconstructed data is then reassembled to obtain reconstructed microscopic measurement data, thus achieving high-quality processing of the microscopic measurement data. Attached Figure Description

[0053] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0054] Figure 1 This is an application scenario diagram of the microscopic measurement data compressed sensing method provided in the embodiments of this application;

[0055] Figure 2 A schematic flowchart of an embodiment of the microscopic measurement data compression sensing method provided in this application.

[0056] Figure 3 A schematic flowchart of Embodiment 2 of the microscopic measurement data compression sensing method provided in this application;

[0057] Figure 4 A schematic flowchart of Embodiment 3 of the microscopic measurement data compression sensing method provided in this application;

[0058] Figure 5 A schematic flowchart of Embodiment 4 of the microscopic measurement data compression sensing method provided in this application;

[0059] Figure 6 A flowchart illustrating Embodiment 5 of the microscopic measurement data compression sensing method provided in this application;

[0060] Figure 7 This is a schematic diagram of the structure of a first embodiment of the microscopic measurement data compression sensing device provided in this application.

[0061] Figure 8 This is a schematic diagram of the structure of a second embodiment of the microscopic measurement data compression sensing device provided in this application.

[0062] Figure 9 A schematic diagram of the structure of the microscopic measurement data compression sensing device according to Embodiment 3 of this application;

[0063] Figure 10 A schematic diagram of the structure of Embodiment 4 of the microscopic measurement data compression sensing device provided in this application;

[0064] Figure 11 A schematic diagram of the structure of Embodiment 6 of the microscopic measurement data compression sensing device provided in this application;

[0065] Figure 12 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application.

[0066] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0067] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0068] Electronic chips and microelectromechanical systems (MEMS) devices are fundamental components in modern aerospace, automotive, and consumer electronics industries. Due to differences in materials and structures, the main defects exhibited by these two types of devices are also different. Therefore, to ensure product quality, it is necessary to measure their surface morphology to obtain microscopic measurement data, and then perform inspection and repair based on this data. Thus, the processing requirements for microscopic measurement data in this field are higher. Current technologies for processing massive amounts of microscopic measurement data generally follow the Nyquist sampling theorem for sampling, and then transmit the data from the data source to the data terminal through one or more data links according to a certain procedure, realizing point-to-point information transmission and exchange. The data terminal then performs defect repair processing based on the received data. However, this method carries the risk of data loss during acquisition, transmission, and storage, and the storage of massive amounts of data also places significant memory demands on storage devices, leading to a substantial reduction in the processing quality of microscopic measurement data.

[0069] To address the aforementioned problems, this application provides a compressed sensing method for microscopic measurement data, achieving high-quality processing of microscopic measurement data. Specifically, traditional Nyquist-based signal acquisition requires maintaining a high sampling rate, which generates a large amount of redundant data, severely consuming network bandwidth and memory. For massive amounts of microscopic measurement data, existing processing methods significantly impact processing quality and also pose a risk of data loss. Considering these issues, the inventors investigated whether it is possible to perform sparse processing on massive amounts of microscopic measurement data by designing a highly adaptable sparse basis, compressing the data by designing a measurement matrix, and finally reconstructing and restoring the data based on a high-quality reconstruction algorithm. This reduces redundant data, improves data transmission efficiency, saves storage space, and ultimately achieves high-quality processing of microscopic measurement data. Based on this, the technical solution of this application is proposed.

[0070] The technical methods of this application embodiment can be applied to fields such as radar, radio communication, and sound acquisition, as well as video surveillance, medical imaging, and autonomous driving. It should also be understood that the technical solutions of this application embodiment can be applied to big data processing scenarios, such as the processing of microscopic measurement data.

[0071] Figure 1 This is an application scenario diagram of the microscopic measurement data compressed sensing method provided in the embodiments of this application, such as... Figure 1 As shown, this scenario includes at least a data acquisition device and a data processor. The data acquisition device can be a confocal microscope, digital camera, spectrometer, 3D scanner, or other device capable of acquiring massive amounts of microscopic measurement data. The data processor can be an electronic device such as a server, smartphone, laptop, or desktop computer capable of processing and storing massive amounts of data. The data acquisition device can compress the acquired massive amounts of microscopic measurement data using compressed sensing technology before transmitting it to the data processor. The data processor can then reconstruct and analyze the compressed data.

[0072] It should be noted that the specific form, model, and quantity of the physical equipment mentioned above are not specifically limited.

[0073] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.

[0074] Figure 2 A schematic flowchart of Embodiment 1 of the microscopic measurement data compression sensing method provided in this application is shown below. Figure 2 As shown, the method includes:

[0075] S201: Classify the microscopic measurement data according to the back propagation (BP) neural network to obtain multiple classification data.

[0076] In this step, in order to improve the processing efficiency of massive amounts of microscopic measurement data, the microscopic measurement data can be preprocessed.

[0077] Specifically, a backpropagation (BP) neural network is used to classify microscopic measurement data into multiple categories based on their frequency. A massive amount of microscopic measurement data is pre-acquired and divided into training and testing sets according to the data's frequency. The basic BP neural network is then trained based on the training and testing sets to obtain the BP neural network.

[0078] For example, an adaptive BP neural network is trained based on a training set. By adjusting the network parameters, the model can better fit the data. After training, the classification accuracy of the trained model is evaluated based on a test set. Then, an accuracy threshold is set. The parameters of the trained model are further optimized based on the accuracy calculated from the test set until the accuracy threshold is reached, at which point the test ends and the target model, i.e., the BP neural network, is obtained. The classification data, i.e., the microscopic measurement data, is then classified through the target model to obtain multiple classification data.

[0079] It should be noted that the above example of using an adaptive BP neural network for training is just one example. Fully connected BP neural networks or recursive BP neural networks can also be used as the base BP neural network for training. This embodiment does not specifically limit the base BP neural network used for training.

[0080] It should be noted that the parameters adjusted during the training of the basic BP neural network include the learning rate, momentum factor, number of iterations, number of hidden layer nodes, and number of layers.

[0081] Optionally, regularization and batch normalization can be performed during model training to improve the model's classification accuracy.

[0082] S202: Based on the sparse transformation basis in the pre-defined sparse transformation basis set, perform sparse transformation on each category of data to obtain multiple sparse matrices corresponding to each category of data.

[0083] In this step, after obtaining multiple classification data, in order to improve data processing efficiency, sparse transformations are performed on each classification data based on multiple pre-designed sparse transformation bases, thereby obtaining multiple sparse matrices corresponding to each classification data.

[0084] Specifically, multiple sparse transformation bases are pre-designed and then aggregated into a sparse transformation base set. Each sparse transformation base is used to perform a linear transformation on each category of data, thereby obtaining multiple sparse matrices corresponding to each category of data.

[0085] Optionally, the set of sparse transform bases includes at least the following sparse transform bases: a first sparse transform base obtained by combining a wavelet base with a Discrete Cosine Transform (DCT) base; a second sparse transform base obtained by combining a wavelet base with a Curvelet base; and a third sparse transform base obtained by combining a wavelet base with a Shearlet base.

[0086] Preferred: A set of wavelet bases can be a combination of the Daubechies wavelet base and the Haar wavelet base. The Daubechies wavelet base is:

[0087]

[0088] Where N is the wavelet basis order, The expression represents the number of combinations, i.e., the number of combinations to be selected from n elements, and θ(t) is the scaling function.

[0089] The Haar wavelet basis is:

[0090]

[0091] The combination of the Daubechies wavelet basis and the Haar wavelet basis yields a set of wavelet bases as follows:

[0092] Ψ(T)=Aφ1(t)+Bφ2(t)

[0093] By combining the Daubechies and Haar wavelet bases mentioned above, a set of wavelet bases can be obtained, which can improve the accuracy and sparsity of signal representation. Since different wavelet bases have different characteristics in the frequency domain, combining them into a set of wavelet bases can better adapt to the frequency characteristics of various signals.

[0094] The first sparse transform basis is obtained by combining the aforementioned set of wavelet bases with the DCT basis, and the DCT basis is expressed as:

[0095]

[0096] Where, x n The original sequence, i.e., categorical data, is given by N, where N is the sequence length and X is the value of X. k These are the coefficients after DCT transformation.

[0097] Then the first sparse transformation basis can be expressed as ψ1=αΨ(T)+βX kIn this equation, α and β are weighting coefficients. The first sparse transform basis can better adapt to non-stationary signals and provide better time-frequency localization characteristics. It can also have good energy concentration for stationary signals, which can concentrate the signal energy on a few coefficients as much as possible, providing a higher signal compression ratio while maintaining good reconstruction quality.

[0098] The second sparse transform basis is obtained by combining the above set of wavelet bases with the Curvelet basis. The Curvelet basis is expressed as:

[0099]

[0100] Where, φ j,l (x,y) is a wavelet basis function with scale j and angle l, B j,l,k (x,y) represents the relationship between φ j,l The function obtained after performing a local Fourier transform, where k represents the position of the function in terms of frequency.

[0101] The second sparse transformation basis can then be expressed as ψ2=αΨ(T)+βB j,l,k (m,n), where α and β are weighting coefficients, the second sparse transform basis can better adapt to non-stationary signals and provide better time-frequency localization characteristics. It can also capture detailed information about the signal at multiple scales, thus providing a more comprehensive signal representation. Furthermore, it can capture not only the frequency characteristics of the signal but also its variations in different directions. It represents directional features such as edges and textures in the signal through the selective concentration of transform coefficients.

[0102] The third sparse transform basis is obtained by combining the above set of wavelet bases with the Shearlet basis. The Shearlet basis is expressed as:

[0103] C j,k,l (x,y)=2 j / 2 C(2 j xk,2 j yl)

[0104] Where j represents the scale, (k, l) represents the position, and C represents the fundamental decomposition function.

[0105] Then the third sparse transformation basis can be expressed as ψ3=αΨ(T)+βC j,k,l (x,y), where α and β are weighting coefficients. The third sparse transform basis can better adapt to non-stationary signals and provide better time-frequency localization characteristics. At the same time, it can also obtain multi-scale and multi-directional representations of signals, thereby better understanding and processing complex signals.

[0106] It should be noted that α and β in the above formulas for the first sparse transformation basis, the second sparse transformation basis, and the third sparse transformation basis are all weight coefficients. The weight coefficients in the three formulas can be the same or different, or they can be weight matrices. The specific values ​​of α and β can be preset according to the actual data characteristics being processed. This embodiment does not impose any specific limitations.

[0107] In one possible implementation, the above set of wavelet bases is just an example. In practical applications, other wavelet bases can be selected and combined according to the characteristics of the data to obtain a set of wavelet bases, such as Symlet wavelet, Coiflet wavelet, etc., and the combination of wavelet bases is not limited to three types.

[0108] S203: For each category of data, determine the target sparse matrix from the multiple sparse matrices corresponding to the category data.

[0109] In this step, in order to save data storage space and improve data transmission efficiency without losing data, after obtaining multiple sparse matrices for each category of data, it is necessary to determine the target sparse matrix from these multiple sparse matrices.

[0110] Specifically, for each category of data, the sparsity of each sparse matrix corresponding to the category data is calculated according to the sparsity calculation formula, and the sparse matrix with the largest sparsity among the multiple sparse matrices corresponding to the category data is determined as the target sparse matrix.

[0111] S204: Compress the target sparse matrix corresponding to each classification data to obtain the target measurement vector.

[0112] In this step, after obtaining the target sparse matrix corresponding to each category of data, in order to achieve high data transmission efficiency and save data storage space, compressed sensing technology is used to compress the target sparse matrix corresponding to each category of data, thereby obtaining the target measurement vector.

[0113] Specifically, firstly, based on the target sparse matrix corresponding to each category of data, the constraint equidistant condition, and the pre-set matching model, the target measurement matrix corresponding to each category of data is determined. Then, the target measurement vector is obtained by performing an inner product calculation on the target sparse matrix corresponding to each category of data and the target measurement matrix corresponding to each category of data.

[0114] S205: Reconstruct the target measurement vector corresponding to each category data using a preset reconstruction algorithm to obtain the target reconstruction data.

[0115] In this step, each category of data is sampled and compressed to obtain the target measurement vector, saving storage space and improving data transmission efficiency. In order to ensure the integrity and accuracy of the data, it is also necessary to reconstruct each category of data. Therefore, a reconstruction algorithm is preset. After obtaining the target measurement vector corresponding to each category of data, the reconstruction calculation is performed through the preset reconstruction algorithm to obtain the target reconstructed data.

[0116] Specifically, for each category of data, based on pre-set constraints, the target measurement vector corresponding to the category data is iteratively calculated using a dictionary learning algorithm to obtain the target dictionary matrix and coefficient matrix. The inner product of the target dictionary matrix and coefficient matrix corresponding to each category of data is then calculated, and the result is the target reconstruction data.

[0117] S206: Reorganize the target reconstruction data corresponding to each category data to obtain reconstructed microscopic measurement data.

[0118] In this step, after obtaining the target reconstruction data corresponding to each classification data, in order to facilitate subsequent image model generation or chip device repair based on the data, it is necessary to reassemble each target reconstruction data to obtain reconstructed microscopic measurement data.

[0119] Specifically, after the microscopic measurement data is classified, each category of data needs to be identified by its location. Then, after obtaining the target reconstruction data corresponding to each category of data, the target reconstruction data corresponding to each category of data is reorganized according to the location number corresponding to each category of data to obtain the reconstructed microscopic measurement data.

[0120] This embodiment provides a compressed sensing method for microscopic measurement data. After classifying the microscopic measurement data into multiple categories using a BP neural network, a sparse transformation is performed on each category of data according to a pre-set sparse transformation basis. A target sparse matrix corresponding to each category of data is determined from the resulting multiple sparse matrices. The target sparse matrix corresponding to each category of data is then compressed to obtain a target measurement vector. Finally, the target measurement vector corresponding to each category of data is reconstructed using a pre-set reconstruction algorithm to obtain target reconstructed data. This reconstructed data is then reassembled to obtain reconstructed microscopic measurement data. This achieves high-quality processing of microscopic measurement data. Furthermore, the BP neural network classification improves data processing efficiency, and the use of different sparse transformation bases for data processing results in higher subsequent reconstruction accuracy.

[0121] Figure 3 A schematic flowchart of Embodiment 2 of the microscopic measurement data compression sensing method provided in this application is shown below. Figure 3As shown, based on the above embodiments, the specific implementation of step S204 may include:

[0122] S301: Determine the target measurement matrix corresponding to each category of data based on the target sparse matrix, constraint isometry, and pre-defined matching model.

[0123] In this step, after obtaining the target sparse matrix corresponding to each category of data, a measurement matrix is ​​designed based on constraints to improve the processing efficiency and quality of the microscopic measurement data. Then, a pre-defined matching model is used to determine the target measurement matrix corresponding to each category of data, thereby improving the accuracy of data reconstruction and the compression rate of the categorical data. The matching model calculates the matching degree between each category of data and its corresponding measurement matrix, and the equidistant constraint is used to filter the target measurement matrix corresponding to each category of data.

[0124] Specifically, the constraints include the correlation coefficient between the target sparse matrix and the measurement matrix, and the constraint of equidistant distance. The correlation coefficient is expressed as:

[0125]

[0126] Where x is the target sparse matrix and y is the measurement matrix.

[0127] The constraint equidistant condition is expressed as:

[0128]

[0129] Where x is the target sparse matrix, y is the measurement matrix, and δ is a constant, then there exists a constant δ such that 0 < δ < 1. k This ensures that the measurement matrix y satisfies the above formula with respect to the target sparse matrix.

[0130] By solving for the correlation coefficient between the target sparse matrix and the measurement matrix corresponding to each category of data approaching 0, the first set of measurement matrices corresponding to each category of data is obtained. Then, the measurement matrices in the first set of measurement matrices corresponding to each category of data are filtered according to the constraint equidistant condition to obtain the second set of measurement matrices corresponding to each category of data. Finally, the second set of measurement matrices corresponding to each category of data is input into the matching model to obtain all matching degrees corresponding to each category of data. For each category of data, the measurement matrix corresponding to the largest matching degree is determined as the target measurement matrix corresponding to the category of data.

[0131] Optionally, after filtering the measurement matrix using the constrained equidistant conditions described above, the measurement matrix can be further optimized using the second-order difference method to reduce its correlation. Specifically, the second-order difference method is used to calculate the difference between adjacent elements in each column of the measurement matrix, and the elements with larger differences are retained. Then, the differenced measurement matrix is ​​normalized so that its elements range between 0 and 1. The normalization method can be implemented using linear transformation or other existing techniques, which will not be elaborated here.

[0132] It should be noted that the normalized difference measurement matrix can be further adjusted. For example, the correlation can be changed by adjusting certain elements of the difference measurement matrix or by performing operations such as translation or rotation on the entire matrix. The above adjustment steps can be repeated multiple times until the actual requirements are met.

[0133] S302: For each category of data, the target measurement vector is obtained by performing an inner product calculation on the target sparse matrix and the target measurement matrix corresponding to the category data.

[0134] In this step, after determining the target measurement matrix corresponding to each category of data through the above-mentioned constraints and matching model calculations, the target sparse matrix corresponding to each category of data is compressed according to the target measurement matrix corresponding to the category data to obtain compressed data, which is the target measurement vector.

[0135] Specifically, for each category of data, the target measurement vector is obtained by performing an inner product calculation between the target measurement matrix corresponding to the category data and the target sparse matrix corresponding to the category data.

[0136] This embodiment provides a compressed sensing method for microscopic measurement data. Based on the target sparse matrix corresponding to each category of data, constrained isometry conditions, and a pre-defined matching model, the target measurement matrix for each category of data is determined. Then, the target measurement vector is obtained by calculating the inner product of the target sparse matrix and the target measurement matrix for each category of data. By jointly solving for the measurement matrix using the pre-defined isometry conditions and the correlation coefficient of the measurement matrix, the compression efficiency of data processing is improved, data storage space is saved, and the target measurement matrix is ​​optimized through the matching model, thus improving the accuracy of subsequent data reconstruction.

[0137] Figure 4 A flowchart illustrating Embodiment 3 of the microscopic measurement data compression sensing method provided in this application is shown below. Figure 4 As shown, based on the foregoing embodiments, the specific implementation of step S301 may include:

[0138] S401: Solve for the correlation coefficient between the target sparse matrix and the measurement matrix corresponding to each classification data point, which approaches 0, to obtain the first set of measurement matrices corresponding to each classification data point.

[0139] In this step, after obtaining the target sparse matrix corresponding to each category of data, in order to improve the data compression rate and not affect the reconstruction quality of subsequent data, it is necessary to determine the measurement matrix. The measurement matrix needs to be uncorrelated with the target sparse matrix. Therefore, for each category of data, it is necessary to solve for the correlation coefficient between the target sparse matrix and the measurement matrix corresponding to the category data to approach 0, so as to obtain the first set of measurement matrices corresponding to the category data. The first set of measurement matrices includes all solutions with correlation coefficients approaching 0.

[0140] Specifically, since compressing the target sparse matrix using the measurement matrix requires that the target sparse matrix and the measurement matrix be uncorrelated, the correlation coefficient formula in the aforementioned embodiment, which approaches 0, can be expressed as:

[0141]

[0142] Substitute the target sparse matrix corresponding to each category of data into x in the above formula, and then solve y. Summarize all the solutions of y into the first measurement matrix set.

[0143] S402: Determine the second set of measurement matrices corresponding to each category of data as all measurement matrices that satisfy the constraint equidistant condition in the first set of measurement matrices corresponding to each category of data.

[0144] In this step, after solving for correlation coefficients approaching 0 through the above steps to obtain the first set of measurement matrices corresponding to each category of data, the measurement matrices also need to satisfy the constraint equidistant condition in order to compress the target sparse matrix and ensure accuracy in subsequent processes. Therefore, for each category of data, the measurement matrix in the first set of measurement matrices corresponding to the category data is substituted into the constraint equidistant condition formula in the aforementioned embodiment to determine whether there exists a constant δ that is 0 < δ < 1. k If the measurement matrix y satisfies this formula with respect to the target sparse matrix, then there exists a constant δ. k The measurement matrices that satisfy this formula are summarized into a second set of measurement matrices corresponding to the categorical data, and there is no constant δ. k The measurement matrix corresponding to this formula is deleted.

[0145] S403: Using the matching model, calculate the matching degree between each category data and each measurement matrix in the corresponding second measurement matrix set to obtain all matching degrees corresponding to each category data.

[0146] In this step, multiple measurement matrices corresponding to each category of data are filtered through the above-mentioned equidistant constraint conditions to obtain a second set of measurement matrices. In order to determine the target measurement matrix in the second set of measurement matrices, for each category of data, the target sparse matrix corresponding to the category data and the corresponding second set of measurement matrices are input into the matching model to calculate the matching degree between the category data and each measurement matrix in the second set of measurement matrices.

[0147] Specifically, a massive amount of microscopic measurement data and measurement matrices are pre-acquired as training data. Based on this training data, the cosine similarity model is pre-trained to obtain a matching model. For each category of data, the target sparse matrix corresponding to the category data and the corresponding second set of measurement matrices are input into the matching model. The pre-similarity between the target sparse matrix corresponding to the category data and each measurement matrix in the second set of measurement matrices is calculated by the matching model, which is the matching degree corresponding to the category data.

[0148] It should be noted that the cosine similarity calculated above ranges from [-1, 1]. The closer the cosine similarity value is to 1, the greater the matching degree.

[0149] Optionally, the above example of pre-training the cosine similarity model based on training data to obtain a matching model is just one example. Alternatively, a matching model can be obtained by training the graph matching model based on the training data, and then the matching degree corresponding to each category of data can be calculated. This embodiment does not specifically limit the matching model.

[0150] S404: Determine the measurement matrix corresponding to the largest matching degree among all matching degrees for each category of data as the target measurement matrix.

[0151] In this step, after calculating all matching degrees for each class of data, for each class of data, the larger the matching degree among all matching degrees, the better the compression effect of the measurement matrix corresponding to that matching degree. Therefore, the measurement matrix corresponding to the largest matching degree among all matching degrees for each class of data is determined as the target measurement matrix.

[0152] This embodiment provides a compressed sensing method for microscopic measurement data. It solves for the correlation coefficient between the target sparse matrix and the measurement matrix corresponding to each category of data, ensuring the correlation coefficient approaches zero, to obtain a first set of measurement matrices for each category. This first set is then filtered using constrained equidistant conditions to determine a second set of measurement matrices. Finally, a matching model is used to calculate all matching degrees for each category of data, and the measurement matrix corresponding to the highest matching degree for each category is determined as the target measurement matrix. This method more accurately determines the target measurement matrix, thereby improving data compression efficiency and the accuracy of subsequent data reconstruction.

[0153] Optionally, based on the above embodiments, after obtaining the target measurement corresponding to each category of data, the target measurement vector corresponding to each category of data can be encoded using a string table compression algorithm (Lempel-Ziv-Welch, LZW) to obtain a new target measurement vector after compression.

[0154] LZW encoding is a lossless data compression algorithm that can efficiently compress text data while ensuring data integrity. Further LZW encoding of the target measurement vector corresponding to each classification data can improve compression efficiency without affecting the reconstruction of subsequent classification data, and can also save storage space and improve transmission efficiency.

[0155] Specifically, LZW encoding compresses text data. This requires first converting the target measurement vector into text format. For example, for the vector [10, 20, 30, 40], it can be converted into the text string "10 20 30 40". Then, LZW encoding is used to gradually build a dictionary and replace repeating sequences, and the compression results are recorded. The final compressed result is in the format of an encoded string. To facilitate subsequent data reconstruction operations, the compressed result is converted into a new format to obtain a new target measurement vector after further compression.

[0156] Accordingly, step S205 includes: reconstructing the new target measurement vector corresponding to each category data using a preset reconstruction algorithm to obtain target reconstruction data.

[0157] This embodiment provides a compressed sensing method for microscopic measurement data. By performing LZW encoding on the target measurement vector corresponding to each classification data, it provides a compression rate without affecting the reconstruction quality, thereby saving more storage space and transmission costs.

[0158] Figure 5 A flowchart illustrating Embodiment 4 of the microscopic measurement data compression sensing method provided in this application is shown below. Figure 5As shown, based on the above embodiment, step S205 specifically includes:

[0159] S501: For each category of data, based on pre-set constraints, the dictionary learning algorithm iteratively calculates the target measurement vector corresponding to the category data to obtain the dictionary matrix and coefficient matrix.

[0160] In this step, after compressing each category of data to obtain the target measurement vector, in order to improve the data processing quality, it is necessary to reconstruct the reconstructed data based on the target measurement vector. The reconstructed data is identical to the category data. Therefore, for each category of data, a dictionary learning algorithm is used to reconstruct the target measurement vector. To improve the reconstruction accuracy, pre-defined constraints are used to optimize the dictionary learning algorithm.

[0161] Specifically, the dictionary learning algorithm is essentially an optimization calculation of the dictionary matrix and the coefficient matrix. The formula for the dictionary learning algorithm is expressed as:

[0162] x'=Da

[0163] Where x' represents the target reconstructed data, D is the dictionary matrix, and a is the coefficient matrix.

[0164] As shown in the above formula, the target reconstruction data can be obtained by calculating the dictionary matrix and coefficient matrix. To obtain high-accuracy target reconstruction data, constraints need to be pre-defined for each category of data. Based on the target measurement vector corresponding to the category data, the dictionary matrix and coefficient matrix are calculated. These constraints include both L0-norm minimization and L1-norm minimization constraints, which jointly constrain dictionary learning. The joint constraint formula is expressed as:

[0165] min‖a‖_1+λ‖a‖_0subject to y=Da

[0166] Where ||a||_1 represents the L1 norm of a, ||a||_0 represents the L0 norm of a, λ is the weight coefficient, y is the target measurement vector, D is the dictionary matrix, and a is the coefficient matrix.

[0167] To solve for the dictionary matrix and sparse matrix using the above constraint formula, the sparse matrix can be solved by fixing the dictionary matrix D. Specific solution methods include: the greedy algorithm Orthogonal Matching Pursuit algorithm, the Basis Pursuit algorithm, or the iterative shrinkage threshold algorithm, etc. This solution process is implemented by existing technology algorithms and can be implemented by the algorithms listed above, but not limited to the algorithms listed above. The specific solution process will not be described in detail in this embodiment.

[0168] Then, the sparse matrix is ​​fixed, and the dictionary matrix is ​​solved. Specific solution methods include: K-SVD algorithm, OMP algorithm combined with dictionary update, or dictionary learning algorithm based on gradient descent, etc. This solution process is implemented by existing technology algorithms, and can be implemented by the algorithms listed above, but not limited to the algorithms listed above. The specific solution process will not be described in detail in this embodiment.

[0169] The calculation is performed iteratively by alternating between the two steps mentioned above until the convergence condition is met, thus obtaining the dictionary matrix and the coefficient matrix.

[0170] It should be noted that the weighting coefficient λ in the above formula is used to balance the L1 norm and L0 norm, and can be set according to the characteristics of the classification data. This embodiment does not impose any specific limitations.

[0171] Optionally, the constraints may also include smoothness constraints, scaling constraints, and nonnegativity constraints. The constraint methods for all of the above are to add corresponding constraint terms to the joint constraint formula. For example, for smoothness constraints, a total variation constraint term or a weighted smoothness constraint term may be added to the joint constraint formula.

[0172] S502: Perform the inner product calculation on the dictionary matrix and coefficient matrix corresponding to each category data to obtain the target reconstructed data.

[0173] In this step, after obtaining the dictionary matrix and coefficient matrix through the dictionary learning algorithm, for each category of data, the reconstruction matrix is ​​obtained by multiplying the dictionary matrix and the corresponding coefficient matrix of the category data by the inner product. Then, data is extracted from the reconstruction matrix to obtain the target reconstruction data.

[0174] This embodiment provides a compressed sensing method for microscopic measurement data. For each category of data, based on pre-set constraints, a dictionary learning algorithm iteratively calculates the target measurement vector corresponding to the category data to obtain a dictionary matrix and a coefficient matrix. The inner product of the dictionary matrix and coefficient matrix for each category of data is then performed to obtain the reconstructed target data. This dictionary learning method improves the reconstruction quality of the target measurement vectors and achieves a high compression ratio without data loss. Optimization of the constraints enhances robustness and generalization ability, and also improves data processing quality.

[0175] Preferably, based on the above embodiments, step S205 further includes: optimizing the dictionary matrix and coefficient matrix corresponding to each category of data through an optimization sparse model to obtain the optimized dictionary matrix and the optimized coefficient matrix.

[0176] Based on the target measurement vector, after calculating the dictionary matrix and coefficient matrix by setting constraints on the dictionary learning algorithm, in order to make full use of the statistical characteristics of the data, improve the quality of the dictionary matrix and coefficient matrix, and thus improve the accuracy of the reconstructed data, the dictionary matrix and coefficient matrix can be further optimized by optimizing the model to obtain the optimized dictionary matrix and optimized coefficient matrix.

[0177] Specifically, the optimization model is obtained by training the Bayesian model based on a large number of pre-acquired sparse matrices and measurement vectors, and is used to further optimize the dictionary matrix and coefficient matrix.

[0178] For example, the optimization process of the optimization model is illustrated by using a sparse Gaussian prior distribution for the coefficient matrix. First, a sparse Gaussian prior distribution is used for the coefficient matrix, meaning that each element a-i of the coefficient matrix a independently follows a Gaussian distribution with a mean of 0 and an accuracy (inverse variance) of β, i.e.:

[0179] a_i~N(0,β ∧ -1)

[0180] Then, an observation model is established, namely, the relational model of the coefficient matrix of the target measurement vector. A linear model is then considered, i.e., y = Da + ε, where D is the dictionary matrix, a is the coefficient matrix, and ε is the noise term, which follows a mean of 0 and an inverse variance of σ. ∧ 2 is a Gaussian distribution, i.e., ε ~ N(0, σ) ∧ 2).

[0181] Finally, by performing posterior distribution and inference, according to Bayes' theorem, the posterior distribution is expressed as:

[0182]

[0183] Where P(y|a) is the likelihood function of the target measurement vector y under the given coefficient matrix a, P(a) is the prior distribution, and P(y) is the normalization constant.

[0184] Considering the Gaussian property of the observation model and the prior distribution, it can be deduced that the posterior distribution P(a|y) is also a Gaussian distribution. Specifically, the posterior distribution can be expressed as:

[0185] P(a|y)=N(a|μ_post,∑_post)

[0186] Where μ_post is the posterior mean vector and ∑_post is the posterior covariance matrix.

[0187] Finally, we define the objective function as maximizing the posterior probability P(a|y), which is equivalent to maximizing the posterior estimate. The maximization of the posterior log-likelihood function is then expressed as:

[0188] argmax_a logP(a|y)=argmax_a logP(y|a)+logP(a)-logP(y)

[0189] Substituting the above observation model and prior distribution into the above formula, we obtain the objective function:

[0190]

[0191] The objective function is maximized, and then the optimized dictionary matrix and the optimized coefficient matrix are calculated.

[0192] Accordingly, step S502 includes: performing an inner product calculation on the optimized dictionary matrix and the optimized coefficient matrix corresponding to each category data to obtain the target reconstructed data.

[0193] This embodiment provides a compressed sensing method for microscopic measurement data. By optimizing the model, the dictionary learning algorithm is further optimized, thereby avoiding overfitting during the calculation of dictionary matrix and sparse matrix, and further improving robustness, generalization ability and data processing quality.

[0194] In one possible implementation, based on the foregoing embodiments, the following operation may be performed before step S202:

[0195] The locations of multiple categories of data are labeled to obtain the location number corresponding to each category of data.

[0196] After classifying the microscopic measurement data, to achieve accurate processing, it is necessary to label the locations of each category, thus obtaining a location number for each category. After processing each category separately, the location labels can be used to reassemble the data, resulting in complete microscopic measurement data.

[0197] Accordingly, step S206 includes: reorganizing the target reconstruction data corresponding to each classification data according to the location number corresponding to each classification data to obtain reconstructed microscopic measurement data.

[0198] After obtaining the target reconstruction data corresponding to each category of data, in order to obtain reconstructed microscopic measurement data that is the same as the microscopic measurement data in a high-efficiency manner, the target reconstruction data corresponding to each category of data can be quickly reorganized according to the position number corresponding to each category of data to obtain the reconstructed microscopic measurement data.

[0199] This embodiment provides a compressed sensing method for microscopic measurement data. By labeling the locations of multiple categorical data points, a location number corresponding to each categorical data point is obtained. Based on the location number corresponding to each categorical data point, the microscopic measurement data is accurately reconstructed during data reconstruction. This method improves the processing efficiency of microscopic measurement data.

[0200] Figure 6 A flowchart illustrating Embodiment 5 of the microscopic measurement data compression sensing method provided in this application is shown below. Figure 6 As shown, based on the foregoing embodiments, step S203 may be specifically implemented in the following ways:

[0201] S601: For each category of data, calculate the sparsity of each sparse matrix corresponding to the category data according to the sparsity calculation formula.

[0202] In this step, in order to improve the high-quality processing of each category of data and thus make the compressed sensing effect better, and in compressed sensing, the sparsity of the sparse matrix affects the sampling and transmission cost of the category data, so for each category of data, a sparsity technique is applied to each sparse matrix corresponding to the category data.

[0203] Specifically, in compressed sensing, the sparsity measure of a sparse matrix is ​​generally represented by its norm, and the formula for calculating sparsity is:

[0204]

[0205] Where A is a sparse matrix, and A' is an m×n matrix.

[0206] Optionally, the above sparsity formula uses the L1 norm; alternatively, the L0 norm can be chosen to measure the sparsity of a sparse matrix, as shown in the formula:

[0207]

[0208] Where A is a sparse matrix, A' is an m×n matrix, [A''''''''''''''''''''''''''''''''''"""","A' ... ij ≠0] means if A ij If it is not 0, it is 1; otherwise, it is 0.

[0209] It should be noted that either the L1 norm or the L0 norm can be used to describe the sparsity of the sparse matrix. The sparsity formula can also be selected according to the characteristics of the classification data. This embodiment does not make specific limitations on the specific formula for sparsity.

[0210] Alternatively, the sparsity of a sparse matrix can be measured by counting the non-zero elements in the matrix and then by the ratio of the number of non-zero elements to the total number of elements in the matrix.

[0211] For example, consider a 4×4 sparse matrix A, where only 3 elements are non-zero and the rest are zero. Then:

[0212] By traversing the entire matrix, the number of non-zero elements is counted: nnz = 3 (the number of non-zero elements);

[0213] By traversing the entire matrix, the total number of elements in the matrix is ​​counted: total = 4 x 4 = 16 (total number of elements in the matrix);

[0214] By calculating the ratio of the number of non-zero elements to the total number of elements in the matrix, we obtain:

[0215] Sparsity = 3 / 16 ≈ 0.1875 (sparseness)

[0216] S602: From the multiple sparse matrices corresponding to the classification data, determine the sparse matrix with the largest sparsity as the target sparse matrix.

[0217] In this step, after calculating the sparsity of multiple sparse matrices corresponding to the classification data, since in compressed sensing, the larger the sparsity of the sparse matrix, the better the data sampling and transmission efficiency and the better the crushing sensing effect, the sparse matrix with the largest sparsity among the multiple sparse matrices corresponding to the classification data is determined as the target sparse matrix.

[0218] In one possible implementation, sparsity is not an absolute rule of "the sparser the better." An excessively sparse sparse matrix may lead to data loss or distortion. In compressed sensing, the appropriate sparsity needs to be determined based on the specific application scenario and signal characteristics. This can be achieved by pre-setting a sparsity threshold based on the characteristics of the categorical data. Among multiple sparse matrices corresponding to the categorical data, the one with the largest sparsity less than the threshold is determined as the target sparse matrix.

[0219] This embodiment provides a compressed sensing method for microscopic measurement data. For each type of data, the sparsity of each sparse matrix corresponding to the data is calculated according to the sparsity calculation formula, and the sparse matrix with the largest sparsity is determined as the target sparse matrix. By selecting the sparse matrix with the largest sparsity as the target sparse matrix, the accuracy of data reconstruction is improved. Furthermore, the sparsity of the sparse matrix results in less redundant information in the classified data in the low-dimensional space, thus achieving efficient compression.

[0220] Figure 7 This is a schematic diagram of the structure of a first embodiment of the microscopic measurement data compression sensing device provided in this application, as shown below. Figure 7 As shown, the microscopic measurement data compression sensing device 700 includes:

[0221] Classification model 701 is used to classify microscopic measurement data based on a BP neural network, resulting in multiple classification data with different frequencies for each classification data.

[0222] The sparse module 702 is used to perform sparse transformation on each category data according to the sparse transformation basis in the pre-set sparse transformation basis set, so as to obtain multiple sparse matrices corresponding to each category data. The sparse transformation basis set includes multiple sparse transformation bases.

[0223] The determination module 703 is used to determine the target sparse matrix from multiple sparse matrices corresponding to each categorical data. The target sparse matrix is ​​the matrix with the largest sparsity among the multiple sparse matrices.

[0224] Compression module 704 is used to compress the target sparse matrix corresponding to each classification data to obtain the target measurement vector.

[0225] The reconstruction module 705 is used to reconstruct the target measurement vector corresponding to each category data using a preset reconstruction algorithm to obtain the target reconstruction data.

[0226] The recombination module 706 is used to recombine the target reconstruction data corresponding to each category of data to obtain reconstructed microscopic measurement data.

[0227] Optionally, the microscopic measurement data compression sensing device 700 also includes:

[0228] The encoding module 707 is used to encode the target measurement vector corresponding to each category data using LZW encoding to obtain a new target measurement vector after compression.

[0229] Correspondingly, the reconstruction module 705 is specifically used to: reconstruct the new target measurement vector corresponding to each category data through a preset reconstruction algorithm to obtain the target reconstruction data.

[0230] In one possible design, the microscopic measurement data compression sensing device 700 also includes:

[0231] The annotation module 708 is used to annotate the locations of multiple categories of data to obtain the location number corresponding to each category of data.

[0232] Correspondingly, the recombination module 706 is specifically used to: recombine the target reconstruction data corresponding to each category data according to the location number corresponding to each category data to obtain reconstructed microscopic measurement data.

[0233] Figure 8 This is a schematic diagram of the structure of a second embodiment of the microscopic measurement data compression sensing device provided in this application, as shown below. Figure 8As shown, based on the above-described device embodiment, the compression module 704 includes:

[0234] The determining unit 801 is used to determine the target measurement matrix corresponding to each category of data based on the target sparse matrix, constraint isometry conditions, and pre-set matching model corresponding to each category of data.

[0235] The computing unit 802 is used to obtain the target measurement vector by performing an inner product calculation on the target sparse matrix corresponding to each category data and the target measurement matrix corresponding to each category data.

[0236] The matching model is a model that calculates the matching degree between each category of data and each corresponding measurement matrix, and the constraint equidistant condition is used to filter the target measurement matrix corresponding to each category of data.

[0237] Figure 9 This is a schematic diagram of the structure of Embodiment 3 of the microscopic measurement data compression sensing device provided in this application, as shown below. Figure 9 As shown, based on the above-described device embodiment, the determining unit 801 includes:

[0238] The solution subunit 901 is used to solve for the correlation coefficient between the target sparse matrix and the measurement matrix corresponding to each classification data approaching 0, and to obtain the first measurement matrix set corresponding to each classification data. The first measurement matrix set includes all measurement matrices obtained by finding the correlation coefficient approaching 0.

[0239] The first determining subunit 902 is used to determine the second set of measurement matrices corresponding to each category data from all measurement matrices in the first set of measurement matrices that satisfy the constraint equidistant condition.

[0240] The calculation subunit 903 is used to calculate the matching degree between each category data and each measurement matrix in the corresponding second measurement matrix set through the matching model, so as to obtain all the matching degrees corresponding to each category data.

[0241] The second determining subunit 904 is used to determine the measurement matrix corresponding to the largest matching degree among all matching degrees corresponding to each category data as the target measurement matrix.

[0242] Figure 10 This is a schematic diagram of the structure of Embodiment 4 of the microscopic measurement data compression sensing device provided in this application, as shown below. Figure 10 As shown, based on the aforementioned embodiments, the reconstruction module 705 includes:

[0243] The first computing unit 1001 is used to iteratively calculate the target measurement vector corresponding to each category data according to the preset constraints using a dictionary learning algorithm, so as to obtain the dictionary matrix and coefficient matrix.

[0244] The second calculation unit 1002 is used to perform inner product calculation on the dictionary matrix and the corresponding coefficient matrix of each category data to obtain the target reconstructed data.

[0245] Among them, the reconstruction algorithm includes the dictionary learning algorithm.

[0246] Optionally, refactoring module 705 also includes:

[0247] The optimization unit 1003 is used to optimize the dictionary matrix and the corresponding coefficient matrix of each classification data through the optimization sparse model to obtain the optimized dictionary matrix and the optimized coefficient matrix. The optimization model is obtained by training the Bayesian sparse model based on the pre-acquired dictionary matrix and coefficient matrix set, and is used to optimize the dictionary matrix and coefficient matrix.

[0248] Correspondingly, the second calculation unit 1002 is specifically used to: perform inner product calculation on the optimized dictionary matrix and the corresponding optimized coefficient matrix for each category of data to obtain the target reconstructed data.

[0249] In another possible implementation, based on the above-described device embodiments, the sparse transform basis set includes at least the following sparse transform bases:

[0250] The first sparse transform basis is obtained by combining wavelet basis and DCT basis;

[0251] The second sparse transform basis obtained by combining the wavelet basis and the Curvelet basis; and...

[0252] The third sparse transform basis is obtained by combining the wavelet basis and the Shearlet basis.

[0253] Figure 11 A schematic diagram of the structure of Embodiment Six of the Microscopic Measurement Data Compression Sensing Device provided in this application is shown below. Figure 11 As shown, based on the above-described device embodiments, module 703 is defined, including:

[0254] The calculation unit 1101 is used to calculate the sparsity of each sparse matrix corresponding to each category data according to the sparsity calculation formula.

[0255] The determining unit 1102 is used to determine the sparse matrix with the largest sparsity from multiple sparse matrices corresponding to the classification data as the target sparse matrix.

[0256] The microscopic measurement data compression sensing devices provided in the above embodiments are used to execute the microscopic measurement data compression sensing method in any of the foregoing method embodiments. Their implementation principles and technical effects are similar and will not be described again here.

[0257] Figure 12 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application, such as... Figure 12 As shown, the electronic device 1200 includes: a memory 1201 and a processor 1202;

[0258] Memory 1201 stores computer-executed instructions.

[0259] Processor 1202 executes computer execution instructions stored in memory to implement the method in any of the method embodiments.

[0260] This application also provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the method in any of the method embodiments.

[0261] The aforementioned computer-readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory, electrically erasable programmable read-only memory, erasable programmable read-only memory, programmable read-only memory, read-only memory, magnetic storage, flash memory, magnetic disk, or optical disk. The readable storage medium can be any available medium accessible to a general-purpose or special-purpose computer.

[0262] Optionally, a readable storage medium can be coupled to a processor, enabling the processor to read information from and write information to the readable storage medium. Alternatively, the readable storage medium can be an integral part of the processor. Both the processor and the readable storage medium can reside in an Application Specific Integrated Circuit (ASIC). Alternatively, the processor and the readable storage medium can exist as discrete components within the device.

[0263] This application also provides a computer program product, which includes a computer program stored in a computer-readable storage medium. At least one processor can read the computer program from the computer-readable storage medium, and when the at least one processor executes the computer program, it can implement the technical solutions provided in any of the above method embodiments.

[0264] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this application are indicated by the following claims.

[0265] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the appended claims.

Claims

1. A method for compressed sensing of microscopic measurement data, characterized in that, include: The backpropagation BP neural network is used to classify the microscopic measurement data, resulting in multiple classification data, each with a different frequency. Based on the sparse transformation basis in the pre-defined sparse transformation basis set, sparse transformation is performed on each category of data to obtain multiple sparse matrices corresponding to each category of data. The sparse transformation basis set includes multiple sparse transformation bases. The sparse transform basis set includes at least the following sparse transform bases: The first sparse transform basis is obtained by combining the wavelet basis with the discrete cosine transform (DCT) basis; The second sparse transform basis obtained by combining the wavelet basis and the Curvelet basis; and... The third sparse transform basis is obtained by combining the wavelet basis and the Shearlet basis; For each category of data, a target sparse matrix is ​​determined from multiple sparse matrices corresponding to the category data. The target sparse matrix is ​​the matrix with the largest sparsity among the multiple sparse matrices. The target sparse matrix corresponding to each category of data is compressed to obtain the target measurement vector, including: Based on the target sparse matrix corresponding to each category of data, the constraint isometry condition, and the pre-defined matching model, the target measurement matrix corresponding to each category of data is determined, including: The correlation coefficient between the target sparse matrix and the measurement matrix corresponding to each category of data is solved until it approaches 0, thus obtaining the first set of measurement matrices corresponding to each category of data. The first set of measurement matrices includes all measurement matrices obtained by solving the correlation coefficient approaching 0. The first set of measurement matrices that satisfy the constraint equidistance condition in the first set of measurement matrices corresponding to each category of data is determined as the second set of measurement matrices corresponding to each category of data. Using the matching model, the matching degree is calculated for each category data and each measurement matrix in the corresponding second measurement matrix set to obtain all matching degrees for each category data. The measurement matrix corresponding to the largest matching degree among all matching degrees for each category of data is determined as the target measurement matrix; For each category of data, the target measurement vector is obtained by performing an inner product calculation on the target sparse matrix and the target measurement matrix corresponding to the category data. The matching model is a model for calculating the matching degree between each category of data and each corresponding measurement matrix, and the constraint equidistant condition is used to filter the target measurement matrix corresponding to each category of data. The target measurement vector corresponding to each category of data is reconstructed using a preset reconstruction algorithm to obtain the target reconstruction data, including: For each category of data, based on pre-set constraints, the target measurement vector corresponding to the category of data is iteratively calculated using a dictionary learning algorithm to obtain a dictionary matrix and a coefficient matrix; The inner product of the dictionary matrix and the coefficient matrix corresponding to each category of data is calculated to obtain the target reconstructed data. The reconstruction algorithm includes the dictionary learning algorithm; The target reconstruction data corresponding to each category of data is reorganized to obtain reconstructed microscopic measurement data.

2. The method according to claim 1, characterized in that, The method further includes: The target measurement vector corresponding to each category data is encoded using the LZW string table compression algorithm to obtain a new target measurement vector after compression. Accordingly, the step of reconstructing the target measurement vector corresponding to each category data using a preset reconstruction algorithm to obtain target reconstruction data includes: The new target measurement vector corresponding to each category of data is reconstructed using the preset reconstruction algorithm to obtain the target reconstruction data.

3. The method according to claim 1, characterized in that, The method further includes: The dictionary matrix and coefficient matrix corresponding to each category of data are optimized by an optimization model to obtain the optimized dictionary matrix and coefficient matrix. The optimization model is obtained by training a Bayesian sparse model based on a pre-acquired set of dictionary matrices and coefficient matrices, and is used to optimize the dictionary matrix and coefficient matrix. Accordingly, the step of performing an inner product calculation on the dictionary matrix and the corresponding coefficient matrix for each category of data to obtain the target reconstructed data includes: The inner product of the optimized dictionary matrix and the optimized coefficient matrix corresponding to each category of data is calculated to obtain the target reconstructed data.

4. The method according to any one of claims 1 to 3, characterized in that, Before performing sparse transformation on each category data according to the sparse transformation basis in the pre-defined sparse transformation basis set to obtain multiple sparse matrices corresponding to each category data, the method further includes: The multiple categories of data are labeled with their locations to obtain the location number corresponding to each category of data. Accordingly, the reorganization of the target reconstruction data corresponding to each classification data to obtain reconstructed microscopic measurement data includes: Based on the location number corresponding to each category of data, the target reconstruction data corresponding to each category of data is reorganized to obtain reconstructed microscopic measurement data.

5. The method according to any one of claims 1 to 2, characterized in that, For each category of data, determining the target sparse matrix from multiple sparse matrices corresponding to the category data includes: For each category of data, the sparsity of each sparse matrix corresponding to the category of data is calculated according to the sparsity calculation formula. From the multiple sparse matrices corresponding to the classification data, the sparse matrix with the largest sparsity is determined as the target sparse matrix.

6. An apparatus corresponding to the microscopic measurement data compression sensing method according to any one of claims 1-5, characterized in that, include: The classification model is used to classify microscopic measurement data based on a backpropagation BP neural network, resulting in multiple classification data, each with a different frequency. The sparse module is used to perform sparse transformation on each category of data according to the sparse transformation basis in the pre-defined sparse transformation basis set, so as to obtain multiple sparse matrices corresponding to each category of data. The sparse transformation basis set includes multiple sparse transformation bases. The determination module is used to determine a target sparse matrix from multiple sparse matrices corresponding to each type of data, wherein the target sparse matrix is ​​the matrix with the largest sparsity among the multiple sparse matrices; The compression module is used to compress the target sparse matrix corresponding to each category of data to obtain the target measurement vector; The reconstruction module is used to reconstruct the target measurement vector corresponding to each category of data using a preset reconstruction algorithm to obtain the target reconstruction data. The reconstruction module is used to reconstruct the target reconstruction data corresponding to each category of data to obtain reconstructed microscopic measurement data.

Images