Mueller quasi-scattering imaging device and method based on sparse representation and dictionary learning

Abstract

This invention relates to the field of detection and imaging technology, specifically to a muon quasi-scattering imaging device and method based on sparse representation and dictionary learning. The device includes a modular muon detector array deployed on one side, an FPGA high-speed data acquisition system, and a computational control unit, and can also integrate a momentum spectrometer module. The method includes offline dictionary preparation and online sparse reconstruction steps. First, a complete dictionary is trained using GEANT4 simulation and the K-SVD algorithm. Then, data is collected to construct an enhanced objective function containing block sparseness and anisotropic total variational regularization. The reconstructed image is obtained by optimizing the solution using the ADMM algorithm. This invention overcomes the limitations of two-sided detection, adapts to single-sided detection scenarios in geotechnical engineering, significantly improves imaging efficiency and quality under low-throughput data, has strong algorithm robustness and a scalable framework, and can accurately identify anomalies such as underground cavities and isolated boulders.

Description

Technical Field

[0001] This invention relates to the field of detection and imaging technology, and more specifically to a muon quasi-scattering imaging device and method based on sparse representation and dictionary learning. Background Technology

[0002] Cosmic ray muon imaging is a novel non-destructive testing technology developed based on the extremely strong penetrating power of muons. It has shown broad application prospects in many fields such as volcanic geological exploration, nuclear material monitoring, large-scale archaeological site detection, and geotechnical engineering testing.

[0003] Traditional muon imaging techniques are mainly divided into two categories: transmission / absorption imaging and scattering imaging. Absorption imaging inverts the density integral along the path by measuring the flux attenuation after a muon penetrates the target; scattering imaging inverts the local density anomaly inside the target by measuring the change in the direction of motion of the muon after penetration (i.e., the scattering angle). Among them, scattering tomography is sensitive to high-density, high-atomic-number materials, but it requires detectors to be deployed on both sides of the target to accurately capture the incident and exit trajectories of the same muon. This limitation makes it difficult to adapt to many practical engineering scenarios, such as the detection of overburden structures in subway tunnels and the advanced geological prediction in front of tunnel boring machines. In these critical geotechnical engineering scenarios, detectors can only be deployed on one side of the target (such as inside the tunnel), which cannot meet the two-sided detection requirements of traditional scattering imaging, thus limiting the application of the technology.

[0004] Meanwhile, the flux of natural cosmic ray muons is extremely low. To obtain data that meets the requirements for imaging reconstruction, traditional techniques require several days or even weeks of cumulative measurement time, and the imaging speed is far from meeting the real-time requirements of geotechnical engineering investigation. Although there have been attempts to improve imaging efficiency by increasing the detector area and optimizing the inversion algorithm, none of them have fundamentally solved the two core technical problems of "dual-sided detection dependence" and "low-throughput data bottleneck".

[0005] Furthermore, existing muon imaging inversion algorithms suffer from insufficient utilization of prior knowledge: most mainstream algorithms rely on the assumption of continuous medium or simple sparsity for inversion calculations, failing to fully incorporate the inherent structural characteristics of geotechnical engineering targets. Targets in geotechnical engineering (such as boulders, cavities, and other anomalous bodies) not only exhibit sparsity in spatial distribution but also possess significant local continuity and blocky aggregation characteristics; moreover, in a uniform or gradually changing geological background scattering environment, the density difference (i.e., differential signal) between the target body and the surrounding background is far more sparsity than the absolute density distribution.

[0006] Existing algorithms ignore the two types of key structured prior knowledge mentioned above, resulting in images that are prone to problems such as blurred target contours and residual background noise. In fact, they cannot effectively identify and locate underground anomalies in engineering test scenarios with low signal-to-noise ratios.

[0007] In summary, there is an urgent need to develop a muon imaging technology and corresponding device that can adapt to unilateral detection conditions, fully explore and utilize the target's structured prior knowledge, and rapidly and accurately reconstruct underground sparse anomalies based on a small amount of muon observation data, so as to meet the actual detection needs of fields such as geotechnical engineering. Summary of the Invention

[0008] To address the aforementioned problems, this invention provides a muon quasi-scattering imaging device and method based on sparse representation and dictionary learning, for use in...

[0009] To achieve the above objectives, the technical solution of the present invention is as follows:

[0010] On the one hand, a muon quasi-scattering imaging device based on sparse representation and dictionary learning is provided, including a muon detector array deployed on one side of the measured area, a high-speed data acquisition system and a computing control unit;

[0011] The muon detector array is used to capture muon quasi-scattering observation data after penetrating the measured region;

[0012] The high-speed data acquisition system is used to capture sparse muon instances;

[0013] The computing and control unit is used to run the imaging algorithm software, control the overall operation of the system, and output reconstructed images of underground anomalies.

[0014] Furthermore, it also includes a momentum spectrometer module integrated into the muon detector array, which is used to measure the momentum of some muons, providing additional information dimensions for imaging.

[0015] Furthermore, the muon detector array adopts a modular design, with each module consisting of a multilayer plastic scintillator strip coupled with a silicon photomultiplier tube to measure the emission position and direction of muons.

[0016] On the other hand, a muon quasi-scattering imaging method based on sparse representation and dictionary learning is provided, including offline dictionary preparation and online sparse reconstruction steps, wherein the offline dictionary preparation provides a trained overcomplete dictionary D for the online sparse reconstruction;

[0017] The offline dictionary preparation includes scene and target definition, Monte Carlo simulation, and dictionary training sub-steps.

[0018] The online sparse reconstruction includes sub-steps such as data acquisition and feature extraction, construction of an enhanced objective function, optimization solution, and image generation and interpretation.

[0019] Furthermore, the specific process for preparing the offline dictionary is as follows:

[0020] S11. Define specific application scenarios, define digital models of typical underground anomalies containing parameters of different locations, sizes, and densities, and form a target model library to be detected;

[0021] S12. Place each target model to be detected in the set geological background model, simulate the penetration of cosmic ray muons through the entire system, record the muon emission direction angle (θ, φ) at the virtual detector position, and form quasi-scattering observation signal instances corresponding to each target model;

[0022] S13. Collect all instances of quasi-scattering observation signals to form a training sample set. Where N is the number of samples, the overcomplete dictionary is obtained by solving the optimization problem using the K-SVD dictionary learning algorithm. The optimization problem is:

[0023]

[0024] in, It is a sparse coefficient matrix; The square of the Frobenius norm; for Norm; This is the threshold for sparsity constraints.

[0025] Furthermore, in the online sparse reconstruction, the specific process of data acquisition and feature extraction is as follows: muon observation data for a fixed time period is collected through a muon detector array; the observation data is statistically processed to generate an observation vector with the same format as the offline dictionary preparation stage. .

[0026] Furthermore, in the online sparse reconstruction, the constructed enhanced objective function is:

[0027]

[0028] in, It is a dictionary matrix; Let be the sparse coefficient vector to be solved; The image to be reconstructed; For block sparse regularization; sparse coefficient vector The middle belongs to the first Subvectors of each group; A tradeoff parameter is used to balance the effects of the data fitting term and the block sparsity regularization term. For total variational regularization; For image The total variation of anisotropy; To weigh the effects of the data fitting term and the total variation regularization term; Let be a linear mapping matrix from sparse coefficients to the image space, used to map the sparse coefficient vector. Density distribution converted to image space From the dictionary Determined by observation geometry.

[0029] Furthermore, the total variation of the anisotropy The calculation formula is:

[0030]

[0031] in, For image The pixel value located in the i-th row and j-th column.

[0032] Furthermore, the optimization solution is performed using the Alternating Direction Multiplier Method (ADMM), and the specific process is as follows:

[0033] S31. Variable Splitting: Introducing Auxiliary Variables and This transforms the constrained optimization problem into an equivalent unconstrained multivariable optimization problem.

[0034] S32, Alternating Update: Update the sparsity coefficients sequentially. , block sparse approximation Image denoising and boundary preservation And update the Lagrange multipliers according to the standard ADMM procedure;

[0035] Among them, update To solve a band Norm-constrained ridge regression problems can be solved quickly using the conjugate gradient method; update For grouped Apply the group Lasso contraction operator for each group Perform soft thresholding, the formula is as follows: ;renew To solve the total variational denoising problem, the ROF model is combined with the Chambolle dual algorithm or the fast gradient projection method.

[0036] S33. Iteration: Repeat step S32 until the original residual and the dual residual satisfy the preset convergence condition, and obtain the converged result. and .

[0037] Furthermore, the specific process of image generation and interpretation is as follows: the converged image is... As the final reconstruction result, the image Connected regions with significantly higher brightness than the background are identified as high-density underground anomalies, while regions with significantly lower brightness than the background are identified as low-density underground anomalies.

[0038] The above approach has the following beneficial effects:

[0039] 1. Abandoning the rigid requirement of traditional muon scattering imaging for the deployment of dual-sided detectors, we innovatively propose the concept of muon quasi-scattering imaging, which only requires detection on one side of the measured area. This successfully solves the industry pain point of not being able to carry out dual-sided detection in narrow and enclosed spaces in key geotechnical engineering scenarios such as subway tunnel overburden detection and advanced geological prediction in front of tunnel boring machines. This extends the application scope of muon imaging technology from general scenarios such as nuclear material monitoring and volcanic geological exploration to essential scenarios such as on-site measurement in geotechnical engineering, achieving a breakthrough expansion in the application fields of the technology.

[0040] 2. Relying on the core technologies of sparse representation and dictionary learning, and combined with the strong prior constraints of block sparsity and total variation (TV), this scheme has the ability to efficiently process muon observation data with low throughput and low signal-to-noise ratio. Under the same imaging accuracy requirements, it can complete high-quality imaging with only 1 / 3 of the muon data accumulation time of traditional algorithms, which greatly shortens the detection cycle of several days or even weeks. It effectively solves the problem of slow imaging speed caused by the low throughput of natural muons and fully meets the actual engineering requirements of geotechnical engineering exploration for real-time performance and high efficiency.

[0041] 3. To address the shortcomings of existing algorithms in utilizing insufficient prior knowledge, this scheme simultaneously mines and incorporates two key structured priors: the spatial sparsity, local continuity, and blocky clustering characteristics of underground anomalies, and the high sparsity of the density difference signal between the target and the background. Block sparsity regularization promotes the grouped activation of dictionary atoms related to the anomaly, avoiding point artifacts and ensuring the continuity and integrity of the target's morphology. Anisotropic total variational regularization smooths the uniform geological background and sharpens the boundaries of anomalies, highlighting density difference information. The final reconstructed image achieves a background with no significant noise, clear target boundaries, and accurate contour morphology, effectively identifying the location, size, and density characteristics of anomalies such as underground cavities and isolated boulders, solving the problems of blurred contours and difficult target recognition in traditional algorithm-reconstructed images.

[0042] Through the dual constraints of block sparsity and fully variational structured regularization, the imaging algorithm of this scheme has stronger anti-interference ability against complex geological scattering backgrounds, environmental noise and data missingness. Even in geotechnical engineering test scenarios with low signal-to-noise ratio, it can still stably achieve effective identification and localization of underground anomalies. The reconstruction results are more in line with the actual geological distribution patterns and conform to geological cognition, which greatly improves the reliability of muon imaging technology in the actual application of complex engineering sites.

[0043] Furthermore, the one-sided detection principle of this invention is not limited to a specific incident direction of the muon. In the detection of overburden layers in subway tunnels, the muon penetrates the soil and rock layers vertically downwards; in geological prediction in front of the tunnel boring machine, the muon penetrates the unexcavated strata in a horizontal or inclined direction. Although the incident directions are different in these two scenarios, they both satisfy the one-sided detection condition that "the detector and the target are on the same side." The quasi-scattering imaging method proposed in this invention can effectively invert the internal structure of the target by analyzing the disturbance of the exit angle distribution after the muon penetration. Attached Figure Description

[0044] Figure 1 This is a schematic diagram of an embodiment of a muon quasi-scattering imaging device based on sparse representation and dictionary learning.

[0045] Figure 2 This is an overall flowchart of an embodiment of the muon quasi-scattering imaging method based on sparse representation and dictionary learning;

[0046] Figure 3 A flowchart of offline dictionary preparation for an embodiment of the muon quasi-scattering imaging method based on sparse representation and dictionary learning;

[0047] Figure 4 This is a flowchart of the online sparse reconstruction process for an embodiment of the muon quasi-scattering imaging method based on sparse representation and dictionary learning.

[0048] Figure 5 This diagram illustrates a comparison of the imaging performance between traditional imaging algorithms and the muon quasi-scattering imaging method based on sparse representation and dictionary learning. Detailed Implementation

[0049] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0050] The following detailed description illustrates the specific implementation method:

[0051] A muon quasi-scattering imaging device based on sparse representation and dictionary learning, such as... Figure 1 As shown, it includes a muon detector array deployed on one side of the measured area, a high-speed data acquisition system, and a computing control unit.

[0052] The muon detector array is used to capture muon observation data (i.e., "quasi-scattering" signals) after penetrating the measured area. Specifically, the muon detector array adopts a modular design, with each module consisting of multi-layer plastic scintillator strips coupled to silicon photomultiplier tubes to accurately measure the emission position and direction of muons. In key geotechnical engineering scenarios such as overburden structure detection in subway tunnels and geological prediction in front of tunnel boring machines, the muon detector array is installed as a whole on the tunnel sidewall or the tunnel boring machine trolley, facing the measured area.

[0053] The high-speed data acquisition system is used to ensure low dead time and effectively capture sparse muon events; specifically, the high-speed data acquisition system is developed based on FPGA to realize multi-channel parallel, low dead time data acquisition and online trigger filtering.

[0054] The computational control unit is used to run the core imaging algorithm; specifically, the computational control unit is a high-performance server that runs the imaging algorithm software, controls the entire system, and outputs reconstructed images.

[0055] In some other embodiments, the above-described muon quasi-scattering imaging device further includes a momentum spectrometer module; the momentum spectrometer module is integrated into the detector array to measure the momentum of some muons, providing additional information dimensions for imaging.

[0056] A muon quasi-scattering imaging method based on sparse representation and dictionary learning, as follows: Figures 2-4 As shown, the method applicable to the above-mentioned muon quasi-scattering imaging device mainly consists of two steps: offline dictionary preparation and online sparse reconstruction, as detailed below:

[0057] Phase 1: Offline Dictionary Preparation

[0058] 1. Scenario and Target Definition: Clearly define the specific application scenario, such as the detection of overlying layers in subway tunnels; define a series of typical target models to be detected, such as spherical cavities of different diameters and cubic boulders, and assign them reasonable density parameters. Among them, the target models to be detected are digital models of typical underground anomalies (such as cavities and boulders) with different attributes (location, size, density).

[0059] 2. Monte Carlo Simulation: This was performed using the GEANT4 simulation tool. For each target model to be detected, it was placed within a defined geological background model. The simulation simulated a large number of cosmic ray muons penetrating the entire system from above or near horizontally. The emission angle (θ, φ) of each muon was recorded at the virtual detector location, forming an instance of the "quasi-scattering" observation signal for that target model. In key geotechnical engineering scenarios such as overburden structure detection in subway tunnels and geological prediction in front of tunnel boring machines, the virtual detector was placed inside the tunnel.

[0060] 3. Dictionary Training: Collect all instances of "quasi-scattering" observation signals corresponding to the target model to be detected, forming a training sample set. , where N is the number of samples.

[0061] The K-SVD dictionary learning algorithm is used to solve the following optimization problem to obtain an overcomplete dictionary D:

[0062]

[0063] in, It is a sparse coefficient matrix; The square of the Frobenius norm, used in dictionaries and sparse coefficient matrix Reconstructing the training sample set The overall error during reconstruction is considered, and the goal is to minimize this reconstruction error. for Norm, representing a vector The number of non-zero elements in the neutron; This is the threshold for sparsity constraints.

[0064] After training, the dictionary Each column vector (atom) represents a typical scattering response pattern learned from the data.

[0065] Phase Two: Online Sparse Reconstruction

[0066] 1. Data Acquisition and Feature Extraction: In a real-world scenario, muon observation data is collected over a fixed time period using a muon detector array. This muon observation data is then statistically analyzed to generate observation vectors consistent with the format used in the offline dictionary preparation stage. (For example, a histogram representing the muon count within a specific angular interval).

[0067] 2. Construct an enhanced objective function: for the observation vector The actual muon observation data collected; the following enhanced objective function is constructed:

[0068]

[0069] in, It is a dictionary matrix; Let be the sparse coefficient vector to be solved, used to represent the observed data. In the dictionary Sparse decomposition under the following conditions; The image to be reconstructed is used to characterize the distribution of subsurface anomalies; it should be noted that the constraints... In the image In algebraic operations, it is treated as a vector. (That is, its matrix is ​​concatenated column-wise as a vector), but when total variation calculations are involved, it is reshaped into matrix form. Similarly, in subsequent ADMM solutions, auxiliary variables... and Both are synonyms, representing the image as a vector, but they need to be reconstructed into a matrix before TV denoising.

[0070] Block sparse regularization term middle: sparse coefficient vector The middle belongs to the first Subvectors of groups, each group corresponding to a spatial support region of a potential anomalous body; A pre-defined set of groups that are adaptively learned from data; This is a tradeoff parameter used to balance the effects of the data fit term and the block sparsity regularization term. If If so, then only total variational regularization is performed.

[0071] Total variational regularization term middle: For image The total variation of anisotropy, ; To balance the effects of the data fitting term, block sparsity regularization term, and total variation regularization term. If Then only block sparse regularization is performed.

[0072] Constraints middle: Let be the linear mapping matrix from sparse coefficients to the image space, derived from the dictionary. Determined by observation geometry.

[0073] 3. Optimized Solution: The above optimization problem can be solved efficiently using the Alternating Direction Multiplier Method (ADMM); specifically, it includes the following steps:

[0074] Variable splitting: introducing auxiliary variables and This transforms the constrained optimization problem into an equivalent unconstrained multivariable optimization problem.

[0075] Alternating updates:

[0076] renew (Sparse coefficient): Solving for a band Norm-constrained ridge regression problems can be solved quickly using the conjugate gradient method.

[0077] renew (Block sparse approximation): For grouped... Apply the group Lasso contraction operator, i.e., for each group Perform soft thresholding:

[0078]

[0079] renew (Image Denoising and Boundary Preservation): Solve a TV denoising problem (in this embodiment, the ROF model is used) using the Chambolle dual algorithm or the fast gradient projection method.

[0080] Update Lagrange multipliers: Update according to standard ADMM procedures.

[0081] Iteration: Repeat the alternating update steps until the convergence condition is met. In this embodiment, the convergence condition is that the original residual and the dual residual are sufficiently small.

[0082] like Figure 5 As shown, traditional L1 regularization ( Figure 5 While neutron maps (c) can promote sparsity, they often lead to dispersed activation of coefficients, resulting in point-like artifacts and incomplete target morphology. In contrast, the block sparsity regularization introduced in this embodiment ( Figure 5 Neutron maps (d) can induce the activation of groups of dictionary atoms associated with the same geological anomaly, thereby reconstructing a continuous and complete target region. Total variational regularization (TV) Figure 5 The neutron map (e) further smooths the uniform background and sharpens the target boundary, highlighting the difference information between the target and the background. Figure 5 The difference images of the subgraphs (f) and (g) visually demonstrate the improvement of block sparsity and total variational regularization over L1 regularization.

[0083] 4. Image Generation and Interpretation: This involves generating and interpreting the image after final convergence. As part of the reconstruction results, connected regions with significantly higher brightness than the background are identified as inferred underground anomalies (such as isolated boulders), while regions with significantly lower brightness than the background are identified as inferred cavities. The specific criteria can be set as follows: using the average pixel value of the image background region as a benchmark, pixels higher than the background average by a certain multiple (e.g., 1.5 times) are defined as high-density anomaly candidate points, and pixels lower than the background average by a certain multiple (e.g., 0.5 times) are defined as low-density anomaly candidate points. Connected component analysis is then used to remove isolated noise, resulting in the final anomaly distribution.

[0084] Based on the above method, verification was conducted through computer simulation: First, a tunnel overburden model was established, containing two sparsely distributed spherical cavities and a high-density boulder. Simulations were performed using muon observation data equivalent to only 1 hour and 24 hours of accumulated data, respectively. Then, the traditional PoCA (Point Nearest Approach) algorithm was used to process the 1-hour data, resulting in an extremely noisy reconstructed image that failed to identify any targets. The method of this embodiment was then used to process the same 1-hour data, clearly reconstructing the positions and approximate shapes of the three targets. Through the above verification, the results of this embodiment's method processing 24-hour data are comparable in quality to the results of the traditional algorithm processing 72-hour data, demonstrating the significant advantage of this embodiment in imaging speed. A comparison diagram is shown below. Figure 5 As shown.

[0085] In another application scenario, this invention can also be used for advanced geological prediction in front of tunnel boring machines. For example... Figure 1 As shown, the muon detector array is installed on the trolley behind the tunnel boring machine (TBM) or on an existing tunnel segment, with the detection direction pointing towards the unexcavated strata in front of the TBM. Cosmic ray muons penetrate the soil and rock layer in a horizontal direction and are captured by the rear detector. The system records the exit position and azimuth angle (θ, φ) of each muon, generating an observation vector. .

[0086] The offline dictionary preparation for this scenario requires simulation data with horizontal incidence: in the GEANT4 simulation, the source of the muon is set to the horizontal direction, the geological background model is a typical stratigraphic structure, and the target anomalies are possible isolated rocks, cavities, etc., in front. After the dictionary training is completed, the online sparse reconstruction process is completely consistent with the aforementioned embodiment: the same enhanced objective function is constructed, solved using the ADMM algorithm, and finally outputs a three-dimensional density image of the strata in front. Tests show that this method can complete the identification of anomalies with a diameter of more than 2 meters within a 50-meter range in front within 24 hours, effectively meeting the geological early warning requirements of shield tunneling construction.

[0087] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A muon quasi-scattering imaging device based on sparse representation and dictionary learning, characterized in that, It includes a muon detector array deployed on one side of the measured area, a high-speed data acquisition system, and a computing and control unit; The muon detector array is used to capture muon quasi-scattering observation data after penetrating the measured region; The high-speed data acquisition system is used to capture sparse muon instances; The computing and control unit is used to run the imaging algorithm software, control the overall operation of the system, and output reconstructed images of underground anomalies.

2. The muon quasi-scattering imaging device based on sparse representation and dictionary learning according to claim 1, characterized in that, It also includes a momentum spectrometer module integrated into the muon detector array, which is used to measure the momentum of some muons, providing additional information dimensions for imaging.

3. The muon quasi-scattering imaging device based on sparse representation and dictionary learning according to claim 1, characterized in that, The muon detector array adopts a modular design, with each module consisting of a multilayer plastic scintillator strip coupled with a silicon photomultiplier tube to measure the emission position and direction of muons.

4. A muon quasi-scattering imaging method based on sparse representation and dictionary learning, applicable to the muon quasi-scattering imaging device based on sparse representation and dictionary learning as described in any one of claims 1-3, characterized in that, The process includes offline dictionary preparation and online sparse reconstruction steps. The offline dictionary preparation provides a trained, overcomplete dictionary for the online sparse reconstruction. ; The offline dictionary preparation includes sub-steps such as scene and target definition, Monte Carlo simulation, and dictionary training. The online sparse reconstruction includes sub-steps such as data acquisition and feature extraction, construction of an enhanced objective function, optimization solution, and image generation and interpretation.

5. The muon quasi-scattering imaging method based on sparse representation and dictionary learning according to claim 4, characterized in that, The specific process for preparing the offline dictionary is as follows: S11. Define specific application scenarios, define digital models of typical underground anomalies containing parameters of different locations, sizes, and densities, and form a target model library to be detected; S12. Place each target model to be detected in the set geological background model, simulate the penetration of cosmic ray muons through the entire system, record the muon emission direction angle (θ, φ) at the virtual detector position, and form quasi-scattering observation signal instances corresponding to each target model; S13. Collect all instances of quasi-scattering observation signals to form a training sample set. Where N is the number of samples, the overcomplete dictionary is obtained by solving the optimization problem using the K-SVD dictionary learning algorithm. The optimization problem is: in, It is a sparse coefficient matrix; The square of the Frobenius norm; for Norm; This is the threshold for sparsity constraints.

6. The muon quasi-scattering imaging method based on sparse representation and dictionary learning according to claim 4, characterized in that, In the online sparse reconstruction, the specific process of data acquisition and feature extraction is as follows: Mun observation data for a fixed time period is acquired through a Mun detector array; the observation data is statistically processed to generate observation vectors with the same format as those in the offline dictionary preparation stage. .

7. The muon quasi-scattering imaging method based on sparse representation and dictionary learning according to claim 4, characterized in that, In the online sparse reconstruction, the enhanced objective function is constructed as follows: in, It is a dictionary matrix; Let be the sparse coefficient vector to be solved; The image to be reconstructed; For block sparse regularization; sparse coefficient vector The middle belongs to the first Subvectors of each group; A tradeoff parameter is used to balance the effects of the data fitting term and the block sparsity regularization term. For total variational regularization; For image The total variation of anisotropy; To weigh the effects of the data fitting term and the total variation regularization term; Let be a linear mapping matrix from sparse coefficients to the image space, used to map the sparse coefficient vector. Density distribution converted to image space From the dictionary Determined by observation geometry.

8. The muon quasi-scattering imaging method based on sparse representation and dictionary learning according to claim 7, characterized in that, The total variation of anisotropy The calculation formula is: in, For image The pixel value located in the i-th row and j-th column.

9. The muon quasi-scattering imaging method based on sparse representation and dictionary learning according to claim 7, characterized in that, The optimization solution is obtained using the Alternating Direction Multiplier Method (ADMM), and the specific process is as follows: S31. Variable Splitting: Introducing Auxiliary Variables and This transforms the constrained optimization problem into an equivalent form. S32, Alternating Update: Update the sparsity coefficients sequentially. , block sparse approximation Image denoising and boundary preservation And update the Lagrange multipliers according to the standard ADMM procedure; Among them, update To solve a band Norm-constrained ridge regression problems can be solved quickly using the conjugate gradient method; update For grouped Apply the group Lasso contraction operator for each group Perform soft thresholding, the formula is as follows: ;renew To solve the total variational denoising problem, the ROF model is combined with the Chambolle dual algorithm or the fast gradient projection method. S33. Iteration: Repeat step S32 until the original residual and the dual residual satisfy the preset convergence condition, and obtain the converged result. and .

10. The muon quasi-scattering imaging method based on sparse representation and dictionary learning according to claim 4, characterized in that, The specific process of image generation and interpretation is as follows: The converged image is then... As the final reconstruction result, the image Connected regions with significantly higher brightness than the background are identified as high-density underground anomalies, while regions with significantly lower brightness than the background are identified as low-density underground anomalies.

Images