High-throughput Raman hyperspectral imaging method and system based on spectral distance modeling

Abstract

The invention discloses a high-throughput Raman hyperspectral imaging method and system based on spectral distance modeling. The method comprises the following steps: S1, representing an initial Raman hyperspectral image as superposition of an optimized Raman hyperspectral image and noise; then, an optimization problem is constructed by taking solving of the optimized Raman hyperspectral image as a target; s2, introducing an auxiliary variable to convert the optimization problem to obtain a converted optimization problem, and then constructing a Lagrange function according to the converted optimization problem; and S3, constructing a plurality of sub-problems according to a Lagrange function, solving the plurality of sub-problems through iterative optimization until a preset condition is met, and ending the iterative optimization so as to obtain an optimized Raman hyperspectral image. According to the method, efficient low-signal-to-noise-ratio image reconstruction is realized, meanwhile, the unsupervised neural network is designed by utilizing the long-range correlation between the spectrum and the space, the imaging quality is remarkably improved, the method does not depend on large-scale training data, the adaptability is high, and the dependence on equipment is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing and spectral information processing, and particularly relates to a high-throughput Raman hyperspectral imaging method and system based on spectral distance modeling. Background Art

[0002] Raman hyperspectral imaging (HSI) is an analytical tool that combines Raman spectroscopy and imaging technology. It can provide molecular vibration information and its distribution in space, and is widely used in biomedical, drug R & D, materials science, and pharmaceuticals. Compared with other chemical imaging methods, such as fluorescence imaging and Fourier transform infrared spectroscopy imaging, Raman hyperspectral imaging technology has the advantages of non-destructive analysis, simple sample preparation, and high spatial and spectral resolution. However, there are still major challenges in improving the quality of Raman hyperspectral imaging. With the research of computational imaging algorithms, a method for reconstructing Raman hyperspectral images with high signal-to-noise ratio (SNR) from low-quality or incomplete measurements is provided based on computational imaging algorithms. For example: Traditional noise reduction algorithms (such as singular value decomposition SVD): Perform SVD decomposition on low-quality Raman spectral images and save their main component signals. However, such methods have limited effects in separating weak Raman signals and complex background noise. Deep learning methods: In recent years, research has shown that deep learning methods can significantly improve imaging quality, but they rely on large-scale paired datasets of low SNR and high SNR Raman spectral images for training. The collection cost of such data is high or even impossible to achieve. In addition, such methods generally adopt a "training-testing" framework. First, the training process requires a large amount of high-quality data, and the training process is complex and time-consuming. In addition, the instrument dependence is strong, and the data characteristics collected by different devices vary greatly. The model needs to be pre-trained for specific devices or data before it can be applied. And the generalization ability of existing algorithms is limited and cannot process spectral data collected by different Raman instruments. Summary of the Invention

[0003] The present invention provides a high-throughput Raman hyperspectral imaging method and system based on spectral distance modeling to solve the technical problems mentioned in the background art.

[0004] To achieve the above object, the technical solution of the present invention is realized as follows: The present invention provides a high-throughput Raman hyperspectral imaging method based on spectral distance modeling, including the following steps: S1. Represent the obtained initial Raman hyperspectral image as the optimized Raman hyperspectral image and noise By superimposing them, a Raman hyperspectral imaging mathematical model is obtained, and then an optimization problem is constructed based on the Raman hyperspectral imaging mathematical model; S2. Introduce auxiliary variables to transform the optimization problem, obtain the transformed optimization problem, and then construct a Lagrangian function based on the transformed optimization problem; S3. Construct multiple sub-problems based on the Lagrangian function, and solve the multiple sub-problems through iterative optimization until the preset conditions are met, ending the iterative optimization, thereby obtaining the optimized Raman hyperspectral image 。

[0005] Furthermore, the specific steps of S1 are as follows: S11. Represent the obtained initial Raman hyperspectral image as the optimized Raman hyperspectral image superimposed with noise to obtain a Raman hyperspectral imaging mathematical model, and the Raman hyperspectral imaging mathematical model is specifically: ; S12. Then, with the goal of solving the optimized Raman hyperspectral image and based on the Raman hyperspectral imaging mathematical model, construct an optimization problem, and the optimization problem is specifically as follows: ; where, respectively represent the optimized Raman hyperspectral image to be solved and the noise to be solved, represents the Frobenius norm; represents the L1 norm; is a regularization term, that is, the prior information constraint on the data; 、 are two different balance coefficients respectively.

[0006] Furthermore, the specific steps of S2 are as follows: S21. Introduce an auxiliary variable to transform the optimization problem in S1, where , and obtain the transformed optimization problem. The transformed optimization problem is specifically as follows: ; where, represents the auxiliary variable to be solved; S22. Construct a Lagrangian function based on the transformed optimization problem. The constructed Lagrangian function is specifically: ; where, Denote with respect to , , , the Lagrangian function, where is the Lagrange multiplier,

[0007] Furthermore, the S3 specifically includes the following steps: S31. Construct multiple sub-problems based on the Lagrangian function, and the multiple sub-problems are respectively sub-problem, sub-problem, sub-problem and sub-problem; S32. Iteratively solve the sub-problem, sub-problem, sub-problem and sub-problem for multiple times; S33. Check whether the following convergence conditions are met, or whether the preset maximum number of iterations is reached. If so, stop the iteration to obtain the optimized Raman hyperspectral image ; where the convergence conditions are specifically as follows: ; where the auxiliary variable obtained in the th iteration; denotes norm; denotes a given extremely small constant.

[0008] Furthermore, the sub-problem is specifically represented by the following mathematical formula: ; where denotes the Raman hyperspectral image obtained in the th iteration; denotes the noise obtained in the th iteration; denotes the auxiliary variable obtained in the th iteration; denotes the Lagrange multiplier obtained in the th iteration; The sub-problem is specifically as follows: ; where denotes the auxiliary variable obtained in the th iteration; The sub - problems are as follows: ; Among them, represents the noise obtained in the th iteration; The sub - problems are as follows: ; Among them, represents the Lagrange multiplier obtained in the th iteration.

[0009] Furthermore, the specific steps of S32 are as follows: S321. Solve and update the sub - problem, sub - problem, sub - problem, and sub - problem multiple times; S322. After the th iteration, obtain , and ; S323. Perform the th iteration on the sub - problem, and input , , and into the sub - problem to obtain the solution of the th iteration sub - problem as: ; Among them, is the identity matrix, and is the scaled identity matrix; S324. Perform the th iteration on the sub - problem, take and as the input, and finally obtain the solution of the sub - problem at the th iteration, that is, ; S325. Take as the input, and update the noise variable through the second objective function; The second objective function is as follows: ; S326. Take , and as input, and update the Lagrange multiplier through the dual ascent method in the augmented Lagrangian method to obtain the Lagrange multiplier after the th iteration ; The Lagrange multiplier after the th iteration is solved as follows: .

[0010] Furthermore, the S324 specifically includes the following steps: S3241. In the th iteration, decompose into the prior image PI and the optimal spectral distance OSD; S3242. Initialize the prior image PI to obtain the initialized prior image , and then use the prior image to initialize the optimal spectral distance OSD respectively to obtain the initialized spectral distance ; The initialized prior image is expressed by the formula as follows: ; where B represents the number of spectral bands, represents the initialized variable i.e., the variable in the first round of training in the th iteration , represents all the values of the b-th band, represents the b-th band; The initialized spectral distance is expressed by the formula as follows: ; S3243. Use the predefined objective function two, regularization term , and the initialized spectral distance to perform multiple rounds of training and update on the weight parameters of the unsupervised neural network UNN to obtain the updated unsupervised neural network UNN; where the objective function two is specifically as follows: ; where, is the network parameter in the t-th round of training in the th iteration; is the spectral distance in the t-th round of training in the th iteration; is the The prior image in the t-th round of training in the i-th iteration; Regularization term The details are as follows: ; Wherein, is the observation value consistency constraint, that is, minimizing the generated image and the initial Raman hyperspectral image 's L1 distance; is the spectral-spatial smoothness constraint; is the balance factor of spectral-spatial smoothness; S3244. Input the spectral distance in the t-th round of training in the i-th iteration into the updated unsupervised neural network UNN to obtain the optimized spectral distance , which is specifically expressed by the formula as follows: ; S3245. Calculate the mean value of all bands of the current reconstructed image as the optimized prior image ; It is specifically expressed by the formula as follows: ; S3246. Then add the optimized spectral distance and the optimized prior image to obtain the reconstructed image , or the auxiliary variable in the i-th iteration.

[0011] Furthermore, the unsupervised neural network UNN includes an encoder and a decoder connected to the encoder; Wherein the encoder includes three identical sub-modules one connected in sequence, and the sub-module one includes a convolutional layer one, an activation function layer one, an SDTrans module, and an activation function layer two connected in sequence; The decoder includes three identical sub-modules two connected in sequence. The sub-module two includes an upsampling layer, a convolutional layer two, an activation function layer three, an SDTrans module, and an activation function layer four connected in sequence. The inputs of the three sub-modules one of the encoder are respectively connected to the outputs of the upsampling layers in the three sub-modules two of the decoder through splicing operations and reverse connection methods.

[0012] Furthermore, the SDTrans module includes a Transformer module and an attention mechanism connected to the Transformer module; ​Among them, the attention mechanism is used to perform pooling operations, matrix multiplication, and row-wise averaging on the output of the Transformer module respectively, so as to obtain the weights of each band.

[0013] On the other hand, the present invention also provides a high-throughput Raman hyperspectral imaging system, including a computer device, which is programmed or configured to execute the above high-throughput Raman hyperspectral imaging method.

[0014] Advantages of the present invention: 1. The present invention discloses a high-throughput Raman hyperspectral imaging method based on spectral distance modeling, which combines an iterative optimization framework to achieve efficient low signal-to-noise ratio image reconstruction. At the same time, an unsupervised neural network is designed using the long-range correlation between spectra and space, significantly improving the imaging quality, not relying on large-scale training data, having strong adaptability, and reducing the dependence on equipment.

[0015] 2. By predicting the spectral distance instead of directly reconstructing the complete spectrum, the present invention significantly reduces the computational complexity. The present invention reduces the amount of calculation during the optimization process, making the image reconstruction process more efficient, especially outstanding in the case of limited computing resources.

[0016] 3. The present invention does not rely on a large-scale training data set, so it can effectively perform high-quality image reconstruction even under low signal-to-noise ratio conditions. This makes the present invention particularly important in practical application scenarios where data acquisition is difficult and can avoid the dependence on a large amount of labeled data.

[0017] 4. The present invention utilizes the long-range correlation between spectra and space to optimize the image reconstruction process, thereby significantly improving the signal-to-noise ratio and imaging quality of the image. This correlation performance helps the model better capture potential image information, reduces noise interference, and can provide a clearer reconstructed image.

[0018] 5. The present invention has high generality and adaptability and can be applied to a variety of complex imaging scenarios, such as high-precision medical imaging, industrial inspection, etc., and has broader application potential.

[0019] 6. The present invention can not only accelerate the imaging process, significantly improve the work efficiency, but also expand the applicability of Raman hyperspectral imaging (HSI) in dynamic sample detection and resource-constrained environments. Especially in scenarios where rapid imaging is required, the present invention can effectively provide real-time simulation results and test data, and has important academic value and broad application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 It is a flowchart of the high-throughput Raman hyperspectral imaging method in the present invention; Figure 2It is the structural block diagram of the unsupervised neural network UNN in the present invention; Figure 3 It is the structural block diagram of the SDTrans module in the present invention. Detailed implementation manners

[0021] To facilitate the understanding of the present invention, the present invention will be described more comprehensively below with reference to the relevant drawings. Preferred embodiments of the present invention are shown in the drawings. However, the present invention can be implemented in many other different forms and is not limited to the embodiments described herein. On the contrary, these embodiments are provided to make the understanding of the disclosure of the present invention more thorough and comprehensive.

[0022] Refer to Figure 1 , an embodiment of the present application provides a high-throughput Raman hyperspectral imaging method based on spectral distance modeling, including the following steps: S1. Represent the acquired initial Raman hyperspectral image as the superimposition of the optimized Raman hyperspectral image and the noise to obtain a Raman hyperspectral imaging mathematical model, and then construct an optimization problem according to the Raman hyperspectral imaging mathematical model; wherein, , , H, W, and B respectively represent the height, width, and spectral band number of the Raman spectral image; represents a vector space; Preferably, the initial Raman hyperspectral image is low signal-to-noise ratio data, that is, a low-quality Raman hyperspectral image ; the optimized Raman hyperspectral image is a high-quality Raman hyperspectral image ; S2. Introduce an auxiliary variable to transform the optimization problem to obtain a transformed optimization problem, and then construct a Lagrangian function according to the transformed optimization problem; S3. Construct multiple sub-problems according to the Lagrangian function, and solve the multiple sub-problems through iterative optimization until a preset condition is met, and end the iterative optimization, so as to obtain a high-quality Raman hyperspectral image .

[0023] The present invention discloses a high-throughput Raman hyperspectral imaging method based on spectral distance modeling, which combines an iterative optimization framework to achieve efficient low signal-to-noise ratio image reconstruction. At the same time, an unsupervised neural network is designed by using the long-range correlation between spectra and space, which significantly improves the imaging quality, does not rely on large-scale training data, has strong adaptability, and reduces the dependence on equipment.

[0024] In some embodiments, the S1 specifically includes the following steps: S11. Represent the acquired low-quality Raman hyperspectral image as a high-quality Raman hyperspectral image and noise superposition to obtain a Raman hyperspectral imaging mathematical model. The Raman hyperspectral imaging mathematical model is specifically: ; S12. Then, taking the solution of the high-quality Raman hyperspectral image as the goal and constructing an optimization problem based on the Raman hyperspectral imaging mathematical model. The optimization problem is specifically as follows: ; Among them, respectively represent the high-quality Raman hyperspectral image to be solved and the noise to be solved , represents the Frobenius norm; represents the L1 norm; is a regularization term, that is, the prior information constraint on the data; , are two different balance coefficients respectively. The

[0025] term helps to reduce the influence of sparse noise interference such as cosmic rays by constraining the sparsity of the noise. In some embodiments, the S2 specifically includes the following steps: S21. Introduce an auxiliary variable to transform the optimization problem in S1. Among them ; Among them, represents the auxiliary variable to be solved ; S22. Construct a Lagrangian function based on the transformed optimization problem. The constructed Lagrangian function is specifically: ; Among them, represents the Lagrangian function with respect to , , , , is the Lagrange multiplier, is the balance factor.

[0026] In some embodiments, the S3 specifically includes the following steps: S31. Construct multiple sub - problems based on the Lagrangian function. The multiple sub - problems are respectively sub - problem, sub - problem, sub - problem, and sub - problem; S32. Iteratively solve the sub - problem, sub - problem, sub - problem, and sub - problem multiple times; S33. Check whether the following convergence conditions are met, or whether the preset maximum number of iterations is reached. If so, stop the iteration to obtain the optimized Raman hyperspectral image ; Among them, the convergence conditions are specifically as follows: ; Among them, the auxiliary variable obtained in the th iteration; represents the norm;

[0027] In some embodiments, the sub - problem is specifically represented by the following mathematical formula: ; Among them, represents the Raman hyperspectral image obtained in the th iteration; represents the noise obtained in the th iteration; represents the auxiliary variable obtained in the th iteration; represents the Lagrange multiplier obtained in the th iteration; The sub - problem is specifically as follows: ; Among them, represents the auxiliary variable obtained in the th iteration; The sub - problem is specifically as follows: ; Among them, represents the noise obtained in the th iteration; The sub - problem is specifically as follows: ; Among them, represents the Lagrange multiplier obtained in the th iteration.

[0028] In some embodiments, S32 specifically includes the following steps: S321. Solve and update the sub-problem, sub-problem, sub-problem, and sub-problem multiple times; Specifically, in the first iteration, that is, when and are initialized to 0, is a low-quality Raman spectrum image ; S322. After the th iteration, , , and are obtained; S323. Perform the th iteration on the sub-problem, and input , , , and into the sub-problem to obtain the solution of the th iteration sub-problem as: ; Among them, is the identity matrix, is the scaled identity matrix; S324. Perform the th iteration on the sub-problem, and use and as the input to finally obtain the solution of the sub-problem at the th iteration, that is, ; S325. Use as the input and update the noise variable through the second objective function; the second objective function is specifically as follows: ; S326. Use , , and As the input, and update the Lagrange multiplier through the dual ascent method in the augmented Lagrangian method to obtain the Lagrange multiplier after the th iteration ; The solution formula of the Lagrange multiplier after the th iteration is specifically as follows: .

[0029] In some embodiments, the S324 specifically includes the following steps: S3241. In the th iteration, decompose into the prior image PI and the optimal spectral distance OSD; Represented by a formula, it is specifically as follows: ; Among them, the prior image (PI) is obtained by taking the average of all bands of the low signal-to-noise ratio Raman spectral image, and is used to provide a preliminary high signal-to-noise ratio image. The optimal spectral distance (OSD) is dynamically optimized and generated by the designed lightweight neural network (UNN), and is used to supplement detailed information; S3242. Initialize the prior image PI to obtain the initialized prior image , and then use the prior image to initialize the optimal spectral distance OSD respectively to obtain the initialized spectral distance ; The initialized prior image is represented by a formula, specifically as follows: ; Among them, B represents the number of spectral bands, represents the initialized variable that is, the variable in the first round of training in the th iteration, represents all values of the bth band, represents the bth band; The initialized spectral distance is represented by a formula, specifically as follows: ; S3243. Use the predefined objective function two and the regularization term , the initialized spectral distance to perform multiple rounds of training and update on the weight parameters of the unsupervised neural network UNN to obtain the updated unsupervised neural network UNN; Among them, the objective function two is specifically as follows: ; Among them, is the network parameter for the t-th round of training in the -th iteration; is the spectral distance for the t-th round of training in the -th iteration; is the prior image for the t-th round of training in the -th iteration; The regularization term is specifically as follows: ; Among them, is the observation value consistency constraint, that is, minimizing the generated image (i.e., the auxiliary variable obtained in the (t + 1)-th iteration) and the initial Raman hyperspectral image 's L1 distance; is the spectral-spatial smoothness constraint; is the balance factor of spectral-spatial smoothness; SSTV is the spectral-spatial total variation (Spectral-Spatial Total Variation); Adopt spectral-spatial total variation to enhance smoothness; S3244. Input the spectral distance for the t-th round of training in the -th iteration into the updated unsupervised neural network UNN to obtain the optimized spectral distance , which is specifically expressed by the formula as follows: ; S3245. Calculate the mean value of all bands of the current reconstructed image (i.e., the auxiliary variable obtained in the t-th iteration) as the optimized prior image ; It is specifically expressed by the formula as follows: ; S3246. Then add the optimized spectral distance and the optimized prior image to obtain the reconstructed image , or the auxiliary variable for the -th iteration.

[0030] The present invention significantly reduces the computational complexity by predicting the spectral distance instead of directly reconstructing the complete spectrum. The present invention reduces the amount of calculation during the optimization process, making the image reconstruction process more efficient, especially outstanding in the case of limited computing resources.

[0031] In some embodiments, referring to Figure 2, the unsupervised neural network UNN includes an encoder and a decoder connected to the encoder; The encoder includes three identical sub-modules one connected in sequence. The sub-module one includes a convolutional layer one, an activation function layer one, an SDTrans module, and an activation function layer two connected in sequence. The convolutional kernel size of the convolutional layer one is 3 and the stride is 2, which is used to extract local features and reduce the resolution of the feature map. The SDTrans module is used to capture spectral correlation and non-local spatial dependence and is the core component for spectral distance modeling.

[0032] The decoder includes three identical sub-modules two connected in sequence. The sub-module two includes an upsampling layer, a convolutional layer two, an activation function layer three, an SDTrans module, and an activation function layer four connected in sequence. The inputs of the three sub-modules one of the encoder are respectively connected to the outputs of the upsampling layers in the three sub-modules two of the decoder through splicing operations and reverse connection methods.

[0033] The upsampling layer is used to restore the spatial dimension of the feature map; The convolutional kernel size of the convolutional layer two is 3 and the stride is 1, which is used to extract local features; In some embodiments, referring to Figure 3 , the SDTrans module includes a Transformer module and an attention mechanism (i.e., Figure 3 the spatial-aware attention in connected to the Transformer module);

[0034] The SDTrans module captures long-range dependencies by introducing the Transformer module to make up for the limitations of traditional convolutional layers. It mainly includes a (multi-head) self-attention mechanism and a convolutional projection layer, and stabilizes the training process through residual connections and normalization. This design fully considers the correlation between spectrum and space, enabling the unsupervised neural network UNN to have efficient and stable modeling capabilities in complex imaging scenarios.

[0035] On the other hand, the present invention also provides a high-throughput Raman hyperspectral imaging system, including a computer device, which is programmed or configured to execute the above high-throughput Raman hyperspectral imaging method.

[0036] As described above, it is only the specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention can easily think of changes or substitutions, which should all be covered within the protection scope of the present invention. Moreover, the technical solutions between various embodiments of the present invention can be combined with each other, but it must be based on the fact that those skilled in the art can implement it. When the combination of technical solutions appears to be contradictory or unable to be implemented, it should be considered that such a combination of technical solutions does not exist and is not within the protection scope required by the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims described above.

Claims

1. A high-throughput Raman hyperspectral imaging method based on spectral distance modeling, characterized in that: The steps include: S1. The initial Raman hyperspectral image acquired Represented as the optimized Raman hyperspectral image With noise The mathematical model of Raman hyperspectral imaging is obtained by superposition, and then the optimization problem is constructed based on the mathematical model of Raman hyperspectral imaging; S2. Introduce auxiliary variables to transform the optimization problem, obtain the transformed optimization problem, and then construct the Lagrangian function based on the transformed optimization problem; S3. Construct multiple sub-problems based on the Lagrangian function and solve them through iterative optimization until the preset conditions are met. Then, the iterative optimization is terminated to obtain the optimized Raman hyperspectral image. .

2. The high-throughput Raman hyperspectral imaging method according to claim 1, characterized in that: The S1 specifically includes the following steps: S11, the initial Raman hyperspectral image acquired Represented as the optimized Raman hyperspectral image With noise The mathematical model of Raman hyperspectral imaging is obtained by superposition. The specific mathematical model of Raman hyperspectral imaging is: ； S12, then solve the optimized Raman hyperspectral image The optimization problem is constructed based on the mathematical model of Raman hyperspectral imaging as the goal. The specific optimization problem is as follows: ； in, Respectively represent the optimized Raman hyperspectral images to be solved and the noise to be solved , represents the Frobenius norm; represents the L1 norm; is the regularization term, i.e., the prior information constraint on the data; , They are two different balance coefficients.

3. The high-throughput Raman hyperspectral imaging method according to claim 2, characterized in that: The S2 specifically includes the following steps: S21. Introducing auxiliary variables Transform the optimization problem in S1 into , we get the optimization problem of conversion, which is as follows: ； in, represents the auxiliary variable Z to be solved; S22. Constructing a Lagrangian function based on the transformed optimization problem. The Lagrangian function is specifically: ; in, Indicates about , , , The Lagrangian function of is the Lagrange multiplier, is the balancing factor.

4. The high-throughput Raman hyperspectral imaging method according to claim 3, characterized in that: The S3 specifically includes the following steps: S31. Construct multiple sub-problems based on the Lagrangian function. The sub-problems are Sub-questions, Sub-questions, Sub-questions and Sub-questions; S32, in turn Sub-questions, Sub-questions, Sub-questions and The sub-problems are solved iteratively multiple times; S33, check whether the following convergence conditions are met, or whether the preset maximum number of iterations is reached, if yes, stop the iteration, so as to obtain the optimized Raman hyperspectral image ; The convergence conditions are as follows: ； in, No. Auxiliary variables obtained by iteration; express norm; Represents a given very small constant.

5. The high-throughput Raman hyperspectral imaging method according to claim 4, characterized in that: Said The sub-problems are expressed mathematically as follows: ； in, Indicates The Raman hyperspectral image obtained by the iteration; Indicates The noise obtained by the iteration; Indicates Auxiliary variables obtained by iteration; Indicates The Lagrange multiplier obtained by the iteration; Said The sub-questions are as follows: ； in, Indicates Auxiliary variables obtained by iteration; Said The sub-questions are as follows: ； in, Indicates The noise obtained by the iteration; Said The sub-questions are as follows: ； in, Indicates The Lagrange multiplier obtained by the iteration.

6. The high-throughput Raman hyperspectral imaging method according to claim 5, characterized in that: The S32 specifically includes the following steps: S321, in turn Sub-questions, Sub-questions, Sub-questions and The sub-problems are solved and updated multiple times; S322, No. After iterations, we get , as well as ; S323, yes Sub-question In iterations, , , as well as Input to In the sub-problem, we get Iterations The solution to the subproblem is: ； in, is the identity matrix, is the scaled identity matrix; S324, will Sub-question In iterations, and As input, we finally get Sub-question The solution of the iteration is ; S325, will As input, the noise variable is updated through objective function 2 ; The specific objective function 2 is as follows: ； S326, will , as well as As input, and update the Lagrange multiplier by the dual ascent method in the augmented Lagrangian method, we get The Lagrange multiplier after iteration ;No. The Lagrange multiplier after iteration The solution is as follows: 。 7. The high-throughput Raman hyperspectral imaging method according to claim 6, characterized in that: The S324 specifically includes the following steps: S3241, in In the iteration, Decomposed into prior image PI and optimal spectral distance OSD; S3242, initialize the prior image PI to obtain the initialized prior image , and then use the prior image Initialize the optimal spectral distance OSD respectively to obtain the initialized spectral distance ; Initialized prior image The formula is as follows: ； Where B represents the number of spectral bands, Indicates the initialized variable , that is, Variables for the first round of training in iteration , represents all values ​​of the b-th band, represents the bth band; Spectral distance after initialization The formula is as follows: ； S3243, using the predefined objective function 2 and regularization term , spectral distance after initialization Weight parameters of unsupervised neural network UNN After multiple rounds of training and updating, an updated unsupervised neural network UNN is obtained; the objective function 2 is as follows: ； in, For the The network parameters for the tth round of training in the iteration; For the The spectral distance of the t-th round of training in the iteration; For the The prior image of the t-th round of training in the iteration; Regularization term The details are as follows: ； in, is the observation consistency constraint, that is, minimizing the generated image Compared with the initial Raman hyperspectral image L1 distance; is the spectral-spatial smoothness constraint; is the balancing factor of spectral-spatial smoothness; S3244, will The spectral distance of the t-th round of training in the iteration Input into the updated unsupervised neural network UNN to obtain the optimized spectral distance , the formula is as follows: ； S3245, calculate the current reconstructed image The mean of all bands is used as the optimized prior image ; The formula is as follows: ； S3246, then the optimized spectral distance and the optimized prior image Add together to get the reconstructed image , or the Auxiliary variables for the iteration.

8. The high-throughput Raman hyperspectral imaging method according to claim 7, characterized in that: The unsupervised neural network UNN comprises an encoder and a decoder connected to the encoder; The encoder includes three identical sub-modules 1 connected in sequence, wherein the sub-module 1 includes a convolutional layer 1, an activation function layer 1, an SDTrans module, and an activation function layer 2 connected in sequence; The decoder includes three sequentially connected and identical sub-modules 2, wherein the sub-module 2 includes an upsampling layer, a convolutional layer 2, an activation function layer 3, an SDTrans module and an activation function layer 4 connected in sequence, and the inputs of the three sub-modules 1 of the encoder are connected to the outputs of the upsampling layers in the three sub-modules 2 of the decoder through splicing operations and reverse connections.

9. The high-throughput Raman hyperspectral imaging method according to claim 8, characterized in that: The SDTrans module includes a Transformer module and an attention mechanism connected to the Transformer module; Among them, the attention mechanism is used to perform pooling operations, matrix multiplication, and row-wise averaging on the output of the Transformer module to obtain the weight of each band.

10. A high-throughput Raman hyperspectral imaging system, comprising a computer device, characterized in that: The computer device is programmed or configured to execute the high-throughput Raman hyperspectral imaging method according to any one of claims 1 to 9.

Images