Spectrum reconstruction method based on variational feature fusion and wavelet dense residual attention

Abstract

According to the spectrum reconstruction method based on variational feature fusion and wavelet dense residual attention provided by the invention, a spectrum reconstruction model based on an attention mechanism is used for reducing the dependence of calculating a reconstruction spectrometer on a correlation coefficient of a micro-nano optical material, and the precision and effect of spectrum reconstruction can be effectively improved; besides, a wavelet dense residual attention module and a variational feature fusion processing module provided by the invention can effectively extract and refine spectral features, eliminate useless features from two aspects of high frequency and low frequency of data, and enhance the fusion effect of deep high-dimensional features, so that the performance of spectrum reconstruction is improved. And the convergence of the deep learning network is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of micro-spectrometers, and in particular relates to a spectrum reconstruction method based on variational feature fusion and wavelet dense residual attention. Background Art

[0002] Spectrometers have been widely used in various industries. However, with the development of the times, traditional desktop spectrometers are no longer suitable for some specific applications, such as deep space exploration and environmental monitoring, which have high requirements for miniaturized spectrometers with high sensitivity, low cost and real-time detection capabilities. For example, in aerospace applications, the launch cost of a miniature spectrometer can be reduced exponentially compared to traditional spectrometers.

[0003] In recent years, a computational reconstruction spectrometer that combines micro-nano optics and computing technology has provided a more compact, efficient, and accurate solution for the miniaturization of spectrometers. Specifically, by carefully setting and selecting micro-nano optical materials such as quantum colloid points, photonic crystals, and metasurfaces to couple unknown spectral signals, and using ridge regression, least squares, and other calculation methods to obtain spectral data. This method can not only greatly reduce the size of the spectrometer, but also improve the accuracy and resolution of spectral data based on the data processing method.

[0004] At present, the research on miniaturized spectrometers is more focused on finding new combinations and arrangements of micro-nano optical materials, and improving the performance of computational reconstruction spectrometers by arranging and combining micro-nano optical materials with the smallest correlation coefficient. However, this method still has certain limitations. For example, when faced with complex and changeable spectral signals, due to its fixed mathematical model and material combination strategy, its adaptability is relatively weak, and it may not be able to accurately reconstruct some special spectral features, thus affecting the accuracy and integrity of spectral data. In contrast, deep learning has shown unique advantages in the field of spectral reconstruction. Deep learning algorithms can automatically learn complex patterns and features in spectral signals without relying on preset fixed models. Through training with a large amount of spectral data, it can build a highly flexible and adaptable model to accurately reconstruct various types of spectral signals. Whether it is a complex mixed spectrum or a weak signal that is easily interfered by noise, the deep learning model has the potential to provide more accurate and stable reconstruction results, opening up new directions and possibilities for improving the performance of spectrometers after miniaturization. It is expected to be widely used in more scenarios with demanding spectral accuracy, and promote the development of spectral analysis technology to a higher level. Summary of the invention

[0005] To solve the above problems, the present invention provides a spectral reconstruction method based on variational feature fusion and wavelet dense residual attention. By designing a high-performance computing method, the dependence of the computational reconstruction spectrometer on the correlation coefficients of micro-nano optical materials is reduced, and the spectral reconstruction accuracy of the computational reconstruction spectrometer is improved.

[0006] A spectrum reconstruction method based on variational feature fusion and wavelet dense residual attention is applied to a computational reconstruction spectrometer. The method comprises the following steps:

[0007] The light signal to be reconstructed with unknown energy is incident on the micro-nano deformable structure in the reconstruction spectrometer, and the light signal to be reconstructed is coupled with the spectral response of the micro-nano deformable structure to obtain a multi-channel spot image;

[0008] The multi-channel spot image is sent to the deep learning network model based on variational feature fusion and wavelet dense residual attention in the reconstruction spectrometer for feature extraction and feature learning to obtain the reconstructed spectrum corresponding to the light signal to be reconstructed.

[0009] Furthermore, the deep learning network model based on variational feature fusion and wavelet dense residual attention includes a shallow feature extraction module, a downsampling and tiling module, a multi-layer perceptron, a three-level wavelet dense residual attention module, an upsampling module, a variational feature fusion module, and a biquadratic difference module;

[0010] The shallow feature extraction module is used to extract the initial shallow features of the multi-channel spot image to obtain a first feature map;

[0011] The downsampling and tiling module is used to reduce the dimension of the first feature map and tile the first feature map after the dimension reduction into a one-dimensional feature vector;

[0012] The multi-layer perceptron is used to perform nonlinear mapping on the one-dimensional feature vector to obtain a one-dimensional low data resolution spectrum;

[0013] The low data resolution spectrum sequentially enters three-level wavelet dense residual attention modules for feature enhancement, and the last level wavelet dense residual attention module outputs a one-dimensional feature spectrum after feature enhancement;

[0014] The up-sampling module is used to improve the resolution of the one-dimensional characteristic spectrum to obtain a one-dimensional high-resolution characteristic spectrum;

[0015] The biquadratic difference module is used to map a one-dimensional low data resolution spectrum into a one-dimensional high data resolution spectrum, and the resolution of the one-dimensional high data resolution spectrum is the same as that of the one-dimensional high resolution characteristic spectrum;

[0016] The variational feature fusion module is used to perform feature fusion on a one-dimensional high data resolution spectrum and a one-dimensional high resolution feature spectrum to obtain a final reconstructed spectrum.

[0017] Furthermore, the wavelet dense residual attention module includes a discrete wavelet transform unit, two parallel residual extraction submodules, a discrete wavelet inverse transform unit, and a channel attention unit, wherein the residual extraction submodule includes three cascaded dense residual units, a channel splicing unit, and a spatial attention unit;

[0018] The discrete wavelet transform unit is used to perform high-pass filtering and low-pass filtering decomposition on the spectrum to be processed that is input into the current-level wavelet dense residual attention module for feature enhancement, so as to obtain high-frequency sub-band features and low-frequency sub-band features;

[0019] The high-frequency sub-band features and the low-frequency sub-band features enter a residual extraction sub-module respectively, wherein the processing process of the sub-band features of any frequency band in the residual extraction sub-module is as follows: the current sub-band features are sequentially extracted through three cascaded dense residual units, the channel splicing unit performs channel splicing on the output features of the three-level dense residual units to obtain the spliced ​​sub-band features, and the spatial attention features perform spatial feature extraction on the spliced ​​sub-band features to obtain the residual features finally output by the residual extraction sub-module;

[0020] The inverse discrete wavelet transform unit is used to fuse the high-frequency residual features and the low-frequency residual features output by the two residual extraction submodules to obtain a fused residual feature;

[0021] The channel attention unit is used to perform deep feature extraction on the fused residual features to obtain the enhanced features finally output by the current level wavelet dense residual attention module.

[0022] Furthermore, the variational feature fusion module includes two parallel variational feature fusion submodules and a spatial attention unit; wherein the two variational feature fusion submodules each include a fractional-order variational unit, three convolution blocks and a weight unit, and the weights of the two weight units are different;

[0023] The one-dimensional high data resolution spectrum and the one-dimensional high resolution characteristic spectrum are respectively input into a variational feature fusion submodule, wherein the processing process of any variational feature fusion submodule on the input spectrum to be processed is as follows: the spectrum to be processed is divided into two paths, the first path of the spectrum to be processed enters the fractional-order variational unit to obtain a variational result; the variational result is further divided into two paths, one of which is subjected to feature extraction by two convolution blocks in turn to obtain an intermediate spectrum feature map, and the intermediate spectrum feature map enters the weight unit and is multiplied element by element with the set weight to obtain a weighted spectrum feature map; the other path of variational result is subtracted element by element from the second path of the spectrum to be processed, and the obtained difference spectrum enters the last convolution block to obtain a difference feature map;

[0024] After the difference feature maps output by the two variational feature fusion submodules are added element by element, the obtained difference fusion feature map enters the spatial attention module for spatial feature extraction to obtain the difference spatial feature map;

[0025] The weighted spectral feature map output by the two variational feature fusion submodules is added element by element to the difference space feature map, and the addition result is used as the reconstructed spectrum finally output by the variational feature fusion module.

[0026] Furthermore, the method for the fractional-order variational unit to convert the spectrum to be processed into a variational result is:

[0027] Solve the following objective function

[0028]

[0029] Where F is the spectrum to be processed, F out is the variational result, μ is the regularization parameter, β∈(1,2) is the order of the fractional order, ρ represents the Gaussian positioning parameter, q represents the q-norm, Indicates calculation F out The gradient of ; where the Gaussian positioning parameter ρ is calculated as follows:

[0030]

[0031] Among them, η is a non-negative setting weight, Gδ is a Gaussian template, δ is the variance of the Gaussian template, Indicates the calculation of the gradient of Gδ*F.

[0032] Furthermore, the bias voltages loaded on the micro-nano deformable structure are different, the spectral responses of the micro-nano deformable structure are different, and the number of channels of the spot image is the same as the number of bias voltages.

[0033] Beneficial effects:

[0034] 1. The present invention provides a spectral reconstruction method based on variational feature fusion and wavelet dense residual attention. The spectral reconstruction model based on the attention mechanism is used to reduce the dependence of the computational reconstruction spectrometer on the correlation coefficient of micro-nano optical materials, which can effectively improve the accuracy and effect of spectral reconstruction. In addition, the wavelet dense residual attention module and variational feature fusion processing module provided by the present invention can effectively extract and refine spectral features, eliminate useless features from both high-frequency and low-frequency aspects of the data, enhance the fusion effect of deep-level high-dimensional features, and thus improve the performance of spectral reconstruction and the convergence of deep learning networks.

[0035] 2. The present invention provides a spectral reconstruction method based on variational feature fusion and wavelet dense residual attention, which performs initial shallow feature extraction on multi-channel spot images, and directly inputs the shallow output features into a multi-layer perceptron to achieve mapping from three-dimensional multi-channel image data to one-dimensional low data resolution spectrum; uses the provided wavelet dense residual attention module stacking to refine and enhance the features; uses upsampling to increase the data resolution of the features; uses jump connections including biquadratic interpolation to map the low data resolution spectrum into a high data resolution spectrum; uses variational feature fusion to fuse the output results to generate a final high data resolution spectrum as the spectral data output by the computational reconstruction spectrometer; through the method of the present invention, the data collected from the photoelectric detection can be directly used to reconstruct a high-quality high data resolution spectrum. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 Schematic diagram of the network architecture of the attention mechanism-based spectral reconstruction deep learning model provided by the present invention.

[0037] Figure 2 A schematic diagram of the architecture of the wavelet dense residual model provided by the present invention.

[0038] Figure 3 A schematic diagram of the architecture of the variational feature fusion processing module provided by the present invention.

[0039] Figure 4 It is the calculation result of the present invention for half of the spectral characteristic peak.

[0040] Figure 5 This is the calculation result of the characteristic peak of the full spectrum of the present invention. DETAILED DESCRIPTION

[0041] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application.

[0042] A spectrum reconstruction method based on variational feature fusion and wavelet dense residual attention is applied to a computational reconstruction spectrometer. The method comprises the following steps:

[0043] The energy of the light signal to be reconstructed is unknown and is incident on the micro-nano deformable structure in the reconstruction spectrometer. The light signal to be reconstructed is coupled with the spectral response of the micro-nano deformable structure to obtain a multi-channel light spot image. It should be noted that the bias voltage loaded on the micro-nano deformable structure is different, the spectral response of the micro-nano deformable structure is different, and the number of channels of the light spot image is the same as the number of bias voltages.

[0044] That is to say, the input data is the result of coupling the unknown spectrum input light of the spectrometer with the response spectrum of the micro-nano deformable structure, which is detected in the photodetector and is a multi-channel image data.

[0045] The multi-channel spot image is sent to the deep learning network model based on variational feature fusion and wavelet dense residual attention in the reconstruction spectrometer for feature extraction and feature learning to obtain the reconstructed spectrum corresponding to the light signal to be reconstructed.

[0046] like Figure 1 As shown, the deep learning network model based on variational feature fusion and wavelet dense residual attention includes a shallow feature extraction module, a downsampling and tiling module, a multi-layer perceptron, a three-level wavelet dense residual attention module, an upsampling module, a variational feature fusion module, and a biquadratic difference module;

[0047] The shallow feature extraction module is used to extract the initial shallow features of the multi-channel spot image to obtain a first feature map; it should be noted that the shallow feature extraction module includes five sequentially cascaded convolution modules stacked together, wherein each convolution module consists of a 7×7 convolution layer, a linear rectification activation function and a global pooling. After shallow feature extraction, the multi-channel spot image is reduced to a two-dimensional first feature map.

[0048] The downsampling and tiling module is used to reduce the dimension of the first feature map and tile the first feature map after the dimension reduction into a one-dimensional feature vector;

[0049] The multi-layer perceptron is used to perform nonlinear mapping on the one-dimensional feature vector to obtain a one-dimensional low data resolution spectrum;

[0050] The low data resolution spectrum sequentially enters three-level wavelet dense residual attention modules for feature enhancement, and the last level wavelet dense residual attention module outputs a one-dimensional feature spectrum after feature enhancement;

[0051] The up-sampling module is used to improve the resolution of the one-dimensional characteristic spectrum to obtain a one-dimensional high-resolution characteristic spectrum;

[0052] The biquadratic difference module is used to map a one-dimensional low data resolution spectrum into a one-dimensional high data resolution spectrum, and the resolution of the one-dimensional high data resolution spectrum is the same as that of the one-dimensional high resolution characteristic spectrum;

[0053] The variational feature fusion module is used to perform feature fusion on a one-dimensional high data resolution spectrum and a one-dimensional high resolution feature spectrum to obtain a final reconstructed spectrum.

[0054] like Figure 2As shown, the wavelet dense residual attention module includes a discrete wavelet transform unit, two parallel residual extraction submodules, a discrete wavelet inverse transform unit, and a channel attention unit, wherein the residual extraction submodule includes three cascaded dense residual units, a channel splicing unit, and a spatial attention unit;

[0055] The discrete wavelet transform unit is used to perform high-pass filtering and low-pass filtering decomposition on the spectrum to be processed that is input into the current-level wavelet dense residual attention module for feature enhancement, so as to obtain high-frequency sub-band features and low-frequency sub-band features;

[0056] The high-frequency sub-band features and the low-frequency sub-band features enter a residual extraction sub-module respectively, wherein the processing process of the sub-band features of any frequency band in the residual extraction sub-module is as follows: the current sub-band features are sequentially extracted through three cascaded dense residual units, the channel splicing unit performs channel splicing on the output features of the three-level dense residual units to obtain the spliced ​​sub-band features, and the spatial attention features perform spatial feature extraction on the spliced ​​sub-band features to obtain the residual features finally output by the residual extraction sub-module;

[0057] The inverse discrete wavelet transform unit is used to fuse the high-frequency residual features and the low-frequency residual features output by the two residual extraction submodules to obtain a fused residual feature;

[0058] The channel attention unit is used to perform deep feature extraction on the fused residual features to obtain the enhanced features finally output by the current level wavelet dense residual attention module.

[0059] That is to say, the input spectrum to be processed is transformed into high-frequency sub-bands and low-frequency sub-bands by discrete wavelet transform; then the high-frequency sub-bands and low-frequency sub-bands are respectively input into three stacked dense residual units, and the features output by the three dense residual units are connected by channels in the form of dense connections; spatial attention is used again to refine the spatial context information of the features; then the features of the high-frequency sub-band and the low-frequency sub-band are connected in the form of channels; finally, channel attention is used to refine the channel information of the features to obtain the output features of wavelet dense residual attention.

[0060] like Figure 3 As shown, the variational feature fusion module includes two parallel variational feature fusion submodules and a spatial attention unit; wherein the two variational feature fusion submodules each include a fractional-order variational unit, three convolution blocks and a weight unit, and the weights of the two weight units are different;

[0061] The one-dimensional high data resolution spectrum and the one-dimensional high resolution characteristic spectrum are respectively input into a variational feature fusion submodule, wherein the processing process of any variational feature fusion submodule on the input spectrum to be processed is as follows: the spectrum to be processed is divided into two paths, the first path of the spectrum to be processed enters the fractional-order variational unit to obtain a variational result; the variational result is further divided into two paths, one of which is subjected to feature extraction by two convolution blocks in turn to obtain an intermediate spectrum feature map, and the intermediate spectrum feature map enters the weight unit and is multiplied element by element with the set weight to obtain a weighted spectrum feature map; the other path of variational result is subtracted element by element from the second path of the spectrum to be processed, and the obtained difference spectrum enters the last convolution block to obtain a difference feature map;

[0062] After the difference feature maps output by the two variational feature fusion submodules are added element by element, the obtained difference fusion feature map enters the spatial attention module for spatial feature extraction to obtain the difference spatial feature map;

[0063] The weighted spectral feature map output by the two variational feature fusion submodules is added element by element to the difference space feature map, and the addition result is used as the reconstructed spectrum finally output by the variational feature fusion module.

[0064] That is to say, the two spectra input by the variational feature fusion processing submodule are sequentially passed through the fractional-order variational unit and two cascaded convolution blocks, and then element-wise multiplied with the hyperparameters α and 1-α in the two weight units, respectively, to obtain two sets of different weighted spectral feature maps. In addition, the two spectra input by the variational feature fusion processing submodule are element-wise subtracted from the variational results output by the fractional-order variational unit, and the two difference spectra obtained after element-wise subtraction are passed through a convolution block and then element-wise added to obtain a set of difference feature maps. Finally, after the three sets of feature maps are added, they are passed through the cascaded convolution blocks to obtain high-quality, high-data-resolution spectra.

[0065] Furthermore, the method for the fractional-order variational unit to convert the spectrum to be processed into a variational result is:

[0066] Solve the following objective function

[0067]

[0068] Where F is the spectrum to be processed, F out is the variational result, μ is the regularization parameter, β∈(1,2) is the order of the fractional order, ρ represents the Gaussian positioning parameter, q represents the q-norm, Indicates calculation F out The gradient of ; where the Gaussian positioning parameter ρ is calculated as follows:

[0069]

[0070] Among them, η is a non-negative setting weight, G δ is the Gaussian template, δ is the variance of the Gaussian template, Indicates the calculation of G δ * Gradient of F.

[0071] In order to verify the effectiveness of the spectral reconstruction deep learning model based on the attention mechanism provided by the present invention, spectral data of Gaussian function, Lorentz function and multi-function combination are used as target spectra, micro-nano deformable structure is used as micro-nano optical material of computational reconstruction spectrometer, and Gaussian noise with zero mean is added to simulate instrument interference. In the present invention, 5000 different spectral data are synthesized, of which 4500 are used for training, 300 are used for verification, and 200 are used for testing. In order to show the fairness of the experiment, the 200 spectra with the smallest average correlation coefficient are selected as test data by traversal method.

[0072] The entire experimental process was carried out on an NVIDIA GeForce RTX 3090 server, and the network was implemented based on the PyTorch framework. During the network training process, the entire network was trained for 50,000 rounds, using the Adam optimizer with an initial learning rate of 0.0001 and a decay of 0.2 every 100 rounds.

[0073] The present invention provides Figure 4 and Figure 5 is the difference between the reconstructed spectrum and the real spectrum, Figure 4 The reconstruction effect of half the spectral peak is given. Figure 5 This is the reconstruction effect of the entire result. It can be seen that the reconstructed result has a high similarity with the real spectrum, with only some errors in some areas.

[0074] In summary, the present invention provides a spectral reconstruction method based on variational feature fusion and wavelet dense residual attention. First, the initial shallow feature extraction is performed on the input unknown spectrum and the multi-channel spot image coupled with the micro-nano deformable structure; secondly, the shallow output features are directly input into the multi-layer perceptron to realize the mapping from three-dimensional multi-channel image data to one-dimensional low data resolution spectrum; furthermore, the provided wavelet dense residual attention module is used to stack and build to refine and enhance the features; then, upsampling is used to increase the data resolution of the features; then, a jump connection containing biquadratic interpolation is used to map the low data resolution spectrum into a high data resolution spectrum; finally, the provided variational feature fusion processing module is used to fuse the results to generate the final high data resolution spectrum as the spectral data output by the computational reconstruction spectrometer. It can be seen that through the method of the present invention, the data collected from the photoelectric detection can be directly used to reconstruct a high-quality high data resolution spectrum.

[0075] Of course, the present invention may have many other embodiments. Without departing from the spirit and essence of the present invention, those skilled in the art may certainly make various corresponding changes and modifications based on the present invention, but these corresponding changes and modifications should all fall within the scope of protection of the claims attached to the present invention.

Claims

1. A spectrum reconstruction method based on variational feature fusion and wavelet dense residual attention, applied to computational reconstruction spectrometer, characterized in that: The method comprises the following steps: The light signal to be reconstructed with unknown energy is incident on the micro-nano deformable structure in the reconstruction spectrometer, and the light signal to be reconstructed is coupled with the spectral response of the micro-nano deformable structure to obtain a multi-channel spot image; The multi-channel spot image is sent to the deep learning network model based on variational feature fusion and wavelet dense residual attention in the reconstruction spectrometer for feature extraction and feature learning to obtain the reconstructed spectrum corresponding to the light signal to be reconstructed.

2. The spectral reconstruction method based on variational feature fusion and wavelet dense residual attention as claimed in claim 1, characterized in that: The deep learning network model based on variational feature fusion and wavelet dense residual attention includes a shallow feature extraction module, a downsampling and tiling module, a multi-layer perceptron, a three-level wavelet dense residual attention module, an upsampling module, a variational feature fusion module, and a biquadratic difference module; The shallow feature extraction module is used to extract the initial shallow features of the multi-channel spot image to obtain a first feature map; The downsampling and tiling module is used to reduce the dimension of the first feature map and tile the first feature map after the dimension reduction into a one-dimensional feature vector; The multi-layer perceptron is used to perform nonlinear mapping on the one-dimensional feature vector to obtain a one-dimensional low data resolution spectrum; The low data resolution spectrum sequentially enters three-level wavelet dense residual attention modules for feature enhancement, and the last level wavelet dense residual attention module outputs a one-dimensional feature spectrum after feature enhancement; The up-sampling module is used to improve the resolution of the one-dimensional characteristic spectrum to obtain a one-dimensional high-resolution characteristic spectrum; The biquadratic difference module is used to map a one-dimensional low data resolution spectrum into a one-dimensional high data resolution spectrum, and the resolution of the one-dimensional high data resolution spectrum is the same as that of the one-dimensional high resolution characteristic spectrum; The variational feature fusion module is used to perform feature fusion on a one-dimensional high data resolution spectrum and a one-dimensional high resolution feature spectrum to obtain a final reconstructed spectrum.

3. The spectral reconstruction method based on variational feature fusion and wavelet dense residual attention as claimed in claim 2, characterized in that: The wavelet dense residual attention module includes a discrete wavelet transform unit, two parallel residual extraction submodules, a discrete wavelet inverse transform unit, and a channel attention unit, wherein the residual extraction submodule includes three cascaded dense residual units, a channel splicing unit, and a spatial attention unit; The discrete wavelet transform unit is used to perform high-pass filtering and low-pass filtering decomposition on the spectrum to be processed that is input into the current-level wavelet dense residual attention module for feature enhancement, so as to obtain high-frequency sub-band features and low-frequency sub-band features; The high-frequency sub-band features and the low-frequency sub-band features enter a residual extraction sub-module respectively, wherein the processing process of the sub-band features of any frequency band in the residual extraction sub-module is as follows: the current sub-band features are sequentially extracted through three cascaded dense residual units, the channel splicing unit performs channel splicing on the output features of the three-level dense residual units to obtain the spliced ​​sub-band features, and the spatial attention features perform spatial feature extraction on the spliced ​​sub-band features to obtain the residual features finally output by the residual extraction sub-module; The inverse discrete wavelet transform unit is used to fuse the high-frequency residual features and the low-frequency residual features output by the two residual extraction submodules to obtain a fused residual feature; The channel attention unit is used to perform deep feature extraction on the fused residual features to obtain the enhanced features finally output by the current level wavelet dense residual attention module.

4. The spectral reconstruction method based on variational feature fusion and wavelet dense residual attention as claimed in claim 2, characterized in that: The variational feature fusion module includes two parallel variational feature fusion submodules and a spatial attention unit; wherein the two variational feature fusion submodules each include a fractional-order variational unit, three convolution blocks and a weight unit, and the weights of the two weight units are different; The one-dimensional high data resolution spectrum and the one-dimensional high resolution characteristic spectrum are respectively input into a variational feature fusion submodule, wherein the processing process of any variational feature fusion submodule on the input spectrum to be processed is as follows: the spectrum to be processed is divided into two paths, the first path of the spectrum to be processed enters the fractional-order variational unit to obtain a variational result; the variational result is further divided into two paths, one of which is subjected to feature extraction by two convolution blocks in turn to obtain an intermediate spectrum feature map, and the intermediate spectrum feature map enters the weight unit and is multiplied element by element with the set weight to obtain a weighted spectrum feature map; the other path of variational result is subtracted element by element from the second path of the spectrum to be processed, and the obtained difference spectrum enters the last convolution block to obtain a difference feature map; After the difference feature maps output by the two variational feature fusion submodules are added element by element, the obtained difference fusion feature map enters the spatial attention module for spatial feature extraction to obtain the difference spatial feature map; The weighted spectral feature map output by the two variational feature fusion submodules is added element by element to the difference space feature map, and the addition result is used as the reconstructed spectrum finally output by the variational feature fusion module.

5. The spectrum reconstruction method based on variational feature fusion and wavelet dense residual attention as claimed in claim 4, characterized in that: The method by which the fractional-order variational unit converts the spectrum to be processed into a variational result is: Solve the following objective function Where F is the spectrum to be processed, F out is the variational result, μ is the regularization parameter, β∈(1,2) is the order of the fractional order, ρ represents the Gaussian positioning parameter, q represents the q norm, Indicates calculation F out The gradient of ; where the Gaussian positioning parameter ρ is calculated as follows: Among them, η is a non-negative setting weight, G δ is the Gaussian template, δ is the variance of the Gaussian template, Indicates the calculation of G δ * Gradient of F.

6. The spectral reconstruction method based on variational feature fusion and wavelet dense residual attention as claimed in claim 1, characterized in that: The bias voltages loaded on the micro-nano deformable structure are different, and the spectral responses of the micro-nano deformable structure are different, and the number of channels of the spot image is the same as the number of bias voltages.

Images