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

Abstract

The present application provides a kind of based on variational feature fusion and wavelet dense residual attention spectrum reconstruction method, based on the attention mechanism of spectral reconstruction model to reduce the dependence of the reconstructed spectrometer on micro-nano optical material correlation coefficient, can effectively improve the precision and effect of spectral reconstruction;In addition, the wavelet dense residual attention module and variational feature fusion processing module provided by the present application can effectively extract and refine spectral features, eliminate useless features from the high frequency and low frequency of data, enhance the fusion effect of deep high-dimensional features, and further improve the performance of spectral reconstruction and the convergence of deep learning network.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of micro spectrometers, and particularly relates to a spectrum reconstruction method based on variational feature fusion and wavelet dense residual attention. BACKGROUND

[0002] Spectrometers have been widely used in various industries, however, with the development of the times, traditional benchtop spectrometers have not been suitable for some specific applications, such as deep space exploration and environmental monitoring, etc. There is a high demand for miniaturized spectrometers with high sensitivity, low cost and real-time detection capability. For example, in space applications, compared with traditional spectrometers, the launch cost of micro spectrometers can be reduced by several times.

[0003] In recent years, a kind of computational reconstruction spectrometer combining micro-nano optics and computing technology provides a more compact, efficient and accurate solution for the miniaturization of spectrometers. Specifically, by carefully setting selected quantum colloidal points, photonic crystals, metasurfaces and other micro-nano optical materials to couple unknown spectral signals, and using ridge regression, least squares and other computing methods to obtain spectral data. This method not only can greatly reduce the volume of the spectrometer, but also can improve the accuracy and resolution of the spectral data according to the data processing method.

[0004] Currently, the research on miniaturized spectrometers focuses more on finding new combination and arrangement methods of micro-nano optical materials, by arranging and combining micro-nano optical materials with the smallest correlation coefficient, to improve the performance of the computational reconstruction spectrometer. However, this method still has certain limitations. For example, when facing complex and variable spectral signals, due to its fixed mathematical model and material combination strategy, the adaptability is relatively weak, and it may not be able to accurately reconstruct some special spectral features, thereby affecting the accuracy and integrity of the 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 pre-set fixed models. Through a large amount of spectral data training, it can build a highly flexible and adaptable model, thereby accurately reconstructing various types of spectral signals, whether it is a complex mixed spectrum or a weak and easily disturbed signal, the deep learning model has the potential to provide more accurate and stable reconstruction results, opening up new directions and possibilities for performance improvement after the miniaturization of spectrometers, and is expected to be widely used in more demanding spectral accuracy scenarios, promoting the development of spectral analysis technology to a higher level. SUMMARY

[0005] To solve the above problems, the present application provides a spectrum reconstruction method based on variational feature fusion and wavelet dense residual attention, which reduces the dependence of the computational reconstruction spectrometer on the correlation coefficient of micro-nano optical materials by designing a high-performance calculation method, and improves the spectrum reconstruction accuracy of the computational reconstruction spectrometer.

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

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

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

[0009] Further, the deep learning network model based on variational feature fusion and wavelet dense residual attention includes a shallow feature extraction module, a down-sampling and tiling module, a multi-layer perceptron, a three-level wavelet dense residual attention module, an up-sampling module, a variational feature fusion module, and a double second difference value module.

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

[0011] The down-sampling and tiling module is used to reduce the dimension of the first feature map and tile the reduced first feature map 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 the three-level wavelet dense residual attention module for feature enhancement, and the last-level wavelet dense residual attention module outputs the one-dimensional feature spectrum after feature enhancement.

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

[0015] The double second difference value module is used to map the one-dimensional low-data-resolution spectrum to 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 feature spectrum.

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

[0017] Further, the wavelet dense residual attention module comprises a discrete wavelet transform unit, two parallel residual extraction sub-modules, a discrete wavelet inverse transform unit, and a channel attention unit, wherein the residual extraction sub-module comprises three cascaded dense residual units, a channel concatenation unit, and a spatial attention unit.

[0018] The discrete wavelet transform unit is configured to perform high-pass filtering decomposition and low-pass filtering decomposition on the input current-level wavelet dense residual attention module 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 of the sub-band features in any frequency band in the residual extraction sub-module is as follows: the current sub-band features sequentially pass through three cascaded dense residual units for feature extraction, the channel concatenation unit concatenates the output features of the three dense residual units to obtain concatenated sub-band features, and the spatial attention unit extracts spatial features from the concatenated sub-band features to obtain residual features output by the residual extraction sub-module.

[0020] The discrete wavelet inverse transform unit is configured to fuse the high-frequency residual features and the low-frequency residual features output by the two residual extraction sub-modules to obtain fused residual features.

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

[0022] Further, the variational feature fusion module comprises two parallel variational feature fusion sub-modules and a spatial attention unit, wherein each of the two variational feature fusion sub-modules comprises a fractional-order variational unit, three convolution blocks, and a weight unit, and the two weight units have different weights.

[0023] The one-dimensional high-data-resolution spectrum and the one-dimensional high-resolution feature spectrum are input into a variational feature fusion sub-module respectively, wherein the processing of the input to-be-processed spectrum in any variational feature fusion sub-module is as follows: the to-be-processed spectrum is divided into two paths, the first path of the to-be-processed spectrum enters the fractional-order variational unit to obtain a variational result; the variational result is divided into two paths again, one path of the variational result sequentially passes through two convolution blocks for feature extraction to obtain an intermediate spectral feature map, and the intermediate spectral feature map enters the weight unit to be multiplied by a set weight element by element to obtain a weighted spectral feature map; the other path of the variational result is subtracted from the second path of the to-be-processed spectrum element by element to obtain a difference spectrum, and the difference spectrum enters the last convolution block to obtain a difference feature map.

[0024] The difference value feature maps output by the two variational feature fusion sub-modules are added element by element, and the difference value fusion feature map obtained is input into a spatial attention module for spatial feature extraction to obtain a difference value spatial feature map.

[0025] The weighted spectral feature maps output by the two variational feature fusion sub-modules are added element by element with the difference value spatial feature map, and the addition result is used as the reconstructed spectrum finally output by the variational feature fusion module.

[0026] Further, the method for converting the to-be-processed spectrum into the variational result by the fractional order variational unit is as follows:

[0027] Solve the following objective function

[0028]

[0029] Wherein, F is a to-be-processed spectrum, F out is a variational result, μ is a regularization parameter, β is an order of a fractional order, and ρ represents a Gaussian positioning parameter, q represents a q-norm, denotes the gradient of F out ; wherein, the calculation method of the Gaussian positioning parameter ρ is as follows:

[0030]

[0031] Wherein, η is a non-negative set weight, Gδ is a Gaussian template, δ is a variance of the Gaussian template, denotes the gradient of Gδ*F.

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

[0033] Beneficial effects:

[0034] 1. The present application provides a spectrum reconstruction method based on variational feature fusion and wavelet dense residual attention, which reduces the dependence of the calculation reconstruction spectrometer on the micro-nano optical material correlation coefficient based on the spectrum reconstruction model of the attention mechanism, and can effectively improve the accuracy and effect of spectrum reconstruction. In addition, the wavelet dense residual attention module and the variational feature fusion processing module provided by the present application can effectively extract and refine spectral features, eliminate useless features from the high frequency and low frequency of data, enhance the fusion effect of deep high-dimensional features, and further improve the performance of spectrum reconstruction and the convergence of the deep learning network.

[0035] 2. The application provides a spectral reconstruction method based on variational feature fusion and wavelet dense residual attention, which extracts initial shallow features from a multi-channel light spot image, directly inputs the shallow output features into a multi-layer perceptron to realize mapping from three-dimensional multi-channel image data to one-dimensional low data resolution spectrum, uses a wavelet dense residual attention module stack to refine and enhance the features, uses up-sampling to increase the data resolution of the features, uses a skip connection containing bilinear 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. BRIEF DESCRIPTION OF DRAWINGS

[0036] Figure 1 The network architecture schematic diagram of the spectral reconstruction deep learning model based on the attention mechanism provided by the application is shown.

[0037] Figure 2 The architecture schematic diagram of the wavelet dense residual model provided by the application is shown.

[0038] Figure 3 The architecture schematic diagram of the variational feature fusion processing module provided by the application is shown.

[0039] Figure 4 The calculation result of the application to half of the spectral feature peak.

[0040] Figure 5 The calculation result of the application to the full spectral feature peak. DETAILED DESCRIPTION

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

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

[0043] An energy-unknown light signal to be reconstructed is incident to a 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 light spot image; it should be noted that the spectral response of the micro-nano deformable structure is different when different bias voltages are loaded on the micro-nano deformable structure, and the number of channels of the light spot image is the same as the number of bias voltages.

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

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

[0046] As shown in Figure 1 The deep learning network model based on variational feature fusion and wavelet dense residual attention includes a shallow feature extraction module, a down-sampling and tiling module, a multi-layer perceptron, a three-level wavelet dense residual attention module, an up-sampling module, a variational feature fusion module, and a double second difference value module.

[0047] The shallow feature extraction module is used to extract initial shallow features of the multi-channel light spot image to obtain a first feature map. It should be noted that the shallow feature extraction module includes five convolution module stacks connected in sequence, wherein each convolution module is composed of a 7x7 convolution layer, a linear rectifier activation function and a global pooling. After shallow feature extraction, the multi-channel light spot image is reduced to a two-dimensional first feature map.

[0048] The down-sampling and tiling module is used to reduce the dimension of the first feature map and tile the reduced first feature map 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 the three-level wavelet dense residual attention module for feature enhancement, and the last-level wavelet dense residual attention module outputs the one-dimensional feature spectrum after feature enhancement.

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

[0052] The double second difference value module is used to map the one-dimensional low-data-resolution spectrum to 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 feature spectrum.

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

[0054] As shown in Figure 2As shown, the wavelet dense residual attention module includes a discrete wavelet transform unit, two parallel residual extraction sub-modules, a discrete wavelet inverse transform unit, and a channel attention unit, wherein the residual extraction sub-module includes three cascaded dense residual units, a channel concatenation unit, and a spatial attention unit.

[0055] The discrete wavelet transform unit is configured to perform high-pass filtering decomposition and low-pass filtering decomposition on the input spectrum to be processed for feature enhancement by the current-stage wavelet dense residual attention module, 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 of the sub-band features of any frequency band in the residual extraction sub-module is as follows: the current sub-band features sequentially pass through three cascaded dense residual units for feature extraction, the channel concatenation unit concatenates the output features of the three dense residual units to obtain concatenated sub-band features, and the spatial attention feature extracts spatial features from the concatenated sub-band features to obtain residual features output by the residual extraction sub-module.

[0057] The discrete wavelet inverse transform unit is configured to fuse the high-frequency residual features and the low-frequency residual features output by the two residual extraction sub-modules to obtain fused residual features.

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

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

[0060] As shown in Figure 3 The variational feature fusion module includes two parallel variational feature fusion sub-modules and a spatial attention unit; wherein each of the two variational feature fusion sub-modules includes 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 input into a variational feature fusion sub-module, wherein the processing procedure of the variational feature fusion sub-module for 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 a fractional order variational unit to obtain a variational result; the variational result is divided into two paths again, one path of the variational result is sequentially subjected to feature extraction through two convolution blocks to obtain an intermediate spectral feature map, and the intermediate spectral feature map enters a weight unit and is multiplied by a set weight element by element to obtain a weighted spectral feature map; another path of the variational result is subtracted from the second path of the spectrum to be processed element by element to obtain a difference spectrum, and the difference spectrum enters a last convolution block to obtain a difference feature map.

[0062] The difference feature maps output by the two variational feature fusion sub-modules are added element by element to obtain a difference fusion feature map, and the difference fusion feature map enters a spatial attention module for spatial feature extraction to obtain a difference spatial feature map.

[0063] The weighted spectral feature maps output by the two variational feature fusion sub-modules and the difference spatial feature map are added element by element, and the addition result is taken as a reconstructed spectrum output by the variational feature fusion module.

[0064] That is, the two spectra input into the variational feature fusion processing sub-module are sequentially subjected to the fractional order variational unit and the two cascaded convolution blocks, and then are multiplied by the hyperparameters α and 1-α in the two weight units element by element to obtain two groups of different weighted spectral feature maps. In addition, the two spectra input into the variational feature fusion processing sub-module are subtracted from the variational result output by the fractional order variational unit element by element, and the two difference spectra obtained after the element subtraction are subjected to a convolution block and then are added element by element, and a group of difference feature maps is obtained after the addition. Finally, the three groups of feature maps are added, and the cascaded convolution block can obtain a high-quality high data resolution spectrum.

[0065] Further, the method for converting the spectrum to be processed into the variational result by the fractional order variational unit is as follows:

[0066] The following objective function is solved

[0067]

[0068] wherein F is the spectrum to be processed, F out is the variational result, μ is a regularization parameter, β ∈ (1, 2) is the order of the fractional order, ρ represents a Gaussian positioning parameter, and q represents a q-norm, denotes the gradient of F out ; wherein the calculation method of the Gaussian positioning parameter ρ is as follows:

[0069]

[0070] wherein, η is a non-negative set weight, G δ is a Gaussian template, and δ is a variance of the Gaussian template, represents calculating the gradient of G δ .

[0071] In order to verify the effectiveness of the spectrum reconstruction deep learning model based on the attention mechanism provided by the present application, Gaussian function, Lorentz function and spectrum data of a combination of multiple functions are used as target spectrum, and micro-nano deformable structure is used as micro-nano optical material of the calculation and reconstruction spectrometer, and 0-mean Gaussian noise is added to simulate instrument interference. In the present application, 5000 different spectrum data are synthesized, of which 4500 data are used for training, 300 spectrum data are used for verification, and 200 spectrum data are used for testing, and in order to show the fairness of the experiment, the method of traversal is used to select the 200 spectrum with the minimum average correlation coefficient as the test data.

[0072] The whole experiment process is carried out on the NVIDIA GeForce RTX 3090 server, and the network is realized based on the PyTorch framework. In the network training process, the whole network is trained for 50000 rounds, and the Adam optimizer with an initial learning rate of 0.0001 and a decay of 0.2 every 100 rounds is used for optimization.

[0073] The spectrum reconstruction method provided by the present application Figure 4 and Figure 5 is a difference map of the reconstructed spectrum and the real spectrum, Figure 4 is the reconstruction effect of a half spectrum peak, Figure 5 is the reconstruction effect of the whole result, and it can be seen that the reconstructed result has high similarity with the real spectrum, and only some errors exist in some areas.

[0074] In summary, the spectrum reconstruction method based on variational feature fusion and wavelet dense residual attention provided by the present application firstly extracts initial shallow features from the input unknown spectrum and the multi-channel light spot image coupled with the micro-nano deformable structure; secondly, the features output by the shallow layer are directly input into the multi-layer perceptron to realize the mapping from the three-dimensional multi-channel image data to the one-dimensional low data resolution spectrum; thirdly, the wavelet dense residual attention module provided is used to stack and build the feature refining and enhancement; then, the data resolution of the feature is up-sampled; then, the low data resolution spectrum is mapped into the high data resolution spectrum by using the skip connection containing the double quadratic interpolation; finally, the result is fused by using the variational feature fusion processing module provided to generate the final high data resolution spectrum as the spectrum data output by the calculation and reconstruction spectrometer. Therefore, by the method of the present application, the high-quality high data resolution spectrum can be reconstructed directly by using the data collected from the photoelectric detection.

[0075] Of course, the present application can have other various embodiments, and those skilled in the art can certainly make various corresponding changes and modifications according to the present application without departing from the spirit and essence of the present application, but these corresponding changes and modifications shall all belong to the protection scope of the claims attached to the present application.

Claims

1. A spectral reconstruction method based on variational feature fusion and wavelet dense residual attention, applied to a computing reconstruction spectrometer, characterized in that, The method comprises the following steps: The energy-unknown light signal to be reconstructed is incident to 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 light spot image; The multi-channel light spot image is sent to a 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 a reconstructed spectrum corresponding to the light signal to be reconstructed; The deep learning network model based on variational feature fusion and wavelet dense residual attention comprises a shallow feature extraction module, a down-sampling and tiling module, a multi-layer perceptron, a three-level wavelet dense residual attention module, an up-sampling module, a variational feature fusion module and a double second difference value module; wherein the wavelet dense residual attention module comprises a discrete wavelet transform unit, two parallel residual extraction sub-modules, a discrete wavelet inverse transform unit and a channel attention unit, wherein the residual extraction sub-module comprises three cascaded dense residual units, a channel concatenation unit and a spatial attention unit.

2. The spectral reconstruction method based on variational feature fusion and wavelet dense residual attention of claim 1, wherein, The shallow feature extraction module is used for extracting initial shallow features of the multi-channel light spot image to obtain a first feature map; The down-sampling and tiling module is used for reducing the dimension of the first feature map and tiling the reduced first feature map into a one-dimensional feature vector; The multi-layer perceptron is used for performing nonlinear mapping on the one-dimensional feature vector to obtain a one-dimensional low-data-resolution spectrum; The low-data-resolution spectrum sequentially enters the three-level wavelet dense residual attention module for feature enhancement, and the last-level wavelet dense residual attention module outputs the one-dimensional feature spectrum after feature enhancement; The up-sampling module is used for improving the resolution of the one-dimensional feature spectrum to obtain a one-dimensional high-resolution feature spectrum; The double second difference value module is used for mapping the 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 feature spectrum; The variational feature fusion module is used for performing feature fusion on the one-dimensional high-data-resolution spectrum and the 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 of claim 2, wherein, The discrete wavelet transform unit is used for performing high-pass filtering decomposition and low-pass filtering decomposition on the to-be-processed spectrum input into the current-level wavelet dense residual attention module for feature enhancement 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 respectively enter a residual extraction sub-module, 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 sequentially pass through three cascaded dense residual units for feature extraction, the channel concatenation unit concatenates the output features of the three-level dense residual units to obtain concatenated sub-band features, and the spatial attention feature extracts spatial features from the concatenated sub-band features to obtain residual features finally output by the residual extraction sub-module; The discrete wavelet inverse transform unit is used for fusing the high-frequency residual features and the low-frequency residual features output by the two residual extraction sub-modules to obtain fused residual features; The channel attention unit is used for deep feature extraction on the fused residual features to obtain enhanced features as the final output of the current level wavelet dense residual attention module.

4. The spectral reconstruction method based on variational feature fusion and wavelet dense residual attention of claim 2, wherein, The variational feature fusion module includes two parallel variational feature fusion sub-modules and a spatial attention unit; each of the two variational feature fusion sub-modules includes a fractional order variational unit, three convolution blocks and a weight unit, and the two weight units have different weights; The one-dimensional high-data-resolution spectrum and the one-dimensional high-resolution feature spectrum are input into a variational feature fusion sub-module, wherein the processing process of any variational feature fusion sub-module 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 again divided into two paths, one of which is sequentially subjected to feature extraction by two convolution blocks to obtain an intermediate spectral feature map, and the intermediate spectral feature map enters the weight unit to be multiplied by a set weight element by element to obtain a weighted spectral feature map; the other path of the variational result is subtracted from the second path of the spectrum to be processed element by element to obtain a difference spectrum, and the difference spectrum enters the last convolution block to obtain a difference feature map; The difference feature maps output by the two variational feature fusion sub-modules are added element by element to obtain a difference fusion feature map, which enters the spatial attention module for spatial feature extraction to obtain a difference spatial feature map; The weighted spectral feature maps output by the two variational feature fusion sub-modules are added element by element with the difference spatial feature map, and the addition result is used as the reconstructed spectrum as the final output of the variational feature fusion module.

5. The spectral reconstruction method based on variational feature fusion and wavelet dense residual attention of claim 4, wherein, The method for converting the spectrum to be processed into a variational result by the fractional order variational unit is as follows: Solving the following objective function wherein, is the spectrum to be processed, is the variational result, μ is the regularization parameter, is the order of the fractional order, denotes a Gaussian localization parameter, q denotes q a norm, denotes the computation of the gradient of ; wherein the Gaussian localization parameter is computed as follows: wherein, η the weight is set to be non-negative, G δ is a Gaussian template, δ is a variance of the Gaussian template, denotes computing a gradient of.

6. The spectral reconstruction method based on variational feature fusion and wavelet dense residual attention of claim 1, wherein, The spectral response of the micro-nano deformable structure is different under different bias voltages, and the number of channels of the spot image is the same as the number of bias voltages.

Images