A hyperspectral imaging method based on spectral redundancy compression

Abstract

The application discloses a hyperspectral imaging method based on spectral redundancy compression, comprising the following steps: 1, building a hyperspectral imaging system, using the hyperspectral imaging system to compress and sample a target scene to obtain observation values and an RGB image, the hyperspectral imaging system being a hyperspectral imaging system based on single-pixel reconstruction; 2, constructing a single-pixel hyperspectral reconstruction network, using the single-pixel hyperspectral reconstruction network to process the observation values and the RGB image obtained in step 1 to obtain a reconstruction result with high spatial resolution and high spectral resolution, the single-pixel hyperspectral reconstruction network comprising a single-pixel reconstruction module based on spectral redundancy compression and a double-dimension stretching module based on progressive super-resolution, so that the reconstruction result with high spatial resolution and high spectral resolution is obtained without increasing the actual compression sampling times.

Description

TECHNICAL FIELD

[0001] The present application relates to a hyperspectral imaging method based on spectral redundancy compression, and belongs to the technical field of optical imaging. BACKGROUND

[0002] Sampling and reconstruction process of traditional compressive sensing network: target scene image , wherein represents the spatial dimension, represents the spectral dimension. For the convenience of explanation, take , that is , first compressively sample the target scene image to obtain the observation value, in the present application, use Hadamard matrix and pass energy ordering to obtain the observation matrix . Make deformation into , and compressively sample by the following formula

[0003]

[0004] In the formula, , the compression sampling rate , and then reconstruct the target scene image by using the traditional compressive sensing reconstruction network.

[0005] When it is applied to hyperspectral reconstruction, that is, the input image is no longer , but , only the process of the above formula is repeated times to obtain . Because the traditional compressive sensing reconstruction network estimates the preliminary reconstruction result from the input compressively sampled value through the full connection network. This means that all units in two adjacent layers in the network will be connected to each other, which will greatly increase the calculation complexity of the training process, and will also cause the result to appear overfitting problem.

[0006] The single-pixel hyperspectral reconstruction network of the prior art usually uses the small block cutting method to reduce the network calculation parameters, and the limitation of the network calculation parameters means that these image blocks will be very small.

[0007] Therefore, a new hyperspectral imaging method based on spectral redundancy compression is needed to solve the above problems. SUMMARY

[0008] The present application aims to provide a hyperspectral imaging method based on spectral redundancy compression to solve the problems raised in the background art.

[0009] A hyperspectral imaging method based on spectral redundancy compression comprises the following steps:

[0010] I. Build a hyperspectral imaging system, use the hyperspectral imaging system to obtain observation values and RGB images by compressively sampling a target scene, and the hyperspectral imaging system is a hyperspectral imaging system based on single-pixel reconstruction;

[0011] II. Construct a single-pixel hyperspectral reconstruction network, use the single-pixel hyperspectral reconstruction network to process the observation values and RGB images obtained in step I to obtain a reconstruction result with high spatial and hyperspectral resolution, and the single-pixel hyperspectral reconstruction network is composed of a single-pixel reconstruction module based on spectral redundancy compression and a double-dimension stretching module based on progressive super-resolution.

[0012] Further, the RGB image in step I is an RGB image with high spatial resolution.

[0013] Further, the hyperspectral imaging system in step I includes an objective lens L1, a flat plate beam splitter BS, a single-pixel camera, a total internal reflection prism TIR, a digital micromirror device DMD, a collimator L2, and a fiber spectrometer, the objective lens L1 is arranged between the target scene and the flat plate beam splitter BS, the single-pixel camera and the total internal reflection prism TIR are arranged on two light paths of the flat plate beam splitter BS respectively, the digital micromirror device DMD is arranged on the direct light path of the total internal reflection prism TIR, the collimator L2 is arranged on the deflected light path of the total internal reflection prism TIR, and the fiber spectrometer receives light rays of the collimator L2 through an optical fiber.

[0014] Further, the digital micromirror device DMD uses a Hadamard template.

[0015] Further, the single-pixel hyperspectral reconstruction network in step II includes the following steps:

[0016] 21. First, use the single-pixel reconstruction module based on spectral redundancy compression to perform compressive sensing reconstruction on the observation values in step I to obtain a preliminary reconstruction result after dimension reduction;

[0017] 22. Second, introduce the RGB image, use the double-dimension stretching module based on progressive super-resolution to process the preliminary reconstruction result in step 21 to obtain a reconstruction result with high spatial and hyperspectral resolution.

[0018] Further, the single-pixel reconstruction module based on spectral redundancy compression in step II uses the following model:

[0019]

[0020] wherein, represents the observation values obtained by compressively sampling the target scene, represents the result after dimension reduction of the observation values by the dimension reduction module, represent reconstruction and enhancement results in the case of dimension reduction, representing a full connection reconstruction module, representing a residual enhancement module, representing a full connection in the continuous spectral dimension for dimension reduction processing of the compressed sampled observation values.

[0021] Further, the two-dimensional stretching module based on progressive super-resolution in step two adopts the following model:

[0022]

[0023] wherein, representing feature band splicing, representing residual connection, representing an RGB image, representing a spectral dimension recovery result, representing a Transformer model, representing a spatial resolution super-resolution result, is a sub-pixel convolution model. : convolution operation with a convolution kernel size of 3x3

[0024] Further, the Transformer model includes the following steps:

[0025] 221, reshape to , and then perform a linear projection operation on using the following formula:

[0026]

[0027] wherein, is the result after projection, is a learnable parameter, represents the spatial resolution of the image, is the number of spectral bands of the image;

[0028] 222, divide into parts along the spectral channel number, i.e. , and the size of each part is , and then calculate the attention mechanism of using the following formula:

[0029]

[0030] wherein, is the transpose of , and is a learnable parameter, Represents the activation function;

[0031] 223. The parts are interconnected and operated using the following formula:

[0032]

[0033] In the formula, For learnable parameters, Composed of two It consists of an activation function. This is the output result.

[0034] This invention utilizes the Transformer model TFM to recover the spectral dimension of an image. To obtain optimized feature vectors, TFM is superimposed with an attention network and a fully connected layer. TFM can add an independent parameter to each dimension, allowing adjustment of the weight relationships between spectral dimensions during the recovery process.

[0035] 2. Furthermore, the loss function of the single-pixel hyperspectral reconstruction network in step two is shown in the following formula:

[0036]

[0037] In the formula, Represents the expected result. This represents the result output by the network, where F denotes the F-norm. This represents the weighting coefficient, and M is the number of elements.

[0038] When selecting the loss function for constrained network learning, considering the requirement to learn three-dimensional information simultaneously, the MSE function based on three-dimensional data was chosen to calculate the loss.

[0039] Beneficial Effects: This invention's hyperspectral imaging method based on spectral redundancy compression, combined with deep learning, proposes a single-pixel hyperspectral reconstruction network. It utilizes spectral redundancy compression for compressed sensing reconstruction at a low dimensionality, reducing the number of parameters requiring computation and enabling end-to-end hyperspectral imaging for larger-scale hyperspectral image reconstruction. A progressive image super-resolution method is employed to sequentially restore the spectral dimension and perform super-resolution of the spatial dimension. By introducing low-cost, high spatial resolution RGB images, high spatial and spectral resolution reconstruction results are achieved without increasing the actual number of compressed samplings. Attached Figure Description

[0040] Figure 1 This is a schematic diagram of the structure of a single-pixel hyperspectral reconstruction network based on spectral redundancy compression and progressive image super-resolution;

[0041] Figure 2is a single-pixel reconstruction based on spectral redundancy compression;

[0042] Figure 3 is a dimension compression module;

[0043] Figure 4 is a compressive sensing reconstruction and enhancement module;

[0044] Figure 5 is a dual-dimension stretching based on progressive super-resolution;

[0045] Figure 6 is a Transformer module;

[0046] Figure 7 is the reconstruction result of the Cave dataset test image (pompoms_ms) at 500nm, 600nm, 700nm waveband;

[0047] Figure 8 is the error map of the Cave dataset test image (pompoms_ms) at 500nm, 600nm, 700nm waveband;

[0048] Figure 9 is the reconstruction result of the Icvl dataset test image (tree_0822-0853) at 500nm, 600nm, 700nm waveband;

[0049] Figure 10 is the error map of the Icvl dataset test image (tree_0822-0853) at 500nm, 600nm, 700nm waveband;

[0050] Figure 11 is the spectral curve of the Cave dataset test image (pompoms_ms), the left side is the RGB image of the scene, and the right side is the spectral curve of the reconstruction result, p1 and p2 respectively correspond to the positions of p1 and p2 on the RGB image;

[0051] Figure 12 is the spectral curve of the Icvl dataset test image (tree_0822-0853), the left side is the RGB image of the scene, and the right side is the spectral curve of the reconstruction result, p1 and p2 respectively correspond to the positions of p1 and p2 on the RGB image;

[0052] Figure 13 is a structural diagram of a hyperspectral imaging system based on single-pixel reconstruction;

[0053] Figure 14 is a representative visual comparison of real scenes and wavebands;

[0054] Figure 15is a representative spectral curve of a real scene, the left side is the RGB image of the scene, and the right side is the spectral curve of the reconstruction result, p1 and p2 correspond to the positions p1 and p2 on the RGB image respectively. DETAILED DESCRIPTION

[0055] The application will be further clarified by the following embodiments, which should not be construed as limiting the scope of the application. After reading the application, those skilled in the art will be able to modify various equivalent forms of the application, which fall within the scope of the appended claims.

[0056] The hyperspectral imaging method based on spectral redundancy compression of the application comprises the following steps:

[0057] I. Build a hyperspectral imaging system, use the hyperspectral imaging system to compress and sample the target scene to obtain observation values and an RGB image, and the hyperspectral imaging system is a hyperspectral imaging system based on single-pixel reconstruction; wherein the RGB image is an RGB image with high spatial resolution.

[0058] Specifically, the hyperspectral imaging system comprises an objective lens L1, a flat beam splitter BS, a single-pixel camera, a total internal reflection prism TIR, a digital micromirror device DMD, a collimator L2, and a fiber spectrometer, the objective lens L1 is arranged between the target scene and the flat beam splitter BS, the single-pixel camera and the total internal reflection prism TIR are arranged on two light paths of the flat beam splitter BS respectively, the digital micromirror device DMD is arranged on the direct light path of the total internal reflection prism TIR, the collimator L2 is arranged on the deflected light path of the total internal reflection prism TIR, and the fiber spectrometer receives the light rays of the collimator L2 through an optical fiber.

[0059] Invention principle: the image of the target scene enters the flat beam splitter BS after passing through the objective lens L1, is reflected into the single-pixel camera on one path, and an RGB image with high spatial resolution (256x256x3) is obtained, and on the other path, it enters the digital micromirror device DMD for modulation, and then is deflected to a direction perpendicular to the incident light by the total internal reflection prism TIR, is received by the collimator L2 into the optical fiber, and finally is collected by the fiber spectrometer. The resolution of the modulated scene is 64x64, and the digital micromirror device DMD adopts a Hadamard template. The model of the single-pixel camera is Balser, acA800-510 uc, the parameters of the digital micromirror device DMD are pixel count: 1024x768, pixel pitch: 13.68 μm, and the model of the fiber spectrometer is AvaSpec-ULS2048CL-EVO.

[0060] II. Constructing a single-pixel hyperspectral reconstruction network, using the single-pixel hyperspectral reconstruction network to process the observation value and the RGB image obtained in step one to obtain a reconstruction result with high spatial and hyperspectral resolution, the single-pixel hyperspectral reconstruction network being composed of a single-pixel reconstruction module based on spectral redundancy compression and a double-dimension stretching module based on progressive super-resolution.

[0061] Referring to Figures 1-6 As shown in the figure, the single-pixel hyperspectral reconstruction network of the application is composed of a single-pixel reconstruction DFR based on spectral redundancy compression and a double-dimension stretching FDS based on progressive super-resolution. The input is the observation value obtained by compressively sampling the scene and the RGB image with high spatial resolution, and the output is the result with high spatial and hyperspectral resolution. The single-pixel hyperspectral reconstruction network not only solves the problem of too many full-connection parameters of the reconstruction network, but also makes the network in the form of end-to-end, so that the hyperspectral data can be quickly and accurately obtained, and the application in the actual scene has advantages.

[0062] The single-pixel hyperspectral reconstruction network in step two includes the following steps:

[0063] 21. First, the single-pixel reconstruction module based on spectral redundancy compression is used to compressively sense and reconstruct the observation value in step one, so as to obtain a preliminary reconstruction result after dimension reduction;

[0064] The single-pixel reconstruction module based on spectral redundancy compression adopts the following model:

[0065]

[0066] Among them, represents the observation value obtained by compressively sampling the target scene, represents the result after dimension reduction of the observation value by using the dimension reduction module, represents the reconstruction and enhancement result under the condition of dimension reduction, represents a full-connection reconstruction module, represents a residual enhancement module, represents that the observation value of the compressive sampling is processed by using the continuous spectral dimension full-connection for dimension reduction.

[0067] Hyperspectral data has the characteristics of nonlinearity, large data volume and high dimension. There are various dimension reduction methods in the prior art to remove redundancy and extract effective information. Among them, when the dimension reduction result is higher than 5, the principal component analysis method PCA can basically recover the original 31-dimensional hyperspectral image without loss in the process of dimension reduction and recovery of the 31-dimensional hyperspectral image. Therefore, in the process of designing the network structure, the single-pixel reconstruction of the hyperspectral image is combined with dimension reduction to obtain a single-pixel reconstruction module DFR based on spectral redundancy compression. By introducing a dimension reduction module in the process of single-pixel hyperspectral reconstruction, the problem of excessive parameters faced by using a full connection structure to estimate the preliminary reconstruction result in the network is solved. The structure is as shown in Figure 2 The compressed and sampled observation values are first subjected to a dimension compression module DDM composed of multiple consecutive spectral dimension full connections, and then subjected to a compressed sensing reconstruction and enhancement module FRM, so as to obtain a preliminary reconstruction result under low dimension.

[0068] The structure of the DDM (Dimension Down Module) is as shown in Figure 3 The dimension reduction processing of the compressed sampling result is performed by using consecutive spectral dimension full connections, which can greatly reduce the number of parameters to be calculated in the subsequent reconstruction process. The dimension compression is divided into 4 steps, and the corresponding dimensions are 31, 20, 15, 20 and 7.

[0069] The structure of the FRM (Fully Reconstruction Module) is as shown in Figure 4 Under the condition of low dimension, a full connection layer is first used to obtain a preliminary reconstruction result, and then consecutive convolution is used in combination with a residual module to realize image enhancement of the preliminary reconstruction result. The structure contains a total of 6 convolution operations, and the size of the convolution kernel of each layer is 11x11, 1x1, 7x7, 11x11, 1x1 and 7x7 respectively, and the feature mapping thereof is 64, 32, 1, 64, 32 and 7 respectively. In order to ensure the invariability of the output resolution, zero is used for padding, and ReLU activation function is used except for the output layer.

[0070] 22. Secondly, an RGB image is introduced, and a double-dimension stretching module based on progressive super-resolution is used to process the preliminary reconstruction result of step 21 to obtain a reconstruction result with high spatial and spectral resolution.

[0071] The double-dimension stretching module based on progressive super-resolution adopts the following model:

[0072]

[0073] Among them, represents feature band splicing, represents residual connection, represents an RGB image, denotes a spectral dimension recovery result, denotes a Transformer model, denotes a spatial resolution super-resolution result, is a sub-pixel convolution model. The sub-pixel convolution module replaces the up-sampling convolution operation in deep learning.

[0074] After achieving the low-dimensional preliminary reconstruction result in the dimension reduction, the application designs a two-dimensional stretching module FDS based on progressive super-resolution, the structure of which is shown in Figure 5 The input end includes the preliminary reconstruction result after dimension reduction and the RGB image with high spatial resolution . The whole can be divided into spectral dimension recovery of low-spectral-resolution images and spatial dimension super-resolution of low-spatial-resolution images, and the high-spatial-resolution RGB image is added for information fusion, and the hyperspectral image directly up-sampled from the preliminary reconstruction result provides the preliminary estimation for network learning. Through the two-dimensional stretching based on progressive super-resolution, the high spatial and high spectral resolution results can be obtained without increasing the actual number of compressed sampling, so as to obtain better visual effect and more rich scene information.

[0075] In the spatial dimension super-resolution process of the low-spatial-resolution image, the sub-pixel convolution module Sub Pixel is used to replace the up-sampling convolution operation in deep learning.

[0076] Preferably, referring to Figure 6 The Transformer model includes the following steps:

[0077] 221, reshape to , and then perform a linear projection operation on using the following formula:

[0078]

[0079] In the formula, is the result after projection, is a learnable parameter;

[0080] 222, divide into parts along the number of spectral channels, that is , and the size of each part is , and then calculate the attention mechanism of using the following formula:

[0081]

[0082] wherein, is the transpose of is a learnable parameter, is a learnable parameter, represents an activation function, which can normalize a numerical vector into a probability distribution vector;

[0083] 223、connect the parts to each other and operate by the following formula:

[0084]

[0085] wherein, is a learnable parameter, consists of two and an activation function, is an output result.

[0086] In addition, the application utilizes a Transformer model TFM to restore the spectral dimension of an image, and in order to obtain an optimized feature vector, the TFM is stacked with an attention network and a full connection layer. The TFM can add an independent parameter for each dimension, and in the process of restoring the spectral dimension of the image, the weight relationship between the spectral dimensions can be adjusted.

[0087] Preferably, the loss function of the single-pixel hyperspectral reconstruction network in step two is as follows:

[0088]

[0089] wherein, represents an expected result, represents a network output result, F represents an F norm, represents a weight coefficient, and M is the number of elements.

[0090] Simulation and result verification:

[0091] Firstly, the single-pixel hyperspectral reconstruction network of the application is verified on the public data sets Cave data set and Icvl data set.

[0092] The application utilizes four methods Tradition, TVAL3, Reconnet and Coast to compare with the method of the application. The Tradition and TVAL3 methods need to be compressed sensing reconstructed by traditional calculation or iterative means, and the Reconnet and Coast methods adopt deep learning as a calculation means.

[0093] The psnr, ssim and sam indexes are selected as the image evaluation indexes, wherein the psnr is a peak signal-to-noise ratio, the ssim is a structural similarity, and the sam is a spectral angle. The reconstructed scene images all have a resolution of 256x256x31. The actual sampling times of the hyperspectral imaging method based on spectral redundancy compression are 64x64x31x , the actual sampling times of other networks are 256x256x31x , wherein, the compression sampling rate, the actual sampling times of the present application are lower, and the average time required for reconstructing a group of hyperspectral images (256x256x31) from the collected data is only 64.1 ms.

[0094] Table 1 is the reconstruction result index of different methods under three compression sampling rates (0.25, 0.1 and 0.01).

[0095] Compared with the other four methods, the reconstruction results of the present application have significant advantages in most cases according to the data of the three indexes of PSNR, SAM and SSIM. The present application introduces low-cost RGB images into the network structure, and the RGB images can provide supplementary information for the missing information in the network learning. With the gradual reduction of the compression sampling rate, the indexes of the present application do not decrease significantly. However, in the Reconnet and Coast methods, since only the compressed sampling data is used as the network input, with the reduction of the compression sampling rate, the indexes of the reconstruction results decrease.

[0096] Table 1 is the average index comparison of the Cave data set and the Icvl data set under different compression sampling rates.

[0097]

[0098] Figure 7 、 Figure 8 、 Figure 9 and Figure 10 show the reconstruction results and error maps of 500nm, 600nm and 700nm of the pompoms_ms scene of the Cave data set and the tree_0822-0853 scene of the Icvl data set under the compression sampling rate of 0.25. From the visual reconstruction results, the present application can display clearer image details and higher fidelity in the reconstruction results. The sam index has certain limitations for the spectral display of the reconstructed images, therefore, the spectral curves of the reconstructed images are shown, as shown in Figure 11 , two regions p1 and p2 are selected on each scene to highlight the spectral curves, and the spectral curves of the hyperspectral data reconstructed by the present application are basically consistent with the true data.

[0099] Please refer to Figure 12 andFigure 13 To verify that the single-pixel hyperspectral reconstruction network of the application is not a simple mapping process from RGB to hyperspectrum, the RGB image at the input end of the network is directly replaced by a grayscale image, other network results remain unchanged, and verification is carried out on the Cave dataset, and Table 2 is the comparison result.

[0100] Table 2 Average indicators of RGB scheme and grayscale scheme under different compression sampling rates on Cave dataset

[0101]

[0102] Experiment and result

[0103] The hyperspectral imaging system based on single-pixel reconstruction of the application uses a fiber spectrometer to replace a single-point detector as a signal acquisition device on the basis of a single-pixel camera. The hyperspectral imaging system is used to collect and train data in an actual scene, and the applicability of the hyperspectral imaging method based on spectral redundancy compression is verified.

[0104] 200 groups of data are made for the experiment, 180 groups of training data and 20 groups of test data, and the number of data is increased by combining random flipping and rotating operations. The network is also trained using the simulation experiment strategy, and the input end is an RGB image (256x256x3) and an observation value obtained by compressed sampling hyperspectral data (64x64x31).

[0105] Table 3 is the network reconstruction result of 3 kinds of compression sampling rates (0.25, 0.1, 0.01) of the real shot data. The PSNR and SSIM are proportional to the sampling rate, and the SAM is inversely proportional to the sampling rate. Figure 14 For a representative scene and visual comparison, Figure 15 For the corresponding spectral curve. In summary, the hyperspectral imaging method based on spectral redundancy compression of the application has excellent imaging effect.

[0106] Table 3 Average indicators of real shot data under different compression sampling rates

[0107]

[0108] The application discloses a hyperspectral imaging method based on spectral redundancy compression, which obtains observation values by designing a hyperspectral imaging system based on single-pixel reconstruction to compress and sample a scene to be processed; an end-to-end single-pixel hyperspectral reconstruction network (CSRSH) based on spectral redundancy compression and progressive image super-resolution is constructed; the spectral dimension of a preliminary reconstruction image is recovered and the spatial resolution of the preliminary reconstruction result is super-resolved through a progressive super-resolution strategy and in combination with an RGB image with high spatial resolution, so that the spatial and spectral resolutions of the reconstruction result are finally obtained without increasing the actual compression sampling times.

[0109] Compared with the prior art single-pixel hyperspectral reconstruction method, the hyperspectral imaging method based on spectral redundancy compression has the following advantages:

[0110] 1. The single-pixel hyperspectral imaging method introduces spectral redundancy compression, solves the problem of excessive network full-connection parameters, and can realize hyperspectral image reconstruction of a larger scene.

[0111] 2. The method of the application realizes recovery of the spectral dimension of an image and super-resolution of the spatial dimension by using the progressive image super-resolution method, so that hyperspectral data with high spatial and spectral resolutions are obtained without increasing the actual sampling times.

[0112] 3. The single-pixel-based hyperspectral reconstruction imaging system of the application obtains observation values by compressing and sampling a target scene, and then reconstructs a hyperspectral image by using a single-pixel hyperspectral reconstruction network. The overall imaging is in an end-to-end mode, and the application can be applied to an actual scene to quickly and accurately obtain hyperspectral data with high spatial and spectral resolutions.

Claims

1. A hyperspectral imaging method based on spectral redundancy compression, characterized by, Comprise the following steps: I. Build a hyperspectral imaging system, use the hyperspectral imaging system to compress sample the target scene to obtain observation values and RGB images, the hyperspectral imaging system is a hyperspectral imaging system based on single-pixel reconstruction; II. Construct a single-pixel hyperspectral reconstruction network, use the single-pixel hyperspectral reconstruction network to process the observation values and RGB images obtained in step I to obtain a reconstruction result with high spatial and hyperspectral resolution, the single-pixel hyperspectral reconstruction network is composed of a single-pixel reconstruction module based on spectral redundancy compression and a double-dimension stretching module based on progressive super-resolution; The single-pixel reconstruction module based on spectral redundancy compression in step II adopts the following model: ； wherein, representing observation values obtained by compressive sampling a target scene, representing results of dimension reduction of the observation values by a dimension reduction module, representing reconstruction and enhancement results in the case of dimension reduction, denoting a full connection reconstruction module, denoting a residual enhancement module, denoting dimension reduction processing of the observation values of the compressive sampling by full connection of consecutive spectral dimensions; The double-dimension stretching module based on progressive super-resolution in step II adopts the following model: ； wherein, denotes characteristic band stitching, denotes residual connection, denotes RGB image, denotes spectral dimension recovery result, denotes Transformer model, denotes spatial resolution super-resolution result, is a sub-pixel convolution model, represents convolution operation; The Transformer model comprises the following steps: 221、the input image is reshaped reshaped into and then linearly projected using the following equation: a linear projection operation: ； wherein, is the result of the projection, is a learnable parameter, represents the spatial resolution of the image, is the number of spectral bands of the image; 222、along the number of spectral channels are divided into parts, i.e. each part has a size of and the attention mechanism of is calculated using the following formula: ； wherein is the transpose of is a learnable parameter, represents an activation function; 223、will be the parts are connected to each other and operate by the following formula: ； In the formula, are learnable parameters, consist of two and an activation function, is the output result.

2. The method for hyperspectral imaging based on spectral redundancy compression as claimed in claim 1, wherein: The RGB image in step I is an RGB image with high spatial resolution.

3. The method for hyperspectral imaging based on spectral redundancy compression as claimed in claim 1, wherein: The hyperspectral imaging system in step I comprises an objective lens L1, a flat beam splitter BS, a single-pixel camera, a total internal reflection prism TIR, a digital micromirror device DMD, a collimator L2 and a fiber spectrometer, the objective lens L1 is arranged between the target scene and the flat beam splitter BS, the single-pixel camera and the total internal reflection prism TIR are arranged on two light paths of the flat beam splitter BS respectively, the digital micromirror device DMD is arranged on the direct light path of the total internal reflection prism TIR, the collimator L2 is arranged on the deflected light path of the total internal reflection prism TIR, and the fiber spectrometer receives light rays of the collimator L2 through an optical fiber.

4. The method for hyperspectral imaging based on spectral redundancy compression as claimed in claim 3, wherein: The digital micromirror device DMD adopts a Hadamard template.

5. The method for hyperspectral imaging based on spectral redundancy compression as claimed in claim 1, wherein: The single-pixel hyperspectral reconstruction network in step II comprises the following steps: First, use the single-pixel reconstruction module based on spectral redundancy compression to perform compressive sensing reconstruction on the observation values in step I to obtain a preliminary reconstruction result after dimension reduction; Second, introduce the RGB image, use the double-dimension stretching module based on progressive super-resolution to process the preliminary reconstruction result in step 21 to obtain a reconstruction result with high spatial and hyperspectral resolution.

6. The method for hyperspectral imaging based on spectral redundancy compression as claimed in claim 1, wherein: The loss function of the single-pixel hyperspectral reconstruction network in step II is as follows: ； wherein represents the desired result, represents the result of the network output, F denotes F norm, represents the weight coefficient, and M is the number of elements.

Images