A method for reconstructing a hyperspectral image based on compressed measurement data, a computer device, a readable storage medium and a program product

Abstract

The application discloses the field of image processing, a kind of method for reconstructing hyperspectral image based on compressed measurement data, computer equipment, readable storage medium and program product, to solve the problem of image reconstruction method is divorced from physical mechanism and the problem of low reconstruction accuracy.It includes: the method for reconstructing hyperspectral image provided by the application, according to local spatial features, three-dimensional space reconstruction accuracy is higher and has high physical interpretability;According to global spectral characteristics, spectral information also has high physical interpretability;Physical credibility is higher, and physical precision is higher;According to the normalized measurement energy distribution, the compressed measurement data is weighted, so that each waveband feature is differentiated, and the physical information distribution recorded in the physical measurement mask is reflected;Considering that local spatial information and spectral information are two different physical information, therefore, the weighted waveband feature is extracted, and then the two features are distinguished, to improve the physical authenticity of subsequent reconstruction process.

Description

Technical Field

[0001] This invention relates to a method, computer device, readable storage medium, and program product for reconstructing hyperspectral images based on compressed measurement data, belonging to the field of image processing technology. Background Technology

[0002] Hyperspectral images contain rich spatial morphology of target scenes and spectral reflectance information in hundreds of continuous bands, playing an irreplaceable role in fields such as precision agriculture, environmental monitoring, and medical diagnosis.

[0003] However, high-dimensional HSI data imposes significant storage and time costs on traditional optical scanning imaging systems. Current techniques compress this data onto a two-dimensional detector for single-shot imaging, greatly improving sampling efficiency. However, reconstructing a three-dimensional hyperspectral image from compressed two-dimensional measurement data is a typical ill-conditioned inverse problem. Traditional model-driven methods, which iteratively solve the problem by introducing manually designed priors, suffer from fatal flaws such as extremely high computational complexity, long processing times, and difficulty in determining hyperparameters. In recent years, end-to-end reconstruction networks based on deep learning have achieved breakthroughs in reconstruction speed. Existing end-to-end deep learning reconstruction methods generally treat the imaging process as a purely "data mapping black box," detached from the physical imaging mechanism of the optical system.

[0004] Therefore, existing hyperspectral image reconstruction methods are divorced from the physical imaging mechanism of optical systems, and the reconstruction accuracy needs to be improved. Summary of the Invention

[0005] The purpose of this application is to overcome the shortcomings of the prior art and provide a method, computer equipment, readable storage medium and program product for reconstructing hyperspectral images based on compressed measurement data with high reconstruction accuracy, based on the physical imaging mechanism of optical systems.

[0006] To achieve the above objectives, this application employs the following technical solution:

[0007] In a first aspect, this application provides a method for reconstructing hyperspectral images based on compressed measurement data, comprising:

[0008] Obtain the physical measurement mask corresponding to the compressed measurement data, and obtain the normalized measurement energy distribution based on the physical measurement mask;

[0009] The compressed measurement data is weighted according to the normalized measurement energy distribution to obtain weighted band features;

[0010] Feature extraction is performed on the weighted band features to obtain local spatial flow features and global spectral flow features;

[0011] The reconstructed hyperspectral image is obtained by weighted reconstruction based on the local spatial flow features and the global spectral flow features.

[0012] Further, the step of weighting the compressed measurement data according to the normalized measurement energy distribution to obtain weighted band features includes:

[0013] The measurement energy of each band is extracted from the normalized measurement energy distribution, and the measurement energy of each band is normalized to obtain the normalized measurement energy of each band.

[0014] Based on the set first energy threshold, the bands whose normalized measured energy is lower than the first energy threshold are selected to obtain the selected bands;

[0015] The normalized measured energy of the selected band is inverted and amplified to obtain the inverted energy, which is obtained by the following formula:

[0016] ,

[0017] In the formula, To reverse energy, The measured energy is normalized;

[0018] The inverted energy is mapped to a dynamic modulation factor using a fully connected mapping network, and the dynamic modulation factor is obtained by the following formula:

[0019] .

[0020] In the formula, For dynamic modulation factor, This represents the Sigmoid activation function. Here is the weight matrix of the fully connected mapping network. This represents the bias parameter vector of a fully connected mapping network. This indicates that the weight matrix and bias parameter vector of a fully connected mapped network are multiplied by matrix multiplication.

[0021] The dynamic modulation factor is combined with the basic band weights, and the combined weights are applied to the compressed measurement data channel by channel to obtain the weighted band features; the weighted band features are calculated using the following formula:

[0022] ,

[0023] In the formula, For weighted band characteristics, This is an element-wise multiplication operation. Based on the basic band weights, The combined weights.

[0024] Further, the step of weighted reconstruction based on the local spatial flow features and the global spectral flow features to obtain the reconstructed hyperspectral image includes:

[0025] The local spatial flow features and the global spectral flow features are concatenated along the channel dimension to obtain the first concatenated features;

[0026] The first concatenated features are processed using a joint convolutional network to obtain dynamic gating corresponding to spatial flow and dynamic gating corresponding to spectral flow; the dynamic gating corresponding to spatial flow and the dynamic gating corresponding to spectral flow are obtained by the following formula:

[0027] ,

[0028] In the formula, For dynamic gating corresponding to spatial flow, For dynamic gating corresponding to spectral flow, For dimensionality reduction convolution operations, This is a dimensionality-upgrading convolution operation with two output parts. It is a linear rectified activation function. It is a local spatial flow characteristic. For global spectral flow characteristics, These are the features after the first splicing;

[0029] The updated spatial flow features and updated spectral flow features are obtained based on the dynamic gating corresponding to the spatial flow and the dynamic gating corresponding to the spectral flow; the updated spatial flow features and updated spectral flow features are obtained by the following formula:

[0030] ,

[0031] In the formula, For the updated spatial flow features, For the updated spectral flow characteristics, This is an unbiased convolution operation that maps spectral features to spatial feature dimensions. This is an unbiased convolution operation that maps spatial features to the spectral feature dimension.

[0032] The updated spatial flow features and the updated spectral flow features are stitched together to obtain fused features, and a reconstructed hyperspectral image is obtained based on the fused features.

[0033] Further, obtaining the reconstructed hyperspectral image based on the fusion features includes:

[0034] The fused features are subjected to pixel-level adaptive band interaction processing without global pooling to obtain the processed features.

[0035] The amplitude-phase decoupled implicit neural decoder is used to decompose the interactively processed features into amplitude features and shape features;

[0036] Based on the amplitude feature and the shape feature, an amplitude reconstruction image and a shape reconstruction image are obtained respectively, and the amplitude reconstruction image and the shape reconstruction image are obtained by the following formula:

[0037] ,

[0038] In the formula, For amplitude reconstruction of the image in the first Pixel value at that location, To reconstruct the shape of the image in the first Pixel value at that location, These are spatial pixel coordinates. Indexing the target hyperspectral band, This is the implicit channel index for fused features. and The first Amplitude and shape offset parameters of the band, For amplitude features in the first eigenvalues ​​at that location For shape features in the first eigenvalues ​​at that location The non-negative amplitude basis is in the th case. The value at that location, For free-form basis in the first The value at that location, This is a summation operation along the implicit channel dimension;

[0039] A preliminary reconstructed hyperspectral image is obtained using the amplitude-reconstructed image and the shape-reconstructed image, and a final reconstructed hyperspectral image is obtained based on the preliminary reconstructed hyperspectral image; the preliminary reconstructed hyperspectral image is obtained by the following formula:

[0040] ,

[0041] In the formula, To initially reconstruct the hyperspectral image, These are learnable equilibrium parameters.

[0042] Furthermore, the method for obtaining the nonnegative amplitude basis and the free-shape basis is as follows:

[0043] Construct a frequency domain coordinate network, and use the frequency domain coordinate network to generate frequency position codes for spectral feature dimensions;

[0044] The nonnegative amplitude basis and the free-shape basis are obtained by the following formula:

[0045] ,

[0046] In the formula, For smooth non-negative activation functions, To generate a multilayer perceptron network for amplitude generation, Generate a multilayer perceptron network for shape generation. Frequency position encoding.

[0047] Further, obtaining the final reconstructed hyperspectral image based on the preliminary reconstructed hyperspectral image includes:

[0048] The residual between the compressed measurement data and the reconstructed hyperspectral image is obtained by the following formula:

[0049] ,

[0050] In the formula, For residuals, To compress measurement data, To initially reconstruct the hyperspectral image, For physical measurement mask, For the physical measurement mask of the first The characteristic of the continuous band, the first physical measurement mask Each band tensor For the projection result, To physically sum and compress the projection results along the spectral band dimension; the projection results

[0051] The normalized physical gradient matrix is ​​obtained from the residual, and the normalized physical gradient matrix is ​​calculated using the following formula:

[0052] ,

[0053] In the formula, For the normalized physical gradient matrix, To prevent extremely small constants with a denominator of zero, Represents the spatial energy distribution of the mask;

[0054] The absolute value of the residual is obtained, and a pixel-level adaptive stride matrix is ​​obtained using a lightweight convolutional network based on the absolute value of the residual and the preliminary reconstructed hyperspectral image.

[0055] The pixel-level adaptive step size matrix is ​​used to perform physical consistency correction on the preliminary reconstructed hyperspectral image to obtain the final reconstructed hyperspectral image; the final reconstructed hyperspectral image is obtained by the following formula:

[0056] ,

[0057] In the formula, For the final reconstructed hyperspectral image, A learnable global step scaling scalar for end-to-end network training. In spatial location The pixel-level adaptive step size matrix at that location.

[0058] Furthermore, a lightweight convolutional network is used to adaptively estimate the spatial location. pixel-level adaptive step size matrix ;

[0059] The method for obtaining the lightweight convolutional network is as follows:

[0060] Obtain unprocessed lightweight convolutional networks

[0061] The smoothing error between the mean values ​​of the target band and its adjacent bands in the final reconstructed hyperspectral image is calculated along the spectral feature dimension. The smoothing error is used as a loss function value and incorporated into the total loss function of the unprocessed lightweight convolutional network. The total loss function is used to optimize the unprocessed lightweight convolutional network to obtain the final lightweight convolutional network.

[0062] The smoothing error is obtained by the following formula:

[0063] ,

[0064] In the formula, To smooth out errors, To initially reconstruct the total number of bands in the hyperspectral image, For smooth absolute value loss function, , and The reconstructed image is the first The, the and the Continuous spectral band tensor.

[0065] In a second aspect, this application also provides a computer device, including a processor and a memory connected to the processor, wherein a computer program is stored in the memory, and when the computer program is executed by the processor, it performs the steps of the method for reconstructing a hyperspectral image based on compressed measurement data as described in any embodiment of the first aspect.

[0066] Thirdly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method for reconstructing a hyperspectral image based on compressed measurement data as described in any embodiment of the first aspect.

[0067] Fourthly, this application also provides a computer program product, including a computer program / instructions, characterized in that, when the computer program / instructions are executed by a processor, they implement the steps of the method for reconstructing a hyperspectral image based on compressed measurement data as described in any embodiment of the first aspect.

[0068] Compared with the prior art, the beneficial effects achieved by this application are as follows:

[0069] The method for reconstructing hyperspectral images based on compressed measurement data provided in this application relies on local spatial features, thus achieving high accuracy and physical interpretability in three-dimensional spatial reconstruction; it also relies on global spectral features, thus providing high physical interpretability for spectral information in space; the reconstruction process as a whole has high physical reliability and accuracy; the compressed measurement data is weighted according to the normalized measurement energy distribution based on the physical measurement mask, and different weights are used to differentiate the features of each band, reflecting the true physical information distribution recorded in the physical measurement mask; considering that local spatial information and spectral information are two different types of physical information, feature extraction is performed on the weighted band features separately, thereby distinguishing between local spatial flow features and global spectral flow features, improving the physical realism of the subsequent reconstruction process. Attached Figure Description

[0070] To more clearly illustrate the technical solutions in this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0071] Figure 1 This is a flowchart illustrating the steps of a method for reconstructing hyperspectral images based on compressed measurement data, as provided in an embodiment of this application.

[0072] Figure 2 This is a schematic diagram of the structure of the dynamic bidirectional coupling gating module provided in the embodiments of this application;

[0073] Figure 3 This is a schematic diagram of the physical mapping principle of the amplitude-phase decoupled implicit neural decoder provided in the embodiments of this application;

[0074] Figure 4 This is the computational flow graph of the differentiable measurement consistency correction layer provided in the embodiments of this application;

[0075] Figure 5 This is a visualization of the reconstruction results of different reconstruction methods on the kaist dataset in the embodiments of this application;

[0076] Figure 6This is a schematic block diagram of the computer device provided in the embodiments of this application;

[0077] Figure 7 This is a flowchart illustrating the steps of another method for reconstructing a hyperspectral image based on compressed measurement data, provided in an embodiment of this application. Detailed Implementation

[0078] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments and specific features in the embodiments are detailed descriptions of the technical solution of the present application, rather than limitations thereof. In the absence of conflict, the embodiments and technical features in the embodiments can be combined with each other.

[0079] In this article, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0080] Example 1:

[0081] Figure 7 This is a flowchart illustrating a method for reconstructing a hyperspectral image based on compressed measurement data, as described in Embodiment 1 of the present invention. This flowchart merely illustrates the logical sequence of the method described in this embodiment. Without conflict, different methods may be used in other possible embodiments of the present invention. Figure 7 Complete the steps shown or described in the order indicated.

[0082] The method for reconstructing hyperspectral images based on compressed measurement data provided in this embodiment can be applied to a terminal and can be executed by a hyperspectral image reconstruction device, such as any smartphone, tablet, or computer device with communication capabilities. See also Figure 7 The method implemented in this way specifically includes the following steps:

[0083] Obtain the physical measurement mask corresponding to the compressed measurement data, and obtain the normalized measurement energy distribution based on the physical measurement mask; the normalized measurement energy distribution contains physical information corresponding to the compressed measurement data, and this physical information has been normalized, providing a basis for image reconstruction using the physical information it contains.

[0084] The compressed measurement data is weighted according to the normalized measurement energy distribution to obtain weighted band features. In the subsequent reconstruction of hyperspectral images, different bands of the weighted band features are expressed differently based on physical evidence, so that the reconstruction process has physical interpretability and high physical accuracy.

[0085] Feature extraction is performed on the weighted band features to obtain local spatial flow features and global spectral flow features. This step aims to extract spatial features from the weighted band features and divide them into local spatial flow and global spectral flow, which is beneficial for using spatial and optical information to perform image reconstruction with high physical basis and high physical accuracy.

[0086] The reconstructed hyperspectral image is obtained by weighted reconstruction based on the local spatial flow features and the global spectral flow features.

[0087] The hyperspectral image reconstructed in this embodiment is based on local spatial features, thus achieving high accuracy and physical interpretability in 3D spatial reconstruction. It also incorporates global spectral features, resulting in high physical interpretability of the spectral information in space. Overall, the reconstruction process exhibits high physical reliability and accuracy. The compressed measurement data is weighted according to the normalized measurement energy distribution based on the physical measurement mask, using different weights to differentiate the features of each band, reflecting the true physical information distribution recorded in the physical measurement mask. Considering that local spatial information and spectral information are two different types of physical information, feature extraction is performed on the weighted band features separately to distinguish between local spatial flow features and global spectral flow features, thereby improving the physical realism of subsequent reconstruction processes.

[0088] Example 2:

[0089] This embodiment provides a method for reconstructing hyperspectral images based on compressed measurement data. This embodiment is an optimization based on Embodiment 1 to improve the technical effect and refine the technical solution. For details not described in this embodiment, please refer to Embodiment 1.

[0090] As one embodiment,

[0091] The step of weighting the compressed measurement data according to the normalized measurement energy distribution to obtain weighted band features includes:

[0092] The measurement energy of each band is extracted from the normalized measurement energy distribution, and the measurement energy of each band is normalized to obtain the normalized measurement energy of each band.

[0093] Based on the set first energy threshold, the bands whose normalized measured energy is lower than the first energy threshold are selected, and the selected bands, i.e. low-energy bands, are obtained.

[0094] To improve the physical reconstruction performance, the normalized measurement energy of the selected band is inverted and amplified to obtain the inverted energy, which is obtained by the following formula:

[0095] ,

[0096] In the formula, To reverse energy, The measured energy is normalized;

[0097] The inverted energy is mapped to a dynamic modulation factor using a fully connected mapping network, and the dynamic modulation factor is obtained by the following formula:

[0098] .

[0099] In the formula, For dynamic modulation factor, This represents the Sigmoid activation function, used to restrict the output to a specific range. Here is the weight matrix of the fully connected mapping network. This represents the bias parameter vector of a fully connected mapping network. This indicates that the weight matrix and bias parameter vector of a fully connected mapped network are multiplied by matrix multiplication.

[0100] The dynamic modulation factor is combined with the basic band weights, and the combined weights are applied to the compressed measurement data channel by channel to obtain the weighted band features; the weighted band features are calculated using the following formula:

[0101] ,

[0102] In the formula, For weighted band characteristics, This is an element-wise multiplication operation. Based on the basic band weights, To initialize as full Learnable feature tensors are used to provide the underlying channel responses. The combined weights, In Indicates and All with the same dimensions Tensor, therefore This can be viewed as an element-wise multiplication operation between tensors.

[0103] As one embodiment, the step of weighted reconstruction based on the local spatial flow features and the global spectral flow features to obtain the reconstructed hyperspectral image includes:

[0104] To facilitate subsequent processing, the local spatial flow features and the global spectral flow features are stitched together along the channel dimension to obtain the first stitched features;

[0105] The first concatenated features are processed using a joint convolutional network to obtain dynamic gating corresponding to spatial flow and dynamic gating corresponding to spectral flow; the dynamic gating corresponding to spatial flow and the dynamic gating corresponding to spectral flow are obtained by the following formula:

[0106] ,

[0107] In the formula, For dynamic gating corresponding to spatial flow, For dynamic gating corresponding to spectral flow, For dimensionality reduction convolution operations, This is a dimensionality-upgrading convolution operation with two output parts. It is a linear rectified activation function. It is a local spatial flow characteristic. For global spectral flow characteristics, The features after the first splicing, This represents the Sigmoid activation function, used to normalize the gate weights;

[0108] Next, the updated modulation amount across the current is calculated in parallel, and bidirectional residual injection is performed using the dynamic gating as the modulation intensity. The updated spatial flow features and updated spectral flow features are obtained based on the dynamic gating corresponding to the spatial flow and the dynamic gating corresponding to the spectral flow. The updated spatial flow features and updated spectral flow features are obtained using the following formula:

[0109] ,

[0110] In the formula, For the updated spatial flow features, For the updated spectral flow characteristics, This is an unbiased convolution operation that maps spectral features to spatial feature dimensions. This is an unbiased convolution operation that maps spatial features to the spectral feature dimension.

[0111] To facilitate reconstruction, the updated spatial flow features and the updated spectral flow features are stitched together to obtain fused features, and the reconstructed hyperspectral image is obtained based on the fused features.

[0112] As one embodiment, obtaining the reconstructed hyperspectral image based on the fusion features includes:

[0113] The fused features are subjected to pixel-level adaptive band interaction processing without global pooling to supplement the physical interaction details between space and spectrum, and the processed features are obtained.

[0114] The decoder decomposes the reconstruction process into amplitude feature paths and free shape feature paths constrained by nonnegative physical priors. An amplitude-phase decoupled implicit neural decoder is used to decompose the interactively processed features into amplitude features and shape features.

[0115] Based on the amplitude feature and the shape feature, an amplitude reconstruction image and a shape reconstruction image are obtained respectively, and the amplitude reconstruction image and the shape reconstruction image are obtained by the following formula:

[0116] ,

[0117] In the formula, For amplitude reconstruction of the image in the first Pixel value at that location, To reconstruct the shape of the image in the first Pixel value at that location, These are spatial pixel coordinates. Indexing the target hyperspectral band, This is the implicit channel index for fused features. and The first Amplitude and shape offset parameters of the band, For amplitude features in the first eigenvalues ​​at that location For shape features in the first eigenvalues ​​at that location The non-negative amplitude basis is in the th case. The value at that location, For free-form basis in the first The value at that location, This is a summation operation along the implicit channel dimension;

[0118] A preliminary reconstructed hyperspectral image is obtained using the amplitude reconstruction image and the shape reconstruction image, which possess rich physical characteristics and high physical interpretability. This preliminary reconstructed hyperspectral image undergoes a series of processing steps to obtain the final reconstructed hyperspectral image. The preliminary reconstructed hyperspectral image is obtained using the following formula:

[0119] ,

[0120] In the formula, To initially reconstruct the hyperspectral image, These are learnable equilibrium parameters.

[0121] As one embodiment, the method for obtaining the non-negative amplitude basis and the free-shape basis is as follows:

[0122] Construct a frequency domain coordinate network, and use the frequency domain coordinate network to generate frequency position codes for spectral feature dimensions;

[0123] The implicit spectral basis functions are physically decoupled into nonnegative amplitude basis functions that characterize the absolute reflectivity of the material. and free-form basis for characterizing high-frequency details of spectral profiles The non-negative amplitude basis and the free-shape basis can be obtained by the following formula. :

[0124] ,

[0125] In the formula, For smooth non-negative activation functions, To generate a multilayer perceptron network for amplitude generation, Generate a multilayer perceptron network for shape.

[0126] As one embodiment, obtaining the final reconstructed hyperspectral image based on the preliminary reconstructed hyperspectral image includes:

[0127] The residual between the compressed measurement data and the reconstructed hyperspectral image is obtained. This residual plays a role in adjusting the difference between the compressed measurement data and the original two-dimensional image during the hyperspectral image reconstruction process. The residual is obtained by the following formula:

[0128] ,

[0129] In the formula, For residuals, To compress measurement data, To initially reconstruct the hyperspectral image, For physical measurement mask, For the physical measurement mask of the first The characteristic of the continuous band, the first physical measurement mask Each band tensor For the projection result, To physically sum and compress the projection results along the spectral band dimension; the projection results

[0130] Next, the residual is calculated. The back-projected gradient is obtained and normalized band-by-band using the energy of the physical measurement mask; the normalized physical gradient matrix is ​​obtained based on the residual, and the normalized physical gradient matrix is ​​calculated by the following formula:

[0131] ,

[0132] In the formula, For the normalized physical gradient matrix, To prevent extremely small constants with a denominator of zero, Represents the spatial energy distribution of the mask;

[0133] Obtaining the absolute value of the residual facilitates the use of a lightweight convolutional network to obtain a pixel-level adaptive step size matrix based on the absolute value of the residual and the preliminary reconstruction of the hyperspectral image.

[0134] The pixel-level adaptive step size matrix is ​​used to perform physical consistency correction on the preliminary reconstructed hyperspectral image to obtain the final reconstructed hyperspectral image; the final reconstructed hyperspectral image is obtained by the following formula:

[0135] ,

[0136] In the formula, For the final reconstructed hyperspectral image, A learnable global step scaling scalar for end-to-end network training. In spatial location The pixel-level adaptive step size matrix at that location.

[0137] As one embodiment, a lightweight convolutional network is used to adaptively estimate the spatial location. pixel-level adaptive step size matrix ;

[0138] The method for obtaining the lightweight convolutional network is as follows:

[0139] Obtain unprocessed lightweight convolutional networks

[0140] The smoothing error between the mean values ​​of the target band and its adjacent bands in the final reconstructed hyperspectral image is calculated along the spectral feature dimension. The smoothing error is used as a loss function value and incorporated into the total loss function of the unprocessed lightweight convolutional network. The total loss function is used to optimize the unprocessed lightweight convolutional network to obtain the final lightweight convolutional network.

[0141] The smoothing error is obtained by the following formula:

[0142] ,

[0143] In the formula, To smooth out errors, To initially reconstruct the total number of bands in the hyperspectral image, For smooth absolute value loss function, , and The reconstructed image is the first The, the and the Continuous spectral band tensor.

[0144] The method for reconstructing hyperspectral images based on compressed measurement data provided in this embodiment overcomes the shortcomings of existing deep learning black-box models that are detached from physical mechanisms. It provides a hyperspectral image reconstruction method guided by physical priors and decoupled from amplitude and phase. By deeply injecting optical measurement priors, non-negative reflectance priors, and locally linear spectral priors into the network front-end, feature fusion end, decoding end, and loss function, it achieves high-fidelity and high-physical-consistency image restoration at extremely low sampling rates.

[0145] To fully verify the physical fidelity and generalization ability of the method provided in this embodiment under complex real-world scenarios, this embodiment underwent rigorous comparative verification on the standard KAIST hyperspectral dataset. (Reference) Figure 5 , Figure 5 The images in the "Label" column are real images, the images in the "pro" column are images super-constructed using the method provided in this application, and "GAP", "DIP", "PnP-GAP" and "RSDN" are other existing image reconstruction methods. The images in the "GAP", "DIP", "PnP-GAP" and "RSDN" columns are images reconstructed using the corresponding methods and the same input data as the "pro" column.

[0146] GAP is a classic pure mathematical and model-driven optimization algorithm, belonging to the baseline method in the field of compressed imaging. DIP (Deep Image Prior) and PnP-GAP (Plug-and-Play GAP) are modern deep learning upgrades to the classic GAP framework. The PnP framework retains the physical projection iteration process of GAP to ensure data fidelity, but directly replaces the original prior / denoising steps that used traditional mathematical operations with a pre-trained deep neural network denoiser. LRSDN (Low-Rank Subspace Deep Network) first uses physical and mathematical methods such as singular value decomposition and SVD to extract a global spectral low-rank orthogonal basis, compressing the high-dimensional data cube into a low-dimensional spatial coefficient map. Then, it uses a self-supervised deep network to guide the spatial attention network DGSAN to be specifically responsible for the reconstruction of implicit spatial features in the spatial dimension.

[0147] Refer to Table 1 - Comparison of Reconstruction Capability Scores for Various Methods. In Table 1, Scene 1 to Scene 10 represent... Figure 5 The data consists of multiple test cases from top to bottom. Rows 1 to 10 represent the scores of each reconstruction method in reconstructing the test case, and average represents the average score from Scene 1 to Scene 10.

[0148] Table 1 - Comparison of Reconstruction Capability Scores for Various Methods

[0149]

[0150] Example 3:

[0151] refer to Figure 1 , Figure 2 , Figure 3 and Figure 4 This embodiment provides a method for reconstructing hyperspectral images based on compressed measurement data. This embodiment is an optimization based on Embodiments 1 and 2 to improve the technical effect and refine the technical solution. For details not described in this embodiment, please refer to Embodiments 1 and 2.

[0152] This embodiment provides a method for reconstructing hyperspectral images based on compressed measurement data:

[0153] Step A: Acquire the compressed measurement data and corresponding physical measurement mask collected by the hyperspectral compressed imaging system, and calculate the normalized measurement energy distribution based on the physical measurement mask;

[0154] Step B: Construct physically guided differentiated band weights using the normalized measured energy distribution, and perform front-end weighting on the initialized multi-band input features to obtain weighted band features;

[0155] Step C: Perform local spatial feature extraction and global frequency domain statistical extraction based on learnable lifting wavelet transform on the weighted band features to obtain local spatial flow features and global spectral flow features;

[0156] Step D: Input the local spatial flow features and the global spectral flow features into a pre-constructed dynamic bidirectional coupling gating module, generate bidirectional modulation gating through data-driven generation, realize synchronous residual update and fusion of the two flow features, and obtain fused features;

[0157] Step E: After performing pixel-level adaptive band interaction processing without global pooling on the fused features, the input is fed into the amplitude-phase decoupled implicit neural decoder; the decoder decomposes the reconstruction process into amplitude feature paths and free-shape feature paths constrained by non-negative physical priors, and performs weighted reconstruction to generate a preliminary reconstructed hyperspectral image;

[0158] Step F: Using the differentiable measurement consistency correction layer constructed based on the physical measurement mask, perform end-to-end physical backprojection and adaptive step size correction on the initially reconstructed hyperspectral image to obtain the final reconstructed hyperspectral image.

[0159] As one embodiment, the physical weighting process in step B is as follows: calculate the mask band energy and normalize it. , To obtain the inversion energy from the normalized measured energy. ; mapped to dynamic modulation factors through network layers The weighted band features are obtained by combining the basic weights with the input features: This allows the low-energy bands to receive greater compensation weight.

[0160] As one embodiment, reference Figure 2 The bidirectional coupling fusion process in step D is as follows: Local spatial flow features... With global spectral flow characteristics The splicing generates independent gating, which corresponds to the dynamic gating of the spatial flow. and dynamic gating corresponding to spectral flow ;pass and Cross-current bidirectional residual injection is performed to achieve dynamic intermodulation of spatial gradient and spectral mode. Figure 2 In this context, Conv represents convolution, and "1×1" represents the kernel size. Figure 4 Similarly, for "4×4", Sigmoid, SiLU, and GELU are all activation functions.

[0161] As one embodiment, the amplitude-phase decoupling and reconstruction process in step E is as follows: based on the spectral position coding, the frequency position coding is obtained. Generating non-negative amplitude basis using Softplus constraints And generate free-form base Decompose and fuse features and reconstruct amplitude reconstruction images by inner product with spectral basis. and shape reconstruction images Utilizing learnable parameters Weighted output for preliminary reconstruction of hyperspectral image .

[0162] In one embodiment, the differentiable correction process in step F is as follows: calculate the residual of the forward projection. The physical gradient matrix is ​​obtained by mask energy normalization. The network adaptively estimates the pixel-level step size. Correcting the image: .

[0163] As one embodiment, a spectral local linear prior loss is introduced during model training, which can be obtained by calculating the smoothing error: The second-order difference of the constrained spectral curve approaches zero.

[0164] The method for reconstructing hyperspectral images based on compressed measurement data provided in this embodiment overcomes the loss of hardware sampling information at the source. Based on a physically guided band weighting mechanism, the CASSI hardware measurement energy is directly converted into network compensation weights, forcing the network to pay more attention to low-energy bands at the source, thus solving the technical problem of low-energy band feature collapse. Furthermore, the method provided in this embodiment conforms to the absolute optical reflectivity properties of materials, meeting the requirements of the amplitude-phase decoupling architecture. The Softplus function forces the amplitude basis to be non-negative, conforming to the physical law that the reflectivity of real materials is always greater than or equal to zero, avoiding non-physical negative solutions caused by pure mathematical fitting, significantly reducing the solution space and improving reconstruction accuracy. The method provided in this embodiment also provides a highly adaptive bidirectional feature modulation module, thus breaking the traditional static fusion mechanism. Through data-driven bidirectional gating, spatial edges can sharpen spectral features, while spectral consistency can smooth spatial noise, achieving true deep spatial-spectral complementarity. The method provided in this embodiment essentially embeds a differentiable physical optimization algorithm. Its DMCL layer networks the traditional ADMM iterative algorithm, not only embedding physical consistency calculations into the forward propagation but also utilizing a neural network to adaptively estimate pixel-level correction step sizes. This automatically applies large step-size corrections in complex regions with large residuals, significantly improving the physical consistency and convergence speed of the reconstruction. Finally, the method in this embodiment reduces the absolute dependence on large-scale labels: it proposes the spectral locally linear prior loss (SLL Loss), introducing the physical prior of locally continuous spectra in nature. This loss function enables the model to possess purely physics-driven self-supervised error correction capabilities during training, improving the model's generalization robustness in unknown scenarios.

[0165] Step A is more specifically: The system first receives two-dimensional compressed measurement data acquired by the optical hardware and a three-dimensional physical measurement mask calibrated by the system. By calculating the mean square value of the transmittance of the physical measurement mask in different spectral bands, the sampling energy of each band is obtained and normalized to form an energy distribution vector reflecting the strength of the hardware physical sampling.

[0166] Step B, more specifically: Based on the normalized measured energy distribution obtained in Step A, the inversion energy of each band is calculated. A neural network is used to map the inversion energy to a dynamic modulation factor, and combined with the base band weights, the initial multi-band input features containing noise are weighted channel-by-channel. This step achieves physical compensation at the forefront of feature extraction, forcing the network to focus on bands with severely lost hardware sampling information that are difficult to recover.

[0167] Step C is more specifically: Local spatial feature extraction and global frequency domain statistical extraction based on learnable lifting wavelet transform are performed on the weighted band features to obtain local spatial flow features and global spectral flow features; the weighted band features are input into a dual-branch feature extraction network: the first branch uses learnable lifting two-dimensional discrete wavelet transform to adaptively extract high and low frequency local spatial flow features of the image; the second branch uses fast Fourier transform in parallel to extract frequency domain statistical information and combines it with a state-space sequence model for long-range modeling to extract global spectral flow features.

[0168] Step D, more specifically: The local spatial flow features and the global spectral flow features are input into a dynamic bidirectional coupling gating module. A bidirectional modulation gating is generated through data-driven processing to achieve synchronous residual update and fusion of the two flow features, resulting in fused features. The two flow features are then jointly input into a generation network, outputting independent spatial and spectral gating. Using the generated gating weights, the transformation results of the spectral flow features are injected into the spatial flow as residuals, and simultaneously, the transformation results of the spatial flow features are injected into the spectral flow as residuals. This step achieves dynamic bidirectional intermodulation of spatial edges and spectral consistency.

[0169] Step E: First, pixel-level adaptive band interaction is performed on the fused features from Step D. This involves abandoning global average pooling and independently calculating band weights for each spatial pixel to refine the features. The results are then fed into an implicit neural decoder to construct a positional encoding based on spectral frequency. The implicit spectral basis functions are physically decoupled into two paths: the amplitude feature path forces a non-negative output value through a smoothed non-negative activation function, approximating the absolute reflectance envelope of the real material; the shape feature path outputs an unconstrained value, fitting high-frequency spectral details. The two paths are reconstructed using inner product and then weighted to output a preliminarily reconstructed hyperspectral image.

[0170] Step F: Using a differentiable measurement consistency correction layer constructed based on the physical measurement mask, end-to-end physical backprojection and adaptive step-size correction are performed on the initially reconstructed hyperspectral image to obtain the final reconstructed hyperspectral image. The residual between the physical forward projection of the initially reconstructed hyperspectral image and the actual compressed measurement data is calculated. The traditional optimization algorithm is converted into a forward differentiable network layer, and the backprojection gradient of the residual is normalized using the energy of the physical measurement mask. Simultaneously, the network adaptively estimates the pixel-level spatial step-size matrix based on the absolute value of the residual, and uses this step-size and physical gradient to perform end-to-end consistency correction on the image, outputting the final hyperspectral image that conforms to the optical hardware measurement constraints.

[0171] The process involves constructing physically guided differentiated band weights using the normalized measured energy distribution, and then performing a front-end weighting on the initialized multi-band input features to obtain weighted band features, including:

[0172] The measured energy of each band in the physical measurement mask is calculated and normalized to a normalized measured energy distribution. The low-energy, difficult-to-recover bands are inverted and amplified to obtain the inverted energy.

[0173] The inversion energy is mapped to a dynamic modulation factor through a fully connected mapping network containing an activation function;

[0174] The dynamic modulation factor is combined with the basic band weights and applied channel by channel to the initialized multi-band input features to obtain the weighted band features.

[0175] The weighted band features are obtained by calculation using the following formula:

[0176]

[0177]

[0178]

[0179] In the formula, To reverse energy, To normalize the measured energy distribution, For dynamic modulation factor, It is the Sigmoid activation function. and These are the weight matrix and bias vector of the mapping network, respectively. For initializing multi-band input features, Based on the basic band weights, For weighted band characteristics, This indicates that the tensor is multiplied element by element.

[0180] As supplemented above, this step enables the network to adaptively amplify the focus on difficult-to-recover bands with low hardware sampling energy and severe information loss at the feature extraction source, thereby effectively compensating for the information loss caused by hardware sampling.

[0181] Specifically, the step of inputting the local spatial flow features and the global spectral flow features into a dynamic bidirectional coupling gating module, generating bidirectional modulation gating through data-driven methods, and achieving synchronous residual update and fusion of the two flow features to obtain fused features includes:

[0182] The local spatial flow features and the global spectral flow features are concatenated along the channel dimension, and a dynamic gating system corresponding to the spatial flow and a dynamic gating system corresponding to the spectral flow are generated through a joint convolutional network.

[0183] Parallel calculation of the updated modulation amount across the flow, and bidirectional residual injection using the dynamic gate as the modulation intensity, to obtain the updated spatial flow features and spectral flow features;

[0184] The updated spatial flow features and spectral flow features are concatenated and the fused features are output through dimensionality reduction using a fusion projection layer.

[0185] The updated spatial flow characteristics and spectral flow characteristics are obtained by calculation using the following formula:

[0186]

[0187]

[0188]

[0189] In the formula, and Dynamic gating for spatial flow and spectral flow, respectively. It is a local spatial flow characteristic. For global spectral flow characteristics, For dimensionality reduction convolution operations, It is a linear rectified activation function. For dimensionality-up convolution operations, It is the Sigmoid activation function. and For the updated features, and These are the unbiased convolution operations for cross-stream mapping.

[0190] As supplemented above, this embodiment breaks away from the traditional static channel splicing and fusion method. When processing complex texture edges of an image, dynamic gating guides the spectral feature flow to enhance the band differentiation expression; when processing smooth areas, dynamic gating guides the spatial flow to utilize spectral consistency for powerful noise reduction.

[0191] Specifically, the decoder decomposes the reconstruction process into an amplitude feature path and a free-shape feature path constrained by nonnegative physical priors, and performs weighted reconstruction to generate a preliminary reconstructed hyperspectral image, including:

[0192] Construct a frequency domain coordinate network to generate frequency position codes in the spectral dimension;

[0193] The implicit spectral basis functions are physically decoupled into a non-negative amplitude basis characterizing the absolute reflectivity of the material and a free-form basis characterizing the high-frequency details of the spectral profile.

[0194] The input fusion features are decomposed into amplitude features and shape features in the channel dimension, and then reconstructed by inner product with the corresponding amplitude basis and shape basis respectively to obtain amplitude reconstructed image and shape reconstructed image;

[0195] The amplitude reconstruction image and the shape reconstruction image are fused together using learnable balance parameters to obtain the preliminary reconstructed hyperspectral image.

[0196] The preliminary reconstructed hyperspectral image is obtained by calculation using the following formula:

[0197]

[0198]

[0199]

[0200] In the formula, Frequency position encoding, It is a non-negative amplitude base. For a smooth non-negative activation function, To generate a multilayer perceptron network for amplitude generation, For free-form basis, Generate a multilayer perceptron for shape. and These are the magnitude reconstructed image and the shape reconstructed image after inner product reconstruction, respectively. To initially reconstruct the hyperspectral image, These are learnable equilibrium parameters.

[0201] As supplemented above, this method forces the amplitude basis to be non-negative, perfectly matching the physical law that the reflectivity of real matter is always greater than or equal to zero. This avoids non-physical negative solutions caused by pure mathematical fitting, significantly reduces the solution space, and improves the reconstruction accuracy.

[0202] Specifically, the step of using a differentiable measurement consistency correction layer constructed based on the physical measurement mask to perform end-to-end physical backprojection and adaptive step-size correction on the initially reconstructed hyperspectral image to obtain the final reconstructed hyperspectral image includes:

[0203] The residual between the two-dimensional compressed measurement data collected by the computing system and the physical forward projection of the preliminary reconstructed hyperspectral image under the condition of truncated neural network gradient;

[0204] The back-projection gradient of the residual is calculated and normalized band by band using the energy of the physical measurement mask to obtain the normalized physical gradient matrix.

[0205] The preliminary reconstructed hyperspectral image is concatenated with the absolute value of the residual, and then input into a lightweight convolutional network to adaptively estimate the pixel-level spatial adaptive stride matrix. Based on this matrix, the preliminary reconstructed hyperspectral image is subjected to physical consistency correction.

[0206] The final reconstructed hyperspectral image is obtained by calculation using the following formula:

[0207]

[0208]

[0209]

[0210] In the formula, For physical measurement residuals, For two-dimensional compressed measurement data, For the initial reconstruction of the image Each band characteristic, For the physical measurement mask of the first Each band tensor For the normalized physical gradient matrix, It is a very small constant. For the final reconstructed hyperspectral image, A learnable global step size scalar, It is a pixel-level spatial adaptive step size matrix.

[0211] As supplemented above, this process transforms the traditional iterative update algorithm into a differentiable network layer. The network can adaptively identify difficult-to-reconstruct regions with large residuals and assign larger spatial step sizes locally, thus achieving intelligent adaptive acceleration of physical consistency constraints.

[0212] Furthermore, during the model training phase, in order to break free from the absolute dependence on large-scale real-world pairwise hyperspectral datasets, this embodiment introduces a self-supervised signal driven by physical laws into the loss function:

[0213] Based on the approximately linear reflectance characteristics of natural materials across continuous hyperspectral bands, the smoothing error between the mean of adjacent bands and the target band is calculated along the spectral dimension of the image reconstructed by the network.

[0214] The smoothing error, as a spectral local linear prior loss function, is calculated using the following formula:

[0215] ,

[0216] In the formula, The value of the spectral local linear prior loss function. This represents the total number of bands in the hyperspectral image. Index the target band. For the reconstruction of the image A continuous band tensor and These are the tensors of adjacent bands, This is a smoothing absolute value loss function.

[0217] As supplemented above, in the process of minimizing this loss, even without the guidance of real paired labels, the network can punish those erroneous spectral points that exhibit abnormally drastic jumps through its inherent physical properties, greatly improving the robustness of the model in real and complex scenarios.

[0218] Example 4:

[0219] This embodiment provides a computer device, including a processor and a memory connected to the processor. The memory stores a computer program, and when the computer program is executed by the processor, it performs the steps of the method for reconstructing a hyperspectral image based on compressed measurement data as provided in Embodiment 1 or 2.

[0220] The computer device may be a server or an electronic terminal, as one embodiment, see reference. Figure 6 The computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage medium. The database stores data acquired and generated in the method for reconstructing hyperspectral images based on compressed measurement data. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communicating with external terminals via a network connection. When the computer program is executed by the processor, it implements the method for reconstructing hyperspectral images based on compressed measurement data provided in Embodiments 1, 2, or 3.

[0221] Those skilled in the art will understand that Figure 6 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0222] The computer device provided in this embodiment has the same technical effects as those in Embodiments 1, 2, or 3, and will not be described again here.

[0223] Example 5:

[0224] This embodiment provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method for reconstructing hyperspectral images based on compressed measurement data provided in Embodiment 1 or Embodiment 2.

[0225] The computer-readable storage medium provided in this embodiment has the same technical effects as those in Embodiments 1, 2, or 3, and will not be described again here.

[0226] Example 6:

[0227] This embodiment provides a computer program product on which a computer program is stored. When executed by a processor, this program implements the steps of the method for reconstructing hyperspectral images based on compressed measurement data provided in Embodiment 1 or Embodiment 2. The computer program product provided in this embodiment can be transmitted, distributed, and downloaded via the Internet in the form of signals.

[0228] The computer program product provided in this embodiment has the same technical effects as those in Embodiments 1, 2, or 3, and will not be described again here.

[0229] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0230] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0231] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1The function specified in one or more boxes.

[0232] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0233] Furthermore, the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined with "first," "second," etc., may explicitly or implicitly include one or more of that feature. In the description of this invention, unless otherwise stated, "a plurality of" means two or more.

[0234] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art will understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0235] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method of reconstructing a hyperspectral image based on compressed measurement data, characterized by, include: Obtain the physical measurement mask corresponding to the compressed measurement data, and obtain the normalized measurement energy distribution based on the physical measurement mask; The compressed measurement data is weighted according to the normalized measurement energy distribution to obtain weighted band features; Feature extraction is performed on the weighted band features to obtain local spatial flow features and global spectral flow features; The reconstructed hyperspectral image is obtained by weighted reconstruction based on the local spatial flow features and the global spectral flow features.

2. The method for reconstructing hyperspectral images based on compressed measurement data according to claim 1, characterized in that, The step of weighting the compressed measurement data according to the normalized measurement energy distribution to obtain weighted band features includes: The measurement energy of each band is extracted from the normalized measurement energy distribution, and the measurement energy of each band is normalized to obtain the normalized measurement energy of each band. Based on the set first energy threshold, the bands whose normalized measured energy is lower than the first energy threshold are selected to obtain the selected bands; The normalized measured energy of the selected band is inverted and amplified to obtain the inverted energy, which is obtained by the following formula: ， In the formula, To reverse energy, The measured energy is normalized; The inverted energy is mapped to a dynamic modulation factor using a fully connected mapping network, and the dynamic modulation factor is obtained by the following formula: 。 3. In the formula, For dynamic modulation factor, This represents the Sigmoid activation function. Here is the weight matrix of the fully connected mapping network. This represents the bias parameter vector of a fully connected mapping network. This indicates that the weight matrix and bias parameter vector of a fully connected mapped network are multiplied by matrix multiplication. The dynamic modulation factor is combined with the basic band weights, and the combined weights are applied to the compressed measurement data channel by channel to obtain the weighted band features; the weighted band features are calculated using the following formula: ， In the formula, For weighted band characteristics, This is an element-wise multiplication operation. Based on the basic band weights, The combined weights.

4. The method for reconstructing hyperspectral images based on compressed measurement data according to claim 2, characterized in that, The step of weighted reconstruction based on the local spatial flow features and the global spectral flow features to obtain the reconstructed hyperspectral image includes: The local spatial flow features and the global spectral flow features are concatenated along the channel dimension to obtain the first concatenated features; The first concatenated features are processed using a joint convolutional network to obtain dynamic gating corresponding to spatial flow and dynamic gating corresponding to spectral flow; the dynamic gating corresponding to spatial flow and the dynamic gating corresponding to spectral flow are obtained by the following formula: ， In the formula, For dynamic gating corresponding to spatial flow, For dynamic gating corresponding to spectral flow, For dimensionality reduction convolution operations, This is a dimensionality-upgrading convolution operation with two output parts. It is a linear rectified activation function. It is a local spatial flow characteristic. For global spectral flow characteristics, These are the features after the first splicing; The updated spatial flow features and updated spectral flow features are obtained based on the dynamic gating corresponding to the spatial flow and the dynamic gating corresponding to the spectral flow; the updated spatial flow features and updated spectral flow features are obtained by the following formula: ， In the formula, For the updated spatial flow features, For the updated spectral flow characteristics, This is an unbiased convolution operation that maps spectral features to spatial feature dimensions. This is an unbiased convolution operation that maps spatial features to the spectral feature dimension. The updated spatial flow features and the updated spectral flow features are stitched together to obtain fused features, and a reconstructed hyperspectral image is obtained based on the fused features.

5. The method for reconstructing hyperspectral images based on compressed measurement data according to claim 3, characterized in that, The process of obtaining the reconstructed hyperspectral image based on the fusion features includes: The fused features are subjected to pixel-level adaptive band interaction processing without global pooling to obtain the processed features. The amplitude-phase decoupled implicit neural decoder is used to decompose the interactively processed features into amplitude features and shape features; Based on the amplitude feature and the shape feature, an amplitude reconstruction image and a shape reconstruction image are obtained respectively, and the amplitude reconstruction image and the shape reconstruction image are obtained by the following formula: ， In the formula, For amplitude reconstruction of the image in the first Pixel value at that location, To reconstruct the shape of the image in the first Pixel value at that location, These are spatial pixel coordinates. Indexing the target hyperspectral band, This is the implicit channel index for fused features. and The first Amplitude and shape offset parameters of the band, For amplitude features in the first eigenvalues ​​at that location For shape features in the first eigenvalues ​​at that location The non-negative amplitude basis is in the th case. The value at that location, For free-form basis in the first The value at that location, This is a summation operation along the implicit channel dimension; A preliminary reconstructed hyperspectral image is obtained using the amplitude-reconstructed image and the shape-reconstructed image, and a final reconstructed hyperspectral image is obtained based on the preliminary reconstructed hyperspectral image; the preliminary reconstructed hyperspectral image is obtained by the following formula: ， In the formula, To initially reconstruct the hyperspectral image, These are learnable equilibrium parameters.

6. The method for reconstructing hyperspectral images based on compressed measurement data according to claim 4, characterized in that, The method for obtaining the non-negative amplitude basis and the free-shape basis is as follows: Construct a frequency domain coordinate network, and use the frequency domain coordinate network to generate frequency position codes for spectral feature dimensions; The nonnegative amplitude basis and the free-shape basis are obtained by the following formula: ， In the formula, For smooth non-negative activation functions, To generate a multilayer perceptron network for amplitude generation, Generate a multilayer perceptron network for shape generation. Frequency position encoding.

7. The method for reconstructing hyperspectral images based on compressed measurement data according to claim 5, characterized in that, The step of obtaining the final reconstructed hyperspectral image based on the preliminary reconstructed hyperspectral image includes: The residual between the compressed measurement data and the reconstructed hyperspectral image is obtained by the following formula: ， In the formula, For residuals, To compress measurement data, To initially reconstruct the hyperspectral image, For physical measurement mask, For the physical measurement mask of the first The characteristic of the continuous band, the first physical measurement mask Each band tensor For the projection result, To physically sum and compress the projection results along the spectral band dimension; the projection results The normalized physical gradient matrix is ​​obtained from the residual, and the normalized physical gradient matrix is ​​calculated using the following formula: ， In the formula, For the normalized physical gradient matrix, To prevent extremely small constants with a denominator of zero, Represents the spatial energy distribution of the mask; The absolute value of the residual is obtained, and a pixel-level adaptive stride matrix is ​​obtained using a lightweight convolutional network based on the absolute value of the residual and the preliminary reconstructed hyperspectral image. The pixel-level adaptive step size matrix is ​​used to perform physical consistency correction on the preliminary reconstructed hyperspectral image to obtain the final reconstructed hyperspectral image; the final reconstructed hyperspectral image is obtained by the following formula: ， In the formula, For the final reconstructed hyperspectral image, A learnable global step scaling scalar for end-to-end network training. In spatial location The pixel-level adaptive step size matrix at that location.

8. The method for reconstructing hyperspectral images based on compressed measurement data according to claim 6, characterized in that, A lightweight convolutional network is used to adaptively estimate the spatial location. pixel-level adaptive step size matrix ; The method for obtaining the lightweight convolutional network is as follows: Obtain unprocessed lightweight convolutional networks The smoothing error between the mean values ​​of the target band and its adjacent bands in the final reconstructed hyperspectral image is calculated along the spectral feature dimension. The smoothing error is used as a loss function value and incorporated into the total loss function of the unprocessed lightweight convolutional network. The total loss function is used to optimize the unprocessed lightweight convolutional network to obtain the final lightweight convolutional network. The smoothing error is obtained by the following formula: ， In the formula, To smooth out errors, To initially reconstruct the total number of bands in the hyperspectral image, For smooth absolute value loss function, , and The reconstructed image is the first The, the and the Continuous spectral band tensor.

9. A computer device, characterized in that, The device includes a processor and a memory connected to the processor, wherein a computer program is stored in the memory, and when the computer program is executed by the processor, it performs the steps of the method for reconstructing a hyperspectral image based on compressed measurement data as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the steps of the method for reconstructing hyperspectral images based on compressed measurement data as described in any one of claims 1 to 7.

11. A computer program product, comprising a computer program / instructions, characterized in that, When executed by a processor, the computer program / instructions implement the steps of the method for reconstructing a hyperspectral image based on compressed measurement data as described in any one of claims 1 to 7.

Images