A hyperspectral compression reconstruction method and system based on an iterative degenerate sensing Mamba model

Abstract

This invention discloses a hyperspectral compression reconstruction method and system based on an iterative degradation-aware Mamba model. Two-dimensional measurement frames are generated using physical mask modulation of real three-dimensional hyperspectral images to form the original dataset, which is then divided into training and test sets. The measurement frames and physical masks are shifted and sliced ​​to generate three-dimensional hyperspectral measurement images and three-dimensional mask data. An iterative degradation-aware Mamba model is constructed, trained using the training set, and validated using the test set. This invention dynamically adjusts the degradation factor and optimizes the interaction between spectral and spatial features during the reconstruction process, achieving a good balance between performance and efficiency.

Description

Technical Field

[0001] This invention belongs to the field of image processing technology and relates to a hyperspectral compression reconstruction method and system based on an iterative degradation-sensing Mamba model, which is applied to the reconstruction of single-exposure snapshot hyperspectral data. Background Technology

[0002] Hyperspectral imaging (HSI) is a three-dimensional data structure consisting of two spatial dimensions and one spectral dimension. Due to its rich spectral information and high-resolution spatial detail, HSI has been widely used in numerous fields. Various imaging methods have been developed to capture the complete HSI cube. Traditional methods typically require long exposure times and multiple scans, making them prone to motion artifacts and unsuitable for dynamic scenes. Snapshot compressed imaging (SCI) employs compressed sensing technology to map the 3D hyperspectral image (HSI) to a 2D measurement in a single snapshot, significantly reducing motion artifacts and making it highly suitable for dynamic scenes. Among these, the coded aperture snapshot spectral imaging (CASSI) system is the most representative.

[0003] Due to the inherent ill-posedness of SCI reconstruction in real-world scenarios, traditional methods are typically model-based, relying on handcrafted priors as regularization terms to estimate the HSI cube. While these methods offer interpretability, they often result in high computational costs and dependence on manually designed prior constraints. With the widespread application of deep learning in vision tasks, deep learning-based HSI reconstruction methods have gradually been adopted by researchers. These methods offer stronger fitting capabilities but lack interpretability.

[0004] Most existing algorithms do not consider the degradation patterns of the entire system, therefore, methods that incorporate degradation information into the network still have great development potential. Convolutional neural network-based algorithms excel at extracting local features but struggle to effectively handle long-term dependencies. In contrast, Transformer-based methods enhance model performance by achieving a global receptive field, but are computationally intensive and face unresolved challenges related to the secondary complexity of the attention mechanism. To address these limitations, the Structured Selective State Space Model (Mamba) has recently emerged as a promising alternative. Mamba models offer a novel approach to vision tasks by transforming 2D features into a 1D array through a cross-scan module, thus preserving the global receptive field while reducing computational complexity. However, existing Mamba models often neglect feature information in the spectral dimension. Flattening 2D feature maps into 1D sequences can lead to a loss of spectral dimensional correlation, thereby impairing the model's ability to capture spectral features and reducing the overall effectiveness of HSI reconstruction.

[0005] Therefore, a method is needed that can combine the degradation patterns of the entire system with low computational complexity, while effectively capturing spatial and spectral information and having a certain degree of interpretability, in order to improve the quality and efficiency of hyperspectral image reconstruction. Summary of the Invention

[0006] The purpose of this invention is to address the difficulty in balancing model accuracy and computational efficiency in existing hyperspectral image (HSI) reconstruction methods. It proposes a hyperspectral compression reconstruction method and system based on an iterative degenerate perceptual Mamba model, suitable for efficient reconstruction of single-exposure snapshot hyperspectral data. This invention significantly improves the reconstruction performance and robustness of hyperspectral images by designing an iterative reconstruction network combining a spectral-spatial enhancement module (SSEM) and a degenerate learning gradient descent module (DLGD).

[0007] In a first aspect, the present invention provides a hyperspectral compression reconstruction method based on an iterative degenerate sensing Mamba model, comprising the following steps:

[0008] Step S1: Use paired 3D hyperspectral real image data to generate 2D measurement frames to form the original dataset, and divide it into training set and test set;

[0009] The generation process of the measurement frame simulates the modulation process of the SD-CASSI system. Specifically, the three-dimensional hyperspectral real image is first modulated by a physical mask, then two-dimensional images of different bands are shifted along the y-axis, and finally compressed into a two-dimensional measurement frame.

[0010] Step S2: Shift and slice the measurement frame and physical mask to generate a three-dimensional hyperspectral measurement image and three-dimensional mask data;

[0011] Step S3: Construct the Iterative Degradation-Aware Mamba Model (IDS) 2 E-Mamba was used for training on the training set and for validation on the test set.

[0012] The Iterative Degradation Perceptive Mamba Model (IDS) 2 E-Mamba includes initializing the network and rebuilding the network using K cascaded connections; K ≥ 1;

[0013] The initialization network is used to receive the three-dimensional hyperspectral measurement image Y and the three-dimensional mask data Φ, and preprocess them to obtain the initial hyperspectral image x0 and the initial hidden feature F0.

[0014] The K cascaded reconstruction networks are used to perform multi-scale reconstruction based on the initialized hyperspectral image x0 and the initialized hidden features F0, to obtain the reconstruction result x. K ;

[0015] Step S4: Use the Iterative Degradation-Aware Mamba Model (IDS) with good reconstruction performance for training and validation. 2 E-Mamba enables the reconstruction of single-exposure snapshot hyperspectral data.

[0016] Preferably, the iterative degradation-sensing Mamba model (IDS) 2 In E-Mamba, the initialization network includes an encoder, a bottleneck layer, and a decoder. The encoder includes a first convolutional layer, a first depthwise convolutional layer, and a second convolutional layer. The decoder includes a deconvolutional layer, a second depthwise convolutional layer, and a third convolutional layer. The bottleneck layer includes a dilated convolutional layer and an average pooling layer.

[0017] Preferably, each reconstruction network includes a Degenerate Learning Gradient Descent (DLGD) module and a denoiser;

[0018] In the k-th reconstruction network, the degenerate learning gradient descent module (DLGD) includes a degenerate learning block (DLB) and a proximal gradient descent block (PGD), k∈[1,K];

[0019] The degradation learning block (DLB) is created by introducing the original 3D hyperspectral measurement image and combining it with the reconstruction results from previous iterations. k-1 and degradation factor Φ k-1 Joint learning is conducted to generate a new round of degradation information Φ k :

[0020] Φ k =DLB(Φ k-1 ,x k-1 ,Y)

[0021] Where Φ0 = Φ, and Y represents the original three-dimensional hyperspectral measurement image;

[0022] The proximal gradient descent block (PGD) is used to process the degradation information Φ k Further gradient descent is performed to obtain the proximal operator v. k ;

[0023] The denoiser is used to modify the hidden features F from previous iterations based on the output of the Degenerate Learning Gradient Descent module (DLGD). k-1 Perform noise reduction;

[0024] Preferably, the specific implementation process of the degenerate learning block (DLB) in the degenerate learning gradient descent module (DLGD) is as follows:

[0025] The original three-dimensional hyperspectral measurement image Y and the reconstructed result x k-1 By taking the difference, we obtain the first result S1;

[0026] The original three-dimensional hyperspectral measurement image Y is processed by channel attention and the first spatial refinement layer to obtain the second result S2;

[0027] The first result S1 and the second result S2 are concatenated to obtain the preliminary fusion result S3.

[0028] The initial fusion result S3 was further optimized through channel attention and convolution operations to obtain shallow features Φ';

[0029] The degradation information from the previous iteration Φ k-1 After being concatenated with shallow features Φ', convolutional operations and a second spatial refinement layer are applied to enhance the accuracy of spatial information, ultimately yielding new degenerate information Φ. k .

[0030] Preferably, the first spatial refinement layer and the second spatial refinement layer have the same structure. Specifically, one branch generates multiplication weights through the first conv3×3, LeakyReLU activation function, the second conv3×3, and the Sigmoid activation function; the other branch generates addition correction terms through the third conv3×3, LeakyReLU activation function, and the fourth conv3×3; finally, the multiplication weights and addition correction terms are fused with the input of the current spatial refinement layer.

[0031] Preferably, the proximal gradient descent block (PGD) is implemented as follows:

[0032] The degradation information Φ output by the Degradation Learning Gradient Descent (DLGD) module k The learning process is further refined using perceptual matrix generation blocks and perceptual matrix transpose generation blocks, respectively, to obtain the degradation factor. Transpose degradation factor

[0033] Degeneration factor The reconstruction result x from the previous iteration k-1 After fusion, the results are obtained through dynamic interaction via the forward sensing block.

[0034] The result The result of subtracting from the original three-dimensional hyperspectral measurement image Y and the transpose degradation factor The data is then fused and dynamically interacted with by the backward sensing block, enabling the network to adaptively adjust the reconstruction process based on the current value of the degradation factor.

[0035] Finally, the learnable step size ρ is used to further enhance the optimization, and the reconstruction results x from the previous iterations are used as a basis. k-1 Obtain the proximal operator v k The specific meaning is as follows:

[0036]

[0037] Preferably, in the proximal gradient descent (PGD) block, the perceptron generation block, the perceptron transpose generation block, the forward perceptron block, and the backward perceptron block have the same structure, all including a first conv1×1, a first conv3×3, a ReLU activation function, a second conv3×3, and a second conv1×1 cascaded in sequence, and a residual connection is introduced.

[0038] Preferably, in the reconstruction network, the denoiser is specifically implemented as follows:

[0039] The denoiser includes an encoder, a bottleneck layer, a decoder, and an output layer;

[0040] The encoder includes multiple cascaded encoding layers;

[0041] The input of the bottleneck layer is connected to the output of the last coding layer.

[0042] The decoder includes multiple cascaded decoding layers. The input of each decoding layer is the result of a convolution operation after fusing the output of the encoder in the same layer with the output of the previous decoding layer or the bottleneck layer.

[0043] The output layer performs a weighted summation of the decoder's output after a convolution operation and the output of the Degenerate Learning Gradient Descent (DLGD) module, outputting a denoised image x. k .

[0044] Preferably, in the denoiser, the coding layer, bottleneck layer, and decoding layer all employ spectral spatial enhancement blocks (SSEMB);

[0045] The Spectral Spatial Enhancement Block (SSEMB) includes two layers of normalization (LN), a spatial spectral enhancement Mamba (SSE-Mamba), and a feedforward network (FFN).

[0046] The specific implementation process of the Spatial Spectral Enhancement Mamba (SSE-Mamba) is as follows:

[0047] 2) Preprocessing: The input data is convolved and divided into two equal parts along the channel dimension, and then each part is processed by a depthwise separable convolution to obtain preprocessed features. and

[0048]

[0049] 2) Spectral Attention: By applying preprocessed features... Pooling is performed along the spatial dimension to extract global spectral information. The pooling result is then input into a shared multilayer perceptron (MLP) to obtain the spectral attention weights w. spe :

[0050]

[0051] w spe =σ(MLP(s) avg )+MLP(s max ))

[0052] Where σ is the Sigmoid activation function.

[0053] Spectral attention weight w spe With preprocessing features Element-wise multiplication generates spectrally enhanced features.

[0054]

[0055] 3) Spatial Attention: By analyzing preprocessed features... Average pooling and max pooling are performed along the channel dimension. The pooling results are then concatenated and fed into a convolutional layer to generate spatial attention weights w. spa :

[0056]

[0057] w spa =σ(Conv([m avg ,m max ]))

[0058] Spatial attention weight w spa With preprocessing features Element-wise multiplication generates spatially augmented features.

[0059]

[0060] 4) State-space model and feature fusion: To further enhance long-range dependency information, the Spatial Spectral Enhancement Mamba uses a state-space model (SSM) to model the global correlation between spectral and spatial features.

[0061]

[0062] In a second aspect, the present invention provides a hyperspectral compression and reconstruction system, comprising:

[0063] The data acquisition module uses the SD-CASSI system to capture measurement frames;

[0064] The data preprocessing module shifts and slices the measurement frames to obtain a three-dimensional hyperspectral measurement image;

[0065] The reconstruction module inputs the 3D hyperspectral measurement image and 3D mask into the trained and validated iterative degradation-aware Mamba model IDS. 2 E-Mamba, obtaining reconstruction results x K .

[0066] The method of this invention dynamically adjusts the degradation factor and optimizes the interaction between spectral and spatial features during the reconstruction process, achieving a good balance between network performance and efficiency. Experimental results show that this invention can significantly improve the reconstruction quality of snapshot hyperspectral computational imaging systems while reducing computational complexity. The experimental results also demonstrate that the method of this invention exhibits superior reconstruction performance on both simulated and real hyperspectral datasets, achieving a good balance between computational overhead and model complexity, providing an efficient and flexible solution for spectral compressed imaging. Attached Figure Description

[0067] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0068] Figure 1 This is to measure the process of frame shifting.

[0069] Figure 2 The iterative degenerate perceptual Mamba model (IDS) of this invention 2 A schematic diagram of the E-Mamba network architecture.

[0070] Figure 3 A schematic diagram of the network architecture for the degenerate learning gradient descent module.

[0071] Figure 4 This is a schematic diagram of the network architecture for the degenerate learning block.

[0072] Figure 5 This is a schematic diagram of the network architecture for the spatial refinement layer.

[0073] Figure 6 This is a schematic diagram of the network architecture for the space-spectral enhancement model.

[0074] Figure 7 This is a schematic diagram of the network architecture for a spatial-spectral enhancement block.

[0075] Figure 8 This is a schematic diagram of the network architecture of a feedforward network.

[0076] Figure 9 A schematic diagram of the network architecture of the Spatial-Spectral Enhanced Mamba.

[0077] Figure 10 This is a schematic diagram of a network architecture for spectral self-attention.

[0078] Figure 11 This is a schematic diagram of a spatial self-attention network architecture.

[0079] Figure 12 To reconstruct the simulated HSI data of scenario 2 using 7 state-of-the-art methods and the method of this invention, 3 of the 28 bands (487.0 nm, 558.5 nm and 614.5 nm) were selected.

[0080] Figure 13 The image is an RGB image of scene 2. The pixels in the white selection box are used to measure the spectral curve.

[0081] Figure 14 The spectral curves for scenario 2 include 7 state-of-the-art (SOTA) methods and the method of this invention.

[0082] Figure 15 To reconstruct the real HSI data of Scenario 1 using 7 state-of-the-art methods and the method of this invention, 3 of the 28 bands (558.5nm, 594.5nm and 614.5nm) were selected. Detailed Implementation

[0083] The technical solution of the present invention will be described in detail below with reference to specific embodiments. The present invention provides a hyperspectral compression reconstruction method based on an iterative degenerate perceptual Mamba model, specifically including the following steps and implementation details:

[0084] 1. Generation and preprocessing of hyperspectral datasets

[0085] Step 1: Using a hyperspectral compressed sensing imaging system (SD-CASSI), acquire pairs of 3D hyperspectral real images and measurement frame data to construct the original dataset. The original dataset is divided into a training set and a test set.

[0086] The generation process of the measurement frame simulates the modulation process in the SD-CASSI system. Specifically, the three-dimensional hyperspectral real image is first modulated by a physical mask, then two-dimensional images of different bands are shifted along the y-axis, and finally compressed into a two-dimensional measurement frame.

[0087] Step 2: Perform shifting and slicing operations on the measurement frame and physical mask. Figure 1To measure the frame shifting process, a three-dimensional hyperspectral measurement image X(:,:,n) and the corresponding three-dimensional mask data M are generated. The formula for initializing the hyperspectral image generation is:

[0088] X(:,:,n)=Y(:,(n-1)d:(n-1)d+h),n∈{1,2,3…,N}

[0089] Where h is the height of the image, n is the band index number, N is the total number of bands, d is the displacement step size, Y represents the measurement frame, X(:,:,n) represents all pixels in the two-dimensional image of the nth band of the hyperspectral image X, and Y(:,(n-1)d:(n-1)d+h) represents all pixels from row (n-1)d to row (n-1)d+h) in the height direction of the two-dimensional measurement frame.

[0090] 2. See Figure 2 Iterative Degradation-Aware Mamba Model (IDS) 2 Building and training E-Mamba

[0091] This invention constructs an iterative degradation-aware Mamba model IDS. 2 E-Mamba includes initializing the network and rebuilding the network using K concatenated nodes; K ≥ 1. The specific implementation steps are as follows:

[0092] A. Initialize the network

[0093] The initialization network adopts a U-shaped network architecture, consisting of an encoder, a bottleneck layer, and a decoder, to generate an initial hyperspectral image x0 and an initial hidden feature F0.

[0094] The encoder employs a two-stage convolutional encoder, consisting of a first convolutional layer, a first depthwise convolutional layer, and a second convolutional layer, for extracting multi-scale feature representations. Both the first and second convolutional layers use Conv1×1, while the first depthwise convolutional layer uses DWConv3×3.

[0095] The bottleneck layer, consisting of dilated convolutional layers and average pooling layers, is used to further compress and encode feature information, generating compact feature representations.

[0096] The decoder employs a two-stage convolutional decoder, consisting of a deconvolutional layer, a second depthwise convolutional layer, and a third convolutional layer, to recover the initial hyperspectral image. The deconvolutional layer uses DeConv2×2, the second depthwise convolutional layer uses DWConv3×3, and the third convolutional layer uses Conv1×1 for all layers.

[0097] B. See also Figure 3 Degenerate Learning Gradient Descent Module (DLGD)

[0098] The degradation learning gradient descent module is responsible for dynamically estimating and optimizing degradation factors Φ and Φ'.T The hyperspectral image is then iteratively updated using a gradient descent mechanism. The specific steps are as follows:

[0099] 1) Fuse the input original 3D hyperspectral measurement image Y, the 3D mask Φ, and the reconstruction result x from the previous round. k-1 Degradation factors are generated through Degradation Learning Blocks (DLBs).

[0100] See Figure 4 The specific implementation process of Degenerate Learning Block (DLB):

[0101] The original three-dimensional hyperspectral measurement image Y and the reconstructed result x k-1 By taking the difference, we obtain the first result S1;

[0102] The original three-dimensional hyperspectral measurement image Y is processed by channel attention and the first spatial refinement layer to obtain the second result S2;

[0103] The first result S1 and the second result S2 are concatenated to obtain the preliminary fusion result S3.

[0104] The initial fusion result S3 was further optimized through channel attention and convolution operations to obtain shallow features Φ';

[0105] The degradation information from the previous iteration Φ k-1 After being concatenated with shallow features Φ', convolutional operations and a second spatial refinement layer are applied to enhance the accuracy of spatial information, ultimately yielding new degenerate information Φ. k .

[0106] Both the first and second spatial refinement layers use the same spatial refinement layer (TransLayer). (See [link / reference]). Figure 5 Specifically:

[0107] A branch generates multiplication weights using a first conv3×3, LeakyReLU activation function, a second conv3×3, and a Sigmoid activation function, calculated using the following formula:

[0108] w k =σ(Conv(v) k ))

[0109] Another branch uses a third conv3×3, LeakyReLU activation function, and a fourth conv3×3 to generate an additive correction term, which is calculated using the following formula:

[0110] b k =ReLU(Conv(v) k ))

[0111] Finally, the multiplication weights, addition correction terms, and inputs of the current spatial refinement layer are fused together.

[0112] 2) Proximal gradient descent block (PGD) for degradation information Φ k Further gradient descent is performed to obtain the proximal operator v. k ; Specific implementation process:

[0113] The degradation information Φ output by the Degradation Learning Gradient Descent (DLGD) module k The learning process is further refined using perceptual matrix generation blocks and perceptual matrix transpose generation blocks, respectively, to obtain the degradation factor. Transpose degradation factor

[0114] Degeneration factor The reconstruction result x from the previous iteration k-1 After fusion, the results are obtained through dynamic interaction via the forward sensing block.

[0115] The result The result of subtracting from the original three-dimensional hyperspectral measurement image Y and the transpose degradation factor The data is then fused and dynamically interacted with by the backward sensing block, enabling the network to adaptively adjust the reconstruction process based on the current value of the degradation factor.

[0116] Finally, the learnable step size ρ is used to further enhance the optimization, and the reconstruction results x from the previous iterations are used as a basis. k-1 Obtain the proximal operator v k The specific meaning is as follows:

[0117]

[0118] Where Ф represents the degradation factor, ρ represents the learning step size, and v k This is the optimized intermediate result.

[0119] The perception matrix generation block, perception matrix transpose generation block, forward perception block, and backward perception block have the same structure, all including a first conv1×1, a first conv3×3, a ReLU activation function, a second conv3×3, and a second conv1×1 concatenated in sequence, and a residual connection is introduced.

[0120] C. See Figure 6 noise reducer

[0121] The denoiser employs a three-stage U-shaped structure, comprising an encoder, a bottleneck layer, a decoder, and an output layer. The encoder consists of multiple cascaded encoding layers. The input to the bottleneck layer is connected to the output of the last encoding layer. The decoder consists of multiple cascaded decoding layers. The input to each decoding layer is the result of convolutional processing of the output of the encoder at the same layer with the output of the previous decoding layer or the bottleneck layer. The output layer performs a weighted summation of the decoder's output after convolution with the output of the Degenerate Learning Gradient Descent (DLGD) module, outputting the denoised image x. k .

[0122] In the denoiser, the coding layer, bottleneck layer, and decoding layer all employ spectral spatial enhancement blocks (SSEMB), see [reference]. Figure 7 The Spatial Enhancement Block (SSEMB) consists of two layers of normalization (LN), a spatial spectral enhancement Mamba (SSE-Mamba), and a feedforward network (FFN).

[0123] See Figure 8 The feedforward network (FFN) consists of two branches. One branch, composed of a first Conv1×1 and a first DWConv3×3, encodes the features. The other branch, composed of a second Conv1×1, a second DWConv3×3, and a GELU activation function, generates a gated signal. Element-wise multiplication is performed on the results of these two branches. Finally, convolutional layers are used to generate the final features.

[0124] See Figure 9 The Spatial Spectrum Enhancement Mamba (SSE-Mamba) is specifically implemented as follows:

[0125] 1) Preprocessing: The input data is convolved and divided into two equal parts along the channel dimension, and then each part is processed by a depthwise separable convolution to obtain preprocessed features. and

[0126]

[0127] 2) Spectral attention, see Figure 10 : By preprocessing features Pooling is performed along the spatial dimension to extract global spectral information. The pooling result is then input into a shared multilayer perceptron (MLP) to obtain the spectral attention weights w. spe :

[0128]

[0129] w spe =σ(MLP(s) avg )+MLP(s max ))

[0130] Where σ is the Sigmoid activation function

[0131] Spectral attention weight w spe With preprocessing features Element-wise multiplication generates spectrally enhanced features.

[0132]

[0133] 3) Spatial attention, see [reference] Figure 11 : By preprocessing features Average pooling and max pooling are performed along the channel dimension. The pooling results are then concatenated and fed into a convolutional layer to generate spatial attention weights w. spa :

[0134]

[0135] w soa =σ(Conv([m avg ,m max ]))

[0136] Spatial attention weight w spa With preprocessing features Element-wise multiplication generates spatially augmented features.

[0137]

[0138] 4) State-space model and feature fusion: To further enhance long-range dependency information, the Spatial Spectral Enhancement Mamba uses a state-space model (SSM) to model the global correlation between spectral and spatial features.

[0139]

[0140] D. Iterative optimization and reconstruction

[0141] The initial hyperspectral image x0 is iteratively updated through K rounds of the reconstruction network, eventually generating the reconstructed hyperspectral image x. K The reconstruction process dynamically adjusts the degradation factor and optimizes the interaction between spectral and spatial features, thereby improving the reconstruction quality of the image.

[0142] 3. Experiment and Results Analysis

[0143] A. Experimental setup:

[0144] We conducted both simulated and real HSI experiments to evaluate the performance of our method. For the simulated experiments, we used the CAVE and KAIST datasets, while real compressed measurements were collected for real-world data testing.

[0145] Datasets: To ensure the robustness and generality of our method, we carefully selected datasets representing different spatial and spectral features. These datasets were split into training and testing sets for effective evaluation. We preprocessed the hyperspectral data by selecting a subset of spectral bands, which is standard practice for hyperspectral image reconstruction tasks.

[0146] 1) Simulated Data: In the simulation experiments, we used the CAVE dataset with a spatial resolution of 512×512 and the KAIST dataset with a spatial resolution of 2704×3376. The CAVE dataset was used for training, while 10 samples randomly selected from KAIST were used for testing. We used spectral interpolation to select 28 bands in the range of 450–650 nm.

[0147] 2) Real-world data: For real-world evaluation, we used compressed measurements from TSA-Net with a spatial size of 660×714. Training was performed using the CAVE and KAIST datasets, and real compressed measurements were used as test data.

[0148] Evaluation Metrics: The performance of the proposed method was evaluated using Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity (SSIM). Detailed Implementation Data: Our method is implemented in PyTorch. The model was trained using the ADAM optimizer and a 300-episode cosine annealing learning rate schedule. The initial learning rate was set to 4 × 10⁻⁶. -4 The batch size was set to 2. To enhance generalization, data augmentation techniques such as random flipping and rotation were applied. All experiments were conducted on a single RTX 3090 GPU.

[0149] B. Result Comparison

[0150] Simulated HSI Reconstruction: To evaluate the performance of our proposed method, we compared it with several state-of-the-art open-source methods, including two model-based methods (GAP-TV, DeSCI), one CNN-based method (DGSMP), and three transformer-based methods (MST, CST, DAUHST). All comparative experiments were conducted in the same settings as DGSMP to ensure fairness. We evaluated our method across 10 test scenarios, using PSNR and SSIM as evaluation metrics. The experimental results shown in Table 1 demonstrate that the deep learning method generally outperforms traditional model-based methods. Specifically, our proposed method, with 10 stages, achieves a PSNR 0.21 dB higher than DAUHST (9 stages). Furthermore, we report the computational overhead (GFLOPS), memory usage (parameters), and mean runtime for each method. Compared to other methods, our model achieves a better balance between parameter count and computational complexity. Figure 12 The reconstruction results of our method and other methods on Scene 2 are presented, including three spectral bands (487.0 nm, 558.0 nm, 614.5 nm). It highlights the block artifacts caused by the oversmoothing phenomenon of traditional model-based methods (GAP-TV) and DeSCI. Our method significantly improves the extraction of spatial context and spectral information by introducing spectral self-attention and spatial self-attention modules, resulting in better spatial smoothness and consistency in the reconstructed HSI. Figure 13 The image is an RGB image of scene 2. The pixels in the white selection box are used to measure the spectral curve. Figure 14 The average spectral curves for a 10×10 region in Scenario 2 are presented, including those from seven state-of-the-art (SOTA) methods and the method of this invention. The spectral correlation coefficients indicate that our method exhibits superior or comparable spectral consistency compared to other methods.

[0151] Real-world HSI reconstruction: Following the TSA-Net approach, we train our proposed model on the CAVE and KAIST datasets using a two-stage training process, injecting 11-bit speckle noise into the data. During testing, measurements obtained from real-world SCI systems are used. Figure 15 The reconstruction results of real HSI data from Scenario 1 are shown, and seven state-of-the-art (SOTA) methods are compared with our proposed method. Results include reconstructions of three spectral bands (558.5 nm, 594.5 nm, and 614.5 nm) selected from a total of 28 bands. Our model provides satisfactory results.

[0152] C. Ablation test

[0153] We first removed the DLGD and SSEM modules from the two-stage model to obtain the baseline model (1) for the ablation experiment. Next, by adding DLGD to model (1), we obtained model (2). The PSNR of model (1) was 31.47 dB, while the PSNR of model (2) was improved by 3.51 dB. This is due to the dynamic learning of degradation information by DLGD, which makes the linear projection more flexible and better able to adapt to the degradation characteristics of the real system. Introducing SSEM into model (1) to form model (3) improved the PSNR by 4.77 dB. This improvement is attributed to the combination of spectral and spatial attention layers in SSEM, which enhances the interaction between channels and sequence consistency. Based on model (3), adding DLGD without the DLB module resulted in model (4), which achieved a performance 1.26 dB higher than model (3). Finally, when the DLB module was added to model (4) to form model (5), the PSNR increased by an additional 0.15 dB. This improvement is attributed to the DLB module, which learns the degradation matrix more accurately through an attention mechanism and a spatial transformation layer, further enhancing the reconstruction quality. The ablation experiment results are shown in Table 2.

[0154] Table 1 presents a quantitative analysis of the simulation results for 10 scenarios (higher PSNR and SSIM are better).

[0155]

[0156]

[0157] Among them, IDS 2 In E-Mamba-2, the 2 represents K=2, meaning the model has two reconstruction networks; IDS 2 E-Mamba-5 has 5 rebuild networks; IDS 2 E-Mamba-10 has 10 rebuild networks.

[0158] Table 2 shows the average PSNR and SSIM obtained from the ablation experiments using simulated data for 10 simulation scenarios.

[0159]

[0160] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A hyperspectral compression reconstruction method based on an iterative degenerate sensing Mamba model, comprising the following steps: Step S1: Use the three-dimensional hyperspectral real image to generate a two-dimensional measurement frame through physical mask modulation, which is used to form the original dataset and divide it into training set and test set; Step S2: Shift and slice the measurement frame and physical mask to generate a three-dimensional hyperspectral measurement image and three-dimensional mask data; Step S3: Construct the Iterative Degradation-Aware Mamba Model (IDS) 2 E-Mamba was used for training on the training set and for validation on the test set. Step S4: Use the Iterative Degradation-Aware Mamba Model (IDS) with good reconstruction performance for training and validation. 2 E-Mamba enables the reconstruction of single-exposure snapshot hyperspectral data; Its characteristic is that the iterative degradation-aware Mamba model IDS 2 E-Mamba includes initializing the network and reconstructing the network using K cascaded connections; K ≥ 1; The initialization network is used to receive the three-dimensional hyperspectral measurement image Y and the three-dimensional mask. The initial hyperspectral image is obtained by preprocessing it. and initializing hidden features ; The K cascaded reconstruction networks are used to initialize the hyperspectral image. and initializing hidden features Multi-scale reconstruction was performed to obtain the reconstruction results. ; Each reconstruction network includes a denoiser; The denoiser includes an encoder, a bottleneck layer, a decoder, and an output layer; the encoder, bottleneck, and decoder layers all employ a spectral spatial enhancement block (SSEMB); wherein the spectral spatial enhancement block (SSEMB) includes two layers of normalized linear network (LN), a spatial spectral enhancement Mamba (SSE-Mamba), and a feedforward network (FFN). The specific implementation process of the Spatial Spectral Enhancement Mamba (SSE-Mamba) is as follows: 1) Preprocessing: Perform convolution on the input data and divide it into two equal parts along the channel dimension. Then, perform a depthwise separable convolution on each part to obtain preprocessed features. and ; 2) Spectral attention: By focusing on preprocessed features Pooling is performed along the spatial dimension to extract global spectral information. The pooling result is then input into a shared multilayer perceptron (MLP) to obtain spectral attention weights. ; Spectral attention weights With preprocessing features Element-wise multiplication generates spectrally enhanced features. ; 3) Spatial attention: through the preprocessed features Average pooling and max pooling are performed along the channel dimension. The pooling results are then concatenated and fed into the convolutional layer to generate spatial attention weights. Spatial attention weights With preprocessing features Element-wise multiplication generates spatially augmented features. ; 4) State-space model and feature fusion: The global correlation between spectral and spatial features is obtained through state-space modeling (SSM). : 。 2. The method according to claim 1, characterized in that, Iterative Degradation-Aware Mamba Model (IDS) 2 In E-Mamba, the initialization network includes an encoder, a bottleneck layer, and a decoder. The encoder includes a first convolutional layer, a first depthwise convolutional layer, and a second convolutional layer. The decoder includes a deconvolutional layer, a second depthwise convolutional layer, and a third convolutional layer. The bottleneck layer includes a dilated convolutional layer and an average pooling layer.

3. The method according to claim 1, characterized in that, Each reconstructed network also includes a degenerate learning gradient descent module (DLGD); In the k-th reconstruction network, the degenerate learning gradient descent module DLGD includes a degenerate learning block DLB and a proximal gradient descent block PGD, k∈[1,K]; The degradation learning block (DLB) is created by introducing the original 3D hyperspectral measurement image and combining it with the reconstruction results from previous iterations. and degradation factors Joint learning is conducted to generate a new round of degradation information. ; The proximal gradient descent block (PGD) is used to analyze degradation information. Further gradient descent is performed to obtain the proximal operator. ; The denoiser is used to denoise the hidden features from previous iterations based on the output of the degenerate learning gradient descent module DLGD. Noise reduction is performed.

4. The method according to claim 3, characterized in that, In the Degenerate Learning Gradient Descent (DLGD) module, the specific implementation process of the Degenerate Learning Block (DLB) is as follows: The original 3D hyperspectral measurement image Y and the reconstruction results from previous iterations are compared. By taking the difference, we obtain the first result S1; The original three-dimensional hyperspectral measurement image Y is processed by channel attention and the first spatial refinement layer to obtain the second result S2; The first result S1 and the second result S2 are concatenated to obtain the preliminary fusion result S3. The initial fusion result S3 was further optimized through channel attention and convolution operations to obtain shallow features. ; Degradation information from previous iterations shallow features After cascading, new degradation information is obtained through convolution operations and a second spatial refinement layer. .

5. The method according to claim 4, characterized in that, The first and second spatial refinement layers have the same structure. Specifically, one branch generates multiplication weights using the first conv3×3, LeakyReLU activation function, the second conv3×3, and the Sigmoid activation function; the other branch generates addition correction terms using the third conv3×3, LeakyReLU activation function, and the fourth conv3×3; finally, the multiplication weights and addition correction terms are fused with the input of the current spatial refinement layer.

6. The method according to claim 3, characterized in that, The specific implementation process of the proximal gradient descent block PGD is as follows: Degradation information output by the Degradation Learning Gradient Descent (DLGD) module The learning process is further refined using perceptual matrix generation blocks and perceptual matrix transpose generation blocks, respectively, to obtain the degradation factor. Transpose degradation factor ; Degeneration factor Reconstruction results from previous iterations After fusion, the results are obtained through dynamic interaction via the forward sensing block. ; The result The result of subtracting from the original three-dimensional hyperspectral measurement image Y and the transpose degradation factor The data is then fused and dynamically interacted with by the backward sensing block, enabling the network to adaptively adjust the reconstruction process based on the current value of the degradation factor. Finally, the learnable step size ρ is used to further enhance the optimization, and the reconstruction results from previous iterations are then used as a basis. Obtain the proximal operator .

7. The method according to claim 6, characterized in that, In the reconstruction network, the denoiser includes an encoder, a bottleneck layer, a decoder, and an output layer; The encoder includes multiple cascaded encoding layers; The input of the bottleneck layer is connected to the output of the last coding layer. The decoder includes multiple cascaded decoding layers. The input of each decoding layer is the result of a convolution operation after fusing the output of the encoder in the same layer with the output of the previous decoding layer or the bottleneck layer. The output layer performs a weighted summation between the decoder's output after a convolution operation and the output of the Degradation Learning Gradient Descent (DLGD) module, resulting in a denoised image. .

8. A hyperspectral compression and reconstruction system for implementing the method of any one of claims 1-7, characterized in that... include: The data acquisition module is responsible for capturing measurement frames; The data preprocessing module is responsible for shifting and slicing the measurement frames to obtain three-dimensional hyperspectral measurement images; The reconstruction module is responsible for inputting the aforementioned 3D hyperspectral measurement image and 3D mask into the trained and validated iterative degradation-aware Mamba model (IDS). 2 E-Mamba, obtained reconstruction results .

Images