A spectral super-resolution reconstruction method based on a bidirectional attention mechanism

Abstract

The present application relates to a kind of based on bidirectional attention mechanism's spectral super-resolution reconstruction method, comprising: establishing spectral super-resolution dataset;The RGB image and hyperspectral image are preprocessed, and the shallow feature of the RGB image is extracted;Again, according to the shallow feature of the RGB image, global information modeling is carried out, and the overall context information of image is extracted, to generate global feature;Spectral dimension is reversed to the feature extracted, and the spectral information feature representation of RGB image is obtained;The spatial information feature representation of the RGB image is obtained;Fusion feature is obtained;Reconstructed hyperspectral image is obtained.The beneficial effects of the present application are: the spectral super-resolution framework constructed by the present application, by adopting bidirectional spectral attention mechanism and independent spatial information extraction branch, the significant improvement of spectral recovery precision is realized.

Description

Technical Field

[0001] This invention relates to the field of remote sensing image processing technology, and more specifically, to a spectral super-resolution reconstruction method based on a bidirectional attention mechanism. Background Technology

[0002] Hyperspectral imagery (HSI), with its characteristic of recording continuous, narrow-band spectral information per pixel, is of great significance in applications such as material identification, target detection, environmental monitoring, agricultural remote sensing, and geological exploration. Through hyperspectral imagery, ground features can be distinguished in detail, enabling precise classification and quantitative analysis. However, due to the technological and hardware cost limitations of hyperspectral sensors, its data acquisition process is typically constrained by the following: traditional hyperspectral imaging systems (such as pushbroom or point-by-point scanning systems) require expensive equipment and have slow acquisition speeds, making them unsuitable for rapid imaging in dynamic scenes; hyperspectral imaging equipment is sensitive to factors such as ambient lighting and sensor calibration, resulting in noise and instability in the actual acquired data; and due to the high acquisition and time costs, hyperspectral data faces problems of insufficient data volume or untimely updates when applied on a large scale.

[0003] To address the numerous challenges of directly acquiring hyperspectral images, researchers have proposed the Spectral Super-Resolution (SSR) method, which reconstructs high-dimensional hyperspectral images from low-dimensional RGB images. SSR techniques mainly fall into two categories. Traditional SSR methods often rely on mathematical tools such as sparse representation, dictionary learning, and spectral basis function reconstruction. For example, sparse dictionary-based methods construct a dictionary of training samples to represent hyperspectral data, but these methods are often limited in representational power and generalization performance. Methods based on maximum a posteriori probability and spectral component bias functions, due to their simple models, struggle to capture complex spectral features, resulting in poor restoration performance. With the widespread application of deep neural networks, more and more research is employing novel architectures such as Convolutional Neural Networks (CNNs) and Transformers to implement SSR. Deep learning methods learn the mapping relationship between RGB and HSI through training on large amounts of data, achieving more accurate restoration results than traditional SSR. However, existing frameworks often struggle to balance spectral reconstruction and spatial detail preservation, and lack effective mechanisms to simulate bidirectional spectral-spatial interactions. These limitations lead to poor performance in high-fidelity reconstruction scenarios requiring spectral correlation and fine-grained spatial structure. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing a spectral super-resolution reconstruction method based on a bidirectional attention mechanism.

[0005] Firstly, a spectral super-resolution reconstruction method based on a bidirectional attention mechanism is provided, including:

[0006] Step 1: Establish a spectral super-resolution dataset, which includes corresponding hyperspectral images and RGB images;

[0007] Step 2: Preprocess the RGB image and hyperspectral image, and extract the shallow features of the RGB image; then perform global information modeling based on the shallow features of the RGB image, extract the overall context information of the image, and generate global features;

[0008] Step 3: Invert the spectral dimension of the extracted features to obtain positive and negative features, and perform self-attention calculation on the spectral dimension to obtain the spectral information feature representation of the RGB image.

[0009] Step 4: Obtain the spatial information feature representation of the RGB image;

[0010] Step 5: Perform multi-scale feature interaction fusion on the spectral information feature representation and the spatial information feature representation to obtain fused features;

[0011] Step 6: Obtain the reconstructed hyperspectral image based on the fusion features.

[0012] As a preferred option, step 7 involves optimization using a hybrid loss function.

[0013] Preferably, step 2, the preprocessing of the RGB image and the hyperspectral image, includes:

[0014] The RGB images are uniformly cropped to a fixed size and then normalized.

[0015] The corresponding hyperspectral images are normalized based on the number of channels in the dataset;

[0016] Data augmentation of RGB and hyperspectral images was performed using random flipping, rotation, and cropping methods.

[0017] Preferably, step 3 includes:

[0018] Step 3.1: Perform spectral dimension self-attention calculation on the positive token sequence;

[0019] Step 3.2: Reverse the token sequence in the spectral dimension to obtain the reverse token sequence, and then perform self-attention calculation in the spectral dimension;

[0020] Step 3.3: Based on the self-attention calculation results, perform splicing, linear transformation and position encoding to form the spectral information feature representation of the RGB image.

[0021] Preferably, step 4 includes:

[0022] Step 4.1: Extract local features from the RGB image to obtain local detail information; the local detail information includes image edges and textures;

[0023] Step 4.2: Fuse features at different scales using U-Net-style downsampling and upsampling operations;

[0024] Step 4.3: Through cross-layer splicing and convolution operations, feature dimensionality reduction and information complementarity are achieved to form a complete spatial information feature representation.

[0025] Preferably, step 6 includes:

[0026] Step 6.1: Perform global information modeling based on the fusion features to capture the overall spectral trend of the image;

[0027] Step 6.2: Adaptively weight the channel and spatial information respectively to highlight key features and suppress redundant information;

[0028] Step 6.3: Calculate the hyperspectral image after spectral super-resolution and add it element-by-element to the shallow features to obtain the reconstructed hyperspectral image.

[0029] In a second aspect, a spectral super-resolution reconstruction system based on a bidirectional attention mechanism is provided for performing any of the methods described in the first aspect, including:

[0030] A module is established to create a spectral super-resolution dataset, which includes corresponding hyperspectral and RGB images.

[0031] The preprocessing module is used to preprocess RGB images and hyperspectral images, and extract shallow features from the RGB images; then, based on the shallow features of the RGB images, global information modeling is performed to extract the overall context information of the images and generate global features.

[0032] The calculation module is used to reverse the spectral dimension of the extracted features to obtain positive and negative features, and perform self-attention calculation on the spectral dimension to obtain the spectral information feature representation of the RGB image.

[0033] The first acquisition module is used to acquire the spatial information feature representation of the RGB image;

[0034] The fusion module is used to perform multi-scale feature interaction fusion on the spectral information feature representation and the spatial information feature representation to obtain fused features;

[0035] The second acquisition module is used to acquire the reconstructed hyperspectral image based on the fusion features.

[0036] Thirdly, a computer storage medium is provided, wherein a computer program is stored therein; when the computer program is run on a computer, the computer causes the computer to perform any of the methods described in the first aspect.

[0037] Fourthly, an electronic device is provided, comprising:

[0038] Memory, used to store computer programs;

[0039] A processor for executing the computer program to implement the method as described in any of the first aspects.

[0040] The beneficial effects of this invention are:

[0041] 1. The spectral super-resolution framework constructed in this invention significantly improves the accuracy of spectral reconstruction by employing a bidirectional spectral attention mechanism and an independent spatial information extraction branch. This mechanism enables the network to capture spectral dependencies from both positive and negative directions, overcoming the information loss problem caused by traditional unidirectional modeling, thus achieving a more accurate spectral reconstruction effect. Simultaneously, the independently designed MSAB branch utilizes the Transformer structure and multi-scale interaction mechanism to efficiently extract and recover spatial details such as edges and textures in the image, ensuring the full preservation of image spatial information.

[0042] 2. This invention also employs a multi-scale feature fusion strategy to achieve synergistic optimization of global and local information, resulting in a high level of quality in both the overall structure and local details of the reconstructed hyperspectral image. Simultaneously, by introducing a hybrid loss function and an end-to-end optimization strategy, not only is the hyperspectral reconstruction error significantly reduced, but the fidelity of image spatial information is also maintained, thereby greatly improving the overall training stability and generalization ability of the network. These technical advantages collectively constitute a significant innovation of this invention in the field of spectral super-resolution. Attached Figure Description

[0043] Figure 1 The flowchart of the spectral super-resolution reconstruction method with bidirectional attention mechanism and the flowchart of each module are shown.

[0044] Figure 2 A comparison chart of experimental results on the ICVL dataset;

[0045] Figure 3 A comparison chart of experimental results on the DFC2018 Houston dataset;

[0046] Figure 4 This is a comparison chart of experimental results on the TG1HRSSC dataset. Detailed Implementation

[0047] The present invention will be further described below with reference to embodiments. The description of the embodiments below is only for the purpose of helping to understand the present invention. It should be noted that those skilled in the art can make several modifications to the present invention without departing from the principle of the present invention, and these improvements and modifications also fall within the protection scope of the claims of the present invention.

[0048] Example 1:

[0049] To achieve high-fidelity spectral super-resolution of remote sensing images, this application provides a spectral super-resolution reconstruction method based on a bidirectional attention mechanism. First, preprocessing and data augmentation operations are performed on both RGB and hyperspectral images. Second, shallow feature extraction is performed on the RGB image. Next, the spectral-spatial fusion interaction module SSFIM achieves collaborative optimization of spectral and spatial information. Finally, the designed reconstruction module adaptively fuses the bidirectional spectral information to obtain the final hyperspectral image.

[0050] Specifically, such as Figure 1 As shown, the method includes:

[0051] Step 1: Establish a spectral super-resolution dataset, which includes the corresponding hyperspectral image and RGB image.

[0052] Step 2: Preprocess the RGB image and hyperspectral image, and extract the shallow features of the RGB image; then perform global information modeling based on the shallow features of the RGB image, extract the overall context information of the image, and generate global features.

[0053] Specifically, preprocessing includes: uniformly cropping the input RGB images to a fixed size (e.g., 256×256) and normalizing the images, scaling pixel values ​​to the [0,1] range; simultaneously, normalizing the corresponding hyperspectral images (HSI) according to the specific number of channels in each dataset (e.g., 31, 50, or 54) to ensure scale consistency among data. To prevent overfitting during training, random flipping, rotation, and random cropping are used to augment the RGB images and HSI. These data augmentation strategies not only expand the training set but also improve the model's robustness to noise and environmental changes.

[0054] Extracting shallow features from the RGB image includes: the preprocessed RGB image first enters the shallow feature extraction module, such as... Figure 1 (a) The main task of this module is to map low-dimensional RGB data to a high-dimensional feature space, laying the foundation for the extraction of spectral and spatial information in subsequent modules.

[0055] Furthermore, applying a 1×1 convolution operation to the input RGB image X maps the original 3 channels to a higher-dimensional feature representation:

[0056] F s =CELU(Conv 1×1 (X))

[0057] Among them, Conv 1×1 This indicates a convolution operation with a 1×1 kernel, and GELU represents the GELU activation function.

[0058] The Vision Mamba module is used to model the shallow features output by 1×1 convolutions globally, extracting the overall contextual information of the image and generating global features F. Vim :

[0059] F Vim =VisionMamba(F s )

[0060] This module captures the global correlation of images, providing rich contextual background for subsequent spectral and spatial branches.

[0061] Step 3: Perform self-attention calculation on the spectral dimension of the forward and reverse token sequences to obtain the spectral information feature representation of the RGB image.

[0062] In step 3, the extracted features are inverted in spectral dimension to obtain forward and reverse features (i.e., forward and reverse token sequences). To fully capture the hidden spectral dependency information in the RGB image, this embodiment constructs a bidirectional spectral attention branch, such as... Figure 1 (b) includes a forward branch (MSEB) and a reverse branch (RMSEB). The specific implementation steps are as follows:

[0063] Step 3.1: The forward branch (MSEB) directly performs self-attention calculation on the spectral dimension of the token sequence.

[0064] Step 3.2, the reverse branch (RMSEB) first reverses the order of the token sequence in the spectral dimension, and then performs the same self-attention calculation.

[0065] The specific implementation steps for steps 3.1 and 3.2 are as follows:

[0066] Q = T·W Q K = T·W K V = T·W V

[0067] Among them, W Q W K and W VLet T be the trainable parameter matrix, and T be the token sequence.

[0068] Step 3.3: Based on the self-attention calculation results, perform splicing, linear transformation and position encoding to form the spectral information feature representation of the RGB image.

[0069] Specifically, within each attention head, the attention weight matrix is ​​calculated as follows:

[0070]

[0071]

[0072]

[0073] And for the value V i Weighted summation is performed to generate the outputs of each head, and then the final spectral attention output S-MSA(X) is formed through splicing, linear transformation, and position encoding. token ).

[0074] In step 3, the two branches share weights to ensure that the positive and negative features are complementary, thus compensating for the shortcomings of unidirectional modeling. Features at various scales are fused through U-Net-style downsampling and upsampling operations.

[0075] Step 4: Obtain the spatial information feature representation of the RGB image.

[0076] Step 4 includes:

[0077] Step 4.1: Extract local features from the RGB image to obtain local detail information.

[0078] Specifically, in order to ensure that spatial details of the image are not ignored during the spectral super-resolution process, this embodiment designs an independent spatial information extraction branch, MSAB, which uses multiple 3×3 convolutional layers, ReLU activation function and max pooling layer to extract local features from RGB images and obtain local detail information such as image edges and textures.

[0079] Step 4.2: Fuse features at different scales using U-Net-style downsampling and upsampling operations.

[0080] Step 4.3: Cross-layer concatenation (Cat operation) and 1×1 convolution further achieve feature dimensionality reduction and information complementarity, forming a complete spatial information feature representation. The specific process is similar to step 3, and will not be described here.

[0081] Step 5: Perform multi-scale feature interaction fusion on the spectral information feature representation and the spatial information feature representation to obtain fused features.

[0082] Specifically, step 5 includes:

[0083] Step 5.1: In the U-Net structure, the features from MSEB, RMSEB and MSAB are downsampled layer by layer to extract low-level details, and then upsampled to recover high-resolution information, ensuring that both global structure and local details are expressed.

[0084] Step 5.2: In the corresponding downsampling and upsampling stages, features at different scales are fused through 1×1 convolution and concatenation operations (Cat operations). The specific fusion process can be found in the following formula:

[0085]

[0086]

[0087]

[0088]

[0089] Where SEB represents the spectral attention module, D represents downsampling of the feature map, U represents upsampling of the feature map, Cat represents the concatenation operation of the input features along the channel dimension, and Conv 1×1 This indicates a convolution operation with a 1×1 kernel. and This represents the feature map generated after the nth SEB and SAB. This is the feature map of the 5th SEB after passing through the inverted spectral attention module branch; and For intermediate features; F SSFIM This is the feature map output by SSFIM.

[0090] Step 6: Based on the fusion features, the final hyperspectral image is generated after further processing in the reconstruction module.

[0091] Step 6 includes:

[0092] Step 6.1: Input the fused features into the reconstruction module. First, use the Vision Mamba module to model global information and capture the overall spectral trend of the image.

[0093] Step 6.2: Introduce the CBAM (Convolutional Block Attention Module) module to adaptively weight the channel and spatial information respectively, highlighting key features and suppressing redundant information.

[0094] Step 6.3: Calculate the hyperspectral image after spectral super-resolution and add it element-by-element to the shallow features to obtain the reconstructed hyperspectral image.

[0095] Specifically, the hyperspectral image output by the reconstruction module is denoted as...

[0096]

[0097] in, This represents the hyperspectral image after spectral super-resolution.

[0098] To further enhance the details, the shallow features F s and By adding elements one by one, the final hyperspectral image Y is obtained:

[0099]

[0100] Example 2:

[0101] Based on Example 1, Example 2 of this application provides a more specific spectral super-resolution reconstruction method based on a bidirectional attention mechanism, including:

[0102] Step 1: Establish a spectral super-resolution dataset, which includes the corresponding hyperspectral image and RGB image.

[0103] Specifically, the base datasets used in Example 2 of this application are four publicly available datasets: the Salinas dataset, which contains images of the Salinas Valley in California, USA. The corrected images contain 204 channels, covering the 0.4-2.5μm spectral range. The ICVL dataset consists of 201 images, each containing 31 spectral channels. We randomly selected 184 images as the training set and 6 images as the test set. Before the experiment, each image was cropped to (31, 256, 256). The DFC2018 Houston dataset contains 50 spectral channels. In this experiment, 92 images were randomly selected as the training set and 6 images as the test set. The TG1HRSSC dataset consists of three classes of data: panchromatic images, visible and near-infrared images with 54 effective bands, and shortwave infrared images with 52 effective bands. In this experiment, we selected visible and near-infrared images, randomly selecting 40 images as the training set and 6 images as the test set. As one embodiment, the present invention uses the computer software PyCharm and the PyTorch framework based on the NVIDIA GeForce RTX 3090 GPU to implement an automated running process.

[0104] Step 2: Preprocess the RGB image and hyperspectral image, and extract the shallow features of the RGB image; then perform global information modeling based on the shallow features of the RGB image, extract the overall context information of the image, and generate global features.

[0105] For example, the input RGB images are uniformly cropped to a fixed size (e.g., 256×256) and normalized, scaling pixel values ​​to the [0,1] range. Simultaneously, the corresponding hyperspectral images (HSI) are normalized according to the specific number of channels in each dataset (e.g., 31, 50, or 54) to ensure scale consistency across data. To prevent overfitting due to insufficient sample size during training, random flipping, rotation, and random cropping are used to augment the RGB images and HSI. This data augmentation strategy not only expands the training set but also improves the model's robustness to noise and environmental changes. The preprocessed data serves as both RGB input and HSI annotation data during training, providing a high-quality data foundation for subsequent network training and feature extraction. Furthermore, a 1×1 convolution operation is applied to the input RGB image X, mapping the original 3 channels to a higher-dimensional feature representation. The Vision Mamba module is used to model the shallow features output by the 1×1 convolution, extracting the overall contextual information of the image and generating global features.

[0106] Step 3: A bidirectional spectral attention branch was constructed. The forward branch (MSEB) directly processes the token sequence along the spectral dimension; while the reverse branch (RMSEB) first reverses the spectral order of the token sequence before performing the same self-attention calculation. The two branches share weights, ensuring that the features in both directions are complementary and compensating for the shortcomings of unidirectional modeling. Features at various scales are fused through U-Net-style downsampling and upsampling operations.

[0107] Step 4: To ensure that spatial details of the image are not ignored during spectral super-resolution, this embodiment designs an independent spatial information extraction branch, MSAB, which uses multiple convolutional layers and the ReLU activation function to extract local features from the RGB image, obtaining local details such as image edges and textures. Features at different scales are fused through U-Net-style downsampling and upsampling operations. Cross-layer stitching (Cat operation) and 1×1 convolution further achieve feature dimensionality reduction and information complementarity, forming a complete spatial information feature representation.

[0108] Step 5: Interact the spatial information extracted by MSAB with the spectral features extracted by MSEB and RMSEB to achieve collaborative recovery of spectral and spatial information.

[0109] Step 6: Obtain the reconstructed hyperspectral image based on the fusion features.

[0110] Step 7: Optimize using a hybrid loss function.

[0111] During training, the model generates high-fidelity hyperspectral images through joint constraints from multiple loss functions. The reconstruction loss ensures that the generated hyperspectral and RGB images are highly consistent with the original hyperspectral and RGB images in overall content, guaranteeing that the reconstructed images maintain realism and detail fidelity in the visible light range.

[0112] Specifically, to ensure optimal network performance in spectral and spatial reconstruction tasks, this embodiment employs the following training and optimization strategies:

[0113]

[0114] L Finall =αL HSI +βL RGB

[0115] Among them, L HSI and L RGB These are the losses for reconstructing the HSI and RGB images, respectively, both constrained by L1 loss. α and β represent adjustable parameters, and L... Finally This represents the total loss function.

[0116] Furthermore, Embodiment 2 of this application experimentally verified the above method:

[0117] exist Figure 2 , Figure 3 and Figure 4 Experimental results were presented on the ICVL, DFC2018 Houston, and TG1HRSSC datasets. The first and third rows represent pseudo-color images generated by HSI, while the second and fourth rows show the error maps between the network-generated images and the labels. In the error maps, bluer colors indicate smaller differences from the real-world situation. The proposed method was compared with spectral super-resolution images from six state-of-the-art methods. Previous methods have limitations in recovering spectral and spatial details, resulting in artifacts or blotches and an inability to accurately reconstruct image edges. Furthermore, their overall recovery quality is insufficient, lacking global consistency and continuity in both space and spectral dimensions. In contrast, the proposed method overcomes these shortcomings, achieving high-quality super-resolution with consistent spectral and spatial continuity, benefiting from more accurate spectral priors and effective supplementation of spatial information.

[0118] It should be noted that the parts in this embodiment that are the same as or similar to those in Embodiment 1 can be referred to each other, and will not be repeated in this application.

[0119] Example 3:

[0120] Based on Example 3, Example 3 of this application provides a spectral super-resolution reconstruction system based on a bidirectional attention mechanism, comprising:

[0121] A module is established to create a spectral super-resolution dataset, which includes corresponding hyperspectral and RGB images.

[0122] The preprocessing module is used to preprocess RGB images and hyperspectral images, and extract shallow features from the RGB images; then, based on the shallow features of the RGB images, global information modeling is performed to extract the overall context information of the images and generate global features.

[0123] The calculation module is used to perform spectral dimension self-attention calculation on the forward and reverse token sequences to obtain the spectral information feature representation of the RGB image;

[0124] The first acquisition module is used to acquire the spatial information feature representation of the RGB image;

[0125] The fusion module is used to perform multi-scale feature interaction fusion on the spectral information feature representation and the spatial information feature representation to obtain fused features;

[0126] The second acquisition module is used to acquire the reconstructed hyperspectral image based on the fusion features.

[0127] Specifically, the system provided in this embodiment is the system corresponding to the method provided in embodiment 2. Therefore, the parts in this embodiment that are the same as or similar to those in embodiment 2 can be referred to each other, and will not be described again in this application.

Claims

1. A method for spectral super-resolution reconstruction based on a bidirectional attention mechanism, characterized in that, The method comprises the following steps: Step 1, establishing a spectral super-resolution data set comprising corresponding hyperspectral images and RGB images; Step 2, pre-processing the RGB images and the hyperspectral images, and extracting shallow features of the RGB images; then modeling global information according to the shallow features of the RGB images, extracting overall context information of the images, and generating global features; Step 3, performing spectral dimension reversal on the extracted features to obtain forward and reverse features, and performing self-attention calculation on the spectral dimensions to obtain spectral information feature representation of the RGB images; Step 3 comprises: Step 3.1, performing self-attention calculation on the spectral dimensions of the forward token sequence; Step 3.2, arranging the token sequence in reverse order in the spectral dimension to obtain a reverse token sequence, and then performing self-attention calculation on the spectral dimension; Step 3.3, splicing, linear transformation and position coding according to the self-attention calculation result to form the spectral information feature representation of the RGB images; Step 4, obtaining spatial information feature representation of the RGB images; Step 4 comprises: Step 4.1, extracting local features of the RGB images to obtain local detail information; the local detail information comprises image edges and textures; Step 4.2, fusing features of different scales through U-Net style downsampling and upsampling operations; Step 4.3, realizing feature dimension reduction and information complementation through cross-layer splicing and convolution operations to form complete spatial information feature representation; Step 5, performing multi-scale feature interaction and fusion on the spectral information feature representation and the spatial information feature representation to obtain fused features; Step 6, obtaining reconstructed hyperspectral images according to the fused features; Step 7, optimizing through a hybrid loss function.

2. The method of claim 1, wherein the method is a bidirectional attention mechanism based spectral super-resolution reconstruction method. In step 2, the pre-processing of the RGB images and the hyperspectral images comprises: uniformly cropping the RGB images to a fixed size, and normalizing the RGB images; normalizing the corresponding hyperspectral images according to the channel number of the data set; performing data enhancement on the RGB images and the hyperspectral images by using random flipping, rotation and random cropping.

3. The method of claim 2, wherein the method is implemented by a computer system. Step 6 comprises: Step 6.1, modeling global information according to the fused features to capture overall spectral trends of the images; Step 6.2, adaptively weighting the channel and spatial information respectively to highlight key features and suppress redundant information; Step 6.3, calculating the hyperspectral images after spectral super-resolution and adding them element by element to the shallow features to obtain reconstructed hyperspectral images.

4. A spectral super-resolution reconstruction system based on a bidirectional attention mechanism, characterized in that, The method for performing any one of claims 1 to 3 comprises: a establishing module for establishing a spectral super-resolution data set comprising corresponding hyperspectral images and RGB images; a pre-processing module for pre-processing the RGB images and the hyperspectral images, and extracting shallow features of the RGB images; then modeling global information according to the shallow features of the RGB images, extracting overall context information of the images, and generating global features; The computing module performs spectral dimension inversion on the extracted features to obtain forward and reverse features, and performs spectral dimension self-attention calculation on the forward and reverse features to obtain spectral information feature representation of the RGB image. The first obtaining module is configured to obtain spatial information feature representation of the RGB image. The fusion module is configured to perform multi-scale feature interactive fusion on the spectral information feature representation and the spatial information feature representation to obtain fusion features. The second obtaining module is configured to obtain a reconstructed hyperspectral image according to the fusion features.

5. A computer storage medium, characterized in that The computer storage medium stores a computer program; when the computer program runs on a computer, the computer program makes the computer execute the method in any one of claims 1 to 3.

6. An electronic device, comprising: The computer program product comprises: a memory configured to save a computer program; a processor configured to execute the computer program to implement the method in any one of claims 1 to 3.

Images