Hyperspectral image super-resolution method and device based on iterative reconstruction and spectral refinement

Abstract

The application provides a hyperspectral image super-resolution method and device based on iterative reconstruction and spectral refinement, and relates to the technical field of hyperspectral images. First, a three-neighbor window cube is slid along the spectral dimension and is mapped into a pseudo-RGB three-channel feature through a lightweight head adapter. Then, a frozen pre-training network is subjected to domain adaptation through LoRA low-rank decomposition, and a corresponding three-band high-resolution prediction is output. Multiple predictions of the same target band generated by different center windows are normalized and fused according to the'spectral position window weight x confidence' to obtain a coarse super-resolution result. Finally, an end-to-end refinement is performed under the joint SAM spectral angle constraint by a spectral reconstruction decoder to obtain a final high-resolution hyperspectral image. The purpose is to nondestructively migrate a natural image pre-training model to a hundred-band hyperspectral scene with a very low parameter increment, and to realize the synchronous leap of spatial details and spectral fidelity.

Description

Technical Field

[0001] This invention relates to the field of hyperspectral imaging technology, and more specifically to a hyperspectral super-resolution method and apparatus based on iterative reconstruction and spectral refinement. Background Technology

[0002] In fields such as remote sensing, precision agriculture, environmental monitoring, and geological exploration, hyperspectral imagery provides rich information through a combination of image and spectral data. However, its spatial resolution has long been limited by sensor technology, signal-to-noise ratio, and satellite platform payload capacity. To acquire higher spatial details in orbit, academia and industry commonly employ "super-resolution" techniques to reconstruct high-resolution hyperspectral data from single or multiple low-resolution hyperspectral cubes. However, hyperspectral images typically contain over a hundred bands, and while adjacent bands are highly correlated, subtle differences exist. When traditional natural image super-resolution techniques are directly applied to this field, a dilemma arises where both "spectral distortion" and "spatial artifacts" coexist.

[0003] Early methods, such as cube interpolation, sparse dictionary learning, or low-rank tensor decomposition, only performed numerical fitting in the image domain. They lacked constraints on the physical and spectral continuity of imaging, resulting in blurred details, jagged edges, and systematically distorted spectral curves after magnification, making it difficult to meet the stringent radiometric accuracy requirements of quantitative remote sensing inversion. With the rise of deep learning, end-to-end convolutional or 3D convolutional networks were used to map low-resolution cubes to the target resolution in one step. Although visual sharpness was improved, the number of model parameters and computational cost increased linearly or even quadratically with band. The demand for paired high- and low-resolution samples during training increased exponentially. When the sensor platform was changed, or the spatial resolution or spectral response function was slightly different, the network needed to be retrained from scratch, making the hardware and time costs prohibitive for users.

[0004] To alleviate data and computing power issues, some researchers have attempted to introduce weights pre-trained on large datasets of natural images to give hyperspectral networks stronger spatial priors through transfer learning. Common approaches include: 1. Fine-tuning the entire pre-trained model. Theoretically, this can preserve priors to the greatest extent, but hyperspectral samples are scarce, making overfitting highly likely, and full backpropagation of billions of parameters is practically impossible to deploy in engineering; 2. Freezing the pre-trained backbone and adding only lightweight decoders or post-processing modules to the periphery. Although the training overhead is drastically reduced, the backbone parameters are "locked," making it impossible to make any adaptive adjustments to hyperspectral statistical characteristics. While spatial details are improved, spectral consistency lacks an explicit feedback channel, leading to a systematic amplification of the spectral angular distance between different ground objects, resulting in a chain amplification of subsequent classification and target detection errors; 3. A more compromise "staged" strategy breaks the process down into "coarse super-resolution - spectral correction" or "spatial-spectral decoupling three stages." First, the frozen network is used to generate a high-resolution intermediate image, and then an independent network performs spectral shaping in the second stage. However, only image-level results are passed between stages, the gradient flow is forcibly truncated, the pre-trained network cannot perceive downstream spectral errors, and a suboptimal balance is formed in each processing; at the same time, the rich information at the feature level is repeatedly compressed in the round-trip quantization, introducing irreversible loss; during inference, more than two sets of models need to be loaded serially, which significantly increases the memory usage and engineering deployment complexity.

[0005] On the other hand, RGB-guided hyperspectral super-resolution has recently become a hot topic. This involves using high-resolution RGB or MSI images of the same scene as spatial priors to improve texture details through feature transfer, matrix factorization, or joint learning. However, the RGB band covers only about three wide channels in the visible light range, creating a huge gap in the spectral dimension compared to the more than one hundred narrow channels of the hyperspectral band. Without domain alignment and spectral fidelity constraints, it is easy to introduce false details in hyperspectral sensitive areas such as vegetation and water bodies, and even change the spectral shape, resulting in the paradox of seemingly clearer but actually more distorted images.

[0006] In summary, current technologies face a spatial-spectral dilemma: either sacrifice spectral fidelity for spatial sharpness, or maintain spectral integrity while struggling to overcome spatial limitations. Simultaneously, the robust spatial capabilities provided by pre-trained models have not yet been efficiently and riskily integrated into the hyperspectral domain, with parameter scale and training costs becoming significant hurdles between industry, academia, and research. Therefore, how to achieve adaptive fine-tuning in the hyperspectral domain without altering the original pre-trained weights; how to simultaneously optimize spatial detail and spectral consistency within an end-to-end framework; and how to achieve robust super-resolution of a 100-band hyperspectral cube with minimal parameter increments, minimal training data, and minimal inference latency remain core challenges that urgently need to be overcome in this field.

[0007] In view of the above, this application is hereby submitted. Summary of the Invention

[0008] This invention provides a hyperspectral image super-resolution method and apparatus based on iterative reconstruction and spectral refinement, which can at least partially improve the above-mentioned problems.

[0009] To achieve the above objectives, the present invention adopts the following technical solution:

[0010] A hyperspectral image super-resolution method based on iterative reconstruction and spectral refinement, comprising:

[0011] Acquiring low-resolution hyperspectral images The adjacent three-band sliding windows are extracted according to the preset spectral dimensions, and the adjacent three-band sliding windows are input into the head adapter to obtain pseudo-RGB three-channel features.

[0012] The parameters of the pre-trained model are frozen, and LoRA (Low-Rank Adaptation) is used to adapt the pre-trained model. The pseudo-RGB three-channel features are input into the adapted pre-trained model to obtain the corresponding three bands. The three bands are then input into the tail adapter to obtain the predicted values ​​corresponding to the three bands. The three bands are three different bands.

[0013] The spectral window weighted fusion module is used to process the three prediction values ​​generated by different center windows for the same target band to obtain the coarse HR result. The coarse HR result is then input into the spectral reconstruction decoder for refinement to obtain the final high-resolution hyperspectral image.

[0014] The present invention also provides a hyperspectral image super-resolution device based on iterative reconstruction and spectral refinement, comprising:

[0015] Extraction unit, used to acquire low-resolution hyperspectral images The adjacent three-band sliding windows are extracted according to the preset spectral dimensions, and the adjacent three-band sliding windows are input into the head adapter to obtain pseudo-RGB three-channel features.

[0016] The adaptation unit is used to freeze the parameters of the pre-trained model, adapt the pre-trained model using LoRA, input the pseudo-RGB three-channel features into the adapted pre-trained model to obtain the corresponding three bands, and input the three bands into the tail adapter to obtain the predicted values ​​corresponding to the three bands respectively. The three bands are three different bands.

[0017] The image generation unit is used to process the three predicted values ​​generated by different center windows for the same target band using the spectral window weighted fusion module to obtain the coarse HR result. The coarse HR result is then input into the spectral reconstruction decoder for refinement to obtain the final high-resolution hyperspectral image.

[0018] In summary, this invention addresses the long-standing spatial-spectral dilemma in the field of hyperspectral image super-resolution, as well as the core pain points such as high transfer costs of pre-trained large models, gradient breakage in staged processes, and lack of explicit constraints on spectral consistency. It proposes a complete technical solution: "sliding three-neighbor window + LoRA low-rank adaptation + spectral window weighted fusion + end-to-end spectral reconstruction." By sliding a three-band cube along a spectral dimension step of 1, and injecting only a small number of LoRA increments into the frozen IPT (Image Processing Transformer) / SwinIR (Swin Transformer for Image Restoration) natural image pre-trained backbone, adaptive fine-tuning in the hyperspectral domain can be achieved, avoiding overfitting and computational disaster caused by full fine-tuning. For the same target band... The predicted values ​​generated by the three types of center windows are normalized and fused using "spectral position window weight × neighboring window consistency confidence" to obtain a coarse super-resolution result. Then, the spectral reconstruction decoder of the cascaded Spectral-Spatial Block performs spatial-spectral joint refinement under the joint supervision of SAM (Spectral Angle Mapper) spectral angle loss, and finally outputs a high-resolution hyperspectral cube that preserves the original radiometric relationship.

[0019] The entire process is end-to-end achievable, requiring only one forward pass for inference and no external RGB assistance. This significantly reduces model size, training time, and GPU memory usage. On multiple public datasets, its spatial detail and spectral fidelity metrics outperform existing full-scale fine-tuning and multi-stage solutions. It can be directly deployed on satellites, airborne, or edge terminals, providing a high-spectral super-resolution solution that balances efficiency, accuracy, and feasibility for applications such as quantitative remote sensing, precision agriculture, environmental monitoring, and mineral mapping. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating the hyperspectral image super-resolution method based on iterative reconstruction and spectral refinement provided in the first embodiment of the present invention.

[0021] Figure 2 This is an overall framework diagram of the hyperspectral image super-resolution method based on iterative reconstruction and spectral refinement provided in the embodiments of the present invention.

[0022] Figure 3 This is a schematic diagram of a coarse super-resolution using a sliding window in the spectral dimension provided in an embodiment of the present invention.

[0023] Figure 4 This is a schematic diagram of the spectral reconstruction decoder provided in an embodiment of the present invention.

[0024] Figure 5 This is a schematic diagram of a module of a hyperspectral image super-resolution device based on iterative reconstruction and spectral refinement provided in the second embodiment of the present invention. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0026] refer to Figure 1 , Figure 2 As shown, the first embodiment of the present invention discloses a hyperspectral image super-resolution method based on iterative reconstruction and spectral refinement, which can be executed by a hyperspectral image super-resolution device based on iterative reconstruction and spectral refinement (hereinafter referred to as the super-resolution device), specifically, by one or more processors within the super-resolution device, to implement the following method:

[0027] S1, Acquire low-resolution hyperspectral image The adjacent three-band sliding windows are extracted according to the preset spectral dimensions, and the adjacent three-band sliding windows are input into the head adapter to obtain pseudo-RGB three-channel features.

[0028] Specifically, step S1 further includes: acquiring low-resolution hyperspectral images. And according to the preset spectral dimensions, with a step size of 1, extract the sliding window of three adjacent bands. , among which, when At that time, the sliding window for three adjacent bands is ,when At that time, the sliding window for three adjacent bands is , This represents a specific band index in a hyperspectral image. For real numbers, For the length of low-resolution hyperspectral images, For the width of low-resolution hyperspectral images, This refers to the number of bands in a low-resolution hyperspectral image. Band Index The corresponding low-resolution hyperspectral image, Band Index The corresponding low-resolution hyperspectral image, Band Index The corresponding low-resolution hyperspectral image;

[0029] The adjacent three-band sliding window is input into the head adapter to map pseudo-RGB three-channel features. The head adapter consists of a Conv block and a GELU activation function.

[0030] In this embodiment, using a low-resolution hyperspectral image X as input, a sliding window of three adjacent bands is first extracted according to a preset spectral dimension and a step size of 1. This sliding method ensures that the first and last bands are also enveloped by contextual information, avoiding spectral jumps at the boundaries. The sliding windows of the three adjacent bands are fed into the same head adapter. This adapter consists of a Conv block and a GELU activation function, with weights initialized with a small perturbation of zero mean; GELU follows the standard form, smoothly retaining negative values, balancing nonlinearity and gradient fluidity. After this mapping, each set of three-band data is converted into pseudo-RGB three-channel features, with the spatial resolution unchanged and the spectral dimension compressed by 3, providing an input distribution consistent with natural images for the subsequently frozen IPT / SwinIR backbone, thus allowing direct reuse of its spatial prior without modifying the pre-trained weights.

[0031] The addition of an adapter simply involves making minor adjustments to the data to adapt it for input to the IPT; the initial reconstruction is performed by the pre-trained (IPT) model.

[0032] Please see Figure 3 S2, freeze the parameters of the pre-trained model, adapt the pre-trained model using LoRA, input the pseudo-RGB three-channel features into the adapted pre-trained model to obtain the corresponding three bands, and input the three bands into the tail adapter to obtain the predicted values ​​corresponding to the three bands respectively. The three bands are three different bands.

[0033] Specifically, step S2 further includes: in the pre-trained linear layer The LoRA adapter is applied, and its formula is as follows: R L Among them, the original parameters Frozen, parameters , Training can be conducted. The original weight matrix for pre-training. It is the right matrix of the low-rank matrix. The left matrix of the low-rank matrix. The output vector of the linear layer. The input vector of the linear layer, For the output dimension, For the input dimension, For scaling, For rank, classically it is taken as 4~16, and zero-perturbation initialization makes the initial... ;

[0034] Low-rank incremental adaptation is performed on the Q-matrix, K-matrix, or V-matrix of the pre-trained model, the attention output, and the linear layer of the MLP. The number of parameters after low-rank decomposition is only a fraction of the total number of parameters. Its parameter count is far lower than the original number. .

[0035] The pseudo-RGB three-channel features are input into the adapted pre-trained model to obtain the corresponding three bands;

[0036] The three bands are input into the tail adapter and mapped to obtain a high-resolution HSI output. ,in, Indicates the band The predicted value for the target band ,by , , When inputting separately around the center, the target band is obtained. The three predicted values , , ,in, For band Target band spectral sliding window centered on the target band The predicted value, For band Target band spectral sliding window centered on the target band The predicted value, For band Target band spectral sliding window centered on the target band The predicted value is obtained by using a tail adapter consisting of a Conv block and a GELU activation function.

[0037] In this embodiment, PEFT (Parameter-Efficient Fine-Tuning) uses LoRA to perform low-rank incremental adaptation of the Q / K / V and attention output of IPT / SwinIR with the linear layer of MLP, freezes the original weights, and trains only the increment.

[0038] Specifically, the pre-trained model uses IPT (Image Processing Transformer) or SwinIR, which have been pre-trained in natural image super-resolution tasks, and all their original parameters are used. Frozen and not involved in gradient updates. LoRA low-rank adaptation is applied to the Q-matrix, K-matrix, V-matrix, attention output projection, and MLP linear layers of the pre-trained model. Specifically, for any pre-trained linear layer, while maintaining the original... Add trainable parameters while keeping them unchanged. and The network output is consistent with the pre-training state. This bypass is applied to the Q, K, and V projection matrices of the IPT, the attention output projection, and two MLP dimensionality-upgrading-down linear layers, forming a "layer-by-layer low-rank incremental" structure. The number of parameters is reduced from the original... Reduced to .

[0039] The pseudo-RGB three-channel features are stacked along the batch dimension and then fed into the adapted IPT. Since the initial weight state is consistent with the pre-training, the first batch of forward results already possesses rich prior knowledge of natural image textures. The IPT output remains three-channel, but the spatial size is enlarged, and the channel order corresponds one-to-one with the input pseudo-RGB. This three-channel result then enters the tail adapter: symmetrical to the head, the tail adapter is also composed of Conv+GELU, and the weight initialization strategy is consistent with the head. Its function is to remap the "pseudo-RGB" high-resolution features output by the IPT back to the hyperspectral three bands, obtaining a high-resolution cubic fragment corresponding to the center wavelength of the input sliding window. Because the sliding windows overlap in the spectral dimension, the same target band... Three independent predictions will be received. The three predictions are independent of each other, but share the same set of IPT and tail adapter weights, which avoids introducing additional parameters and provides complementary information for subsequent fusion through different spectral contexts.

[0040] Please see Figure 4 S3, the three predicted values ​​are processed by the spectral window weighted fusion module to obtain the coarse HR result, and the coarse HR result is input into the spectral reconstruction decoder for refinement to obtain the final high-resolution hyperspectral image.

[0041] Specifically, step S3 further includes: using a spectral window weighted fusion module to fuse the three predicted values ​​to obtain a coarse HR result for the entire image. ;

[0042] Among them, the target band In coordinates Rough HR results The formula is , , , , For spectral sliding window , , , Extremely small amount It is the standard deviation parameter of the Gaussian function. For pixel coordinates, For the first A band-centered spectral sliding window for the target band In coordinates The predicted pixel value at that location, The value is , , It is important to note that M k (t,i,j) Representing pixel coordinates, M can be represented for the entire image. k (t).

[0043] The coarse HR result is input into the spectral reconstruction decoder for refinement to obtain the final high-resolution hyperspectral image. Specifically, the coarse HR result is input into the spectral reconstruction decoder for refinement. The image is input into a spectral reconstruction decoder for joint spatial-spectral refinement, resulting in the final high-resolution hyperspectral image. , SRD stands for spectral reconstruction decoder, which consists of a set of cascaded SSBs; specifically, it includes 1. spectral convolution: kernel 1. Expanding the spectral receptive field; 2. Spatial convolution: kernel 3. Depthwise separable spatial convolution; 4. Pointwise blending: Point convolution fusion channel.

[0044] In this embodiment, after three independent predictions, a two-stage refinement process of "spectral window weighted fusion - spectral reconstruction" is initiated. The purpose is both to smooth out minor spectral fluctuations caused by different window centers and to further amplify spatial details while maintaining radiometric consistency. Then, the fusion output is performed according to the formula to obtain the coarse HR result. In the spectral window weighted fusion, the confidence weight is obtained from one or more of the neighbor window consistency factors and normalized together with the spectral position window weight.

[0045] Subsequently, the coarse HR result is input into the spectral reconstruction decoder (SRD) for joint spatial-spectral refinement, yielding the final high-resolution hyperspectral image. Through cascaded processing by the SSB, the coarse HR result is further improved in terms of spatial detail and spectral continuity, and the output is the final high-resolution hyperspectral image, completing the coarse-to-fine super-resolution reconstruction.

[0046] In this invention, the same target band is reconstructed multiple times using a spectral sliding window input, and then normalized and fused using "spectral position window weight × confidence level", and finally the HR is obtained through spectral reconstruction and constraint.

[0047] Preferably, it further includes: the pre-trained model and the spectral reconstruction decoder are subject to spectral consistency constraints during training and fine-tuning, including at least one of SAM constraints or spectral smoothing regularization; and end-to-end joint optimization is employed, with the corresponding loss function formula being: , , The loss weight is set to 0.01. For true hyperspectral images, It is an L1 norm. For H×W, The length of the image, The width of the image, For position The predicted spectral vector at that location, For position The true spectral vector at that location, It is an L2 norm.

[0048] Specifically, in this embodiment, under the premise that the overall network is differentiable, the pre-trained model and the spectral reconstruction decoder are simultaneously fine-tuned, rather than trained in stages. The fine-tuning stage introduces at least spectral consistency constraints to curb the inter-band shape drift common in hyperspectral images. In practice, in addition to the pixel-level L1 loss, SAM (SpectralAngle Mapper) constraints are added; spectral smoothing regularization can also be superimposed depending on the data characteristics, but the existence of the SAM term is guaranteed at a minimum, thereby forming an end-to-end joint optimization objective.

[0049] In summary, this invention transforms the 100-band hyperspectral super-resolution problem into an end-to-end pipeline of three-band sliding, low-rank increment, spectral weighted fusion, and joint consistency constraints. Without altering the original pre-trained weights, it seamlessly injects natural image priors into the hyperspectral domain via LoRA bypass, achieving domain adaptation by compressing the number of parameters to an extremely low proportion of the backbone network. The three-neighbor window sliding mechanism provides three complementary predictions for the same target band, which are then normalized and fused using spectral position weights multiplied by neighbor window consistency weights. This suppresses outliers while preserving edge sharpness, resulting in a smooth spectral curve in the coarse result. Subsequently, a cascaded SSB decoder further refines the image through alternating spectral-spatial convolutions, supervised throughout by SAM angular loss, ensuring the radiation shape matches the ground truth. The entire scheme completes inference in a single forward pass, requiring no external RGB assistance, significantly reducing training and deployment costs. It simultaneously outperforms traditional full-scale fine-tuning and multi-stage methods in both spatial detail and spectral fidelity, providing a feasible, controllable, and scalable technical path for real-time acquisition of high-quality hyperspectral images on satellites, airborne systems, or at the edge.

[0050] Please see Figure 5 A second embodiment of the present invention provides a hyperspectral image super-resolution device based on iterative reconstruction and spectral refinement, comprising:

[0051] Extraction unit 101 is used to acquire low-resolution hyperspectral images. The adjacent three-band sliding windows are extracted according to the preset spectral dimensions, and the adjacent three-band sliding windows are input into the head adapter to obtain pseudo-RGB three-channel features.

[0052] The adaptation unit 102 is used to freeze the parameters of the pre-trained model. The pre-trained model is adapted using LoRA, and the pseudo-RGB three-channel features are input into the adapted pre-trained model to obtain the corresponding three bands. The three bands are then input into the tail adapter to obtain the predicted values ​​corresponding to the three bands respectively. The three bands are three different bands.

[0053] The image generation unit 103 is used to process the three prediction values ​​generated by different center windows for the same target band using the spectral window weighted fusion module to obtain the coarse HR result, and input the coarse HR result into the spectral reconstruction decoder for refinement to obtain the final high-resolution hyperspectral image.

[0054] 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 image super-resolution method based on iterative reconstruction and spectral refinement, characterized in that, include: Acquiring low-resolution hyperspectral images The system extracts adjacent three-band sliding windows according to a preset spectral dimension, and inputs these windows into the head adapter to obtain pseudo-RGB three-channel features, specifically: Acquiring low-resolution hyperspectral images And according to the preset spectral dimensions, with a step size of 1, extract the sliding window of three adjacent bands. , among which, when At that time, the sliding window for three adjacent bands is ,when At that time, the sliding window for three adjacent bands is , This represents a specific band index in a hyperspectral image. For real numbers, For the length of low-resolution hyperspectral images, For the width of low-resolution hyperspectral images, This represents the number of bands in a low-resolution hyperspectral image. Band Index The corresponding low-resolution hyperspectral image, Band Index The corresponding low-resolution hyperspectral image, Band Index The corresponding low-resolution hyperspectral image; The adjacent three-band sliding window is input into the head adapter to map pseudo-RGB three-channel features. The head adapter consists of a Conv block and a GELU activation function. The parameters of the pre-trained model are frozen, and the pre-trained model is adapted using LoRA. The pseudo-RGB three-channel features are input into the adapted pre-trained model to obtain the corresponding three bands. The three bands are then input into the tail adapter to obtain the predicted values ​​corresponding to the three bands respectively. The three bands are three different bands. The spectral window weighted fusion module is used to process the three prediction values ​​generated by different center windows for the same target band to obtain the coarse HR result. The coarse HR result is then input into the spectral reconstruction decoder for refinement to obtain the final high-resolution hyperspectral image.

2. The hyperspectral image super-resolution method based on iterative reconstruction and spectral refinement according to claim 1, characterized in that, Freeze the parameters of the pre-trained model and adapt the pre-trained model using LoRA, specifically as follows: In the pre-trained linear layer The LoRA adapter is applied, and its formula is as follows: R L Among them, the original parameters Frozen, parameters , Training can be conducted. The original weight matrix for pre-training. It is the right matrix of the low-rank matrix. The left matrix of the low-rank matrix. The output vector of the linear layer. The input vector of the linear layer, For the output dimension, For the input dimension, For scaling, For rank; Low-rank incremental adaptation is performed on the Q-matrix, K-matrix, or V-matrix of the pre-trained model, the attention output, and the linear layer of the MLP. The number of parameters after low-rank decomposition is only a fraction of the total number of parameters. Its parameter count is far lower than the original number. .

3. The hyperspectral image super-resolution method based on iterative reconstruction and spectral refinement according to claim 1, characterized in that, The pseudo-RGB three-channel features are input into the adapted pre-trained model to obtain the corresponding three bands. These three bands are then input into the tail adapter to obtain the predicted values ​​corresponding to these three bands. Specifically, the three bands are three different bands: The pseudo-RGB three-channel features are input into the adapted pre-trained model to obtain the corresponding three bands; The three bands are input into the tail adapter and mapped to obtain a high-resolution HSI output. ,in, Indicates the band The predicted value for the target band ,by , , When inputting separately around the center, the target band is obtained. The three predicted values , , ,in, For band Centered spectral sliding window for target band The predicted value, For band Centered spectral sliding window for target band The predicted value, For band Centered spectral sliding window for target band The predicted value is obtained by using a tail adapter consisting of a Conv block and a GELU activation function.

4. The hyperspectral image super-resolution method based on iterative reconstruction and spectral refinement according to claim 1, characterized in that, The spectral window weighted fusion module is used to process the three predicted values ​​of the same target band generated by different center windows to obtain the coarse HR result, as follows: A spectral window weighted fusion module is used to fuse the three predicted values ​​to obtain a coarse HR result for the entire image. ; Among them, the target band In coordinates Rough HR results The formula is , , , , For spectral sliding window , , , Extremely small amount It is the standard deviation parameter of the Gaussian function. () represents pixel coordinates. For the first A band-centered spectral sliding window for the target band In coordinates The predicted pixel value at that location, The value is , , .

5. The hyperspectral image super-resolution method based on iterative reconstruction and spectral refinement according to claim 1, characterized in that, The coarse HR result is input into the spectral reconstruction decoder for refinement to obtain the final high-resolution hyperspectral image. Specifically, the coarse HR result is input into the spectral reconstruction decoder for refinement. The image is input into a spectral reconstruction decoder for joint spatial-spectral refinement, resulting in the final high-resolution hyperspectral image. , SRD stands for spectral reconstruction decoder, which consists of a set of cascaded SSBs.

6. The hyperspectral image super-resolution method based on iterative reconstruction and spectral refinement according to claim 5, characterized in that, Also includes: During training and fine-tuning, the pre-trained model and the spectral reconstruction decoder are subject to spectral consistency constraints, including at least one of the SAM constraint or spectral smoothing regularization. And it adopts end-to-end joint optimization, the corresponding loss function formula is as follows: , , To lose weight, For true hyperspectral images, It is an L1 norm. For H×W, The length of the image, The width of the image, For position The predicted spectral vector at that location, For position The true spectral vector at that location, It is an L2 norm.

7. A hyperspectral image super-resolution device based on iterative reconstruction and spectral refinement, characterized in that, include: Extraction unit, used to acquire low-resolution hyperspectral images The system extracts adjacent three-band sliding windows according to a preset spectral dimension, and inputs these windows into the head adapter to obtain pseudo-RGB three-channel features, specifically: Acquiring low-resolution hyperspectral images And according to the preset spectral dimensions, with a step size of 1, extract the sliding window of three adjacent bands. , among which, when At that time, the sliding window for three adjacent bands is ,when At that time, the sliding window for three adjacent bands is , This represents a specific band index in a hyperspectral image. For real numbers, For the length of low-resolution hyperspectral images, For the width of low-resolution hyperspectral images, This represents the number of bands in a low-resolution hyperspectral image. Band Index The corresponding low-resolution hyperspectral image, Band Index The corresponding low-resolution hyperspectral image, Band Index The corresponding low-resolution hyperspectral image; The adjacent three-band sliding window is input into the head adapter to map pseudo-RGB three-channel features. The head adapter consists of a Conv block and a GELU activation function. The adaptation unit is used to freeze the parameters of the pre-trained model, adapt the pre-trained model using LoRA, input the pseudo-RGB three-channel features into the adapted pre-trained model to obtain the corresponding three bands, and input the three bands into the tail adapter to obtain the predicted values ​​corresponding to the three bands respectively. The three bands are three different bands. The image generation unit is used to process the three predicted values ​​generated by different center windows for the same target band using the spectral window weighted fusion module to obtain the coarse HR result. The coarse HR result is then input into the spectral reconstruction decoder for refinement to obtain the final high-resolution hyperspectral image.

Images