Hyperspectral image super-resolution method and system based on natural image prior

Abstract

This invention proposes a hyperspectral image super-resolution method and system based on natural image priors, belonging to the field of hyperspectral image processing technology. First, a spatial texture prior model is pre-trained using a large-scale natural image dataset. Then, the spectral response function and a spatially guided fusion network are jointly optimized. During optimization, the parameters of the spatial texture prior model are frozen, focusing on learning spectral correlations. The parameters of the spatial texture prior model are then unfrozen, and the network is jointly optimized while learning spatial-spectral correlation knowledge to obtain the final spatially guided super-resolution network parameters. The learned network parameters are then used to process the input image to obtain a high-resolution hyperspectral image after super-resolution. The system includes a spectral response function module, a spatial texture prior module, and a spatially guided fusion module. This invention effectively solves the problems of poor restoration results and the inability to reuse spatial texture knowledge in existing methods, and can provide higher-quality hyperspectral image data.

Description

Technical Field

[0001] This invention relates to a hyperspectral image super-resolution method and system, specifically a hyperspectral image super-resolution method and system based on natural image priors, belonging to the field of hyperspectral image processing technology. Background Technology

[0002] Hyperspectral images can capture image data of tens to hundreds of narrow bands on the surface of an object, providing a wider range of spectral information from visible light to near-infrared and far-infrared bands. Therefore, hyperspectral images are widely used in many fields such as remote sensing, agriculture, medical diagnosis, and environmental monitoring, providing a more accurate and detailed data analysis foundation for various applications.

[0003] Hyperspectral imaging technology has developed rapidly, with common methods including grating-based beam splitting, acousto-optic tunable filtering beam splitting, prism beam splitting, and chip coating. However, due to the physical limitations of sensors, existing hyperspectral imaging devices cannot simultaneously achieve both spatial and spectral resolution; many cameras inevitably involve a trade-off between the two. To capture more spectral information, some hyperspectral imaging systems sacrifice temporal and spatial resolution to obtain higher spectral resolution, but the lower spatial resolution also limits the application of hyperspectral images in downstream tasks.

[0004] To address this issue, hyperspectral image super-resolution technology has received widespread attention. The aim of this technology is to recover hyperspectral images with higher spatial resolution from hyperspectral images with lower spatial resolution. This technology can compensate for the shortcomings of existing hyperspectral imaging equipment and obtain higher-quality hyperspectral images.

[0005] Based on whether additional spatial image information is used, existing hyperspectral image super-resolution techniques can be divided into fusion-based methods and single-image-based methods. Fusion-based methods generally achieve better results and support greater resolution upscaling due to the introduction of additional spatial information. The main drawback of these methods is that they require one or more additional high-resolution cameras to capture the extra spatial information and introduce new alignment coupling problems between different cameras. Single-image-based methods recover spatial resolution from a single hyperspectral image. These methods have lower hardware requirements and are applicable to various types of hyperspectral imaging equipment, thus having wider applications in real-world scenarios.

[0006] Traditional single-image super-resolution methods are mainly based on dictionary learning and low-rank decomposition, which rely heavily on hand-crafted priors. However, with increasing demands for algorithm performance, these methods require increasingly longer time to create hand-crafted priors and can no longer meet the requirements for algorithm time and performance.

[0007] In recent years, the rapid development of deep learning has brought new possibilities to hyperspectral image super-resolution methods. Deep learning, a branch of machine learning, uses multi-layered neural network models to learn from data and extract features. Through multi-layered nonlinear transformations, it learns more abstract and high-level representations from the original training data. Deep learning has achieved significant breakthroughs and applications in many fields, such as computer vision, speech recognition, natural language processing, and recommender systems. Recent research has demonstrated the application potential of deep learning in computer vision fields such as image and video processing, image classification, and object detection. Deep learning technology is continuously driving improvements in the efficiency and quality of hyperspectral imaging.

[0008] Currently, existing deep learning-based hyperspectral image super-resolution methods primarily learn spatial correlations and spectral similarities simultaneously from limited hyperspectral image datasets. Since super-resolution is a typical unrestricted problem, these methods are largely limited by the size of the dataset. However, collecting large-scale spectral image datasets is time-consuming and resource-intensive, and large-scale hyperspectral image datasets remain scarce.

[0009] Recently, some researchers have noticed this problem and attempted to use additional large-scale color image datasets during the training phase to help the network learn spatial correlations. A representative work is the "DeepShare" framework proposed by Li et al. at ETH Zurich in 2022, which improves the performance of hyperspectral image super-resolution by using color image super-resolution as an auxiliary task. Although such methods can learn spatial correlation information from color images, directly using color images for training greatly increases the training time of the model and cannot reuse the learned spatial correlation knowledge.

[0010] Therefore, from the perspective of improving the super-resolution effect of hyperspectral images and optimizing the design of algorithm frameworks, there is an urgent need for a method that can fully utilize the prior information of spatial texture in color images to assist in the super-resolution of hyperspectral images, and a hyperspectral image super-resolution method and system that can adapt to various downstream tasks such as natural images and remote sensing images. Summary of the Invention

[0011] The purpose of this invention is to address the shortcomings and defects of existing hyperspectral image super-resolution technologies by creatively proposing a hyperspectral image super-resolution method and system based on natural image priors.

[0012] This invention is the first to propose using natural image priors to recover spatial information from hyperspectral images. Under the premise that the algorithm has a fast processing speed and a small number of parameters, it significantly improves the quality of the recovered hyperspectral images. At the same time, combined with a new spectral response function optimization method, it further improves the quality of hyperspectral image recovery.

[0013] This method includes a training phase and an application phase.

[0014] During the training phase, the parameter distribution of the network is trained using a low-resolution / high-resolution paired hyperspectral image dataset and a large-scale color image dataset, and the learned network parameters are saved. The spectral response function is jointly optimized during training to simultaneously synthesize the best image for both the spatial texture prior network and the spatially guided super-resolution network. The spatial texture prior network employs a state-of-the-art color super-resolution network model and is pre-trained on a large-scale color image dataset to learn spatial texture prior information.

[0015] A spatial texture prior network is used to restore the spatial resolution of hyperspectral images, and a spectral response function is introduced for joint optimization to improve the quality of the restored hyperspectral images. The model first uses the spectral response function to transform the hyperspectral image into a natural image domain, then uses the spatial texture prior network to restore the spatial information of the hyperspectral image, and finally uses a spatially guided super-resolution network to fuse the restored spatial information and the original spectral information to restore the hyperspectral image.

[0016] The system implementation process includes both training and usage phases. This allows for more effective transfer of prior knowledge of spatial texture from color images to recover spatial information from hyperspectral images. The spatially guided super-resolution network is designed to be simple and effective, focusing on learning spectral correlation modeling. Using a more advanced network as the spatial texture prior model can improve the system's performance without any further modifications. During the usage phase, a network model is created based on the model structure and parameters saved during training. The acquired raw hyperspectral images are input into the model for computation, and the recovered high-resolution hyperspectral images are stored. If real high-resolution images exist in the image data used, the system's built-in evaluation method can be used to assess the quality of the recovered images to determine the effectiveness of the model's image recovery.

[0017] To achieve the above objectives, the present invention employs the following technical solutions.

[0018] On the one hand, this invention proposes a hyperspectral image super-resolution method based on natural image priors, comprising the following steps:

[0019] Step 1: Pre-train a spatial texture prior model using a large-scale natural image dataset.

[0020] Among them, spatial texture prior models can employ various different color image super-resolution models. More advanced models have better spatial texture recovery capabilities.

[0021] Step 2: Jointly optimize the spectral response function and the spatially guided fusion network, freezing the parameters of the spatial texture prior model during optimization, and focusing on learning spectral correlations.

[0022] Step 3: Unfreeze the parameters of the spatial texture prior model, jointly optimize the network and learn spatial-spectral correlation knowledge to obtain the final spatially guided super-resolution network parameters.

[0023] Step 4: Using the three learned network parameters, process the input image to obtain a high-resolution hyperspectral image after super-resolution.

[0024] On the other hand, this invention proposes a hyperspectral image super-resolution system based on natural image priors, including a spectral response function module, a spatial texture prior module, and a spatial guided fusion module.

[0025] The spectral response function module is used to fuse spatial information from hyperspectral images to synthesize natural images.

[0026] The spatial texture prior module is used to recover the spatial texture information of hyperspectral images.

[0027] The spatially guided fusion module is used to fuse and reconstruct the final high-resolution hyperspectral image.

[0028] The connection relationships between the above-mentioned components are as follows:

[0029] The output of the spectral response function module is connected to the input of the spatial texture prior module, and the output of the spatial texture prior module is connected to the input of the spatial guided fusion module.

[0030] Beneficial effects

[0031] Compared with the prior art, the present invention has the following advantages:

[0032] This invention introduces natural image priors for hyperspectral image super-resolution, effectively solving problems such as poor restoration results and the inability to reuse spatial texture knowledge in existing hyperspectral image super-resolution methods. It provides higher-quality hyperspectral image data for applications in remote sensing, agriculture, medical diagnosis, and environmental monitoring. Simultaneously, it offers a new paradigm for hyperspectral super-resolution algorithms and a novel approach to the field of hyperspectral image restoration. Attached Figure Description

[0033] Figure 1 This is a general framework diagram of the method of the present invention;

[0034] Figure 2 This is a schematic diagram of the model training process described in the method of the present invention;

[0035] Figure 3This is a diagram of the spatially guided fusion network structure proposed in this invention;

[0036] Figure 4 This is a diagram illustrating the effect of the system of the present invention. Detailed Implementation

[0037] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments.

[0038] like Figure 1 As shown, this method utilizes large-scale natural image data to train and obtain a high-performance spatial texture prior, which is used to recover the spatial texture information of hyperspectral images. This is beneficial for addressing the problem of insufficient hyperspectral training data.

[0039] like Figure 2 As shown, a hyperspectral image super-resolution method based on natural image priors includes the following steps:

[0040] Step 1: Pre-train a spatial texture prior model using a large-scale natural image dataset.

[0041] Specifically, it includes the following steps:

[0042] First, select the network results of the spatial texture prior model.

[0043] Currently, commonly used networks are mainly based on convolutional neural networks (CNNs) and self-attention-based Transformers, each with its own characteristics. Due to the spatial translation invariance of CNNs, they are easier to train, and their model inference is more suitable for hardware-specific optimization. Self-attention-based Transformers, on the other hand, have a more flexible structure, thus requiring more training data and achieving better results. Different performance levels of spatial texture prior models will affect the effectiveness of the proposed method; using a higher-performance model can significantly improve the quality of the super-resolution restored image without any other improvements to the proposed method.

[0044] Let H STPN This represents a spatial texture prior network that takes a low-resolution natural image Y as input. l Restored to a high-resolution natural image Y s The process is represented as:

[0045] Y s =H STPN (Y l (1)

[0046] Spatial texture prior models need to be pre-trained on large-scale color natural image datasets to learn spatial texture correlations in natural images. Natural image datasets are typically much larger and have richer spatial textures than hyperspectral image datasets.

[0047] Then, the paired low-resolution and high-resolution natural images are input into the spatial texture prior network, and the output is the recovered high-resolution image Y. s The L1 loss function is used to constrain the distance between the reconstructed image and the real high-resolution image, and the network parameters are iteratively optimized to narrow the gap between the reconstructed image and the real image, so as to learn prior knowledge of spatial texture.

[0048] Finally, the parameters trained by the spatial texture prior network are stored for use in subsequent training.

[0049] Step 2: Jointly optimize the spectral response function and the spatially guided fusion network, freezing the parameters of the spatial texture prior model during optimization, and focusing on learning spectral correlations.

[0050] Because hyperspectral images capture image information across multiple narrow spectral bands, while color images typically capture information from the red, green, and blue channels, there is a significant domain difference between the two. For example... Figure 3 As shown, the spectral response function can map a multi-channel hyperspectral image into a three-channel natural image, and the generated natural image can be applied to a pre-trained natural image restoration model.

[0051] The spectral response function is modeled as a linear mapping function, implemented as a 1×1 convolution or matrix multiplication. Let X represent the hyperspectral image and Y represent the color natural image, the function is described as:

[0052] Y = RX(2)

[0053] Where R represents the parameter matrix of the spectral response function. The above equation is used to generate low-resolution natural images from low-resolution hyperspectral images.

[0054] The spectral response function has a significant impact on the final fusion and reconstruction process. Current research has explored methods for optimizing the optimal spectral response function for spectral reconstruction or fusion-based spectral super-resolution methods. However, the optimized spectral response curves obtained through these methods will be used to guide the design of hardware cameras. Therefore, additional regularization terms, such as non-negativity constraints and smoothing constraints, are added to the spectral response function. Smoothing constraints have been shown to have a positive effect on the final result, while non-negativity constraints, due to excessive reduction in the resolution space, can lead to a decline in performance. Furthermore, research on the human eye's color perception curve indicates that the human eye exhibits negative responses to some colors, reflecting the contribution of negative response values ​​to the true color.

[0055] This method is not intended to guide the design of a hardware camera; therefore, nonnegativity constraints are removed to obtain a larger solution space and better final optimization results. Different spatial texture prior models also have different optimal spectral response functions. Jointly optimizing the spectral response function requires trade-offs among several different objectives. Multiple loss functions are used to describe the final optimization of the spectral response function.

[0056] Specifically, the loss function L corresponding to the spectral response function srf Represented as:

[0057]

[0058] in, Represents a low-resolution hyperspectral image, Y i G represents the corresponding target low-resolution color image, G represents the first-order variational matrix used for smoothing constraints, and N is the number of hyperspectral images used for training.

[0059] The spatially guided fusion network fuses the original low-resolution hyperspectral image and the recovered high-resolution color image to obtain a high-resolution hyperspectral image. Its structure is as follows: Figure 3 As shown, since a high-performance spatial texture prior is used to recover spatial information, this module can be designed to focus more on spectral information modeling. In comparison, the spectral dimension has a lower data dimensionality than the spatial dimension; therefore, the network model can have fewer parameters and requires fewer parameters to train this part of the network.

[0060] Restored Y s It possesses high-quality spatial texture. Therefore, a single convolutional layer is first used to extract texture features F. tex The process is represented as:

[0061] F tex =H tex (Y s (4)

[0062] Among them, Y s H represents the high-resolution natural image recovered from the spatial texture prior. tex This indicates a convolutional layer.

[0063] To enhance the spatial representation of the recovered hyperspectral image, F tex It is shared by all space-guided fusion modules.

[0064] Due to the original hyperspectral image X l With lower resolution, nearest-neighbor upsampling is used to align the spatial resolution of the image while avoiding the introduction of additional upsampling errors. Similarly, a single convolutional layer is used to extract shallow features F0 from the spectral image; this process is described as follows:

[0065] F0 = H SF (X l (5)

[0066] Among them, H SF X represents a convolutional layer. l This indicates a low-resolution hyperspectral image.

[0067] To fuse the recovered high-frequency texture information, a set of spatially guided fusion modules is used to fuse the extracted deep features F1, F2, ..., F K and the recovered texture features F tex The process is described as follows:

[0068] F i =H i SGFB (F i ,F tex (6)

[0069] Among them, H i SGFB This represents the i-th spatial guided fusion module. Since deep features primarily focus on recovering high-frequency information, long links are added to incorporate low-frequency information.

[0070] Finally, a convolutional layer is used to recover the features to obtain the final reconstructed high-resolution hyperspectral image. This process is described as follows:

[0071] X s =H tail (F K +F0)(7)

[0072] Among them, X s H represents the reconstructed high-resolution hyperspectral image. tail F represents the final reconstructed convolutional layer. K This represents the deep features extracted by the Kth layer of the network.

[0073] It should be noted that the spatially guided fusion module is a residual module built upon the spatially guided fusion layer. For a given input feature F... {i,0} This indicates that the i-th spatially guided fusion module uses L spatially guided fusion layers to extract hidden features F. i,1 ,F i,2 ,…,F i,L The process is represented as:

[0074] F i,j =H SGFPi,j (F i,j-1 ,F tex (8)

[0075] Among them, Fi,j H represents the feature extracted by the j-th spatial guidance fusion layer of the i-th spatial guidance fusion module. SGFPI,j This represents the j-th spatial guidance fusion layer of the i-th spatial guidance fusion module.

[0076] Then, a residual connection is added before the input of subsequent modules, which is represented as:

[0077] F i,out =F i,L +F i,0 (9)

[0078] Among them, F i,out F represents the output of the i-th spatial guidance fusion module. i,L This represents the features extracted by the Lth guiding fusion layer of the i-th spatial guiding fusion module.

[0079] The spatially guided fusion layer uses texture features to extract and enhance deep features. First, two convolutional layers are used to transform the texture features to the latent space; this process is represented as follows:

[0080] F hidden =H hidden (F tex (10)

[0081] Among them, F hidden H represents the extracted hidden features. hidden This represents the feature mapping function.

[0082] Then, the deep features and hidden features are concatenated along the feature channels, and two additional convolutional layers are used to fuse the features. This process is represented as follows:

[0083] F fuse =H fuse (stack(F hidden ,F i,j+1 )+F i,j-1 (11)

[0084] Among them, F fuse H represents the features after fusion. fuse This represents the feature fusion function.

[0085] Both of the mapping functions above use 1×1 convolution to focus more on learning spectral dimension features and avoid learning spatial information.

[0086] Finally, spatial information is aggregated using 3×3 convolutions, a process represented as follows:

[0087] F i,j =H out (F fuse (12)

[0088] Among them, F i,j H indicates. out This represents the spatial information aggregation function for output.

[0089] Although the proposed network model contains many modules, these modules are all differentiable, allowing for end-to-end training. The L1 pixel loss effectively penalizes smaller recovery errors and exhibits better training stability; therefore, this loss function is used to evaluate the model's error. pixel Represented as:

[0090]

[0091] Among them, X i X i s These represent the real hyperspectral image and the hyperspectral image recovered by the network, respectively.

[0092] To balance the loss from the spectral response curve and the final reconstruction loss, a coefficient μ is introduced, and the loss function L of the entire algorithm is... total Represented as:

[0093] L total =L srf (R)+μL pixel (14)

[0094] Among them, L srf (R) represents the loss function term of the spectral response function.

[0095] During training, the parameters of the spatial texture prior model are frozen to prevent the forgetting of texture prior information from natural images in the early stages of training. Simultaneously, the model trained in this phase primarily learns spectral correlations, without focusing on learning the spatial information of the hyperspectral image, thus validating the effectiveness of the proposed transfer of natural image priors for hyperspectral image super-resolution.

[0096] Step 3: Unfreeze the parameters of the spatial texture prior model, jointly optimize the network and learn spatial-spectral correlation knowledge to obtain the final trained parameters of the spatially guided super-resolution network.

[0097] The goal of step 3 is to jointly learn spatial spectral correlation knowledge simultaneously. Since hyperspectral images and natural images may capture completely different scenes, such as remotely sensed images and natural images, there are differences in texture distribution between them. Therefore, it is necessary to jointly learn and optimize the network's spatial spectral correlation knowledge on hyperspectral datasets.

[0098] Unfreezing the parameters of the spatial texture prior model and continuing to train the model with a small learning rate can further improve the hyperspectral image restoration performance of the proposed method.

[0099] Step 4: Using the three learned network parameters, process the input image to obtain a high-resolution hyperspectral image after super-resolution.

[0100] In this stage, the three learned network parameters are used to reconstruct high-resolution hyperspectral images by feeding the high-resolution remote sensing images into the three network modules. The network parameters are fixed in this stage, eliminating the need for further learning and updating.

[0101] A hyperspectral image super-resolution system based on natural image priors includes a spectral response function module, a spatial texture prior module, and a spatial guided fusion module.

[0102] The spectral response function module is used to fuse spatial information from hyperspectral images to synthesize natural images.

[0103] The spatial texture prior module is used to recover the spatial texture information of hyperspectral images.

[0104] The spatially guided fusion module is used to fuse and reconstruct the final high-resolution hyperspectral image.

[0105] The connection relationships between the above-mentioned components are as follows:

[0106] The output of the spectral response function module is connected to the input of the spatial texture prior module, and the output of the spatial texture prior module is connected to the input of the spatial guided fusion module.

[0107] The output of the spatially guided fusion module is the reconstructed high-resolution hyperspectral image.

[0108] Reconstructed hyperspectral images such as Figure 4 As shown, the quality of the hyperspectral image reconstructed by this method is significantly better than that of the comparison method.

Claims

1. A hyperspectral image super-resolution method based on natural image priors, characterized in that, Includes the following steps: Step 1: Pre-train a spatial texture prior model using a natural image dataset; Step 2: Jointly optimize the spectral response function and the spatially guided fusion network, freezing the parameters of the spatial texture prior model during optimization, and focusing on learning spectral correlations; The spectral response function is modeled as a linear mapping function, which is implemented as follows: Convolution or matrix multiplication; with Represents a hyperspectral image. Representing a color natural image, the function is described as follows: in, The parameter matrix represents the spectral response function, and the above formula is used to generate low-resolution natural images from low-resolution hyperspectral images; Loss function corresponding to the spectral response function Represented as: in, This represents a low-resolution hyperspectral image. This represents the corresponding low-resolution color image of the target. This indicates that the first-order variational matrix is ​​used for smoothing constraints. The number of hyperspectral images used for training; First, a convolutional layer is used to extract texture features. The process is represented as: in, This represents a high-resolution natural image reconstructed from a spatial texture prior. Indicates a convolutional layer; To enhance the spatial representation of the recovered hyperspectral image, Shared by all space-guided fusion modules; Use a single convolutional layer to extract shallow features from the spectral image. The process is described as follows: in, Indicates a convolutional layer. Represents low-resolution hyperspectral images; A set of spatially guided fusion modules is used to fuse the extracted deep features. and the recovered texture features The process is described as follows: in, This represents the i-th spatial guidance fusion module; Finally, a convolutional layer is used to recover the features to obtain the final reconstructed high-resolution hyperspectral image. This process is described as follows: in, This represents the reconstructed high-resolution hyperspectral image. This represents the final reconstructed convolutional layer. Indicates the first The deep features extracted by the layer network; The spatially guided fusion module is a residual module built upon the spatially guided fusion layer; for given input features... Indicates the first Each space guides the fusion module, using Each spatial guide fusion layer extracts hidden features. The process is represented as: in, Indicates the first The first spatial guidance fusion module Features extracted by the spatial guidance fusion layer Indicates the first The first space-guided fusion module A spatial guidance fusion layer; Then, a residual connection is added before the input of subsequent modules, which is represented as: in, Indicates the first The output of the spatial guidance fusion module Indicates the first The first spatial guidance fusion module Features extracted by the guided fusion layer; The spatially guided fusion layer uses texture features to extract and enhance deep features; firstly, two convolutional layers are used to transform the texture features into the hidden space, a process represented as follows: in, This represents the extracted hidden features. Represents the feature mapping function; Then, the deep features and hidden features are concatenated along the feature channels, and two additional convolutional layers are used to fuse the features. This process is represented as follows: in, Indicates the characteristics after fusion. Indicates the feature fusion function; Both of the above mapping functions use convolution; Finally, use Convolution aggregates spatial information; this process can be represented as: in, express; This represents the spatial information aggregation function for the output. Using loss function The error of the model is evaluated by the function expressed as: in, , These represent the real hyperspectral image and the hyperspectral image reconstructed by the network, respectively. Introducing coefficients The loss function of the entire algorithm Represented as: in, The loss function term represents the spectral response function; During training, the parameters of the spatial texture prior model are frozen. At the same time, the model trained in this stage learns spectral correlations and verifies the effectiveness of the proposed transfer of natural image priors for hyperspectral image super-resolution. Step 3: Unfreeze the parameters of the spatial texture prior model, jointly optimize the network and learn spatial-spectral correlation knowledge to obtain the final spatially guided super-resolution network parameters; Step 4: Use the learned network parameters to process the input image to obtain a high-resolution hyperspectral image after super-resolution.

2. The hyperspectral image super-resolution method based on natural image priors as described in claim 1, characterized in that, In step 1: First, select the network result of the spatial texture prior model, and let... This represents a spatial texture prior network, which takes a low-resolution natural image as input. Restored to high-resolution natural images The process is represented as: The spatial texture prior model is pre-trained on a color natural image dataset to learn spatial texture correlation knowledge of natural images; Then, the paired low-resolution and high-resolution natural images are input into the spatial texture prior network, and the output is the recovered high-resolution image. The L1 loss function is used to constrain the distance between the reconstructed image and the real high-resolution image, and the network parameters are iteratively optimized to narrow the gap between the reconstructed image and the real image, and to learn prior knowledge of spatial texture.

3. A hyperspectral image super-resolution system based on natural image priors for implementing the method of claim 1, characterized in that, This includes a spectral response function module, a spatial texture prior module, and a spatial guided fusion module; Among them, the spectral response function module is used to fuse the spatial information of hyperspectral images to synthesize natural images; The spatial texture prior module is used to recover the spatial texture information of hyperspectral images; The spatially guided fusion module is used to fuse and reconstruct the final high-resolution hyperspectral image; The connection relationships between the above-mentioned components are as follows: The output of the spectral response function module is connected to the input of the spatial texture prior module, and the output of the spatial texture prior module is connected to the input of the spatial guided fusion module.

Images