Hyperspectral image blind super-resolution reconstruction method based on element-transfer learning

Abstract

The invention relates to a hyperspectral image blind super-resolution reconstruction method based on meta-transfer learning, which effectively transfers spatial knowledge of a natural image domain to a hyperspectral image domain by utilizing a pre-training model of a large number of natural images, and improves the initial performance of the model. Besides, a meta-transfer learning strategy is adopted, so that the model can quickly adapt to different degradation forms, and the robustness and generalization ability in practical application are enhanced. Meanwhile, through the knowledge edge distillation compression model, the parameter quantity and the calculation overhead of the network are reduced, and the efficiency of the model is improved. And the expression capability of the model on image details is enhanced through the transmission of the multi-stage attention feature map. And finally, the model is finely adjusted by using the block recursion characteristic, so that the precision of high-resolution reconstruction is further improved, and the edges and details of the generated image are clearer.

Description

Technical Field

[0001] The present invention belongs to the field of computer vision image restoration and relates to a hyperspectral image blind super-resolution reconstruction method based on meta-transfer learning. Background Art

[0002] Hyperspectral images have been widely used in remote sensing, environmental monitoring, medical diagnosis and other fields due to their ability to capture a large number of spectral bands. Hyperspectral images can record detailed spectral information for each pixel, making it possible to identify and distinguish tiny spectral features. However, existing spectral imaging systems often need to sacrifice spatial resolution while improving spectral resolution, resulting in hyperspectral images lacking in spatial detail. In addition, complex external environments and equipment limitations during the imaging process may introduce image degradation, further affecting the accurate description of the target object by hyperspectral images. These factors pose great challenges to the accurate interpretation of hyperspectral images.

[0003] In order to overcome this problem, hyperspectral image super-resolution technology came into being. Its core idea is to reconstruct high-resolution images from low-resolution hyperspectral images, thereby improving the image's ability to express spatial details and enhancing the accuracy of the description of object information. According to whether additional auxiliary image information is used, the existing hyperspectral image super-resolution methods can be roughly divided into two categories: auxiliary image-based fusion methods and single hyperspectral image super-resolution methods. Auxiliary image-based fusion methods improve the spatial resolution of hyperspectral images by fusing high-resolution images (such as RGB images or multispectral images) registered with hyperspectral images. However, due to external environmental factors in practical applications, it is often very difficult to obtain auxiliary images that are accurately registered with hyperspectral images, which greatly limits the practicality of such methods. In contrast, the super-resolution method of a single hyperspectral image does not require additional auxiliary information. It achieves super-resolution reconstruction by analyzing and extracting the spectral and spatial features inside the hyperspectral image, so it is more popular in practical applications. The existing single hyperspectral image super-resolution methods are mainly divided into traditional methods and deep learning-based methods. Traditional methods include low-order approximation, dictionary learning, and sparse representation techniques, which use manual priors to capture image features, but often have limited expressive power and are unable to recover complex image details. In recent years, deep learning has made significant progress in the field of image super-resolution, especially in an end-to-end training framework, which effectively improves the image reconstruction performance by implicitly learning the mapping function. However, directly applying deep learning to hyperspectral image super-resolution may lead to spectral distortion because these methods usually ignore the correlation between hyperspectral image bands.

[0004] To this end, in recent years, some emerging network structure designs have begun to take into account both spectral and spatial features, and achieved better reconstruction effects through techniques such as three-dimensional convolution, but this inevitably increases the computational complexity. In addition, these methods often rely on specific degradation models to generate low-resolution images. When the actual degradation model does not match the preset, the reconstruction effect of the model will be significantly reduced, limiting its adaptability and robustness. In order to overcome these limitations, blind super-resolution methods for hyperspectral images have gradually become a research hotspot. Such methods achieve super-resolution reconstruction without relying on specific degradation models. However, most of the existing blind super-resolution methods still rely on high-resolution auxiliary images, such as RGB images or multispectral images, which is difficult to achieve in practical applications. Therefore, the development of lightweight single-frame blind super-resolution models that do not require additional auxiliary images remains an important direction of current research. Summary of the invention

[0005] Technical issues to be solved

[0006] In order to avoid the shortcomings of the prior art, the present invention proposes a blind super-resolution reconstruction method for hyperspectral images based on meta-transfer learning to improve the super-resolution reconstruction capability of hyperspectral images.

[0007] The method of the present invention is a single hyperspectral image blind super-resolution method based on multi-stage transfer learning of knowledge distillation. The main idea is: by multi-stage training of the network, combining the similarity between natural images and HSI, using meta-learning strategy to quickly adapt to new degradation forms, and effectively compressing the model through knowledge distillation technology, the reconstruction performance of hyperspectral images is improved. The method of the present invention includes the following four steps: pre-training the model by using a large amount of low-resolution-high-resolution (LR-HR) data from natural images, transferring the spatial knowledge of the natural image domain to the hyperspectral image domain, and obtaining an initial solution suitable for the hyperspectral image super-resolution task. In order to adapt to the complex situation of different degradation models, a meta-learning mechanism is adopted so that the network can quickly adapt to the new environment through a small number of iterations. On this basis, the knowledge distillation technology is used to compress the model, and the attention feature map (AFM) is extracted from the teacher network to guide the training of the student network, thereby ensuring the performance while reducing the complexity of the model. The generalization ability of the model is further improved by fine-tuning the block recursive characteristics of the network.

[0008] Technical Solution

[0009] A hyperspectral image blind super-resolution reconstruction method based on meta-transfer learning, characterized by the following steps:

[0010] Step 1: Construct an image training set: perform blur processing and data enhancement on high-resolution images to obtain low-resolution LR1 and high-resolution HR1 image pairs:

[0011]

[0012] Where: X H and Y H is the LR1 and HR1 data in the HSI domain, and B represents the number of bands corresponding to the LR1 and HR1 images;

[0013] Step 2: Pre-train network model: Generate LR2 images by bicubic downsampling the HR images of the natural image dataset DIV2K, and form image pairs with the HR2 images:

[0014]

[0015] Among them, X R and Y R denote LR2 and HR2 images respectively, p and P (p<<P) denote the number of pixels of the image;

[0016] The EDSR basic network is trained using image pairs formed by LR2 images and HR2 images to obtain a pre-trained three-channel network model;

[0017] Step 3: Migrate the pre-trained model to HSI SR: H The pseudo-RGB image in is cyclically input into the three-channel network model, and the input order follows the band index order inherent in HSI; then, all reconstructed bands are concatenated along the input spectral dimension to obtain the final SR HIS;

[0018] Step 4: Use meta-training to train the teacher network: H Divide into training set and test set They correspond to the support set and query set in meta-learning respectively;

[0019] In the degradation space p(k) and high-resolution hyperspectral image samples Y H Select n types of degradation for meta-training, and generate corresponding low-resolution images for each degradation type, where the low-resolution images generated by the training set and Matching, test set generation Then match;

[0020] The degradation process of the degenerate space p(k) is expressed as:

[0021] =Y*k)↓ s +N,

[0022] Among them, * represents the fuzzy operation, k is the fuzzy kernel, ↓ srepresents the downsampling operation with a scale factor of s, and N represents the noise added during the degradation process;

[0023] Step 6: Student network distillation: compress the model using knowledge distillation technology; first randomly select a degradation type from the degradation space and apply it to Y H Generate X H , then X H Input into the knowledge distillation network;

[0024] Secondly, convolutional neural networks are used to extract low-level, mid-level, and high-level features of HSI and calculate the spatial attention feature maps of the teacher network and the student network at different stages.

[0025] AFM is expressed as:

[0026]

[0027] Among them, C represents the number of channels in the feature map, Represents the feature graph in the network;

[0028] The overall loss L in the knowledge distillation stage KD for:

[0029]

[0030] Among them, L K3 The output of the student network and Y H The reconstruction loss between them; the distillation loss L K1 To constrain the difference between the attention feature maps of the teacher network and the student network; use the output of the teacher network as a label to constrain the difference between the output of the teacher network and the output of the student network, forming a distillation loss L K2 ; L K3 The output of the student network and Y H The reconstruction loss parameters between them; λ1, λ2, and λ3 are used to balance the importance of these three losses;

[0031] Step 7: Use meta-test to reconstruct the image: H The degraded image X obtained by bicubic downsampling H , generate the downsampled X H ; Then use the downsampled X H and X H Perform n gradient updates on the model parameters; the model fully exploits the internal image information and fine-tunes itself to adapt to the specific degradation; finally, H The fine-tuned model is used to obtain the final SR image F s (X H ;θ).

[0032] The blurring process in step 1 is to blur the high-resolution image using a Gaussian blur kernel.

[0033] The data augmentation in step 1 is to downsample the blurred image using bicubic interpolation to obtain the initial data pair of the image, i.e., the low-resolution and high-resolution image pair. Then, the training set data is augmented by flipping and rotating to obtain the final low-resolution LR and high-resolution HR image pair.

[0034] The includes the B band, and the t-th (1 < t < B) band and its adjacent and bands are regarded as a pseudo RGB image.

[0035] The process of step 3 is expressed as:

[0036]

[0037] In the formula, f(·) and F(·) are the implicit mapping expressions of the natural image and HSI's LR-HR respectively, w represents the weight matrix, and b is the bias term.

[0038] When training the teacher network in step 5, the model parameters of each task are initialized with the initial solution θ0 obtained through pre-training. Subsequently, N iterations are performed on the image pairs in the support set to gradually optimize the training loss function through gradient descent

[0039]

[0040] where F t (·) represents the mapping function of the teacher network, and the task-specific parameter θ is updated accordingly i :

[0041]

[0042] After completing the training on the support set, the model performance is evaluated using the updated task-specific parameters on the image pairs in the query set, and the validation loss is The initial parameters θ0 of the model are updated through the average loss of the query set, and finally the model parameters θ are adjusted by the meta learning rate β

[0043]

[0044] : T :

[0045]

[0046] It enables it to generalize better in future tasks. i is one of the n degradation class types selected from the degradation space, α is the task-level learning rate and β is the meta-learning rate.

[0047] The constraint is the difference between the attention feature maps of the teacher network and the student network. K1 Described as:

[0048]

[0049] in, denotes the attention feature maps of the teacher network and the student network respectively. The index j denotes the position at different stages of the network; M denotes the number of images in a training batch, and e is a hyperparameter with a very small value to prevent zero division errors.

[0050] The distillation loss L constrains the difference between the output of the teacher network and the output of the student network K2 Described as:

[0051]

[0052] Among them, θ T and θ S Represent the model parameters of the teacher network and the student network respectively, F s (·) represents the mapping function of the student network; F t (·) and F s (·) represent the mapping functions corresponding to the teacher network and the student network respectively.

[0053] The output of the student network is H The reconstruction loss L between K3 Described as:

[0054]

[0055] A computer program product, characterized in that it includes computer executable instructions, which are used to implement the hyperspectral image blind super-resolution reconstruction method based on meta-transfer learning when executed.

[0056] Beneficial Effects

[0057] The present invention proposes a blind super-resolution reconstruction method for hyperspectral images based on meta-transfer learning, which adopts a strategy combining meta-transfer learning with knowledge edge distillation to perform blind super-resolution reconstruction on hyperspectral images. By utilizing a large number of pre-trained models of natural images, the spatial knowledge of the natural image domain is effectively transferred to the hyperspectral image domain, thereby improving the initial performance of the model. In addition, the present invention adopts a meta-transfer learning strategy, so that the model can quickly adapt to different forms of degradation, thereby enhancing the robustness and generalization ability in practical applications. At the same time, through the knowledge edge distillation compression model, the number of network parameters and computational overhead are reduced, and the efficiency of the model is improved. The model's ability to express image details is also enhanced through the transfer of multi-stage attention feature maps. Finally, the block recursive feature is used to fine-tune the model, which helps to further improve the accuracy of high-resolution reconstruction, making the generated image clearer at the edges and details. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 Specific flow chart for the implementation of the present invention

[0059] Figure 2 This is the structure diagram of the blind super-resolution model based on meta-transfer learning. DETAILED DESCRIPTION

[0060] The present invention will now be further described with reference to the embodiments and the accompanying drawings:

[0061] The present invention proposes a hyperspectral image blind super-resolution reconstruction method based on meta-transfer learning, which mainly includes the following steps:

[0062] 1. Construct an image training set. The hyperspectral image dataset is divided into a training set and a test set. For the training set, the high-resolution image is first blurred using a Gaussian blur kernel, and then downsampled by bicubic interpolation to obtain the initial training set (low-resolution and high-resolution image pairs). The training set data is then enhanced by flipping and rotating.

[0063] 2. Pre-trained network model. The network is pre-trained by using a large number of low-resolution-high-resolution (LR-HR) image pairs from natural images, and the pre-trained model is applied to the hyperspectral image super-resolution algorithm (HSI SR) through transfer learning. Unlike the traditional HSI SR algorithm, which usually converts RGB images into YCbCr color space and only extracts the brightness channel to train a single-channel super-resolution model, this paper directly uses a three-channel model, which can fully consider the spectral correlation of hyperspectral images;

[0064] 3. Meta-transfer learning adaptation module. To cope with the complexity of different degradation models, the present invention adopts a meta-transfer learning strategy. Through meta-learning, the network can quickly adapt to new degradation forms through a small number of iterations in practical applications, thereby improving the model's adaptability to various real-world scenarios.

[0065] 4. Knowledge distillation model compression module. During the training process, the attention feature map (AFM) in the teacher network is extracted through knowledge distillation technology and used as knowledge to guide the training of the student network. This is not a simple feature map extraction, but multi-scale attention features are extracted from different stages of the network, thereby effectively compressing the model and reducing the computational cost;

[0066] 5. Model fine-tuning and image reconstruction. Inspired by the block recursive characteristics, the model is fine-tuned using low-resolution hyperspectral images (LRHSI) to further improve the reconstruction effect. Through this step, a high-resolution hyperspectral image (SR HSI) is finally generated. After the model training is completed, the low-resolution image of the test is input into the model for blind super-resolution reconstruction and output of high-quality images.

[0067] In order to more clearly illustrate the technical solution implemented by the present invention, the various modules required in the embodiment description are briefly introduced below. Obviously, the accompanying drawings are only flow charts of the present invention. For ordinary technicians in this field, they can also expand based on the accompanying drawings and obtain other accompanying drawings without paying creative labor. Specific implementation method:

[0069] Reference Figure 1 The super-resolution method of the present invention is implemented in the following steps:

[0070] Step 1: Construct an image dataset. First, use a Gaussian blur kernel to blur the high-resolution image, and downsample it through bicubic interpolation to obtain the initial image data pair (low-resolution and high-resolution image pair). Then, enhance the training set data by flipping and rotating to obtain the final low-resolution (LR) and high-resolution (HR) image pair:

[0071]

[0072] Where X H and Y H is the LR and HR data in the HSI domain, and B represents the number of bands corresponding to the LR and HR images.

[0073] Step 2: Pre-training network model. This paper selects a simple network, namely the baseline version of EDSR. The natural image dataset DIV2K is used to pre-train the model, where HR images are generated into LR images by bicubic downsampling to obtain LR and HR image pairs for pre-training, namely

[0074]

[0075] Among them, X R and Y R respectively represent the LR and HR images, and p and P (p << P) represent the number of pixels of the images.

[0076] Generate the LR image by bicubic downsampling of the HR image of the natural image dataset DIV2K, and form an image pair with the HR image:

[0077]

[0078] Among them, X R and Y R respectively represent the LR and HR images, and p and P (p << P) represent the number of pixels of the images.

[0079] Train the EDSR basic network with the LR image and the HR image to form an image pair, and obtain a pre-trained network model;

[0080] Step 3, transfer the pre-trained model to HSI SR. Directly apply the pre-trained three-channel model to HSI SR. Specifically, including the B band, regard the t-th (1 < t < B) band and its adjacent and bands as a pseudo-RGB image. These pseudo-RGB images are cyclically input into the three-channel model, and the input order follows the inherent band index order of HSI. Then, all the reconstructed bands are concatenated along the input spectral dimension to obtain the final SR HSI. This process can be expressed as:

[0081]

[0082] In the formula, f(·) and F(·) are the LR-HR implicit mapping expressions of natural images and HSI respectively, w represents the weight matrix, and b is the bias term.

[0083] Step 4, design the degradation space p(k). The degradation process can be expressed as:

[0084] X = (Y * k)↓ s + N,

[0085] where * represents the blurring operation, k is the blur kernel, ↓ s represents the downsampling operation with a scale factor of s, and N represents the noise added during the degradation process.

[0086] Step 5, use meta-training to train the teacher network. In the degradation space p(k) and the high-resolution hyperspectral image sample YH Select n degradation types for meta-training. First, transform Y H Divide into training set and test set They correspond to the support set and query set in meta-learning, respectively. The support set is used for model training and parameter update, while the query set is used to evaluate the performance and generalization ability of the model. Next, corresponding low-resolution images are generated for each degradation type, where the low-resolution images generated by the training set and Matching, test set generation Then Matching. The model parameters of each task are initialized using the initial solution θ0 obtained through pre-training, and then the image pairs in the support set are N iterations are performed to gradually optimize the training loss function through gradient descent

[0087]

[0088] And use this to update each task-specific parameter θ i :

[0089]

[0090] After training on the support set, these updated task-specific parameters are used to train the query set image pairs. The model performance is evaluated on , and the validation loss is

[0091]

[0092] The initial parameters θ0 of the model are updated by the average loss of the query set, and finally the model parameters θ are adjusted by the meta-learning rate β T :

[0093]

[0094] It enables it to generalize better in future tasks. i is one of the n degradation class types selected from the degradation space, α is the task-level learning rate and β is the meta-learning rate.

[0095] Step 6: Student network distillation. While achieving good performance, the computational burden of large deep learning models must be considered. Therefore, knowledge distillation techniques are introduced to compress the model while maintaining as much performance as possible. First, a degradation type is randomly selected from the degradation space and applied to Y H Generate X H , then X HInput into the knowledge distillation network. Secondly, the convolutional neural network is used to extract the low-level features, mid-level features and high-level features of HSI. On this basis, the spatial attention feature maps of the teacher network and the student network at different stages are also calculated. AFM can be expressed as:

[0096]

[0097] Among them, C represents the number of channels in the feature map, Represents the feature map in the network. Using distillation loss L K1 To constrain the difference between the attention feature maps of the teacher network and the student network, thereby guiding the student network to learn more potential feature extraction capabilities. K1 It can be described as:

[0098]

[0099] in, Represent the attention feature maps of the teacher network and the student network respectively. The index j represents the position at different stages of the network. M represents the number of images in a training batch, and e is a hyperparameter with a very small value to prevent zero division errors. In addition, the output of the teacher network is used as a label to constrain the difference between the output of the teacher network and the output of the student network, forming the distillation loss L K2 This approach guides the student network to learn the implicit feature mapping capabilities of the meta-network. K2 It can be described as:

[0100]

[0101] Among them, θ T and θ S Represent the model parameters of the teacher network and the student network respectively, F t (·) and F s (·) represent the mapping functions corresponding to the teacher network and the student network respectively.

[0102] L K3 The output of the student network and Y H The reconstruction loss between is:

[0103]

[0104] The overall loss L in the knowledge distillation stage KD It can be described as:

[0105]

[0106] Among them, L K3 The output of the student network and Y HThe parameters λ1, λ2, and λ3 are used to balance the importance of these three losses.

[0107] Step 7: Use meta-test to reconstruct the image. H The degraded image X obtained by bicubic downsampling H , and process it in the same way to generate the downsampled X H Then use the downsampled X H and X H Perform n gradient updates on the model parameters. The model can make full use of internal image information and fine-tune itself to adapt to the specific degradation. Finally, in X H The fine-tuned model is used to obtain the final SR image F s (X H The effect of the present invention can be further illustrated by the following simulation experiment.

[0108] 1. Simulation conditions

[0109] The present invention is that the central processing unit is The simulation was performed on an i7-6800K 3.40GHz CPU, NVIDIA GeForce GTX3090GPU, and Ubuntu operating system using Python software and PyTorch deep learning framework. The data used in the simulation are the hyperspectral public datasets CAVE and Harvard. The CAVE dataset contains 32 images, each with a size of 31×145×145; the Harvard dataset contains 71 images, each with a size of 31×1392×1040. 80% of the samples in each dataset are selected as training sets, and the rest are used as test sets.

[0110] 2. Simulation content

[0111] In order to demonstrate the effectiveness of the hyperspectral image blind super-resolution reconstruction method based on meta-transfer learning of knowledge distillation in the present invention, the peak signal-to-noise ratio (PSNR), structural similarity index (SSIM) and spectral angle mapper (SAM) were selected as evaluation criteria, and tests were carried out on three public datasets CAVE, Harvard and Houston-2018. Six typical super-resolution (SR) methods, including Residual Dense Network (RDN), Multi-Scale Convolutional Network (MCNet), Spatial Feature Collaborative Super-Resolution (SFCSR), Coarse Super-Resolution Network (CoarSR), Deep Convolutional Neural Network (DCNN) and Multi-Scale Residual Dense Attention Network (MRDA), were selected as comparison algorithms. The RDN algorithm is introduced in detail in the paper "Y. Zhang, Y. Tian, ​​Y. Kong, B. Zhong and Y. Fu, "Residual dense network for image super-resolution," IEEE / CVF Conf. Comput. Vis. Pattern Recognit., pp. 2472-2481, 2018."; the MCNet algorithm is introduced in detail in the paper "Q. Li, Q. Wang and X. Li, "Mixed 2D / 3D convolutional network for hyperspectral image super-resolution," Remote Sens., vol. 12, no. 10, pp. 1660, 2020."; the SFCSR algorithm is introduced in detail in the paper "Q. Wang, Q. Li, and X. Li, "Hyperspectral image superresolution using spectrum and feature context," IEEE Trans.Ind.Electron.,vol.68,no.11,pp.11276–11285,2021."; the comparison results are shown in Table 1:

[0112] Table 1 Comparison results of different algorithms on CAVE and Harvard datasets

[0113]

[0114] Generally, the larger the values ​​of PSNR and SSIM, the smaller the value of SAM, and the better the reconstructed image. As can be seen from Table 1, on the three data sets, the method proposed in the present invention is significantly better than other algorithms in terms of the three indicators.

Claims

1. A hyperspectral image blind super-resolution reconstruction method based on meta-transfer learning, characterized in that Here are the steps: Step 1: Construct an image training set: blur and enhance the high-resolution image to obtain low-resolution LR1 and high-resolution HR1 image pairs: Where: X H and Y H is the LR1 and HR1 data in the HSI domain, and B represents the number of bands corresponding to the LR1 and HR1 images; Step 2: Pre-train network model: Generate LR2 images by bicubic downsampling the HR images of the natural image dataset DIV2K, and form image pairs with the HR2 images: Among them, X R and Y R respectively represent the LR2 and HR2 images, and p and P (p << P) represent the number of pixels of the images; The EDSR basic network is trained using image pairs formed by LR2 images and HR2 images to obtain a pre-trained three-channel network model; Step 3: Migrate the pre-trained model to HSI SR: H The pseudo-RGB image in is cyclically input into the three-channel network model, and the input order follows the band index order inherent in HSI; then, all reconstructed bands are concatenated along the input spectral dimension to obtain the final SR HIS; Step 4: Use meta-training to train the teacher network: H Divide into training set and test set They correspond to the support set and query set in meta-learning respectively; In the degradation space p(k) and high-resolution hyperspectral image samples Y H Select n types of degradation for meta-training, and generate corresponding low-resolution images for each degradation type, where the low-resolution images generated by the training set and Matching, test set generation Then match; The degradation process of the degenerate space p(k) is expressed as: =Y*k)↓ s +N, Among them, * represents the blur operation, k is the blur kernel, ↓s represents the downsampling operation with a scale factor of s, and N represents the noise added during the degradation process; Step 6: Student network distillation: compress the model using knowledge distillation technology; first randomly select a degradation type from the degradation space and apply it to Y H Generate X H , then X H Input into the knowledge distillation network; Secondly, convolutional neural networks are used to extract low-level, mid-level, and high-level features of HSI and calculate the spatial attention feature maps of the teacher network and the student network at different stages. AFM is expressed as: Among them, C represents the number of channels in the feature map, Represents the feature graph in the network; The overall loss L in the knowledge distillation stage KD for: Among them, L K3 The output of the student network and Y H The reconstruction loss between them; the distillation loss L K1 To constrain the difference between the attention feature maps of the teacher network and the student network; use the output of the teacher network as a label to constrain the difference between the output of the teacher network and the output of the student network, forming a distillation loss L K2 ; L K3 The output of the student network and Y H The reconstruction loss parameters between them; λ1, λ2, and λ3 are used to balance the importance of these three losses; Step 7: Use meta-test to reconstruct the image: H The degraded image X obtained by bicubic downsampling H , generate the downsampled X H ; Then use the downsampled X H and X H Perform n gradient updates on the model parameters; the model fully exploits the internal image information and fine-tunes itself to adapt to the specific degradation; finally, H The fine-tuned model is used to obtain the final SR image F s (X H ;θ).

2. The hyperspectral image blind super-resolution reconstruction method based on meta-transfer learning according to claim 1, characterized in that: The blur processing in step 1 is to blur the high-resolution image using a Gaussian blur kernel.

3. The method for blind super-resolution reconstruction of hyperspectral images based on meta-transfer learning according to claim 1, characterized in that: The data enhancement in step 1 is to downsample the blurred image using bicubic interpolation to obtain the initial data pair of the image, i.e., the low-resolution and high-resolution image pair, and then enhance the training set data by flipping and rotating to obtain the final low-resolution LR and high-resolution HR image pair.

4. The hyperspectral image blind super-resolution reconstruction method based on meta-transfer learning according to claim 1, characterized in that: The said includes the B band, and regards the t-th (1 < t < B) band and its adjacent and bands as a pseudo RGB image.

5. The method for blind super-resolution reconstruction of hyperspectral images based on meta-transfer learning according to claim 1, characterized in that: The process of step 3 is expressed as follows: Where f(·) and F(·) are the LR-HR implicit mapping expressions for natural images and HSI, respectively, w represents the weight matrix, and b is the bias term.

6. The method for blind super-resolution reconstruction of hyperspectral images based on meta-transfer learning according to claim 1, characterized in that: When the teacher network of step 5 is trained, the model parameters of each task are initialized with the initial solution θ0 obtained through pre-training, and then the image pairs of the support set are trained. N iterations are performed to gradually optimize the training loss function through gradient descent Among them, F t (·) represents the mapping function of the teacher network, which is used to update each task-specific parameter θ i : After training on the support set, the updated task-specific parameters are used to train the image pairs in the query set. The model performance is evaluated on , and the validation loss is The initial parameters θ0 of the model are updated by the average loss of the query set, and finally the model parameters θ are adjusted by the meta-learning rate β T : It enables it to generalize better in future tasks. i is one of the n degradation class types selected from the degradation space, α is the task-level learning rate and β is the meta-learning rate.

7. The method for blind super-resolution reconstruction of hyperspectral images based on meta-transfer learning according to claim 1, characterized in that: The constraint is the difference between the attention feature maps of the teacher network and the student network. K1 Described as: in, denotes the attention feature maps of the teacher network and the student network respectively. The index j denotes the position at different stages of the network; M denotes the number of images in a training batch, and e is a hyperparameter with a very small value to prevent zero division errors.

8. The method for blind super-resolution reconstruction of hyperspectral images based on meta-transfer learning according to claim 1, characterized in that: The distillation loss L constrains the difference between the output of the teacher network and the output of the student network K2 Described as: Among them, θ T and θ S Represent the model parameters of the teacher network and the student network respectively, F s (·) represents the mapping function of the student network; F t (·) and F s (·) represent the mapping functions corresponding to the teacher network and the student network respectively.

9. The method for blind super-resolution reconstruction of hyperspectral images based on meta-transfer learning according to claim 1, characterized in that: The output of the student network is H The reconstruction loss L between K3 Described as:

10. A computer program product, characterized in that The invention comprises computer executable instructions, which are used to implement the hyperspectral image blind super-resolution reconstruction method based on meta-transfer learning as described in claims 1 to 9 when being executed.

Images