A remote sensing image reconstruction model construction and training method and system

Abstract

The present invention belongs to the field of image processing technology, and specifically relates to a construction and training method and system of a remote sensing image reconstruction model. The remote sensing image reconstruction model is guided by prior features, extracts local features and global features, and performs adaptive feature fusion to reconstruct low-resolution images, wherein the prior features provide additional information for the structure, texture, and edge of the image, so that the model can more accurately reconstruct the details of the high-resolution image; and when extracting local features and global features, the prior features and the global contextual relationship are fully utilized, and the local features and the global features are fused by adaptive feature fusion processing, thereby enhancing the performance of the model in super-resolution reconstruction tasks.

Description

Technical Field

[0001] The present invention belongs to the technical field of image processing, and in particular relates to a construction and training method and system for a remote sensing image reconstruction model. Background Art

[0002] Remote sensing images are pictures of the earth's surface obtained by sensors on satellites or high-altitude aircraft, which can capture information that cannot be directly observed on the ground. High-resolution remote sensing images play an important role in tasks such as land cover classification, change detection, object recognition, and image segmentation. However, due to the influence of sensor performance and complex environment, the quality of remote sensing images is sometimes poor and cannot fully reflect detailed information. Therefore, remote sensing image super-resolution technology has received widespread attention.

[0003] With the advancement of technology, deep learning methods have been widely used in the super-resolution technology of remote sensing images. They can be mainly divided into three categories: methods based on convolutional neural networks (CNN), methods based on generative adversarial networks (GAN), and methods based on diffusion probability models (DDPM). However, these three deep learning methods have their own shortcomings in the super-resolution technology of remote sensing images. The methods based on convolutional neural networks (CNN) are difficult to capture global structural information because the convolution operation is mainly local perception. The limited receptive field and loss of detail information of multi-layer convolution also limit its performance. The methods based on generative adversarial networks (GAN) face the problems of training instability, difficulty in balancing the generator and the discriminator, and require a lot of computing resources. Although the methods based on diffusion probability models (DDPM) have potential, they have long training time, complex model structure, high consumption of computing resources, and strong dependence on high-quality data. Summary of the invention

[0004] In order to solve the problems raised by the background technology, the present invention provides a method and system for constructing and training a remote sensing image reconstruction model.

[0005] The technical solution of the present invention is as follows:

[0006] The present invention provides a method for constructing and training a remote sensing image reconstruction model, comprising the following steps:

[0007] S1: Obtain high-resolution remote sensing image data, downsample it to obtain low-resolution image data, pair the high-resolution remote sensing image data with the low-resolution image data, and construct a remote sensing image dataset;

[0008] S2: After the high-resolution remote sensing image data and the low-resolution image data are merged along the channel dimension, they are processed by convolution to obtain the first prior feature;

[0009] The low-resolution image data is processed by convolution to obtain conditional features;

[0010] The first prior feature is forward denoised to obtain pure noise, and based on the conditional feature, the pure noise is iteratively denoised to generate the second prior feature;

[0011] S3: The low-resolution image data is fused with the second prior feature to obtain a first feature to be processed; the first feature to be processed is processed by local self-attention to obtain an extracted local feature;

[0012] The extracted local features are fused with the second prior features to obtain the second features to be processed; the second features to be processed are processed by recursive generalized self-attention to obtain the extracted global features;

[0013] The extracted local features and the extracted global features are processed by adaptive feature fusion to obtain fusion features;

[0014] After the fusion features are processed by the first convolution, pixel rearrangement, and the first convolution in sequence, a reconstructed remote sensing image is obtained, completing the first training of the remote sensing image reconstruction model;

[0015] S4: Use the high-resolution remote sensing image data and low-resolution image data in the remote sensing image dataset to execute S2 and S3, train the remote sensing image reconstruction model, and obtain the trained remote sensing image reconstruction model according to the loss function.

[0016] In step S3, the first feature to be processed is processed by local self-attention to obtain the extracted local feature, which is specifically:

[0017] The first feature to be processed is divided into several windows. The features of each window are linearly projected to obtain the query matrix, key matrix and value matrix. The attention score matrix of the features of each window is calculated and weighted with the value matrix to obtain the processed features of each window. The processed features of all windows are concatenated to obtain the extracted local features.

[0018] In step S3, the second feature to be processed is processed by recursive generalized self-attention to obtain the extracted global feature, which is specifically:

[0019] The second feature to be processed is processed by the third convolution to obtain a representative feature map.

[0020] The second feature to be processed is linearly projected to obtain a query matrix, and the representative feature map is linearly projected to obtain a key matrix and a value matrix; the attention score matrix is ​​calculated and weighted with the value matrix to obtain weighted features, and the extracted global features are obtained through feature fusion and reshaping.

[0021] In step S3, the extracted local features and the extracted global features are subjected to adaptive feature fusion processing to obtain fused features, specifically:

[0022] The extracted local features are multiplied by the learnable parameters and then added to the extracted global features; or the extracted global features are multiplied by the learnable parameters and then added to the extracted local features, and after the first convolution processing, they are added element by element with the low-resolution image data to obtain the fused features.

[0023] In step S3, the low-resolution image data is fused with the second priori feature to obtain the first feature to be processed, which is specifically:

[0024] The low-resolution image data is processed by the first convolution to obtain initial features;

[0025] The second prior feature is linearly processed to obtain the third prior feature; the initial feature is normalized, the normalized initial feature is element-wise multiplied with the third prior feature, and then added to the third prior feature to obtain the first feature to be processed.

[0026] In step S2, based on the conditional features, pure noise is iteratively denoised to generate a second priori feature, which is based on the formula:

[0027] , perform iterative denoising;

[0028] In the formula, is the generated result of the current time step t-1; is the generated result at time step t; is the noise control coefficient at time step t; is the total noise variance; Remove the predicted values ​​from the noise for each iteration; is the conditional feature; t is the time step; is the random noise added during sampling, and , I is a unit matrix, representing the standard normal distribution of independent distribution; t=1, 2, 3, ..., T ;

[0029] Pure noise T After iterative denoising, the second prior feature is generated.

[0030] In step S2, the high-resolution remote sensing image data and the low-resolution image data are combined along the channel dimension and then processed by convolution to obtain the first prior feature, which is specifically:

[0031] After the high-resolution remote sensing image data and the low-resolution image data are merged along the channel dimension, the pixels are rearranged and the first convolution process is performed to obtain the first feature map;

[0032] After the first feature map is processed by several second convolutions, a second feature map is obtained;

[0033] The second feature map is processed by the first convolution, average pooling, linear processing, activation function processing, linear processing, and activation function processing in sequence to obtain the first prior feature.

[0034] The second convolution processing is to sequentially perform the first convolution processing, the activation function processing, and the first convolution processing on the feature map to be processed to obtain a processed feature map.

[0035] The present invention also provides a remote sensing image reconstruction model construction and training system, comprising:

[0036] Remote sensing image data set construction module: used to obtain high-resolution remote sensing image data, downsample it to obtain low-resolution image data, pair the high-resolution remote sensing image data with the low-resolution image data, and construct a remote sensing image data set;

[0037] The second prior feature generation module is used to combine the high-resolution remote sensing image data and the low-resolution image data along the channel dimension and perform convolution processing to obtain the first prior feature;

[0038] The low-resolution image data is processed by convolution to obtain conditional features;

[0039] The first prior feature is forward denoised to obtain pure noise, and based on the conditional feature, the pure noise is iteratively denoised to generate the second prior feature;

[0040] Remote sensing image reconstruction module: used to fuse the low-resolution image data with the second prior feature to obtain the first feature to be processed; the first feature to be processed is processed by local self-attention to obtain the extracted local feature;

[0041] The extracted local features are fused with the second prior features to obtain the second features to be processed; the second features to be processed are processed by recursive generalized self-attention to obtain the extracted global features;

[0042] The extracted local features and the extracted global features are processed by adaptive feature fusion to obtain fusion features;

[0043] After the fusion features are processed by the first convolution, pixel rearrangement, and the first convolution in sequence, a reconstructed remote sensing image is obtained, completing the first training of the remote sensing image reconstruction model;

[0044] Remote sensing image reconstruction model training module: used to use the high-resolution remote sensing image data and low-resolution image data in the remote sensing image data set to enter the second prior feature generation module and the remote sensing image reconstruction module for processing, to train the remote sensing image reconstruction model, and to obtain the trained remote sensing image reconstruction model based on the loss function.

[0045] Beneficial Effects

[0046] The remote sensing image reconstruction model provided by the present invention is guided by prior features, extracts local features and global features and performs adaptive feature fusion to reconstruct low-resolution images, wherein the prior features provide additional information for the structure, texture and edge of the image, so that the model can more accurately reconstruct the details of the high-resolution image; and when extracting local features and global features, the prior features and the global contextual relationship are fully utilized, and the local features and the global features are fused by adaptive feature fusion processing, thereby enhancing the performance of the model in super-resolution reconstruction tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 A schematic diagram of a process for constructing and training a remote sensing image reconstruction model is provided in an embodiment. DETAILED DESCRIPTION

[0048] The following examples are intended to illustrate the present invention rather than to further limit the present invention.

[0049] The present invention provides a method for constructing and training a remote sensing image reconstruction model. Figure 1 As shown, the following steps are included:

[0050] S1: Obtain high-resolution remote sensing image data, downsample it to obtain low-resolution image data, pair the high-resolution remote sensing image data with the low-resolution image data, and construct a remote sensing image dataset.

[0051] Among them, downsampling can adopt a bicubic interpolation method, which calculates the interpolation result by considering the values ​​of 16 adjacent pixels, thereby generating a smooth image and retaining more detail information.

[0052] S2: After the high-resolution remote sensing image data and the low-resolution image data are merged along the channel dimension, they are processed by convolution to obtain the first prior feature Z. Specifically:

[0053] After the high-resolution remote sensing image data and the low-resolution image data are merged along the channel dimension, the pixels are rearranged to rearrange the spatial information, which helps to reduce the amount of data while maintaining the main information of the image. The rearranged image is processed by the first convolution, such as a 3×3 convolution, to initially extract features and obtain the first feature map;

[0054] After the first feature map is processed by several second convolutions, a second feature map is obtained;

[0055] The second feature map is processed by the first convolution, average pooling, linear processing, activation function processing, linear processing, and activation function processing in sequence to obtain the first prior feature Z. The first convolution is used to further integrate the extracted features; the average pooling is used for downsampling to further compress the data volume; the linear processing and activation function processing are combined twice to further enhance the representation capability of the features, and the activation function can be a Leaky ReLU activation function.

[0056] Preferably, the second convolution processing is used to further extract and process features, and the feature map to be processed is sequentially subjected to the first convolution processing, the Leaky ReLU activation function processing, and the first convolution processing to obtain a processed feature map.

[0057] The first prior feature Z is obtained based on high-resolution remote sensing image data and low-resolution image data, and is an actual prior feature.

[0058] The low-resolution image data is processed by convolution to obtain the conditional feature D.

[0059] The convolution processing is used to compress the low-resolution image into the potential feature space, which is similar to the process of obtaining the first prior feature Z: the low-resolution image data undergoes the first convolution processing to initially extract features, and then undergoes several second convolution processings, and then successively undergoes the first convolution processing, average pooling processing, linear processing, activation function processing, linear processing, and activation function processing to obtain the conditional feature D.

[0060] The input features are diffused in a highly compressed latent feature space, which significantly reduces the inference iteration steps and computational cost of the model.

[0061] The first prior feature Z is forward denoised to obtain pure noise, and based on the conditional feature, the pure noise is iteratively denoised to generate the second prior feature Z´.

[0062] Among them, the forward noise addition process is:

[0063] through T The Markov diffusion process of iterative steps transforms the first prior feature Z into Gaussian pure noise , follow distribution.

[0064] Then, the reverse denoising process starts from completely random Gaussian noise Initially, the noise is iteratively removed until the predicted a priori feature is restored. Specifically, the conditional feature D is integrated into each reverse denoising step to generate the best predicted a priori feature, that is, the second a priori feature Z'. Preferably, the pure noise is iteratively denoised based on the conditional feature to generate the second a priori feature Z', which is according to the formula:

[0065] , perform iterative denoising;

[0066] In the formula, is the generated result of the current time step t-1; is the generated result at time step t, that is, the input image feature at the current time step t-1; is the noise control coefficient of time step t, which is usually used to control the amount of noise in the denoising process and reflects the degree of denoising at each time step; is the total noise variance; Remove the predicted values ​​from the noise for each iteration; is the conditional feature; t is the time step, which represents the stage from the initial noise state to the final image in the image generation process; is the random noise added during sampling, and some noise is retained during the denoising process to avoid overfitting, and , I is a unit matrix, representing the standard normal distribution of independent distribution, represents a Gaussian distribution with a mean of 0 and a variance of 1; t=1, 2, 3, ..., T ;

[0067] Pure noise T After iterative denoising, the second prior feature Z´ is generated.

[0068] After step S2, a predicted prior feature, namely the second prior feature Z', can be generated from the low-resolution image to guide the reconstruction of the remote sensing image.

[0069] S3: The low-resolution image data is fused with the second prior feature Z´ to obtain the first feature to be processed Specifically:

[0070] The low-resolution image data is processed by the first convolution to obtain the initial features ; This processing is mainly used to extract features from low-resolution image data and convert them into features suitable for subsequent processing.

[0071] The second prior feature Z' is processed linearly to obtain the third prior feature L(Z'); the initial feature Normalization is performed, and the initial features after normalization are After multiplying the third prior feature element by element, add it to the third prior feature to get the first feature to be processed . It can be calculated by the formula:

[0072] , and obtain the first feature to be processed.

[0073] In the formula, is the first feature to be processed; is the second prior feature; is the third a priori feature; For linear processing; is the initial feature after normalization; For normalization processing.

[0074] The first feature to be processed After local self-attention processing, the extracted local feature X is obtained. Specifically:

[0075] The first feature to be processed Divide into several windows, for example, a series of windows of size p×p. Assume that each window consists of N pixels, where N=p×p, and the channel dimension of each pixel is C. Assuming that the windows do not overlap, there will be a total of windows, of which , Represents height and width respectively. The spatial dimension of each window can be expressed as , where i represents the index of the window and 1≤i≤M.

[0076] Characteristics of each window After linear projection, we get the query matrix, key matrix and value matrix, expressed as,

[0077] , , ;

[0078] In the formula, is the query matrix; is the key matrix; is the value matrix; , , are the learnable linear projection matrices of the query matrix, key matrix, and value matrix, respectively, and , , ; The characteristics of each window.

[0079] The attention score matrix of the features of each window is calculated, and after weighting with the value matrix, the processing features of each window are obtained. After the processing features of all windows are concatenated, the extracted local features X are obtained.

[0080] Among them, regarding the attention score, according to the formula:

[0081] ,get.

[0082] In the formula, The characteristics of each window The attention score matrix of represents the similarity of each pixel pair, and ; is a scaling factor used to prevent gradients from vanishing or exploding.

[0083] Regarding weighted summation, it is done through the formula:

[0084] , get the processing characteristics of each window.

[0085] In the formula, is the processing feature of each window, and ; The characteristics of each window The attention score matrix of is a value matrix.

[0086] Regarding splicing, it is done through the formula:

[0087] , and obtain the extracted local feature X.

[0088] Where X is the extracted local feature, and ; is the processing feature of each window; M represents the number of windows; Indicates that the processed features of all windows are concatenated according to their positions in the input image.

[0089] The first feature to be processed After local self-attention processing, the extracted local features X are obtained, and then the global features are extracted, including:

[0090] The extracted local feature X is fused with the second prior feature Z' to obtain the second feature to be processed X'. This can be done by the formula:

[0091] , and obtain the second feature to be processed.

[0092] In the formula, is the second feature to be processed, and ; is the second prior feature; is the prior feature after linear processing; For linear processing; is the feature after normalization; For normalization processing.

[0093] The second feature to be processed X´ is processed by recursive generalized self-attention to obtain the extracted global feature Y; specifically:

[0094] The second feature to be processed X' is processed by the third convolution, for example, first by a deep convolution The features are processed recursively to generate a rough aggregate feature X´´, and ,in, is the convolution kernel stride, is a constant, and ; After depth convolution and 1×1 point convolution, the representative feature map Xr is obtained, and , , is the adjustment factor.

[0095] The second feature to be processed X´ is linearly projected to obtain a query matrix, and the representative feature map Xr is linearly projected to obtain a key matrix and a value matrix; the attention score matrix is ​​calculated and weighted with the value matrix to obtain a weighted feature, and after feature fusion and reshaping, the extracted global feature Y is obtained.

[0096] Among them, according to the formula:

[0097] , linearly project the second feature X' to be processed to obtain the query matrix;

[0098] In the formula, is the query matrix, and ; is the learnable linear projection matrix of the query matrix.

[0099] According to the formula:

[0100] , linearly project the second feature X' to be processed to obtain the key matrix;

[0101] , linearly project the second processed feature X´ to obtain a value matrix;

[0102] In the formula, is the key matrix, and ; is a value matrix, and ; , are the learnable linear projection matrices of the key matrix and value matrix respectively.

[0103] According to the formula:

[0104] , by querying the matrix and key matrix The dot product between them calculates the attention score matrix A;

[0105] In the formula, is the scaling factor used for global feature extraction, It is used to normalize the dot product results to ensure that the sum of the elements in each row of the attention score matrix A is 1.

[0106] According to the formula:

[0107] , the attention score matrix With value matrix Multiply them together to get the weighted feature representation, and then pass it through the projection matrix Perform feature fusion, , and obtain the final cross-attention output features.

[0108] The cross-attention output features are reshaped to obtain the extracted global features Y. .

[0109] Since the extraction of local features and the extraction of global features are cross-appearing, in order to effectively utilize global information, adaptive feature fusion processing is adopted, and input features are weighted using learnable parameters to achieve adaptive adjustment. The extracted local feature X and the extracted global feature Y are subjected to adaptive feature fusion processing to obtain fused features. Specifically:

[0110] The extracted local feature X is multiplied by the learnable parameter and then added to the extracted global feature Y; or the extracted global feature Y is multiplied by the learnable parameter and then added to the extracted local feature X, and after the first convolution processing, the feature is added to the low-resolution image data to obtain the fused feature.

[0111] After the fusion features are processed by the first convolution, pixel rearrangement, and the first convolution in sequence, a reconstructed remote sensing image is obtained, completing one training of the remote sensing image reconstruction model.

[0112] S4: Use the high-resolution remote sensing image data and low-resolution image data in the remote sensing image dataset to execute S2 and S3, train the remote sensing image reconstruction model, and obtain the trained remote sensing image reconstruction model according to the loss function.

[0113] During training, firstly, we learn how to generate predicted prior features that are consistent with the actual prior feature distribution through forward denoising and reverse denoising, and secondly, we extract local features and global features and perform adaptive feature fusion based on the predicted prior features to reconstruct low-resolution images to learn how to generate reconstructed remote sensing images that are consistent with high-resolution remote sensing images. Therefore, the goal of training is to reduce the deviation between the second prior feature Z´ (predicted prior feature) and the first prior feature Z (actual prior feature), and to reduce the loss between the reconstructed remote sensing image and the high-resolution remote sensing image. After joint training, the reconstruction performance of the remote sensing image reconstruction model can be improved.

[0114] That is, the loss function includes two parts. One is to use L1-norm to calculate the sum of the absolute value difference between the first prior feature Z and the second prior feature Z', which is used to measure the difference between the two features. By minimizing this part of the loss, the model can gradually narrow the gap between the two features, so that the second prior feature Z' (predicted prior feature) is closer to the first prior feature Z (actual prior feature), which helps the model maintain the accuracy of the features during the training process.

[0115] The second is to use the L1-norm to calculate the sum of the absolute value differences between the pixel values ​​of the high-resolution remote sensing image and the reconstructed remote sensing image, which is used to measure the difference between the reconstructed remote sensing image and the high-resolution remote sensing image. By minimizing this part of the loss, the model can be encouraged to generate a reconstructed remote sensing image that is closer to the high-resolution real image by reducing the pixel-level difference.

[0116] For the remote sensing image to be reconstructed, the remote sensing image to be reconstructed is introduced into the trained remote sensing image reconstruction model, compressed into conditional feature D through convolution processing, and Gaussian noise is randomly sampled. Based on the conditional feature D, the Gaussian noise is iteratively denoised to generate the second prior feature Z´ (predicted prior feature); guided by the second prior feature Z´, local features and global features are extracted and adaptive feature fusion is performed to reconstruct the low-resolution image and obtain a reconstructed remote sensing image with high resolution.

[0117] The remote sensing image reconstruction model provided by the present invention is guided by prior features, extracts local features and global features and performs adaptive feature fusion to reconstruct low-resolution images, wherein the prior features provide additional information for the structure, texture and edge of the image, so that the model can more accurately reconstruct the details of the high-resolution image; and when extracting local features and global features, the prior features and the global contextual relationship are fully utilized, and the local features and the global features are fused by adaptive feature fusion processing, thereby enhancing the performance of the model in super-resolution reconstruction tasks.

[0118] The present invention also provides a remote sensing image reconstruction model construction and training system, comprising:

[0119] Remote sensing image data set construction module: used to obtain high-resolution remote sensing image data, downsample it to obtain low-resolution image data, pair the high-resolution remote sensing image data with the low-resolution image data, and construct a remote sensing image data set;

[0120] The second prior feature generation module is used to combine the high-resolution remote sensing image data and the low-resolution image data along the channel dimension and perform convolution processing to obtain the first prior feature;

[0121] The low-resolution image data is processed by convolution to obtain conditional features;

[0122] The first prior feature is forward denoised to obtain pure noise, and based on the conditional feature, the pure noise is iteratively denoised to generate the second prior feature;

[0123] Remote sensing image reconstruction module: used to fuse the low-resolution image data with the second prior feature to obtain the first feature to be processed; the first feature to be processed is processed by local self-attention to obtain the extracted local feature;

[0124] The extracted local features are fused with the second prior features to obtain the second features to be processed; the second features to be processed are processed by recursive generalized self-attention to obtain the extracted global features;

[0125] The extracted local features and the extracted global features are processed by adaptive feature fusion to obtain fusion features;

[0126] After the fusion features are processed by the first convolution, pixel rearrangement, and the first convolution in sequence, a reconstructed remote sensing image is obtained, completing the first training of the remote sensing image reconstruction model;

[0127] Remote sensing image reconstruction model training module: used to use the high-resolution remote sensing image data and low-resolution image data in the remote sensing image data set to enter the second prior feature generation module and the remote sensing image reconstruction module for processing, to train the remote sensing image reconstruction model, and to obtain the trained remote sensing image reconstruction model based on the loss function.

Claims

1. A method for constructing and training a remote sensing image reconstruction model, characterized in that: The following steps are involved: S1: Obtain high-resolution remote sensing image data, downsample it to obtain low-resolution image data, pair the high-resolution remote sensing image data with the low-resolution image data, and construct a remote sensing image dataset; S2: After the high-resolution remote sensing image data and the low-resolution image data are merged along the channel dimension, they are processed by convolution to obtain the first prior feature; The low-resolution image data is processed by convolution to obtain conditional features; The first prior feature is forward denoised to obtain pure noise, and based on the conditional feature, the pure noise is iteratively denoised to generate the second prior feature; S3: The low-resolution image data is fused with the second prior feature to obtain a first feature to be processed; The first feature to be processed is processed by local self-attention to obtain the extracted local feature; The extracted local features are fused with the second prior features to obtain the second features to be processed; the second features to be processed are processed by recursive generalized self-attention to obtain the extracted global features; The extracted local features and the extracted global features are processed by adaptive feature fusion to obtain fusion features; After the fusion features are processed by the first convolution, pixel rearrangement, and the first convolution in sequence, a reconstructed remote sensing image is obtained, completing the first training of the remote sensing image reconstruction model; S4: Use the high-resolution remote sensing image data and low-resolution image data in the remote sensing image dataset to execute S2 and S3, train the remote sensing image reconstruction model, and obtain the trained remote sensing image reconstruction model according to the loss function.

2. The method for constructing and training a remote sensing image reconstruction model according to claim 1, characterized in that: In step S3, the first feature to be processed is processed by local self-attention to obtain the extracted local feature, which is specifically: The first feature to be processed is divided into several windows. The features of each window are linearly projected to obtain the query matrix, key matrix and value matrix. The attention score matrix of the features of each window is calculated and weighted with the value matrix to obtain the processed features of each window. The processed features of all windows are concatenated to obtain the extracted local features.

3. The method for constructing and training a remote sensing image reconstruction model according to claim 1, characterized in that: In step S3, the second feature to be processed is processed by recursive generalized self-attention to obtain the extracted global feature, which is specifically: The second feature to be processed is processed by the third convolution to obtain a representative feature map. The second feature to be processed is linearly projected to obtain a query matrix, and the representative feature graph is linearly projected to obtain a key matrix and a value matrix; The attention score matrix is ​​calculated and weighted with the value matrix to obtain the weighted features. After feature fusion and reshaping, the extracted global features are obtained.

4. The method for constructing and training a remote sensing image reconstruction model according to claim 1, characterized in that: In step S3, the extracted local features and the extracted global features are subjected to adaptive feature fusion processing to obtain fused features, specifically: The extracted local features are multiplied by the learnable parameters and then added to the extracted global features; or the extracted global features are multiplied by the learnable parameters and then added to the extracted local features, and after the first convolution processing, they are added element by element with the low-resolution image data to obtain the fused features.

5. The method for constructing and training a remote sensing image reconstruction model according to claim 1, characterized in that: In step S3, the low-resolution image data is fused with the second priori feature to obtain the first feature to be processed, which is specifically: The low-resolution image data is processed by the first convolution to obtain initial features; The second prior feature is linearly processed to obtain the third prior feature; the initial feature is normalized, the normalized initial feature is element-wise multiplied with the third prior feature, and then added to the third prior feature to obtain the first feature to be processed.

6. The method for constructing and training a remote sensing image reconstruction model according to claim 1, characterized in that: In step S2, based on the conditional features, pure noise is iteratively denoised to generate a second priori feature, which is based on the formula: , perform iterative denoising; In the formula, is the generated result of the current time step t-1; is the generated result at time step t; is the noise control coefficient at time step t; is the total noise variance; Remove the predicted values ​​from the noise for each iteration; is a conditional feature; t is the time step; is the random noise added during sampling, and , I is a unit matrix, representing the standard normal distribution of independent distribution; t=1, 2, 3, ..., T ; Pure noise T After iterative denoising, the second prior feature is generated.

7. The method for constructing and training a remote sensing image reconstruction model according to claim 1, characterized in that: In step S2, the high-resolution remote sensing image data and the low-resolution image data are combined along the channel dimension and then processed by convolution to obtain the first prior feature, which is specifically: After the high-resolution remote sensing image data and the low-resolution image data are merged along the channel dimension, the pixels are rearranged and the first convolution process is performed to obtain the first feature map; After the first feature map is processed by several second convolutions, a second feature map is obtained; The second feature map is processed by the first convolution, average pooling, linear processing, activation function processing, linear processing, and activation function processing in sequence to obtain the first prior feature.

8. The method for constructing and training a remote sensing image reconstruction model according to claim 7, characterized in that: The second convolution processing is to sequentially perform the first convolution processing, the activation function processing, and the first convolution processing on the feature map to be processed to obtain a processed feature map.

9. A remote sensing image reconstruction model construction and training system, characterized in that: include: Remote sensing image data set construction module: used to obtain high-resolution remote sensing image data, downsample it to obtain low-resolution image data, pair the high-resolution remote sensing image data with the low-resolution image data, and construct a remote sensing image data set; The second prior feature generation module is used to combine the high-resolution remote sensing image data and the low-resolution image data along the channel dimension and perform convolution processing to obtain the first prior feature; The low-resolution image data is processed by convolution to obtain conditional features; The first prior feature is forward denoised to obtain pure noise, and based on the conditional feature, the pure noise is iteratively denoised to generate the second prior feature; Remote sensing image reconstruction module: used to fuse low-resolution image data with the second prior feature to obtain the first feature to be processed; The first feature to be processed is processed by local self-attention to obtain the extracted local feature; The extracted local features are fused with the second prior features to obtain the second features to be processed; the second features to be processed are processed by recursive generalized self-attention to obtain the extracted global features; The extracted local features and the extracted global features are processed by adaptive feature fusion to obtain fusion features; After the fusion features are processed by the first convolution, pixel rearrangement, and the first convolution in sequence, a reconstructed remote sensing image is obtained, completing the first training of the remote sensing image reconstruction model; Remote sensing image reconstruction model training module: used to use the high-resolution remote sensing image data and low-resolution image data in the remote sensing image data set to enter the second prior feature generation module and the remote sensing image reconstruction module for processing, to train the remote sensing image reconstruction model, and to obtain the trained remote sensing image reconstruction model based on the loss function.

Images