Remote sensing image arbitrary-scale super-resolution reconstruction method based on AGFI and CFUM

Abstract

The invention belongs to the technical field of computer vision, and particularly relates to a remote sensing image arbitrary-scale super-resolution reconstruction method based on AGFI and CFUM, and the method comprises the steps: obtaining low-resolution remote sensing image data, and training a remote sensing image arbitrary-scale super-resolution network model; the trained remote sensing image arbitrary-scale super-resolution network model is adopted to realize remote sensing image arbitrary-scale super-resolution reconstruction; the remote sensing image arbitrary scale super-resolution network model comprises a feature extraction module, a feature interaction module and a two-stage up-sampling module; the remote sensing image reconstruction precision can be improved.

Description

Technical Field

[0001] The invention belongs to the technical field of computer vision, and in particular relates to an arbitrary-scale super-resolution reconstruction method for remote sensing images based on AGFI and CFUM. Background Art

[0002] Remote sensing image (RSI) super-resolution (SR) technology is an indispensable part of computer vision research. It aims to reconstruct high-resolution (HR) remote sensing images from low-resolution (LR) remote sensing images. By using a large amount of training data and deep neural networks, the high-frequency details and structural information of the image are learned, and the lost detail information is effectively restored, so as to reconstruct the HR remote sensing image more accurately. However, LR remote sensing images with less spatial information are usually input into deep neural networks to extract deep features. As the scale factor increases, the loss of structural information will be aggravated. In addition, due to the wide coverage of remote sensing images, bicubic downsampling is used to obtain the corresponding LR images from HR remote sensing images. As the scale factor increases, this resolution conversion will cause more high-frequency information to be lost. Therefore, how to effectively solve the scale limitation of structural information is a challenge.

[0003] Traditional remote sensing image super-resolution technology has achieved remarkable results under common scale factors. However, image reconstruction for a specific magnification can no longer meet the needs of practical applications. Arbitrary scale super-resolution (ASSR) technology, with its advantage of being able to magnify images at different scales, provides greater flexibility for obtaining arbitrary high-resolution remote sensing images and expands the application scope of SR technology. ASSR aims to construct a mapping relationship between the image coordinate domain and the signal domain based on the scale factor, and realize the reconstruction of low-resolution images into arbitrary high-resolution images.

[0004] To achieve the above mapping, some ASSR tasks use discrete representations and use multi-layer perceptrons for discrete and local learning, which requires dynamic prediction of the filter weights corresponding to each scale factor. However, for scale factors outside the training scale range, the network performance will be severely degraded and a large amount of computing resources and memory space will be consumed. Some ASSR tasks use continuous representations and use implicit neural functions for continuous and global learning. They can directly map continuous coordinates to signal values, relying on the nearest neighbor latent code corresponding to the query coordinates, but directly operating on the generated latent code usually cannot further guarantee the integrity of the overall structure. Therefore, it is particularly important to explore new strategies to improve the mapping accuracy between the coordinate domain and the signal domain.

[0005] In addition, the existing ASSR networks usually use dynamically predicted filter weights to directly convert coordinate values ​​and signal values ​​after extracting the required deep features, or use feature expansion and local integration to achieve the mapping of latent codes and RGB values ​​at query coordinates through implicit neural functions. They do not fully consider the impact of multi-level contextual information interaction on the reconstruction results, especially remote sensing images with rich surface information require reconstructed images with clear boundaries. Summary of the invention

[0006] In order to solve the above problems, the present invention provides a remote sensing image arbitrary scale super-resolution reconstruction method based on AGFI and CFUM, obtains low-resolution remote sensing image data to train a remote sensing image arbitrary scale super-resolution network model, and uses the trained remote sensing image arbitrary scale super-resolution network model to realize remote sensing image arbitrary scale super-resolution reconstruction; the remote sensing image arbitrary scale super-resolution network model includes a feature extraction module, a feature interaction module (AGFI) and a two-stage upsampling module (CFUM);

[0007] The training process of the remote sensing image arbitrary scale super-resolution network model includes the following steps:

[0008] S1. Input the LR remote sensing image into the feature extraction module to obtain structural features, deep features and detail features;

[0009] S2. Input the structural features, deep features and detail features into the feature interaction module to obtain the discriminative latent code;

[0010] S3. Input the discriminant latent code into a two-stage upsampling module to obtain a reconstructed high-resolution image;

[0011] S4. Calculate the loss based on the reconstructed high-resolution image and the high-resolution remote sensing image corresponding to the LR remote sensing image, and optimize the model parameters through the loss until the parameters converge.

[0012] Furthermore, the feature extraction module includes a parallel structure extraction module, a common encoder module and a shallow feature extraction module; the structure extraction module includes a Laplacian convolution layer and a splicing layer, and the convolution kernel size of the Laplacian convolution layer is 3×3; the shallow feature extraction module includes two 3×3 convolution layers; step S1 inputs the LR remote sensing image into the feature extraction module. The processing specifically includes:

[0013] S11. Inputting the LR remote sensing image into the structure extraction module to obtain the structure feature;

[0014] S12. Input the LR remote sensing image into the common encoder module to obtain deep features;

[0015] S13. Input the LR remote sensing image into the shallow feature extraction module to obtain detail features.

[0016] Furthermore, the feature interaction module includes a first high-order channel attention module, a second high-order channel attention module and a third high-order channel attention module which are cascaded in sequence, and the three high-order channel attention modules have the same structure; step S2 specifically includes:

[0017] S21. The structural feature F SI With the deep feature F LR Fusion gets enhanced features F EF ;

[0018] S22. The feature dilation function with a convolution kernel size of 3×3 and a dilation rate of 1 is used to respectively SI , deep features F LR Detailed features F O and enhanced feature F EF Perform feature expansion to obtain the structural feature F′ SI , deep features F′ LR , detail feature F′ O and enhanced feature F′ EF ;

[0019] S23. The deep feature F′ LR and enhanced feature F′ EF Input the first high-order channel attention module to obtain the intermediate variable F1;

[0020] S24. The intermediate variable F1 and the structural feature F′ SI Input the second high-order channel attention module to obtain the intermediate variable F2;

[0021] S25. Combine the intermediate variable F2 and the detailed feature F′ O Input the third high-order channel attention module to obtain the discriminant latent code F dis .

[0022] Furthermore, each high-order channel attention module receives and processes two input features to obtain an output feature. The specific processing process includes:

[0023] S211. concatenate an input feature with another input feature after passing through the channel attention layer to obtain a concatenated feature;

[0024] S212. The concatenated features are sequentially passed through the convolution layer and the ReLU activation function layer to obtain the channel attention features;

[0025] S213. Input the channel attention feature into the High-Order Attention Block to obtain the attention weight;

[0026] S214. Multiply the channel attention feature and the attention weight and pass them through the convolution layer and the ReLU activation function layer to obtain the output feature.

[0027] Furthermore, step S3 specifically includes:

[0028] S31. Construct an implicit neural function that reflects the mapping relationship between the coordinate domain and the signal domain;

[0029] S32. Obtain the normalized coordinates of each latent code vector in the discriminant latent code Normalize the query coordinate x in the HR space to the range of [-B, +B] to obtain the normalized coordinate In discriminating the latent code F dis Get the distance normalized coordinates The position coordinates of the latent code vector in the four nearest directions;

[0030] S33. According to the latent code vectors in four directions relative to the normalized coordinates The position coordinates of the structural feature F′ SI Get the fine-tuning factor from

[0031] S34. According to the latent code vectors in four directions relative to the normalized coordinates The position coordinates of the latent code F dis Get the target latent code vector from

[0032] S35. Input the fine-tuning factor, target latent vector and position offset into the implicit neural function to obtain the normalized coordinates The corresponding enhanced discriminant latent code;

[0033] S36. Perform mean operation on the enhanced discriminant latent code to obtain normalized coordinates The corresponding first stage latent code;

[0034] S37. The first-order predicted latent code is passed through a structure weighted filter to obtain a modulated latent code;

[0035] S38. Predict the RGB value of each pixel according to the modulation latent code to obtain a reconstructed high-resolution image.

[0036] Furthermore, the loss function of step S4 is expressed as

[0037]

[0038] Among them, I SR (i) represents the reconstructed SR image (i.e., the reconstructed high-resolution image), I HR (i) represents the high-resolution remote sensing image corresponding to the LR remote sensing image, and n represents the number of images.

[0039] Beneficial effects of the present invention:

[0040] The present invention uses a three-branch feature extractor to obtain structural information, deep features and detail information to obtain multi-level context information; the HOCA module in the AGFI module is used to fully interact the deep features with the multi-level context information to improve the network's ability to reconstruct details; CFUM uses the complementarity of discrete representation and continuous representation, and uses improved implicit neural functions and structural weighted filters to achieve mapping between coordinate domain and signal domain, thereby improving the generalization of the network and the overall reconstruction capability of remote sensing images; at the same time, through the universal design of the encoder in the feature extraction module, it can be used as a plug-and-play plug-in network behind the existing SR backbone network to improve model performance and applicability. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 This is a training flow chart of the remote sensing image arbitrary scale super-resolution network model of the present invention;

[0042] Figure 2 This is a schematic diagram of the overall structure of the remote sensing image arbitrary scale super-resolution network model of the present invention;

[0043] Figure 3 Schematic diagram of the structure of the high-order channel attention module (HOCA) of the present invention;

[0044] Figure 4 It is a schematic diagram of the structure of a two-stage upsampling module (CFUM) of the present invention;

[0045] Figure 5 It is a result comparison diagram of the present invention;

[0046] Figure 6 This is a comparison chart of the results of the present invention and other networks. DETAILED DESCRIPTION

[0047] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0048] A remote sensing image arbitrary scale super-resolution reconstruction method based on AGFI and CFUM, which obtains low-resolution remote sensing image data to train a remote sensing image arbitrary scale super-resolution network model, and uses the trained remote sensing image arbitrary scale super-resolution network model to achieve remote sensing image arbitrary scale super-resolution reconstruction; the remote sensing image arbitrary scale super-resolution network model includes a feature extraction module, a feature interaction module (AGFI) and a two-stage upsampling module (CFUM).

[0049] like Figure 1 , Figure 2 As shown in Figure 1, the training process of the remote sensing image arbitrary scale super-resolution network model includes the following steps:

[0050] S1. Input the LR remote sensing image into the feature extraction module to obtain structural features, deep features and detail features.

[0051] Specifically, Figure 2 As shown, the feature extraction module includes a parallel structure extraction module, a common encoder module and a shallow feature extraction module; the structure extraction module includes a Laplacian convolution layer and a splicing layer, and the convolution kernel size of the Laplacian convolution layer is 3×3; the shallow feature extraction module includes two 3×3 convolution layers.

[0052] Step S1: The processing process of inputting the LR remote sensing image into the feature extraction module specifically includes:

[0053] S11. LR remote sensing image I LR Input structure extraction module E SEM (·) Get the structural feature F SI ∈R h×w×l , which can be expressed as

[0054] F SI =E SEM (I LR ;γ)=Concat(Laplacin_kernel(I LR_R ,I LR_G ,I LR_B ;γ))

[0055] Among them, I LR ∈R H×W×3 represents the LR image, H and W represent the height and width of the LR image, and F SI There are h×w feature vectors, where h and w are the height and width of the feature map, l is the number of feature channels, Laplacin_kernel(·) represents the Laplacian convolution layer with an initialized convolution kernel size of 3×3, and γ represents the Laplacian convolution layer parameter; I LR_R ,I LR_G ,I LR_B They represent the red channel image, green channel image, and blue channel image of the LR remote sensing image respectively, and Concat(·) represents the channel concatenation operation.

[0056] S12. LR remote sensing image I LR Input common encoder module E Encoder (·) Get the deep feature F LR ∈R h×w×l, which can be expressed as

[0057] F LR =E Encoder (I LR ; α)

[0058] Among them, the commonly used encoder module E Encoder (·) includes an encoder, which is RCAN, or RDN, or EDSR, so that it can be used as a plug-and-play plug-in network behind the existing SR backbone network. α represents the encoder parameter.

[0059] S13. LR remote sensing image I LR Input shallow feature extraction module to obtain detail feature F o ∈R h×w×l , which can be expressed as

[0060] F o =Conv3(Conv3(I LR ; β))

[0061] Among them, Conv3(·) represents a 3×3 convolutional layer, and β represents the convolutional layer parameters.

[0062] S2. Input the structural features, deep features and detail features into the feature interaction module together to obtain the discriminative latent code.

[0063] Preferably, the present invention constructs an attention-guided feature interaction module (AGFI), such as Figure 2 As shown, the feature interaction module includes three high-order channel attention modules (HOCA). SI , deep features F LR and detail features F O Input attention-guided feature interaction module (AGFI) to fuse deep features F LR and structural features F SI Get the enhanced feature F EF , and use high-order channel attention (HOCA) for deep feature F LR Interact with multi-level context information to obtain the discriminative latent code F dis .

[0064] Specifically, step S2 specifically includes:

[0065] S21. The structural feature F SI With the deep feature F LR Fusion gets enhanced features F EF ∈R h×w×l , which can be expressed as

[0066] F EF =F LR +FSI

[0067] S22. The feature dilation function with a convolution kernel size of 3×3 and a dilation rate of 1 is used to respectively SI , deep features F LR Detailed features F O and enhanced feature F EF Perform feature expansion to obtain the structural feature F′ SI ∈R h×w×9l , deep features F′ LR ∈R h×w×9l , detail feature F′ O ∈R h×w×9l and enhanced feature F′ EF ∈R h×w×9l .

[0068] Specifically, for the structural feature F SI , deep features F LR Detailed features F O and enhanced feature F EF Feature expansion is performed separately so that each feature vector can use the information of the surrounding 8 feature vectors as context supplement, enrich the representation of each feature vector, and realize channel-wise expansion, which can be expressed as

[0069] F mn =Concat({F m+u,n+v} u,v∈{-1,0,1} )

[0070] Among them, F mn represents the expanded feature, m and n represent the logical position of the feature vector, and u and v represent the range of 3×3 convolution.

[0071] The present invention uses a 3×3 convolutional layer to achieve the above feature expansion, which can be expressed as

[0072] F′ SI =F_unfold k＝3,p＝1 (F SI )

[0073] F′ LR =F_unfold k＝3,p＝1 (F LR )

[0074] F′ O =F_unfold k＝3,p＝1 (F o )

[0075] F′ EF =F_unfold k＝3,p＝1 (F EF )

[0076] Among them, F_unfold k＝3,p＝1 (·) represents the feature dilation function with a convolution kernel size of 3×3 and a dilation rate of 1.

[0077] S23. The deep feature F′ LR and enhanced feature F′ EF Input the first high-order channel attention module to obtain the intermediate variable F1.

[0078] S24. The intermediate variable F1 and the structural feature F′ SI Input the second high-order channel attention module to obtain the intermediate variable F2.

[0079] S25. Combine the intermediate variable F2 and the detailed feature F′ O Input the third high-order channel attention module to obtain the discriminant latent code F dis .

[0080] Specifically, each high-order channel attention module receives and processes two input features and obtains an output feature, such as Figure 3 As shown, the specific processing process includes:

[0081] S211. concatenate an input feature with another input feature after passing through the channel attention layer to obtain a concatenated feature;

[0082] S212. The concatenated features are sequentially passed through the convolution layer and the ReLU activation function layer to obtain channel attention features (CAFature Maps);

[0083] S213. Input the channel attention feature into the High-Order Attention Block to obtain the attention weight;

[0084] S214. Multiply the channel attention feature and the attention weight and pass them through the convolution layer and the ReLU activation function layer to obtain the output feature.

[0085] In this embodiment, the deep feature F′ is first LR and enhanced feature F′ EF Input the first high-order channel attention module to obtain the intermediate variable F1, which can be expressed as

[0086] f ca_1 =ReLU(Conv(Concat(CA(F LR ),F EF )))

[0087]

[0088] F1=HOCA(F LR ,FEF )=ReLU(Conv(W HOAB_1 ·f ca_1 ))

[0089] Among them, f ca_1 represents the channel attention feature of F1, CA(·) represents channel attention, ReLU(·) represents the ReLU activation function, and W HOAB_1 represents the attention weight of F1, Sigmoid(·) represents the Sigmoid activation function, and the intermediate variable F1 represents the discriminative attention feature supplemented with semantic level context information; represents the j-th level feature of the i-th order, represented by f ca_1 Obtained through 1×1 convolution.

[0090] Then the intermediate variable F1 and the structural feature F′ SI Input the second high-order channel attention module to obtain the intermediate variable F2, which can be expressed as

[0091] f ca_2 =ReLU(Conv(Concat(CA(F1),F SI )))

[0092]

[0093] F2=HOCA(F1,F SI )=ReLU(Conv(W HOAB_2 ·f ca_2 ))

[0094] Among them, f ca_2 represents the channel attention feature of F2, W HOAB_2 represents the attention weight of F2, represents the j-th level feature of the i-th order, represented by f ca_2 Obtained through a 1×1 convolution, the intermediate variable F2 represents the discriminative attention feature supplemented with structure-level contextual information.

[0095] Finally, the intermediate variable F2 and the detail feature F′ O Input the third high-order channel attention module to obtain the discriminant latent code F dis ∈R h×w×9l , which can be expressed as

[0096] f ca_3 =ReLU(Conv(Concat(CA(F2),F o )))

[0097]

[0098] F dis=HOCA(F2,F o )=ReLU(Conv(W HOAB_3 ·f ca_3 ))

[0099] Among them, f ca_3 Indicates F dis The channel attention feature, W HOAB_3 Indicates F dis The attention weight, represents the j-th level feature of the i-th order, represented by f ca_3 Obtained through 1×1 convolution, F dis Represents the discriminative latent code supplemented with detail-level contextual information.

[0100] S3. Input the discriminative latent code into the two-stage upsampling module to obtain the reconstructed high-resolution image.

[0101] Specifically, Figure 4 As shown, step S3 specifically includes:

[0102] S31. Construct an implicit neural function that reflects the mapping relationship between the coordinate domain and the signal domain.

[0103] In this embodiment, an implicit neural function is used to establish a mapping between the coordinate domain X and the signal domain S, and the image is represented as a function:

[0104] s=f(z,x)

[0105] where f(·) represents the implicit neural function, z represents a latent vector in the discriminant latent code, x = (i, j) ∈ X represents the query coordinates of the image in the HR space, and s = (s r ,s g ,s b )∈S represents the signal value corresponding to the query coordinate x of the image in the HR space.

[0106] S32. Define the 2D coordinate x′ of each latent code vector in the discriminant latent code, and normalize the 2D coordinate x′ to the range of [-B, +B] to obtain the normalized coordinate Normalize the query coordinate x in the HR space (i.e., the coordinate obtained from the HR remote sensing image corresponding to the LR remote sensing image) to the range of [-B, +B] to obtain the normalized coordinate Where B represents the normalized range, which is usually set to 1; dis Get the distance normalized coordinates The position coordinates of the latent code vector in the nearest four directions. In this embodiment, the nearest neighbor function Nt is used to determine the latent code F dis Get the nearest latent code in four directions relative to The position coordinates can be expressed as

[0107]

[0108] Wherein, Nt(·) represents the nearest neighbor function, r represents the scale factor, and t=00, 01, 10, 11 represent the four directions of the query pixel: upper left, upper right, lower left, and lower right, respectively.

[0109] S33. According to the nearest latent vector in four directions relative to the normalized coordinates The position coordinates of the structural feature F′ SI Get the fine-tuning factor from

[0110]

[0111] Among them, z′ t Represents normalized coordinates In the structural feature F′ SI The fine-tuning factors of the latent code vectors in the four directions are obtained as z′ t ={z′ 00 ,z′ 01 ,z′ 10 ,z′ 11}.

[0112] S34. According to the nearest latent vector in four directions relative to the normalized coordinates The position coordinates of the latent code F dis The target latent code vector is obtained from

[0113]

[0114] in, Represents normalized coordinates In discriminating the latent code F dis The target latent code vectors in the four directions are obtained

[0115] S35. Input the fine-tuning factor, target latent vector and position offset into the implicit neural function to obtain the normalized coordinates The corresponding enhanced discriminant latent code can be expressed as

[0116]

[0117] in, Represents normalized coordinates The corresponding enhanced discriminant latent code, Represents normalized coordinates and the position offset between the corresponding coordinates in the LR space, τ = (τ x ,τ y)=(H / hr,H / wr), indicating normalized coordinates The shape information of the query pixel, τ x represents the height of the query pixel, τ y Indicates the width of the query pixel.

[0118] S36. Perform mean operation on the enhanced discriminant latent code to obtain normalized coordinates The corresponding first-stage latent code can be expressed as

[0119]

[0120] in, represents the first-order prediction latent code, S t′ is a weight, t′ is the diagonal element of t, S = ∑ t S t .

[0121] S37. The first-order prediction latent code is further optimized through the structure weighted filter to obtain the modulation latent code, which can be expressed as

[0122]

[0123] Me=M′+M′*θ

[0124] Among them, θ represents the structural weighted filter, W i ,i∈{1,2,…,K} represents the prior convolution kernel, K represents the number of prior kernels, ω i is the fusion weight obtained by passing τ through two fully connected layers (FC), and Me represents the modulated latent code.

[0125] S38. Use scale-aware weights and modulation latent codes to directly predict the RGB value of each pixel:

[0126]

[0127] μ=Concat({Δx t} t∈{00,01,10,11} ,τ)

[0128] in, Represents normalized coordinates The final RGB value at , δ(·) represents a weight prediction function that generates scale-aware weights based on the position offset and shape information τ. represents the position offset and μ represents the scale-aware weight.

[0129] S4. Calculate the loss based on the reconstructed high-resolution image and the high-resolution remote sensing image corresponding to the LR remote sensing image, and optimize the model parameters through the loss until the parameters converge.

[0130] Specifically, this embodiment uses L1 loss to constrain the optimization direction of the network:

[0131]

[0132] Among them, I SR (i) represents the reconstructed SR image (i.e., reconstructed high-resolution image), I HR (i) represents the high-resolution remote sensing image corresponding to the LR remote sensing image, and n represents the number of images.

[0133] In one embodiment, the method of the present invention is compared with other prior arts on different data sets, and the results are as follows: Figure 5 , Figure 6 shown.

[0134] In the present invention, unless otherwise clearly stipulated and limited, the terms such as "installation", "setting", "connection", "fixation" and "rotation" should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral one; it can be a mechanical connection or an electrical connection; it can be directly connected or indirectly connected through an intermediate medium; it can be the internal connection of two elements or the interaction relationship between two elements. Unless otherwise clearly defined, ordinary technicians in this field can understand the specific meanings of the above terms in the present invention according to the specific circumstances.

[0135] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A remote sensing image arbitrary scale super-resolution reconstruction method based on AGFI and CFUM, characterized in that: Acquire low-resolution remote sensing image data to train a remote sensing image arbitrary-scale super-resolution network model, and use the trained remote sensing image arbitrary-scale super-resolution network model to achieve arbitrary-scale super-resolution reconstruction of remote sensing images; the remote sensing image arbitrary-scale super-resolution network model includes a feature extraction module, a feature interaction module and a two-stage upsampling module; The training process of the remote sensing image arbitrary scale super-resolution network model includes the following steps: S1. Input the LR remote sensing image into the feature extraction module to obtain structural features, deep features and detail features; S2. Input the structural features, deep features and detail features into the feature interaction module to obtain the discriminative latent code; S3. Input the discriminant latent code into a two-stage upsampling module to obtain a reconstructed high-resolution image; S4. Calculate the loss based on the reconstructed high-resolution image and the high-resolution remote sensing image corresponding to the LR remote sensing image, and optimize the model parameters through the loss until the parameters converge.

2. The method for super-resolution reconstruction of remote sensing images at any scale based on AGFI and CFUM according to claim 1, characterized in that: The feature extraction module includes a parallel structure extraction module, a common encoder module and a shallow feature extraction module; the structure extraction module includes a Laplacian convolution layer and a splicing layer, and the convolution kernel size of the Laplacian convolution layer is 3×3; The shallow feature extraction module includes two 3×3 convolutional layers; the processing process of step S1 inputting the LR remote sensing image into the feature extraction module specifically includes: S11. Inputting the LR remote sensing image into the structure extraction module to obtain the structure feature; S12. Input the LR remote sensing image into the common encoder module to obtain deep features; S13. Input the LR remote sensing image into the shallow feature extraction module to obtain detail features.

3. The method for super-resolution reconstruction of remote sensing images at any scale based on AGFI and CFUM according to claim 1, characterized in that: The feature interaction module includes a first high-order channel attention module, a second high-order channel attention module and a third high-order channel attention module which are cascaded in sequence, and the three high-order channel attention modules have the same structure; step S2 specifically includes: S21. The structural feature F SI With the deep feature F LR Fusion gets enhanced features F EF ; S22. The feature dilation function with a convolution kernel size of 3×3 and a dilation rate of 1 is used to respectively SI , deep features F LR Detailed features F O and enhanced feature F EF Perform feature expansion to obtain the structural feature F′ SI , deep features F′ LR , detail feature F′ O and enhanced feature F′ EF ; S23. The deep feature F′ LR and enhanced feature F′ EF Input the first high-order channel attention module to obtain the intermediate variable F1; S24. The intermediate variable F1 and the structural feature F′ SI Input the second high-order channel attention module to obtain the intermediate variable F2; S25. Combine the intermediate variable F2 and the detailed feature F′ O Input the third high-order channel attention module to obtain the discriminant latent code F dis .

4. The method for super-resolution reconstruction of remote sensing images at any scale based on AGFI and CFUM according to claim 3, characterized in that: Each high-order channel attention module receives and processes two input features to obtain an output feature. The specific processing process includes: S211. concatenate an input feature with another input feature after passing through the channel attention layer to obtain a concatenated feature; S212. The concatenated features are sequentially passed through the convolution layer and the ReLU activation function layer to obtain the channel attention features; S213. Input the channel attention feature into the High-Order Attention Block to obtain the attention weight; S214. Multiply the channel attention feature and the attention weight and pass them through the convolution layer and the ReLU activation function layer to obtain the output feature.

5. The method for super-resolution reconstruction of remote sensing images at any scale based on AGFI and CFUM according to claim 1, characterized in that: Step S3 specifically includes: S31. Construct an implicit neural function that reflects the mapping relationship between the coordinate domain and the signal domain; S32. Obtain the normalized coordinates of each latent code vector in the discriminant latent code Normalize the query coordinate x in the HR space to obtain the normalized coordinate In discriminating the latent code F dis Get the distance normalized coordinates The position coordinates of the latent code vector in the four nearest directions; S33. According to the latent code vectors in four directions relative to the normalized coordinates The position coordinates of the structural feature F′ SI Get the fine-tuning factor from S34. According to the latent code vectors in four directions relative to the normalized coordinates The position coordinates of the latent code F dis Get the target latent code vector from S35. Input the fine-tuning factor, target latent vector and position offset into the implicit neural function to obtain the normalized coordinates The corresponding enhanced discriminant latent code; S36. Perform mean operation on the enhanced discriminant latent code to obtain normalized coordinates The corresponding first stage latent code; S37. The first-order predicted latent code is passed through a structure weighted filter to obtain a modulated latent code; S38. Predict the RGB value of each pixel according to the modulation latent code to obtain a reconstructed high-resolution image.

6. The method for super-resolution reconstruction of remote sensing images at any scale based on AGFI and CFUM according to claim 1, characterized in that: The loss function of step S4 is expressed as Among them, I SR (i) represents the reconstruction of high-resolution image, I HR (i) represents the high-resolution remote sensing image corresponding to the LR remote sensing image, and n represents the number of images.

Images