A method for arbitrary-scale super-resolution reconstruction of remote sensing images based on AGFI and CFUM

Abstract

This invention belongs to the field of computer vision technology, specifically relating to a method for arbitrary-scale super-resolution reconstruction of remote sensing images based on AGFI and CFUM. The method includes acquiring low-resolution remote sensing image data to train an arbitrary-scale super-resolution network model for remote sensing images, and using the trained model to achieve arbitrary-scale super-resolution reconstruction of remote sensing images. The arbitrary-scale super-resolution network model includes a feature extraction module, a feature interaction module, and a two-stage upsampling module. This invention can improve the accuracy of remote sensing image reconstruction.

Description

Technical Field

[0001] This invention belongs to the field of computer vision technology, specifically relating to a method for arbitrary-scale super-resolution reconstruction of remote sensing images based on AGFI and CFUM. Background Technology

[0002] Remote sensing image super-resolution (SR) technology, an indispensable part of computer vision research, aims to reconstruct high-resolution (HR) remote sensing images from low-resolution (LR) remote sensing images. By using large amounts of training data and deep neural networks, it learns high-frequency details and structural information from the images, effectively recovering lost details and thus reconstructing HR remote sensing images more accurately. However, LR remote sensing images, which typically carry less spatial information, are often input into deep neural networks to extract deep features. As the scale factor increases, the loss of structural information worsens. Furthermore, due to the wide coverage of remote sensing images, using bicubic downsampling to obtain corresponding LR images from HR remote sensing images results in further loss of high-frequency information with increasing scale factors. Therefore, effectively addressing the scale limitations of structural information remains a challenge.

[0003] Traditional super-resolution techniques for remote sensing images have achieved significant results at common scale factors. However, the limitation of image reconstruction to only specific magnifications no longer meets 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 acquiring arbitrarily 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 scale factors, enabling the reconstruction of low-resolution images into arbitrarily high-resolution images.

[0004] To achieve the above mapping, some ASSR tasks employ discrete representations, utilizing multilayer perceptrons for discrete and local learning, requiring dynamic prediction of filter weights for each scale factor. However, scale factors outside the training scale range lead to a severe degradation in network performance and consume significant computational resources and memory. Some ASSR tasks employ continuous representations, utilizing implicit neural functions for continuous and global learning, directly mapping continuous coordinates to signal values. This relies on querying the nearest neighbor latent code corresponding to the coordinates, but directly operating on the generated latent code often fails to further guarantee the integrity of the overall structure. Therefore, exploring new strategies to improve the mapping accuracy between the coordinate and signal domains is particularly important.

[0005] Furthermore, existing ASSR networks typically extract the required deep features and then directly transform coordinate and signal values ​​using dynamically predicted filter weights, or use feature dilation and local integration to map the latent code and RGB values ​​at the query coordinates through implicit neural functions. They do not fully consider the impact of multi-level contextual information interaction on the reconstruction results, especially since remote sensing images with rich surface information require reconstructed images with clear boundaries. Summary of the Invention

[0006] To address the aforementioned issues, this invention provides a method for arbitrary-scale super-resolution reconstruction of remote sensing images based on AGFI and CFUM. The method involves acquiring low-resolution remote sensing image data to train an arbitrary-scale super-resolution network model, and then using this trained model to achieve arbitrary-scale super-resolution reconstruction of the remote sensing image. The arbitrary-scale super-resolution network model includes a feature extraction module, a feature interaction module (AGFI), and a two-level upsampling module (CFUM).

[0007] The training process of a super-resolution network model for remote sensing images at any scale 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 detailed features into the feature interaction module to obtain the discriminant latent code;

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

[0011] S4. Calculate the loss based on the reconstructed high-resolution image and the corresponding high-resolution remote sensing image of 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 convolutional layer and a stitching layer, with the Laplacian convolutional layer having a 3×3 kernel size; the shallow feature extraction module includes two 3×3 convolutional layers; step S1, which inputs the LR remote sensing image into the feature extraction module, specifically includes:

[0013] S11. Input the LR remote sensing image into the structure extraction module to obtain structural features;

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

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

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

[0017] S21. Structural feature F SI With deep features F LR The fusion yields enhanced feature F EF ;

[0018] S22. Using a feature dilation function with a convolution kernel size of 3×3 and a dilation rate of 1, respectively, the structural features F are applied... SI Deep features F LR Detail features F O and enhanced features F EF Perform feature dilation to obtain structural features F′ SI Deep features F′ LR Detail features F′ O and enhanced features F′ EF ;

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

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

[0021] S25. Combine intermediate variable F2 and detail feature F′ O Inputting the third higher-order channel attention module yields the discriminant latent code F. dis .

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

[0023] S211. After passing one input feature through a channel attention layer, it is concatenated with another input feature to obtain a concatenated feature;

[0024] S212. Pass the concatenated features sequentially through a convolutional layer and a ReLU activation function layer to obtain channel attention features;

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

[0026] S214. Multiply the channel attention features and attention weights, and then pass them through a convolutional layer and a ReLU activation function layer to obtain the output features.

[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 discrimination code. The query coordinate x in the HR space is normalized to the range [-B, +B] to obtain the normalized coordinate. In determining the latent code F dis Obtain distance-normalized coordinates The position coordinates of the latent code vectors in the four nearest directions;

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

[0031] S34. Based on the latent code vectors in the four directions relative to the normalized coordinates The position coordinates are used to determine the latent code F. dis The target latent code vector is obtained from the data.

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

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

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

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

[0036] Furthermore, the loss function in step S4 is expressed as follows:

[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] The beneficial effects of this invention are:

[0040] This invention uses a three-branch feature extractor to acquire structural information, deep features, and detail information, obtaining multi-level contextual information. The HOCA module within the AGFI module fully interacts with the deep features and multi-level contextual information, improving the network's ability to reconstruct details. CFUM leverages the complementarity of discrete and continuous representations, utilizing improved implicit neural functions and structure-weighted filters to map the coordinate and signal domains, enhancing the network's generalization and overall remote sensing image reconstruction capabilities. Simultaneously, through the general design of the encoder in the feature extraction module, it can serve as a plug-and-play network behind existing SR backbone networks, improving model performance and applicability. Attached Figure Description

[0041] Figure 1 This is a flowchart of the training process for the arbitrary-scale super-resolution network model for remote sensing images in this invention.

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

[0043] Figure 3 This is a schematic diagram of the High-Order Channel Attention Module (HOCA) structure of the present invention;

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

[0045] Figure 5 This is a comparison chart of the results of the present invention;

[0046] Figure 6 This is a comparison chart of the results of this invention with those of other networks. Detailed Implementation

[0047] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0048] A method for arbitrary-scale super-resolution reconstruction of remote sensing images based on AGFI and CFUM is proposed. The method involves acquiring low-resolution remote sensing image data to train an arbitrary-scale super-resolution network model for remote sensing images, and then using the trained arbitrary-scale super-resolution network model to achieve arbitrary-scale super-resolution reconstruction of remote sensing images. The arbitrary-scale super-resolution network model for remote sensing images includes a feature extraction module, a feature interaction module (AGFI), and a two-level upsampling module (CFUM).

[0049] like Figure 1 , Figure 2 As shown, the training process of a super-resolution network model for remote sensing images at any scale 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, such as 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 convolutional layer and a splicing layer, with the Laplacian convolutional layer having a kernel size of 3×3; the shallow feature extraction module includes two 3×3 convolutional layers.

[0052] Step S1, which involves inputting the LR remote sensing image into the feature extraction module, specifically includes the following processes:

[0053] S11. Transfer the LR remote sensing image I LR Input structure extraction module E SEM (·) Obtain structural feature F SI ∈R h×w×l , can be represented 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 This represents an LR image, where 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 an initialized Laplacian convolutional layer with a kernel size of 3×3, and γ represents the Laplacian convolutional layer parameters; LR_R I LR_G I LR_B These represent the red, green, and blue channels of an LR remote sensing image, respectively, with Concat(·) representing the channel stitching operation.

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

[0057] F LR =E Encoder (I LR ;α)

[0058] Among them, the commonly used encoder module E Encoder (·) includes an encoder, which can be RCAN, RDN, or EDSR, to facilitate its use as a plug-and-play network behind an existing SR backbone, where α represents the encoder parameters.

[0059] S13. Transfer the LR remote sensing image I LR Input the shallow feature extraction module to obtain the detailed features F o ∈R h×w×l , can be represented as

[0060] F o =Conv3(Conv3(I LR ;β))

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

[0062] S2. Input the structural features, deep features, and detailed features into the feature interaction module to obtain the discriminant 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). The structural feature F... SI Deep features F LR and detailed features F O Input attention-guided feature interaction module (AGFI) fuses deep features F LR and structural features F SI Obtain enhanced feature F EF And utilize High-Order Channel Attention (HOCA) for deep feature F LR By interacting with multi-level context information, the discriminant latent code F is obtained. dis .

[0064] Specifically, step S2 includes:

[0065] S21. Structural feature F SI With deep features F LR The fusion yields enhanced feature F EF ∈R h×w×l , can be represented as

[0066] F EF =F LR +FSI

[0067] S22. Using a feature dilation function with a convolution kernel size of 3×3 and a dilation rate of 1, respectively, the structural features F are applied... SI Deep features F LR Detail features F O and enhanced features F EF Perform feature dilation to obtain structural features F′ SI ∈R h×w×9l Deep features F′ LR ∈R h×w×9l Detail features F′ O ∈R h×w×9l and enhanced features F′ EF ∈R h×w×9l .

[0068] Specifically, regarding structural feature F SI Deep features F LR Detail features F O and enhanced features F EF Feature dilation is performed separately, allowing each feature vector to utilize information from its eight surrounding feature vectors as contextual supplementation, enriching the representation of each feature vector and achieving channel-oriented dilation, which can be represented as follows:

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

[0070] Among them, F mn The dilated features are represented by m and n, which represent the logical positions of the feature vectors, and u and v represent the range of the 3×3 convolution.

[0071] This invention uses a 3×3 convolutional layer to achieve the aforementioned feature dilation, which can be represented as follows:

[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 kernel size of 3×3 and a dilation rate of 1.

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

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

[0079] S25. Combine intermediate variable F2 and detail feature F′ O Inputting the third higher-order channel attention module yields the discriminant latent code F. dis .

[0080] Specifically, each higher-order channel attention module receives and processes two input features to obtain one output feature, such as... Figure 3 As shown, the specific processing steps include:

[0081] S211. After passing one input feature through a channel attention layer, it is concatenated with another input feature to obtain a concatenated feature;

[0082] S212. Pass the concatenated features sequentially through a convolutional layer and a ReLU activation function layer to obtain channel attention features (CAFature Maps);

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

[0084] S214. Multiply the channel attention features and attention weights, and then pass them through a convolutional layer and a ReLU activation function layer to obtain the output features.

[0085] In this embodiment, the deep feature F′ is first... LR and enhanced features F′ EF Inputting the first higher-order channel attention module yields the intermediate variable F1, which can be represented as follows:

[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 Denotes the channel attention feature of F1, CA(·) represents channel attention, ReLU(·) represents the ReLU activation function, and W HOAB_1 The F1 value represents the attention weights of F1, Sigmoid(·) represents the Sigmoid activation function, and the intermediate variable F1 represents the discriminative attention features that supplement semantic-level contextual information. Denotes the i-th order and j-th level feature, given by f ca_1 Obtained through 1×1 convolution.

[0090] Then the intermediate variable F1 and the structural feature F′ SI Inputting the second higher-order channel attention module yields the intermediate variable F2, which can be represented 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 W represents the channel attention feature of F2. HOAB_2 This represents the attention weights for F2. Denotes the i-th order and j-th level feature, given by f ca_2 Obtained through 1×1 convolution, the intermediate variable F2 represents the discriminative attention feature that supplements structural-level contextual information.

[0095] Finally, the intermediate variable F2 and the detail feature F′ are... O Inputting the third higher-order channel attention module yields the discriminant latent code F. dis ∈R h×w×9l , can be represented 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 F represents dis Channel attention features, W HOAB_3 F represents dis Attention weights Denotes the i-th order and j-th level feature, given by f ca_3 F is obtained through 1×1 convolution. dis This indicates a discriminant latent code that supplements detailed contextual information.

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

[0101] Specifically, such as 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, representing the image as a function:

[0104] s = f(z,x)

[0105] Where f(·) represents the implicit neural function, z represents a latent code 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 in the HR space of the image.

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

[0107]

[0108] Where 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: top left, top right, bottom left, and bottom right, respectively.

[0109] S33. Based on the nearest latent code vectors in the four directions relative to the normalized coordinates The position coordinates of the structural feature F′ SI The fine-tuning factor is obtained from the data, and is expressed as follows:

[0110]

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

[0112] S34. Based on the nearest latent code vectors in the four directions relative to the normalized coordinates The position coordinates are used to determine the latent code F. dis The target latent code vector obtained from this can be represented as:

[0113]

[0114] in, Represents normalized coordinates In determining 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 code vector, and position offset into the implicit neural function to obtain normalized coordinates. The corresponding enhanced discriminative latent code can be represented as:

[0116]

[0117] in, Represents normalized coordinates The corresponding enhanced discriminative latent code, Represents normalized coordinates And the position offset between the corresponding coordinates in LR space, τ=(τ x ,τ y) = (H / hr, H / wr), representing normalized coordinates The shape information, that is, the shape information at the query pixel, τ x τ represents the height of the queried pixel. y This indicates the width in pixels.

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

[0119]

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

[0121] S37. The first-order predictive latent code is further optimized using a structure-weighted filter to obtain the modulated latent code, which can be expressed as:

[0122]

[0123] Me = M′ + M′*θ

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

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

[0126]

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

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

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

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

[0131]

[0132] 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.

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

[0134] In this invention, unless otherwise explicitly specified and limited, the terms "installation," "setting," "connection," "fixing," "rotation," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal connection of two components or the interaction between two components. Unless otherwise explicitly limited, those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

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

Claims

1. A method for arbitrary-scale super-resolution reconstruction of remote sensing images based on AGFI and CFUM, characterized in that, A remote sensing image arbitrary scale super-resolution network model is trained by acquiring low-resolution remote sensing image data, and the trained remote sensing image arbitrary scale super-resolution reconstruction is realized by using the remote sensing image arbitrary scale super-resolution network model; the remote sensing image arbitrary scale super-resolution network model includes a feature extraction module, a feature interaction module AGFI, and a two-level upsampling module CFUM. The training process of a super-resolution network model for remote sensing images at any scale 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 detailed features into the feature interaction module to obtain the discriminant latent code; The feature interaction module includes a first higher-order channel attention module, a second higher-order channel attention module, and a third higher-order channel attention module cascaded in sequence. The three higher-order channel attention modules have the same structure. Step S2 specifically includes: S21. Structural feature F SI With deep features F LR The fusion yields enhanced feature F EF ; S22. Using a feature dilation function with a convolution kernel size of 3×3 and a dilation rate of 1, respectively, the structural features F are applied... SI Deep features F LR Detail features F O and enhanced features F EF Perform feature dilation to obtain structural features F′ SI Deep features F′ LR Detail features F′ O and enhanced features F′ EF ; S23. Transfer deep features F′ LR and enhanced features F′ EF Input the first higher-order channel attention module to obtain the intermediate variable F1; S24. Combine intermediate variable F1 and structural feature F′ SI Input the second higher-order channel attention module to obtain the intermediate variable F2; S25. Combine intermediate variable F2 and detail feature F′ O Inputting the third higher-order channel attention module yields the discriminant latent code F. dis ; S3. Input the discriminant latent code into the two-stage upsampling module to obtain the reconstructed high-resolution image; 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 discrimination code. The query coordinate x in the HR space will be normalized to obtain the normalized coordinate. In determining the latent code F dis Obtain distance-normalized coordinates The position coordinates of the latent code vectors in the four nearest directions; S33. Based on the latent code vectors in the four directions relative to the normalized coordinates The position coordinates of the structural feature F′ SI Obtain the fine-tuning factor from the middle; S34. Based on the latent code vectors in the four directions relative to the normalized coordinates The position coordinates are used to determine the latent code F. dis The target latent code vector is obtained from the data. S35. Input the fine-tuning factor, target latent code vector, and position offset into the implicit neural function to obtain normalized coordinates. The corresponding enhanced discriminative latent code; S36. Perform a mean operation on the enhanced discriminative latent code to obtain normalized coordinates. The corresponding first-stage latent code; S37. The first-order predictive latent code is passed through a structure-weighted filter to obtain the modulation latent code; S38. Predict the RGB value of each pixel based on the modulation latent code to obtain the reconstructed high-resolution image; S4. Calculate the loss based on the reconstructed high-resolution image and the corresponding high-resolution remote sensing image of the LR remote sensing image, and optimize the model parameters through the loss until the parameters converge.

2. The method for arbitrary-scale super-resolution reconstruction of remote sensing images 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 convolutional layer and a splicing layer, and the kernel size of the Laplacian convolutional layer is 3×3; The shallow feature extraction module includes two 3×3 convolutional layers; step S1, which involves inputting the LR remote sensing image into the feature extraction module, specifically includes: S11. Input the LR remote sensing image into the structure extraction module to obtain structural features; S12. Input the LR remote sensing image into a common encoder module to obtain deep features; S13. Input the LR remote sensing image into the shallow feature extraction module to obtain detailed features.

3. The method for arbitrary-scale super-resolution reconstruction of remote sensing images based on AGFI and CFUM according to claim 1, characterized in that, Each higher-order channel attention module receives and processes two input features to obtain one output feature. The specific processing includes: S211. After passing one input feature through a channel attention layer, it is concatenated with another input feature to obtain a concatenated feature; S212. Pass the concatenated features sequentially through a convolutional layer and a ReLU activation function layer to obtain channel attention features; S213. Input the channel attention features into the High-Order Attention Block to obtain the attention weights; S214. Multiply the channel attention features and attention weights, and then pass them through a convolutional layer and a ReLU activation function layer to obtain the output features.

4. The method for arbitrary-scale super-resolution reconstruction of remote sensing images based on AGFI and CFUM according to claim 1, characterized in that, The loss function for step S4 is expressed as follows: Among them, I SR (i) indicates that a high-resolution image is reconstructed, 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