Remote sensing image super-resolution method based on lightweight partial attention convolution

Abstract

The invention discloses a remote sensing image super-resolution method based on lightweight partial attention convolution, and belongs to the technical field of remote sensing image super-resolution. According to the method, a partial attention convolution module PAConv is introduced, so that the problem that the receptive field of a lightweight convolutional layer is limited is effectively solved, and exploration of non-local information is enhanced; compared with other common attention modules, the multi-level aggregation space attention module MASA based on the PAConv module systematically coordinates multi-scale structure information while reducing the calculation complexity and the parameter quantity, and effectively improves the reconstruction index of the super-resolution of the remote sensing image; the method successfully relieves the problems of large number of parameters, high memory utilization rate, high calculation complexity and the like of a complex remote sensing image super-resolution reconstruction method, realizes good balance between remote sensing image reconstruction performance and calculation efficiency, is more suitable for small edge equipment with limited resources, and improves the reconstruction speed of the remote sensing image super-resolution.

Description

Technical Field

[0001] The present invention relates to the technical field of remote sensing image super-resolution, and in particular to a remote sensing image super-resolution method based on lightweight partial attention convolution. Background Art

[0002] Remote sensing image super-resolution technology is a key component of remote sensing information processing, aiming to reconstruct high-resolution details from low-resolution remote sensing images. Improving the resolution of remote sensing images is crucial for tasks such as object detection, scene classification, and image fusion. However, due to the limitations of optical systems and sensor technology, as well as the high costs associated with hardware upgrades, images acquired by Earth observation satellites often fail to meet the spatial resolution standards required for practical applications.

[0003] To address these challenges, research based on deep neural networks has made significant progress in remote sensing image super-resolution, and researchers are developing increasingly complex remote sensing image super-resolution models. Although these models improve the quality of remote sensing image reconstruction, their complex architectures and high computational costs conflict with resource-constrained small edge devices.

[0004] To address these issues, most lightweight super-resolution networks employ attention mechanisms to improve performance and design lightweight convolutional layers to reduce model parameters. However, the introduction of attention mechanisms often results in an increase in the number of parameters. Furthermore, lightweight convolutional layers have a limited receptive field and cannot effectively capture long-range dependencies. To address this issue, we propose a remote sensing image super-resolution method based on lightweight partial attention convolution. Summary of the Invention

[0005] The technical problem to be solved by the present invention is how to achieve a good balance between remote sensing image reconstruction performance and computational efficiency, and provides a remote sensing image super-resolution method based on lightweight partial attention convolution.

[0006] The present invention solves the above technical problems through the following technical solutions, which include the following steps:

[0007] S1: Image preprocessing

[0008] Input low-resolution image, low-resolution image I LR Replicate multiple times and perform image stitching along the channel dimension. The low-resolution image is the remote sensing image.

[0009] S2: Shallow Feature Extraction

[0010] Map the image obtained in the preprocessing stage to the feature space to obtain the shallow feature F0;

[0011] S3: Deep Feature Extraction

[0012] The shallow feature F0 is input into multiple stacked attention progressive feature distillation blocks APFDB for deep feature extraction. The intermediate features generated in the deep feature extraction process are connected to capture information at different levels, realizing multi-layer feature fusion and obtaining the fused feature F d ;

[0013] S4: Image reconstruction

[0014] Based on shallow feature F0 and fusion feature F d Perform image reconstruction to obtain the corresponding super-resolution image I SR .

[0015] Furthermore, in step S1, the image stitching formula is as follows:

[0016]

[0017] Where Concat(·) represents the cascade operation along the channel dimension, and n represents the low-resolution image I to be cascaded. LR The number of

[0018] Furthermore, in step S2, the calculation formula of the shallow feature F0 is as follows:

[0019]

[0020] in, It is a 3×3 blueprint separable convolutional layer.

[0021] Furthermore, in step S3, the calculation formula of the intermediate feature is as follows:

[0022]

[0023] in, represents the kth attention progressive feature distillation block APFDB, F k-1 and F k They represent the input and output features of the kth attention progressive feature distillation block APFDB, respectively, and m represents the number of stacked attention progressive feature distillation blocks APFDB.

[0024] Furthermore, in the attention progressive feature distillation block APFDB, the specific processing process is as follows:

[0025] S301: Feature Extraction

[0026] By introducing a variant of the partial attention convolution module PAConv, the input feature F is gradually optimized and refined. in , the formula is as follows:

[0027]

[0028] Among them, D i (·), PAConv(·) represents the i-th distillation layer and refinement layer, and Represents the corresponding distillation features and refinement features, PAConv s 、PAConv m 、PAConv l Represents a variant of the partial attention convolution module PAConv;

[0029] S302: Feature Fusion

[0030] Use a 1×1 standard convolution layer to fuse all distilled features. The formula is as follows:

[0031] F fusion =Conv 1×1 (Concate(F d1 ,F d2 ,F d3 ,F d4 ))

[0032] Among them, F fusion is the fusion feature, Conv 1×1 Represents a 1×1 standard convolutional layer;

[0033] S303: Feature Enhancement

[0034] Fusion feature F fusion Enhanced by the multi-level aggregation spatial attention module MASA, the formula is as follows:

[0035]

[0036] Among them, F enhanced Represents enhanced features, Represents the multi-level aggregate spatial attention module MASA;

[0037] S304: Feature Conversion

[0038] Use pixel normalization operation to enhance the feature F enhanced To perform the conversion, the formula is as follows:

[0039] F trans =PixelNorm(F enhance )

[0040] Among them, F trans is the transformed feature, PixelNorm(·) represents the pixel normalization operation;

[0041] S305: Feature Output

[0042] The output features are obtained by connecting the residual within the block to the input, as follows:

[0043] F APFDB =F trans +F in

[0044] Among them, F APFDB It is the output feature of the attention progressive feature distillation block APFDB, that is, the intermediate feature.

[0045] Furthermore, in step S301, the specific processing process of the partial attention convolution module PAConv is as follows:

[0046] S3011: For input features Adaptive max pooling is used to capture low-frequency components; then a deep convolutional layer with a kernel size of k×k is used to generate the attention map. The formula is as follows:

[0047] X o ,X p =Split(X)

[0048]

[0049] Where Split(·) represents the channel splitting operation, represents the adaptive maximum pooling operation, i represents the downsampling factor, DWConv k×k (·) is a k×k depth convolutional layer, X o ,X p is the intermediate feature after segmentation;

[0050] S3012: Using Attention Map X map Adaptively adjust the input features X p , to extract representative structural information The formula is as follows:

[0051]

[0052] Among them, φ(·) represents the GELU activation function, represents the nearest neighbor upsampling operation, ⊙ represents the element-wise product;

[0053] S3011: Use a 1×1 standard convolutional layer to further integrate the interaction in the channel dimension and enhance feature representation. The formula is as follows:

[0054]

[0055] Among them, Concate(·) represents the concatenation operation along the channel dimension, X out Represents the output features of the partial attention convolution module PAConv;

[0056] Three variants of the partial attention convolution module PAConv are designed based on different downsampling factors and convolution kernel sizes: Approximate Self-Attention Module PAConv s , multi-window feature aggregation module PAConv m , large core attention module PAConv l ; Based on the partial attention convolution module PAConv, approximate self-attention module PAConv s The downsampling factor is 8, the convolution kernel size is 1×1; the multi-window feature aggregation module PAConv m The downsampling factor is 4, the convolution kernel size is 3×3; the large kernel attention module PAConv l Only a 5×5 convolution kernel is used.

[0057] Furthermore, in step S303, the specific processing process of the multi-level aggregation spatial attention module MASA is as follows:

[0058] S3031: Use a 1×1 standard convolution layer to compress the channel to reduce computational complexity. The formula is as follows:

[0059] F cr =Conv 1×1 (F in )

[0060] Among them, F cr Represents the compressed intermediate features;

[0061] S3032: Integrate multi-level structural information by stacking partial attention convolution modules PAConv variants. The formula is as follows:

[0062] F s =PAConv s (F cr )

[0063] F m =PAConv m (F cr )

[0064] F l =PAConv l (F cr )

[0065] F r =Conv 1×1 (F cr )

[0066] Among them, F i Represents the corresponding multi-level structural information, i = s, m, l, r;

[0067] S3033: Use the GELU activation function and a 3×3 blueprint separable convolution layer to process structural information at different levels, accumulate the processed information, and use a 1×1 standard convolution layer to expand it to the original dimension. The formula is as follows:

[0068]

[0069] in, Represents the multi-level structural information after processing by the blueprint separable convolutional layer, F map Represents the generated attention map to adaptively adjust the input features;

[0070] S3034: Attention map F processed using sigmoid function map Adaptively adjust the input features F in , and then output the feature map F MASA , the formula is as follows:

[0071] F MASA =F in ⊙Sigmoid(F map )+F in

[0072] Where Sigmoid(·) represents the sigmoid activation function.

[0073] Furthermore, in step S3, the multi-layer feature fusion formula is as follows:

[0074]

[0075] in, represents the feature fusion module, F d represents fusion features, and Concate(·) represents feature concatenation operation.

[0076] Furthermore, the feature fusion module fuses the intermediate features through a 1×1 standard convolution layer, then activates it through a GELU function, and finally smoothes the processed features through a blueprint separable convolution layer to obtain the fused feature F d .

[0077] Furthermore, in step S4, the image reconstruction formula is as follows:

[0078]

[0079] in, It is an image reconstruction module, which includes a 3x3 standard convolution layer and a sub-pixel convolution layer to achieve image reconstruction.

[0080] Compared with the prior art, the present invention has the following advantages:

[0081] 1. It successfully alleviates the problems of large number of parameters, high memory usage and high computational complexity in super-resolution reconstruction methods of complex remote sensing images. It introduces partial attention convolution PAConv to achieve a good balance between remote sensing image reconstruction performance and computational efficiency. It is more suitable for small edge devices with limited resources and improves the reconstruction speed of remote sensing image super-resolution.

[0082] 2. By introducing the partial attention convolution module (PAConv), we effectively address the limited receptive field of lightweight convolutional layers and enhance the exploration of non-local information. Compared to other common attention modules, the multi-level aggregate spatial attention module (MASA) based on the PAConv module reduces computational complexity and parameter count while systematically coordinating multi-scale structural information, effectively improving the reconstruction performance of remote sensing image super-resolution. BRIEF DESCRIPTION OF THE DRAWINGS

[0083] Figure 1 1 is a flow chart of a remote sensing image super-resolution method based on lightweight partial attention convolution in an embodiment of the present invention;

[0084] Figure 2 is an architecture diagram of a remote sensing image super-resolution model based on lightweight partial attention convolution in an embodiment of the present invention;

[0085] Figure 3 2 is a schematic diagram of the structure of the partial attention convolution module PAConv in an embodiment of the present invention;

[0086] Figure 4 2 is a schematic diagram of the structure of the attention progressive feature distillation block APFDB in an embodiment of the present invention;

[0087] Figure 5 It is a structural diagram of the multi-level aggregation spatial attention module MASA in an embodiment of the present invention. DETAILED DESCRIPTION

[0088] The following is a detailed description of an embodiment of the present invention. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process. However, the protection scope of the present invention is not limited to the following embodiment.

[0089] like Figure 1 As shown, this embodiment provides a technical solution: a remote sensing image super-resolution method based on lightweight partial attention convolution, comprising the following steps:

[0090] Step 1: Preprocessing

[0091] Input low-resolution image, low-resolution image I LR Replicate multiple times and perform image stitching along the channel dimension. The low-resolution image is the remote sensing image.

[0092] Step 2: Shallow feature extraction process: Shallow feature extraction aims to capture shallow features such as edges, textures, colors, etc. from the input remote sensing image (low-resolution image), and expand the original image into a shallow feature map with 32 channels.

[0093] Step 3: Deep feature extraction process: Deep features refer to extracting finer deep features from shallow features by stacking deep modules (Attention Progressive Feature Distillation Block APFDB) to enhance the network's expressive power.

[0094] Step 4: Multi-layer feature fusion process: Multi-layer feature fusion refers to connecting the intermediate features generated in the deep feature extraction process through skip connections and feature splicing to capture different levels of information;

[0095] Step 5: Image reconstruction process: Global residual learning is enhanced through long skip connections to obtain high-resolution remote sensing images. During the reconstruction process, both the initial feature information and the advantages of multi-layer fusion features are considered to improve the accuracy and quality of remote sensing image reconstruction.

[0096] For the above steps, if Figure 2 As shown in the figure, this paper introduces in detail a remote sensing image super-resolution model based on lightweight partial attention convolution. The specific construction and implementation are as follows:

[0097] First, the preprocessing stage converts the original input image I LR By replicating n times and concatenating these images along the channel dimension, the first stage can be expressed as:

[0098]

[0099] Concat(·) represents the cascade operation along the channel dimension, and n represents the number of I to be cascaded. LR The number of

[0100] The next shallow feature extraction stage uses convolutional layers Map the image obtained in the preprocessing stage to the feature space to obtain the shallow feature F0, such as:

[0101]

[0102] in, It is a 3×3 blueprint separable convolution layer (BSConv), which contains a 1×1 standard convolution layer and a depth convolution layer.

[0103] Subsequently, the deep feature extraction stage inputs F0 into m stacked attention progressive feature distillation blocks APFDB to extract finer deep features and gradually refine the features in the process. The process can be described as:

[0104]

[0105] in, represents the kth APFDB, F k-1 and F k They represent the input and output features of the kth APFDB respectively, and m represents the number of stacked APFDBs.

[0106] The next step is to use skip connections and feature concatenation to combine the intermediate features F generated during deep feature extraction k Connected to capture different levels of information. The process of multi-layer feature fusion can be described as:

[0107]

[0108] in, represents the feature fusion module, F d Indicates fusion features.

[0109] The feature fusion module fuses the intermediate features through a 1×1 standard convolution layer and activates a GELU function, and finally smoothes the processed features through a blueprint separable convolution.

[0110] Finally, the model adopts global skip connections to enhance residual learning, which helps to capture high-frequency information and obtain super-resolution images through the image reconstruction process. SR In the reconstruction process, the initial feature information is considered and the advantages of multi-layer fusion features are utilized to improve the accuracy and quality of remote sensing image reconstruction. The above process can be expressed as:

[0111]

[0112] in, Represents the image reconstruction module, which includes a 3x3 standard convolution layer and a sub-pixel convolution layer to achieve fast image reconstruction.

[0113] During the model optimization process, the network is trained by using a combined loss function that integrates the mean absolute error loss and the frequency domain loss function based on fast Fourier transform. The total loss function of the model is expressed as:

[0114]

[0115] Among them, I HR is a high-quality real image, ∥·∥1 is the L1-norm, represents the fast Fourier transform, and λ is a weight parameter set to 0.05.

[0116] The core basic component of the remote sensing image super-resolution model in this invention is the partial attention convolution module PAConv (subsequently, its variants constitute the attention progressive feature distillation block APFDB and the multi-level aggregation spatial attention module MASA), and its architecture is as follows: Figure 2 As shown, the specific construction and implementation methods are as follows:

[0117] First, for the input features Selectively process part of these features to reduce computational cost. Specifically, an adaptive max pooling operation is used to capture the low-frequency components containing key information, and then a depthwise convolution with a kernel size of k×k is used to generate an attention map. Where i represents the downsampling factor. The process can be expressed as:

[0118] X o ,X p =Split(X)

[0119]

[0120] Where Split(·) represents the channel splitting operation, Represents the adaptive maximum pooling operation, DWConv k×k (·) is a k×k depthwise convolutional layer.

[0121] Second, we use the attention map X map Adaptively adjust the input features X p , to extract representative structural information as follows:

[0122]

[0123] Among them, φ(·) represents the GELU activation function, represents the nearest neighbor upsampling operation, and ⊙ represents the element-wise product.

[0124] Finally, a 1×1 convolutional layer is used to further integrate the interactions in the channel dimension and enhance the feature representation, as described below:

[0125]

[0126] This paper designs three PAConv variants based on different downsampling factors and convolution kernel sizes: approximate self-attention module, multi-window feature aggregation module, and large kernel attention module. Specifically, the approximate self-attention module PAConv s The downsampling factor is 8, and the 1×1 convolution kernel is combined to imitate the self-attention mechanism to effectively capture long-distance dependencies. On the other hand, the multi-window feature aggregation module PAConv m A downsampling factor of 4 and a 3×3 convolution kernel are used to effectively integrate salient features from different windows. Finally, the large kernel attention module PAConv l Only a 5×5 convolution kernel is used to significantly enhance the receptive field.

[0127] like Figure 3 As shown in Figure 3, in the deep feature extraction process, the attention progressive feature distillation block APFDB mainly consists of four stages: feature extraction, feature fusion, feature enhancement, and feature conversion.

[0128] In the feature extraction stage, different variants of PAConv are introduced to gradually optimize and refine the input feature F in The process can be described as follows:

[0129]

[0130] Among them, D i (·), PAConv(·) represents the i-th distillation layer and refinement layer, and Indicates the corresponding distillation characteristics and refining characteristics.

[0131] In the feature fusion stage, a 1×1 standard convolutional layer is used to fuse all distilled features as follows:

[0132] F fusion =Conv 1×1 (Concate(F d1 ,F d2 ,F d3 ,F d4 )) (11)

[0133] Among them, F fusion is the fusion feature, Conv 1×1 Represents a 1×1 convolutional layer.

[0134] Fusion feature F fusion It is enhanced by the multi-level aggregate spatial attention module MASA as follows:

[0135]

[0136] Among them, Fenhanced Represents enhanced features, Represents Multi-Level Aggregate Spatial Attention MASA.

[0137] The features are then transformed using a pixel normalization operation as follows:

[0138] F trans =PixelNorm(F enhance ) (13)

[0139] Among them, F trans is the converted feature, and PixelNorm(·) represents the pixel normalization operation. Finally, by connecting the residual in the block to the input and adopting long skip connections to enhance the residual learning ability of the model, the output F of the attention progressive feature distillation block APFDB is APFDB It can be expressed as:

[0140] F APFDB =F trans +F in (14)

[0141] like Figure 4 As shown in the figure, in the feature enhancement stage, the multi-level aggregation spatial attention module MASA effectively integrates multi-scale structural information by stacking PAConv variants. The specific construction and implementation are as follows:

[0142] First, a 1×1 standard convolutional layer is used to compress the channel to reduce computational complexity, as shown below:

[0143] F cr =Conv 1×1 (F in ) (15)

[0144] Secondly, we enhance the model’s ability to capture multi-scale information by stacking PAConv variants, as follows:

[0145]

[0146] The next step is to use the GELU activation function and a 3×3 BSConv to process the structural information at different levels. The processed information is accumulated and expanded to the original dimension using a 1×1 standard convolution layer. The above process can be expressed as:

[0147]

[0148] Finally, the attention map F processed by the sigmoid function map Adaptively adjust the input features F in , then output F MASA It can be expressed as:

[0149] F MASA =F in ⊙Sigmoid(F map )+F in (18)

[0150] Model training relies on datasets and training algorithms, while model performance is measured using multiple evaluation metrics. The following details the datasets, evaluation metrics, and implementation details involved in this invention.

[0151] Datasets: To conduct a fair comparative analysis, this example selects two publicly available benchmark remote sensing image datasets: UCMerced and RSSCN7, and generates LR images (low-resolution images) by applying bicubic downsampling to HR images (high-resolution images). The present invention adopts a consistent strategy to process the UCMerced and RSSCN7 datasets to construct and divide them into training sets, validation sets, and test sets. Specifically, each dataset is evenly divided into two halves, one half is designated as the training set and the other half is designated as the test set. Subsequently, an additional 20% of the data in the training set is set aside as a validation set.

[0152] Evaluation Metrics: This example uses metrics such as Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM), Spatial Correlation Coefficient (SCC), and Spectral Angle Mapper (SAM) to evaluate the quality of the restored image in the RGB color space. Higher values ​​of PSNR, SSIM, and SCC, and lower values ​​of SAM, indicate higher quality of the reconstructed image.

[0153] Implementation details: The PAFAN model (remote sensing image super-resolution model) proposed in this paper consists of 8 APFDBs, and the number of feature channels in each module is set to 36. In this embodiment, the batch size is set to 64 and the image patch size is set to 48×48. To ensure the stability of the training process, the present invention introduces an exponential moving average (EMA) with a parameter set to 0.999. The widely used Adam optimizer is used for training, and its β1 and β2 are configured to 0.9 and 0.99 respectively. The learning rate is initialized to 2×10 -3 and gradually reduced to 1×10 in 100K iterations using a cosine annealing scheme. -7 .

[0154] In summary, the remote sensing image super-resolution method based on lightweight partial attention convolution in the above embodiment designs a new lightweight basic module called the partial attention convolution module PAConv, and develops three PAConv variants with different receptive fields to collaboratively utilize non-local information. To fully utilize the advantages of PAConv, the present invention arranges PAConv variants in a progressive and iterative manner to form an attention progressive feature distillation block (APFDB), which aims to gradually optimize and refine the extracted features. Furthermore, by stacking PAConv variants, a multi-level aggregate spatial attention module (MASA) is constructed to systematically coordinate multi-scale structural information, achieving a good balance between remote sensing image reconstruction performance and computational efficiency.

[0155] Although the embodiments of the present invention have been shown and described above, it will be understood that the above embodiments are illustrative and are not to be construed as limitations on the present invention. A person skilled in the art may change, modify, replace and modify the above embodiments within the scope of the present invention.

Claims

1. A remote sensing image super-resolution method based on lightweight partial attention convolution, characterized by: The following steps are involved: S1: Image preprocessing Input low-resolution image, low-resolution image I LR Replicate multiple times and perform image stitching along the channel dimension. The low-resolution image is the remote sensing image. S2: Shallow Feature Extraction Map the image obtained in the preprocessing stage to the feature space to obtain the shallow feature F0; S3: Deep Feature Extraction The shallow feature F0 is input into multiple stacked attention progressive feature distillation blocks APFDB for deep feature extraction. The intermediate features generated in the deep feature extraction process are connected to capture information at different levels, realizing multi-layer feature fusion and obtaining the fused feature F d ; S4: Image reconstruction Based on shallow feature F0 and fusion feature F d Perform image reconstruction to obtain the corresponding super-resolution image I SR .

2. A remote sensing image super-resolution method based on lightweight partial attention convolution according to claim 1, characterized in that In step S1, the image stitching formula is as follows: Where Concat(·) represents the concatenation operation along the channel dimension, and n represents the low-resolution image I to be concatenated. LR The number of 3. The remote sensing image super-resolution method based on lightweight partial attention convolution according to claim 1, characterized in that In step S2, the calculation formula of the shallow feature F0 is as follows: in, It is a 3×3 blueprint separable convolutional layer.

4. The remote sensing image super-resolution method based on lightweight partial attention convolution according to claim 1, characterized in that In step S3, the calculation formula of the intermediate feature is as follows: in, represents the kth attention progressive feature distillation block APFDB, F k-1 and F k They represent the input and output features of the kth attention progressive feature distillation block APFDB, respectively, and m represents the number of stacked attention progressive feature distillation blocks APFDB.

5. The remote sensing image super-resolution method based on lightweight partial attention convolution according to claim 4 is characterized in that In the attention progressive feature distillation block APFDB, the specific processing process is as follows: S301: Feature Extraction By introducing a variant of the partial attention convolution module PAConv, the input feature F is gradually optimized and refined. in , the formula is as follows: Among them, D i (·), PAConv(·) represents the i-th distillation layer and refinement layer, and Represents the corresponding distillation features and refinement features, PAConv s 、PAConv m 、PAConv l Represents a variant of the partial attention convolution module PAConv; S302: Feature Fusion Use a 1×1 standard convolution layer to fuse all distilled features. The formula is as follows: F fusion =Conv 1×1 (Concate(F d1 ,F d2 ,F d3 ,F d4 )) Among them, F fusion is the fusion feature, Conv 1×1 Represents a 1×1 standard convolutional layer; S303: Feature Enhancement Fusion feature F fusion Enhanced by the multi-level aggregation spatial attention module MASA, the formula is as follows: Among them, F enhanced Represents enhanced features, Represents the multi-level aggregate spatial attention module MASA; S304: Feature Conversion Use pixel normalization operation to enhance the feature F enhanced To perform the conversion, the formula is as follows: F trans =PixelNorm(F enhance ) Among them, F trans is the transformed feature, PixelNorm(·) represents the pixel normalization operation; S305: Feature Output The output features are obtained by connecting the residual within the block to the input, as follows: F APFDB =F trans +F in Among them, F APFDB It is the output feature of the attention progressive feature distillation block APFDB, that is, the intermediate feature.

6. The remote sensing image super-resolution method based on lightweight partial attention convolution according to claim 5, characterized in that In step S301, the specific processing process of the partial attention convolution module PAConv is as follows: S3011: For input features Adaptive max pooling is used to capture low-frequency components; then a deep convolutional layer with a kernel size of k×k is used to generate the attention map. The formula is as follows: X o ,X p =Split(X) Where Split(·) represents the channel splitting operation, represents the adaptive maximum pooling operation, i represents the downsampling factor, DWConv k×k (·) is a k×k depth convolutional layer, X o ,X p is the intermediate feature after segmentation; S3012: Using Attention Map X map Adaptively adjust the input features X p , to extract representative structural information The formula is as follows: Among them, φ(·) represents the GELU activation function, represents the nearest neighbor upsampling operation, ⊙ represents the element-wise product; S3011: Use a 1×1 standard convolutional layer to further integrate the interaction in the channel dimension and enhance feature representation. The formula is as follows: Among them, Concate(·) represents the concatenation operation along the channel dimension, X out Represents the output features of the partial attention convolution module PAConv; Three variants of the partial attention convolution module PAConv are designed based on different downsampling factors and convolution kernel sizes: Approximate Self-Attention Module PAConv s , multi-window feature aggregation module PAConv m , large core attention module PAConv l ; Based on the partial attention convolution module PAConv, approximate self-attention module PAConv s The downsampling factor is 8, the convolution kernel size is 1×1; the multi-window feature aggregation module PAConv m The downsampling factor is 4, the convolution kernel size is 3×3; the large kernel attention module PAConv l Only a 5×5 convolution kernel is used.

7. The remote sensing image super-resolution method based on lightweight partial attention convolution according to claim 6, characterized in that In step S303, the specific processing process of the multi-level aggregation spatial attention module MASA is as follows: S3031: Use a 1×1 standard convolution layer to compress the channel to reduce computational complexity. The formula is as follows: F cr =Conv 1×1 (F in ) Among them, F cr Represents the compressed intermediate features; S3032: Integrate multi-level structural information by stacking partial attention convolution modules PAConv variants. The formula is as follows: F s =PAConv s (F cr ) F m =PAConv m (F cr ) F l =PAConv l (F cr ) F r =Conv 1×1 (F cr ) Among them, F i Represents the corresponding multi-level structural information, i = s, m, l, r; S3033: Use the GELU activation function and a 3×3 blueprint separable convolution layer to process structural information at different levels, accumulate the processed information, and use a 1×1 standard convolution layer to expand it to the original dimension. The formula is as follows: in, Represents the multi-level structural information after processing by the blueprint separable convolutional layer, F map Represents the generated attention map to adaptively adjust the input features; S3034: Attention map F processed using sigmoid function map Adaptively adjust the input features F in , and then output the feature map F MASA , the formula is as follows: F MASA =F in ⊙Sigmoid(F map )+F in Where Sigmoid(·) represents the sigmoid activation function.

8. The remote sensing image super-resolution method based on lightweight partial attention convolution according to claim 4, characterized in that In step S3, the multi-layer feature fusion formula is as follows: in, represents the feature fusion module, F d represents the fusion feature, and Concate(·) represents the concatenation operation along the channel dimension.

9. The remote sensing image super-resolution method based on lightweight partial attention convolution according to claim 8, characterized in that The feature fusion module fuses the intermediate features through a 1×1 standard convolution layer, then activates it through a GELU function, and finally smoothes the processed features through a blueprint separable convolution layer to obtain the fused feature F d .

10. The remote sensing image super-resolution method based on lightweight partial attention convolution according to claim 1, characterized in that In step S4, the image reconstruction formula is as follows: in, It is an image reconstruction module, which includes a 3x3 standard convolution layer and a sub-pixel convolution layer to achieve image reconstruction.

Images