A remote sensing image restoration network fusing frequency domain channel attention and global-local dual branch

Abstract

The application belongs to the technical field of remote sensing image processing and computer vision, and particularly relates to a remote sensing image restoration network fusing a frequency domain channel attention and a global-local double-branch, which comprises shallow feature extraction modules, deep feature extraction modules and high-resolution reconstruction modules connected in sequence; the deep feature extraction module is composed of N feature aggregation Mamba groups stacked in sequence, each feature aggregation Mamba group is composed of M feature aggregation Mamba blocks connected in series, and N and M are positive integers greater than or equal to 2; the application solves the defects of a single Mamba architecture, such as a bias towards global low frequency and over-smoothed reconstruction results, by means of a double-branch design of channel splitting, a global Mamba branch is responsible for capturing long-distance structure dependence of large-scale ground objects, and a local detail enhancement branch focuses on strengthening high-frequency texture and edge information, and the two branches are complementary at the feature level, thereby significantly improving the restoration degree of texture details of remote sensing images.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing image processing and computer vision technology, and specifically relates to a remote sensing image restoration network that integrates frequency domain channel attention and global-local bi-branch. Background Technology

[0002] High-resolution remote sensing images have irreplaceable application value in fields such as urban planning, agricultural resource monitoring, national defense reconnaissance, and ground target detection. However, due to limitations such as the physical limitations of satellite optical systems, atmospheric transmission attenuation, and imaging platform jitter, directly acquiring ultra-high-resolution remote sensing images is extremely costly and difficult. Image super-resolution (SR) technology reconstructs high-resolution images from low-resolution images using algorithms, and is a cost-effective and efficient image quality enhancement solution that has become a research hotspot in the field of remote sensing image processing.

[0003] Existing super-resolution techniques are mainly divided into three categories: CNN-based methods, Transformer-based methods, and Mamba-based methods.

[0004] 1. CNN-based methods: Represented by EDSR, RCAN, and RDN, these methods have achieved good results due to their powerful local feature extraction and nonlinear modeling capabilities. However, CNNs have a limited effective receptive field, making it difficult to capture long-distance dependencies of large-scale features such as roads, rivers, and mountains in remote sensing images, and resulting in poor structural reconstruction of complex scenes.

[0005] 2. Transformer-based methods: Represented by SwinIR and SRFormer, these methods capture long-range dependencies through self-attention mechanisms, significantly improving global modeling capabilities. However, the computational complexity of self-attention is quadratic with the input size. Remote sensing images are generally large in size, leading to extremely high memory consumption and computational overhead, making them difficult to deploy on resource-constrained platforms.

[0006] 3. Mamba-based methods: Represented by MambaIR and FMSR, these methods achieve long sequence modeling with linear complexity based on selective state-space models, outperforming Transformers in computational efficiency. However, directly applying existing Mamba models to remote sensing image super-resolution still has significant drawbacks: Single-scale scanning mechanisms are difficult to adapt to the characteristics of multi-scale ground features coexisting in remote sensing images, and cannot simultaneously take into account large-scale structure modeling and fine-grained texture restoration. Mamba is biased towards modeling global low-frequency information and is not good at capturing local high-frequency texture details. The reconstruction results are prone to oversmoothing and blurring of ground object edges and textures. Traditional spatial domain channel attention mechanisms are poorly adapted to the Mamba architecture, failing to effectively distinguish high-frequency details from noise, and making it difficult to enhance effective textures while suppressing noise.

[0007] Furthermore, while some existing solutions incorporate frequency domain processing to aid image restoration, they often directly perform frequency domain transformations on the entire image without deeply integrating them with Mamba's state space modeling. They also lack multi-band adaptive attention mechanisms tailored to the frequency distribution characteristics of remote sensing images, resulting in limited performance improvements. Therefore, there is an urgent need to design a Mamba super-resolution network that adapts to the characteristics of remote sensing images and balances reconstruction quality with computational efficiency. Summary of the Invention

[0008] The purpose of this invention is to provide a remote sensing image restoration network that integrates frequency domain channel attention and a global-local dual-branch architecture. This network overcomes the shortcomings of existing technologies, such as insufficient multi-scale feature capture, severe loss of local high-frequency details, poor compatibility between traditional attention and Mamba, and overly smooth reconstruction results that are difficult to balance between computational efficiency and reconstruction quality when applied to remote sensing image super-resolution. The invention provides a remote sensing image super-resolution restoration network and control method that integrates frequency domain channel attention and a global-local dual-branch architecture. While maintaining the advantages of Mamba's efficient long-range modeling, it enhances the ability to model multi-scale land features and restore local details. At the same time, it achieves a balance between noise suppression and detail enhancement through frequency domain attention, thereby realizing high-quality and high-efficiency super-resolution reconstruction of remote sensing images.

[0009] To achieve the above objectives, the present invention provides the following technical solution: a remote sensing image restoration network that integrates frequency domain channel attention and global-local dual branches, comprising a shallow feature extraction module, a deep feature extraction module, and a high-resolution reconstruction module connected in sequence; The deep feature extraction module consists of N stacked feature aggregation Mamba groups, and each feature aggregation Mamba group consists of M concatenated feature aggregation Mamba blocks, where N and M are both positive integers greater than or equal to 2. The feature aggregation Mamba block includes a channel splitting unit, a global multi-scale Mamba branch, a local detail enhancement branch, and a frequency domain channel attention unit. The channel splitting unit is used to divide the input feature map into two features along the channel dimension, which are respectively input into the global multi-scale Mamba branch and the local detail enhancement branch. The global multi-scale Mamba branch employs a multi-scale visual state space model, comprising an original resolution single-scan branch and a downsampled multi-directional scan branch. The original resolution single-scan branch performs unidirectional serialization and flattening on the original size feature map, and extracts long-distance global dependencies through the S6 selective state space module. The downsampled multi-directional scan branch performs spatial downsampling on the feature map, serializing and flattening it along the horizontal, vertical, and diagonal directions respectively. After processing by the shared S6 selective state space module, the two-dimensional features are restored and upsampled to the original resolution. The output features of the two branches are added element-wise to obtain the global multi-scale features. The local detail enhancement branch uses a depthwise separable convolutional architecture to extract local spatial texture features, which are then processed by pointwise convolution and non-linear activation mapping to obtain local enhancement features. After fusing global multi-scale features and local enhancement features, the fused features are input into the frequency domain channel attention unit after layer normalization. The frequency domain channel attention unit decomposes the features into multiple frequency components covering low, medium, and high frequencies based on two-dimensional discrete cosine transform. For each frequency component, three types of statistical features are extracted: average pooling, max pooling, and min pooling. Channel attention weights are generated through convolution mapping, and adaptive recalibration of the channel dimension is performed on the fused features. The recalibrated features are added to the unfused features through residual connections, and the result is the output of the feature aggregation Mamba block. The high-resolution reconstruction module fuses the shallow features output by the shallow feature extraction module and the depth features output by the depth feature extraction module through residual connection. After being amplified to the target resolution by the pixel rearrangement upsampling layer, the reconstructed high-resolution remote sensing image is output.

[0010] As a remote sensing image restoration network that integrates frequency domain channel attention and global-local dual-branch architecture according to the present invention, preferably, the specific processing flow of the multi-scale visual state space model is as follows: the input feature map is fed into two parallel computing branches respectively: The first branch passes through a linear transformation layer, a depthwise separable convolutional layer, a SiLU activation function, a multi-scale two-dimensional scanning module, and a layer normalization layer in sequence, and outputs global semantic features. The second branch passes through a linear transformation layer and a SiLU activation function in sequence, and then multiplies the global semantic features output by the first branch element by element. After another linear transformation, the final global multi-scale features are output. The multi-scale two-dimensional scanning module integrates the original resolution single-scan branch and the downsampling multi-directional scanning branch.

[0011] As a remote sensing image restoration network that integrates frequency domain channel attention and global-local dual branches in this invention, preferably, the downsampling multi-directional scanning branch uses a depthwise separable convolution with a stride of 2 to downsample the feature map by 2 times, resulting in a dimensionality-reduced feature map with a sequence length of 1 / 4 of the original sequence. Perform row-first horizontal scanning, column-first vertical scanning, and main diagonal scanning on the dimensionality-reduced feature map to obtain three independent one-dimensional sequences, which share the same S6 selective state space module for sequence modeling. After modeling, the three sets of sequences are inversely transformed into two-dimensional feature maps. After element-wise addition, they are upsampled to the original resolution through bilinear interpolation and then added element-wise with the output features of the single-scan branch at the original resolution to complete multi-scale feature fusion.

[0012] As a remote sensing image restoration network that integrates frequency domain channel attention and global-local dual branches according to the present invention, preferably, the structure of the local detail enhancement branch is as follows: a 3×3 depth convolutional layer, a first 1×1 pointwise convolutional layer, a GELU activation function layer, and a second 1×1 pointwise convolutional layer. The 3×3 depth convolutional layer is used to independently extract local spatial texture features in each channel, and the two pointwise convolutional layers are used to realize feature transformation and nonlinear mapping in the channel dimension, thereby enhancing the expression of high-frequency details.

[0013] As a remote sensing image restoration network that integrates frequency domain channel attention and global-local bi-branch architecture according to the present invention, preferably, the processing steps of the frequency domain channel attention unit are as follows: S1: Perform global average pooling on the input feature map to compress the spatial dimension and obtain global statistical features in the channel dimension; S2: Select K preset two-dimensional discrete cosine transform frequency bases, project the global statistical features onto each frequency base, and obtain K channel feature vectors corresponding to different frequency bands; S3: Perform average pooling, max pooling, and min pooling on the channel feature vector of each frequency component to obtain three complementary frequency domain statistical features; S4: Aggregate the average statistical features of all frequency components to obtain three types of global frequency domain statistical features. Input these features into the attention generation network consisting of two 1×1 convolutional layers and output the three types of channel attention weights. S5: Add the three types of channel attention weights element by element to obtain the final channel attention vector, and multiply it element by element with the original input feature map to complete the channel recalibration of the features.

[0014] As a remote sensing image restoration network that integrates frequency domain channel attention and global-local bi-branch in this invention, preferably, the number K of the two-dimensional discrete cosine transform frequency basis is 16, covering low-frequency contour components, mid-frequency structure components and high-frequency detail components; in the attention generation network, the ReLU activation function is applied after the first convolution layer, and the Sigmoid activation function is applied after the second convolution layer to normalize the channel weights to the 0-1 range.

[0015] As a remote sensing image restoration network that integrates frequency domain channel attention and global-local bi-branch in this invention, preferably, the global multi-scale Mamba branch has residual connections: the input features are first processed by layer normalization, and then input into the multi-scale visual state space model. The model output is added element-wise to the input features before normalization to output global multi-scale features.

[0016] As a remote sensing image restoration network that integrates frequency domain channel attention and global-local bi-branch in this invention, preferably, the shallow feature extraction module is a 3×3 convolutional layer used to map the input three-channel low-resolution remote sensing image into a shallow feature map of the C channel. The deep feature extraction module has a 3×3 convolutional layer at the end, which is used to refine and integrate the features output by the stacked feature aggregation Mamba group. The high-resolution reconstruction module sequentially includes a 3×3 convolutional layer, a pixel rearrangement upsampling layer, and a 3×3 output convolutional layer. The upsampling factor of the pixel rearrangement upsampling layer is 2, 3, or 4.

[0017] As a remote sensing image restoration network that integrates frequency domain channel attention and global-local bi-branch architecture according to the present invention, preferably, the S6 selective state space module adopts an input-dependent dynamic state update mechanism, and its state update and output formulas are as follows: ; ; in, This is the hidden state from the previous moment. Let the input vector be the input vector at the current time. The state transition matrix is ​​a fixed, shared matrix. For dynamic step size parameters, , The input projection matrix and output projection matrix are dynamically changing with the input. This is the output vector at the current moment.

[0018] As a remote sensing image restoration network that integrates frequency domain channel attention and global-local bi-branching according to the present invention, preferably, in the deep feature extraction module, a cross-block residual connection is set between the input and output of each feature aggregation Mamba group; after the input features of the feature aggregation Mamba group are processed by all feature aggregation Mamba blocks in the group, they are added element-wise with the original input features and then passed to the next feature aggregation Mamba group to enhance gradient propagation and shallow feature reuse.

[0019] Compared with the prior art, the beneficial effects of the present invention are: This invention employs a dual-branch design with channel splitting. The global Mamba branch is responsible for capturing the long-distance structural dependencies of large-scale features, while the local detail enhancement branch focuses on strengthening high-frequency textures and edge information. The two complement each other at the feature level, solving the shortcomings of a single Mamba architecture that emphasizes global low frequencies and results in overly smooth reconstructions, and significantly improving the restoration of texture details in remote sensing images.

[0020] This invention proposes a multi-scale visual state space model employing a dual-path design of "preserving details at original resolution + improving efficiency through downsampling": the original resolution branch retains fine-grained spatial information, while the downsampling branch reduces the sequence length to one-quarter of its original length, significantly reducing the computational overhead and peak memory usage of state space scanning. Simultaneously, multi-directional scanning enhances the ability to capture multi-dimensional global dependencies. Compared to the single-scale Mamba model, this approach improves the accuracy of multi-scale ground feature reconstruction while reducing computational resource consumption.

[0021] This invention utilizes multi-frequency decomposition attention based on 2DDCT to distinguish low-frequency contours, mid-frequency structures, and high-frequency details in the frequency domain. It integrates three types of pooling statistics to comprehensively capture the frequency domain feature distribution and adaptively adjusts the weights of each channel. Compared to traditional spatial domain attention, it is more suitable for the frequency distribution characteristics of remote sensing images, effectively distinguishing high-frequency details from noise. It complements Mamba's global modeling, improving the structural fidelity of the reconstructed image.

[0022] This invention, through a combination of multi-scale scanning and channel splitting, enables the network to improve reconstruction accuracy while reducing the number of parameters and peak memory consumption compared to existing mainstream Mamba super-resolution models. The computational complexity maintains a linear growth, making it more suitable for large-size remote sensing image processing and resource-constrained deployment scenarios, thus achieving a good balance between reconstruction quality and computational efficiency. Attached Figure Description

[0023] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 A schematic diagram of the overall architecture of a remote sensing image restoration network that integrates frequency domain channel attention and global-local bi-branch architecture.

[0024] Figure 2 A visualization of the feature aggregation Mamba block (FAMB) of a remote sensing image restoration network that integrates frequency domain channel attention with global-local bi-branch architecture.

[0025] Figure 3 A schematic diagram of the structure of a multi-scale visual state space model (MVSSM) for remote sensing image restoration networks that integrate frequency domain channel attention with global-local bi-branch architecture.

[0026] Figure 4 A feature visualization of the frequency domain channel attention (FDCA) mechanism in a remote sensing image restoration network that integrates frequency domain channel attention with global-local bi-branch architecture.

[0027] Figure 5 This image shows a comparison of the reconstruction performance of a remote sensing image restoration network that integrates frequency domain channel attention and global-local bi-branch architecture with existing models on remote sensing datasets. Detailed Implementation

[0028] 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.

[0029] Please see Figure 1-5 The present invention provides the following technical solution: a remote sensing image restoration network that integrates frequency domain channel attention and global-local bi-branch, including the following embodiments: Example 1: Overall Network Architecture like Figure 1 As shown, the remote sensing image super-resolution restoration network that integrates frequency domain channel attention and global-local dual branches provided in this embodiment is divided into three stages: shallow feature extraction, deep feature extraction, and high-resolution reconstruction.

[0030] 1. Shallow Feature Extraction Stage Input low-resolution remote sensing images Where H and W represent the height and width of the image, respectively. A shallow feature map is obtained by performing shallow feature extraction through a 3×3 convolutional layer. C represents the number of feature channels; in this embodiment, C=180. The process can be represented as: ; 2. Deep Feature Extraction Stage Shallow feature map The input consists of a deep feature extraction module comprising six sequentially stacked Feature Aggregates Mamba Groups (FAMGs), each FAMG containing six concatenated Feature Aggregates Mamba Blocks (FAMBs). After feature extraction through multiple levels of FAMBs, a 3×3 convolutional layer is used for feature refinement to obtain the deep feature map. This process can be represented as: ; 3. High-resolution reconstruction stage depth features shallow features Residual connections are achieved through element-wise addition, enhancing the stability of feature propagation. Subsequently, the image passes through a 3×3 convolutional layer, a pixel rearrangement upsampling layer (upsampling factor of 4), and a 3×3 output convolutional layer to obtain the final high-resolution image. This process can be represented as: ; Example 2: Feature Aggregation Mamba Block (FAMB) like Figure 2 As shown, each FAMB consists of four parts: a channel splitting unit, a global multi-scale Mamba branch, a local detail enhancement branch, and a frequency domain channel attention unit. The specific execution flow is as follows: 1. Channel splitting Input deep features First, a 1×1 convolution is applied to expand the channels, and then the channels are evenly divided into two paths: Input global multi-scale Mamba branches, Input a local detail enhancement branch.

[0031] ; 2. Global branch handling The global branch uses a residual structure: first, for Layer normalization (LN) is performed, and then a multi-scale visual state space model (MVSSM) is input for global modeling. The output is compared with that before normalization. By adding elements one by one, we can obtain the global features. .

[0032] ; 3. Local branch processing The local branches sequentially pass through 3×3 depthwise convolution, 1×1 pointwise convolution, GELU activation function, and 1×1 pointwise convolution to extract and enhance local texture details, resulting in local enhanced features. .

[0033] ; 4. Feature fusion and frequency domain calibration The fused feature is obtained by adding global and local features element by element. After layer normalization, the input is given to the frequency domain channel attention unit (FDCA) for recalibration, and the output is the same as... The residuals are summed to obtain the final output of FAMB. .

[0034] ; Example 3: Multi-scale Visual State Space Model (MVSSM) like Figure 3 As shown, the MVSSM adopts a gated dual-branch structure, with the core being a multi-scale two-dimensional scanning (MS2D) module. The specific process is as follows: 1. Two parallel branches Input features Divided into two routes: First branch: undergoes linear transformations sequentially. 1×1 depthwise separable convolution, SiLU activation function, MS2D module, layer normalization, output features .

[0035] The second branch: undergoes linear transformations sequentially. The SiLU activation function is used to output gated features. The two features are multiplied element-wise and then subjected to a linear transformation. Output the final global features .

[0036] ; ; in This indicates element-wise multiplication.

[0037] 2. MS2D Multiscale Scanning Module Includes original resolution branches and downsampling multi-directional branches: Original resolution branch: The original size feature map is obtained by using a depthwise convolution with a stride of 1. After flattening the data into a one-dimensional sequence by row, input it into the S6 module for modeling, and then restore it to a two-dimensional feature. .

[0038] Downsampling multi-directional branch: A depthwise convolution with a stride of 2 is used to obtain a 2x downsampling feature map. The sequences are flattened into three sets of one-dimensional sequences along the horizontal, vertical, and diagonal directions, respectively. They share the same S6 module for modeling, are restored to two-dimensional features, and then added element by element. Finally, they are upsampled to the original resolution using bilinear interpolation to obtain downsampled fused features. The features of the two branches are added element by element to obtain the output of MS2D.

[0039] ; This design reduces the sequence length of traditional four full-resolution scans from 4L to 1.75L, significantly reducing computational load and memory consumption, while retaining fine-grained details and multi-directional global modeling capabilities.

[0040] Example 4: Frequency Domain Channel Attention Unit (FDCA) like Figure 4As shown, FDCA implements channel attention in the frequency domain based on 2DDCT. The specific steps are as follows: 1. Global Pooling and Frequency Domain Decomposition For input features Global average pooling is performed to compress the spatial dimension. Sixteen pre-defined 2DDCT frequency bases (covering low, medium, and high frequencies) are selected, and the pooled features are projected onto each frequency base to obtain channel feature vectors for 16 frequency components. K=1,2,...,16.

[0041] ; in Let $\mathbf{k}$ be the basis function for the k-th DCT, and $\mathbf{a}$ be the average pooling function.

[0042] 2. Extraction of multiple statistics For each frequency component Average pooling, max pooling, and min pooling are performed respectively to obtain three types of frequency domain statistical features, comprehensively capturing frequency domain distribution information.

[0043] 3. Attention Weight Generation: The statistical features of all frequency components of the same type are averaged and aggregated to obtain three categories of global statistical features. Each category of features is then processed through two 1×1 convolutional layers (the first layer followed by ReLU, and the second layer by Sigmoid) to generate three categories of channel attention weights. The three weights are then added element-wise to obtain the final channel attention vector. .

[0044] 4. Feature recalibration involves element-wise multiplying the attention vector M with the original input features F to obtain the recalibrated features. This completes the channel enhancement for frequency domain sensing.

[0045] .

[0046] Summary of the process and principles of the invention control method The core process principle of this invention can be summarized as "shallow initial extraction - dual-branch global-local joint modeling - frequency domain adaptive calibration - residual upsampling reconstruction": Convolutional layers are used to quickly extract shallow spatial features of images, providing a foundation for subsequent depth modeling; Parallel decoupled modeling of global and local features is achieved through channel splitting: the global branch adopts a multi-scale state space model to efficiently capture multi-scale long-distance dependencies in a way that "preserves original details and improves efficiency through downsampling"; the local branch enhances high-frequency texture details through depthwise separable convolution, and the two complement each other to solve the Mamba oversmoothing problem. By introducing frequency domain channel attention, channel recalibration of fused features is performed from the frequency dimension, which adaptively enhances object features, suppresses noise, and improves reconstruction fidelity. By leveraging residual connections to fuse features from both shallow and deep layers, and through pixel rearrangement to achieve efficient upsampling, a high-resolution remote sensing image is ultimately output.

[0047] While maintaining the advantages of Mamba's linear computational complexity, the overall process specifically addresses the pain points of difficulty in multi-scale ground feature modeling and easy loss of high-frequency details in remote sensing images, achieving a dual improvement in quality and efficiency.

[0048] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A remote sensing image restoration network that integrates frequency domain channel attention and global-local bi-branch architecture, characterized in that, It includes a shallow feature extraction module, a deep feature extraction module, and a high-resolution reconstruction module connected in sequence; The deep feature extraction module consists of N stacked feature aggregation Mamba groups, and each feature aggregation Mamba group consists of M concatenated feature aggregation Mamba blocks, where N and M are both positive integers greater than or equal to 2. The feature aggregation Mamba block includes a channel splitting unit, a global multi-scale Mamba branch, a local detail enhancement branch, and a frequency domain channel attention unit. The channel splitting unit is used to divide the input feature map into two features along the channel dimension, which are respectively input into the global multi-scale Mamba branch and the local detail enhancement branch. The global multi-scale Mamba branch employs a multi-scale visual state space model, comprising an original resolution single-scan branch and a downsampled multi-directional scan branch. The original resolution single-scan branch performs unidirectional serialization and flattening on the original size feature map, and extracts long-distance global dependencies through the S6 selective state space module. The downsampled multi-directional scan branch performs spatial downsampling on the feature map, serializing and flattening it along the horizontal, vertical, and diagonal directions respectively. After processing by the shared S6 selective state space module, the two-dimensional features are restored and upsampled to the original resolution. The output features of the two branches are added element-wise to obtain the global multi-scale features. The local detail enhancement branch uses a depthwise separable convolutional architecture to extract local spatial texture features, which are then processed by pointwise convolution and non-linear activation mapping to obtain local enhancement features. After fusing global multi-scale features and local enhancement features, the fused features are input into the frequency domain channel attention unit after layer normalization. The frequency domain channel attention unit decomposes the features into multiple frequency components covering low, medium, and high frequencies based on two-dimensional discrete cosine transform. For each frequency component, three types of statistical features are extracted: average pooling, max pooling, and min pooling. Channel attention weights are generated through convolution mapping, and adaptive recalibration of the channel dimension is performed on the fused features. The recalibrated features are added to the unfused features through residual connections, and the result is the output of the feature aggregation Mamba block. The high-resolution reconstruction module fuses the shallow features output by the shallow feature extraction module and the depth features output by the depth feature extraction module through residual connection. After being amplified to the target resolution by the pixel rearrangement upsampling layer, the reconstructed high-resolution remote sensing image is output.

2. The remote sensing image restoration network that integrates frequency domain channel attention and global-local bi-branch as described in claim 1, characterized in that: The specific processing flow of the multi-scale visual state space model is as follows: the input feature map is fed into two parallel computing branches: The first branch passes through a linear transformation layer, a depthwise separable convolutional layer, a SiLU activation function, a multi-scale two-dimensional scanning module, and a layer normalization layer in sequence, and outputs global semantic features. The second branch passes through a linear transformation layer and a SiLU activation function in sequence, and then multiplies the global semantic features output by the first branch element by element. After another linear transformation layer, the final global multi-scale features are output. The multi-scale two-dimensional scanning module integrates the original resolution single-scan branch and the downsampling multi-directional scanning branch.

3. The remote sensing image restoration network that integrates frequency domain channel attention and global-local bi-branch as described in claim 2, characterized in that: The downsampling multi-directional scanning branch uses a depthwise separable convolution with a stride of 2 to downsample the feature map by 2 times, resulting in a dimensionality-reduced feature map with a sequence length of 1 / 4 of the original sequence. Perform row-first horizontal scanning, column-first vertical scanning, and main diagonal scanning on the dimensionality-reduced feature map to obtain three independent one-dimensional sequences, which share the same S6 selective state space module for sequence modeling. After modeling, the three sets of sequences are inversely transformed into two-dimensional feature maps. After element-wise addition, they are upsampled to the original resolution through bilinear interpolation and then added element-wise with the output features of the single-scan branch at the original resolution to complete multi-scale feature fusion.

4. The remote sensing image restoration network fusing frequency domain channel attention and global-local bi-branch as described in claim 1, characterized in that: The structure of the local detail enhancement branch is as follows: a 3×3 depth convolutional layer, a first 1×1 pointwise convolutional layer, a GELU activation function layer, and a second 1×1 pointwise convolutional layer. The 3×3 depth convolutional layer is used to independently extract local spatial texture features in each channel, and the two pointwise convolutional layers are used to realize feature transformation and nonlinear mapping in the channel dimension to enhance the expression of high-frequency details.

5. The remote sensing image restoration network fusing frequency domain channel attention and global-local bi-branch as described in claim 1, characterized in that: The processing steps of the frequency domain channel attention unit are as follows: S1: Perform global average pooling on the input feature map to compress the spatial dimension and obtain the global statistical features of the channel dimension. S2: Select K preset two-dimensional discrete cosine transform frequency bases, project the global statistical features onto each frequency base, and obtain K channel feature vectors corresponding to different frequency bands; S3: Perform average pooling, max pooling and min pooling on the channel feature vector of each frequency component respectively to obtain three types of complementary frequency domain statistical features. S4: Aggregate the average statistical features of all frequency components to obtain three types of global frequency domain statistical features. Input these features into an attention generation network consisting of two 1×1 convolutional layers to output three types of channel attention weights. S5: Add the three types of channel attention weights element by element to obtain the final channel attention vector. Multiply the vector element by element with the original input feature map to complete the channel recalibration of the features.

6. The remote sensing image restoration network fusing frequency domain channel attention and global-local bi-branch as described in claim 5, characterized in that: The number of frequency bases K in the two-dimensional discrete cosine transform is 16, covering low-frequency contour components, mid-frequency structural components and high-frequency detail components; in the attention generation network, the ReLU activation function is applied after the first convolution, and the Sigmoid activation function is applied after the second convolution to normalize the channel weights to the 0-1 range.

7. The remote sensing image restoration network fusing frequency domain channel attention and global-local bi-branch as described in claim 1, characterized in that: The global multi-scale Mamba branch has residual connections: the input features are first processed by layer normalization, then input into the multi-scale visual state space model, and the model output is added element by element to the input features before normalization to output global multi-scale features.

8. The remote sensing image restoration network fusing frequency domain channel attention and global-local bi-branch as described in claim 1, characterized in that: The shallow feature extraction module is a 3×3 convolutional layer used to map the input three-channel low-resolution remote sensing image into a shallow feature map of the C channel. The deep feature extraction module has a 3×3 convolutional layer at the end, which is used to refine and integrate the features output by the stacked feature aggregation Mamba group. The high-resolution reconstruction module sequentially includes a 3×3 convolutional layer, a pixel rearrangement upsampling layer, and a 3×3 output convolutional layer. The upsampling factor of the pixel rearrangement upsampling layer is 2, 3, or 4.

9. The remote sensing image restoration network fusing frequency domain channel attention and global-local bi-branch as described in claim 1, characterized in that: The S6 selective state space module adopts an input-dependent dynamic state update mechanism, and its state update and output formulas are as follows: ； ； in, This is the hidden state from the previous moment. Let the input vector be the input vector at the current time. The state transition matrix is ​​a fixed, shared matrix. For dynamic step size parameters, , The input projection matrix and output projection matrix are dynamically changing with the input. This is the output vector at the current moment.

10. The remote sensing image restoration network fusing frequency domain channel attention and global-local bi-branch as described in claim 1, characterized in that: In the deep feature extraction module, a cross-block residual connection is set between the input and output of each feature aggregation Mamba group. After the input features of the feature aggregation Mamba group are processed by all feature aggregation Mamba blocks in the group, they are added element by element to the original input features and then passed to the next feature aggregation Mamba group to enhance gradient propagation and shallow feature reuse.

Images