A light-weight single image super-resolution reconstruction method based on local-global feature collaborative enhancement perception

Abstract

The application discloses a kind of light-weight single image super-resolution reconstruction methods based on local-global feature collaborative enhancement perception, to solve the problem of insufficient use of high-low frequency clues, low feature fusion efficiency and difficult to balance structure consistency and visual fidelity of reconstructed image in single image super-resolution.The network uses double-flow heterogeneous architecture: local branch uses multiple types of differential convolution to explicitly encode image edge and texture prior, and enhances the ability of detail representation;Global branch effectively models long-range dependence and low-frequency semantic information by integrating local, cross-regional and global multi-level spatial self-attention mechanism.The frequency perception fusion module proposed in the application generates channel-specific spatial weights based on frequency characteristics, and adaptively fuses double-branch features, accurately balances structure preservation and detail enhancement requirements during feature fusion process, and effectively reduces feature redundancy.The super-resolution network of the application significantly reduces the model complexity, while maintaining the lightweight design, can generate high-resolution images with consistent structure, visual fidelity and rich details, and provides an effective solution for efficient and high-performance lightweight single image super-resolution reconstruction model design.

Description

Technical Field

[0001] This invention belongs to the field of image processing and imaging technology, and specifically relates to a lightweight single-image super-resolution reconstruction method based on local and global feature collaborative enhancement perception. Background Technology

[0002] In today's digital age, the demand for image data processing and optimization is growing. Image super-resolution reconstruction technology, as a key means to improve image quality, has received widespread attention. Traditional image super-resolution methods mainly rely on interpolation algorithms, such as bilinear interpolation and bicubic interpolation. Although these methods are computationally simple, they often lead to problems such as blurred image edges and loss of details, making it difficult to meet the needs of high-quality image reconstruction.

[0003] With the development of deep learning technology, Convolutional Neural Networks (CNNs) have gradually become the mainstream method in the field of image super-resolution reconstruction. CNNs can effectively extract local features of images by learning from a large number of low-resolution and high-resolution image pairs, thus achieving good super-resolution results. However, CNNs have certain limitations in processing global information, and their capture of long-range dependencies and global structural information in images is insufficient, which to some extent limits further performance improvement. In recent years, the Transformer architecture has achieved great success in the field of natural language processing, and its powerful self-attention mechanism can effectively capture global information and long-range dependencies. This advantage has prompted researchers to introduce Transformers into the field of image super-resolution reconstruction to compensate for the shortcomings of CNNs in global information processing. Transformer-based image super-resolution reconstruction methods model images through self-attention mechanisms, which can recover better global structural and detailed information of images, thus significantly improving super-resolution results. However, with the improvement of model performance, the complexity and computational cost of the model also increase dramatically. In practical applications, especially for resource-constrained scenarios such as mobile devices and embedded systems, these complex models face enormous challenges. Therefore, lightweight network design has become an important research direction in the field of image super-resolution reconstruction. Lightweight networks aim to enable models to run efficiently on resource-constrained devices by optimizing network structure, reducing the number of parameters and computational cost, without significantly reducing model performance. While current super-resolution reconstruction techniques have made some progress in CNN and Transformer-based model design, they still fall short in balancing high performance and lightweight design. Summary of the Invention

[0004] The main objective of this invention is to overcome the shortcomings of some existing technologies mentioned above and propose a lightweight single-image super-resolution reconstruction method based on a CNN-Transformer collaborative mechanism. It aims to combine the advantages of CNN and Transformer, while achieving a balance between high performance and low computational cost through innovative network structure design and optimization strategies, thus meeting the image super-resolution reconstruction needs of different application scenarios. This method can effectively utilize the high- and low-frequency cues implicit in low-resolution images to achieve efficient feature fusion and ensure the structural consistency and visual realism of the reconstruction results. This innovative method provides a new solution for the field of single-image super-resolution reconstruction, effectively balancing detail restoration and global structural rationality, and has broad application prospects and significant technical value.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] A lightweight single-image super-resolution reconstruction method based on local and global feature collaborative enhancement perception includes the following steps:

[0007] Step 1: Acquire low-resolution image

[0008] Step 2: Shallow Feature Extraction: Convolution operations are used to process the low-resolution image to extract shallow features. This process captures the image's basic texture and structural information, laying the foundation for subsequent deep feature extraction. The specific formula is:

[0009]

[0010] in, For the input low-resolution image, This is a 3×3 convolution operation. This is a shallow feature map.

[0011] Step 3: Deep Feature Extraction: Using A cascaded residual local-global cross-integration module and a convolutional operation are used to extract deep features from the shallow feature maps. This process aims to further extract detailed information from the image, especially high-frequency details, to improve the quality of the reconstructed image. The specific extraction method includes the following sub-steps:

[0012] Step 3.1: First, split the input feature map into the first sub-feature map along the channel dimension. Second sub-feature map Then on Simultaneously, 3×3 convolution and multiple difference convolution are performed collaboratively, and the results are summed and output after passing through an activation function. The multiple difference convolution collaborative computation includes parallel operations of central difference convolution, axial difference convolution, vertical difference convolution, and horizontal difference convolution. During the inference phase, 3×3 convolution and multiple difference convolution are combined into a single convolution computation to reduce computational overhead. Subsequently, [the process is repeated]. Max pooling, 1×1 convolution, and activation function processing are performed sequentially. Finally, the outputs of steps (2) and (3) are concatenated, and after 1×1 convolution, they are added to the original input feature map to enhance the expressive power of the features.

[0013] Step 3.2: To enhance the model's ability to interact across windows, this invention borrows the idea of ​​the Swing Transformer and introduces a Hybrid Attention Module (HAB) in the Local-Global Cross-Integration Module (LGCIM). The HAB works collaboratively through parallel window multi-head self-attention (W-MSA) and shifted window multi-head self-attention (SW-MSA) branches to explicitly capture the structural similarity of local regions and extract global information from the feature maps. Given input features, the HAB first divides them into two parts along the channel dimension, inputting them into the W-MSA branch and the SW-MSA branch for processing, respectively. The output feature maps of the two branches are then concatenated along the channel dimension, layer normalized (LN), and added to the original input (i.e., adding residual connections) to obtain intermediate features. .then, The data is fed into a multilayer perceptron (MLP) layer. The MLP output is then normalized (LN) and compared with... The summation (i.e., adding residual connections again) finally yields the output feature map of HAB. .in, and These represent the portions of the input feature map after partitioning into the W-MSA and SW-MSA branches, respectively. This design of the present invention can significantly improve the model's ability to perceive global information while maintaining computational efficiency, thereby achieving better performance in image super-resolution reconstruction tasks.

[0014] Step 3.3: Global Information-Guided Grid Attention Module (GG-GAB) or Structure-Guided Global Attention Module (SG-GAB): These modules aim to explicitly model the similarity across image regions with linear computational complexity, thereby improving the model's ability to perceive global information while maintaining computational efficiency. The GG-GAB module focuses on global information through sparsely sampled grid points, while the SG-GAB module utilizes adaptive average pooling to map the structural information of image patches to a low-dimensional space, where self-attention is computed. The following is a detailed description of these two modules:

[0015] The GG-GAB module focuses on global information through sparsely sampled grid points, its core being the efficient capture of similarities between different regions in an image. The specific steps are as follows: First, layer normalization is performed on the input: The input feature map is processed by layer normalization to stabilize the training process and improve the model's generalization ability. Next, grid partitioning is performed: The feature map is divided into multiple small blocks according to a preset grid interval, each block containing information about a local region. Then, global interaction features are generated; for each grid block, its maximum and average values ​​are extracted, and these features are concatenated and passed through a convolutional layer to generate global interaction features. This process captures global information between different grid blocks. Subsequently, multi-head self-attention is calculated. Based on the global interaction features, the attention weights of each grid block are calculated, and these weights are used to perform a weighted summation of the value vectors to obtain the attention features of each grid block. This process allows the model to pass information between different grid blocks, enhancing global perception. Finally, grid inverse operation and output are performed: The attention features of all grid blocks are restored to their original size and residually connected with the input feature map. Finally, the output of the module is obtained through further processing by a multilayer perceptron (MLP). The GG-GAB module focuses on global information through sparse sampling, avoiding the problem of computational complexity increasing with the square of the image size in traditional global self-attention mechanisms. This significantly improves the model's ability to perceive the overall structure of the image while maintaining high computational efficiency.

[0016] The SG-GAB module maps the structural information of image patches to a low-dimensional space using adaptive average pooling, and calculates self-attention within this space. Its core function is to efficiently capture global structural similarity. The specific steps are as follows: First, layer normalization is performed on the input feature map to stabilize the training process and improve the model's generalization ability. Next, window partitioning is performed, dividing the feature map into multiple non-overlapping window blocks, each containing information about a local region. Then, multi-scale structural feature extraction is performed; for each window block, multi-scale structural features are extracted using adaptive average pooling. These features reflect the structural information of the window block at different scales. Then, these multi-scale features are concatenated and compressed to obtain the structural statistics of each window block. Global attention is calculated based on the compressed structural statistics, calculating global attention weights, and using these weights to perform a weighted summation of the value matrix to obtain the global attention features. This process allows the model to efficiently capture global structural similarity in a low-dimensional space. Finally, the output is obtained: the calculated global attention features are residually connected to the input feature map, and then further processed by a multilayer perceptron (MLP) to obtain the module's output. The SG-GAB module efficiently computes global attention in a low-dimensional space through adaptive average pooling and multi-scale feature extraction, thereby significantly improving the model's ability to perceive the global structure of an image while maintaining computational efficiency.

[0017] Step 3.4: Based on the high-frequency feature map and low-frequency feature map obtained in Steps 3.1 and 3.2, first, the high-frequency feature map... Low-frequency characteristics Addition generates initial features Then on Split along the channel into For each subgraph, a two-dimensional discrete cosine transform (2D-DCT) is performed to extract frequency components. The frequency components are then concatenated and processed through a fully connected layer and an activation function to generate channel weights. At the same time, for Perform global average pooling and max pooling on the channel, and then concatenate the results to generate spatial weights via convolutional layers and activation functions. .Will and Generate an initial spatial importance graph by multiplying channel by channel. Subsequently, on and Perform channel shuffling and grouped convolution to generate the final spatial weights. Finally, the modulation feature is calculated according to the formula = Weighted summation. The modulation characteristics are then compared with... , After addition, the features are output through convolution.

[0018] Step 4: Image Reconstruction: Shallow and deep feature maps are fused through convolution and subpixel convolution operations and upsampled to obtain the super-resolution reconstruction result. This process integrates the extracted features to generate the final high-resolution image. Advantages and positive effects of the present invention

[0019] This invention achieves lightweight single-image super-resolution reconstruction by synergistically enhancing local and global features. Specifically, the EDHFB module effectively extracts high-frequency detail features, the HAB module captures local and global information, the GG-GAB and SG-GAB modules explicitly model cross-region similarity of the image with linear computational complexity, and the FAFM module dynamically adjusts the spatial weights of each channel, thereby enhancing high-frequency details while maintaining low-frequency structural consistency. The synergistic effect of these modules enables this invention to effectively improve the quality of image super-resolution reconstruction while maintaining a lightweight model, especially excelling in the recovery of high-frequency details, and thus possessing high practical value. Attached Figure Description

[0020] Figure 1 The overall architecture diagram of the lightweight single-image super-resolution reconstruction method based on local and global feature collaborative enhancement perception provided by the present invention is shown below.

[0021] Figure 2This is a schematic diagram of the enhanced detail high-frequency feature extraction module constructed according to the present invention;

[0022] Figure 3 A schematic diagram of the hybrid attention module constructed in this invention;

[0023] Figure 4 A schematic diagram of the globally information-guided grid attention structure constructed for this invention;

[0024] Figure 5 A schematic diagram of the structure-guided global attention structure constructed for this invention;

[0025] Figure 6 A schematic diagram of the globally information-guided grid attention calculation constructed for this invention;

[0026] Figure 7 A schematic diagram of the structure-guided global attention calculation constructed for this invention;

[0027] Figure 8 A schematic diagram of the frequency-driven channel-specific attention mechanism structure in the frequency-aware fusion module constructed in this invention;

[0028] Figure 9 This is the high-resolution image reconstructed by super-resolution as constructed in this invention. Detailed Implementation

[0029] The embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0030] like Figure 1 As shown, the present invention provides a single-image super-resolution reconstruction method, comprising the following steps:

[0031] The main objective of this invention is to overcome the shortcomings of some existing technologies mentioned above and propose a lightweight single-image super-resolution reconstruction method based on a CNN-Transformer collaborative mechanism. It aims to combine the advantages of CNN and Transformer, while achieving a balance between high performance and low computational cost through innovative network structure design and optimization strategies, thus meeting the image super-resolution reconstruction needs of different application scenarios. This method can effectively utilize the high- and low-frequency cues implicit in low-resolution images to achieve efficient feature fusion and ensure the structural consistency and visual realism of the reconstruction results. This innovative method provides a new solution for the field of single-image super-resolution reconstruction, effectively balancing detail restoration and global structural rationality, and has broad application prospects and significant technical value.

[0032] To achieve the above objectives, the present invention adopts the following technical solution:

[0033] A lightweight single-image super-resolution reconstruction method based on local and global feature collaborative enhancement perception includes the following steps:

[0034] Step 1: Acquire low-resolution image

[0035] Step 2: Shallow Feature Extraction: Convolution operations are used to process the low-resolution image to extract shallow features. This process captures the image's basic texture and structural information, laying the foundation for subsequent deep feature extraction. The specific formula is:

[0036]

[0037] in, For the input low-resolution image, This is a 3×3 convolution operation. This is a shallow feature map.

[0038] Step 3: Deep Feature Extraction: Using A cascaded residual local-global cross-integration module and a convolutional operation are used to extract deep features from the shallow feature maps. This process aims to further extract detailed information from the image, especially high-frequency details, to improve the quality of the reconstructed image. The specific extraction method includes the following sub-steps:

[0039] Step 3.1: The enhanced detail high-frequency feature extraction module of this embodiment is as follows: Figure 2 As shown, the input feature map is first split into a first sub-feature map along the channel dimension. Second sub-feature map Then on Simultaneously, 3×3 convolution and multiple difference convolution are performed collaboratively, and the results are summed and output after passing through an activation function. The multiple difference convolution collaborative computation includes parallel operations of central difference convolution, axial difference convolution, vertical difference convolution, and horizontal difference convolution. During the inference phase, 3×3 convolution and multiple difference convolution are combined into a single convolution computation to reduce computational overhead. Subsequently, [the process is repeated]. Max pooling, 1×1 convolution, and activation function processing are performed sequentially. Finally, the outputs of steps (2) and (3) are concatenated, and after 1×1 convolution, they are added to the original input feature map to enhance the expressive power of the features.

[0040] Step 3.2: The hybrid attention module in this embodiment is as follows: Figure 3As shown, to enhance the model's ability to interact across windows, this invention borrows the idea of ​​the Swing Transformer and introduces a Hybrid Attention Module (HAB) in the Local-Global Cross-Integration Module (LGCIM). The HAB works collaboratively through parallel window multi-head self-attention (W-MSA) and shifted window multi-head self-attention (SW-MSA) branches to explicitly capture the structural similarity of local regions and extract global information from the feature maps. Given input features, the HAB first divides them into two parts along the channel dimension, inputting them into the W-MSA and SW-MSA branches for processing, respectively. The output feature maps of the two branches are then concatenated along the channel dimension, layer normalized (LN), and added to the original input (i.e., adding residual connections) to obtain intermediate features. .then, The data is fed into a multilayer perceptron (MLP) layer. The MLP output is then normalized (LN) and compared with... The summation (i.e., adding residual connections again) finally yields the output feature map of HAB. .in, and These represent the portions of the input feature map after partitioning into the W-MSA and SW-MSA branches, respectively. This design of the present invention can significantly improve the model's ability to perceive global information while maintaining computational efficiency, thereby achieving better performance in image super-resolution reconstruction tasks.

[0041] Step 3.3: Global Information-Guided Grid Attention Module (GG-GAB) or Structure-Guided Global Attention Module (SG-GAB): These modules aim to explicitly model the similarity across image regions with linear computational complexity, thereby improving the model's ability to perceive global information while maintaining computational efficiency. The GG-GAB module focuses on global information through sparsely sampled grid points, while the SG-GAB module utilizes adaptive average pooling to map the structural information of image patches to a low-dimensional space, where self-attention is computed. The following is a detailed description of these two modules:

[0042] The structure of the globally information-guided grid attention module in this embodiment is as follows: Figure 4 As shown, this attention mechanism is as follows Figure 6 As shown, the GG-GAB module focuses on global information through sparsely sampled grid points, and its core function is to efficiently capture the similarity between different regions in an image.

[0043] The specific steps are as follows: First, layer normalization is performed on the input: The input feature map is processed by layer normalization to stabilize the training process and improve the model's generalization ability. Next, grid partitioning is performed: The feature map is divided into multiple small blocks according to a preset grid interval, each block containing information about a local region. Then, global interaction features are generated; that is, for each grid block, its maximum and average value features are extracted, and these features are concatenated and passed through a convolutional layer to generate global interaction features. This process can capture global information between different grid blocks. Subsequently, multi-head self-attention is calculated. Based on the global interaction features, the attention weights of each grid block are calculated, and these weights are used to perform a weighted summation of the value vectors to obtain the attention features of each grid block. This process allows the model to pass information between different grid blocks, enhancing global perception capabilities. Finally, grid inverse operation and output are performed: The attention features of all grid blocks are restored to their original size and residually connected with the input feature map. Finally, the output of the module is obtained through further processing by a multilayer perceptron (MLP). The GG-GAB module focuses on global information through sparse sampling, avoiding the problem of computational complexity increasing with the square of the image size in traditional global self-attention mechanisms. This significantly improves the model's ability to perceive the overall structure of the image while maintaining high computational efficiency.

[0044] The structure of the structure-guided global attention module in this embodiment is as follows: Figure 5 As shown, this attention mechanism is as follows Figure 7 As shown. The SG-GAB module maps the structural information of image patches to a low-dimensional space through adaptive average pooling, and calculates self-attention in this space. Its core is to efficiently capture global structural similarity.

[0045] The specific steps are as follows: First, layer normalization is performed on the input feature map to stabilize the training process and improve the model's generalization ability. Next, window partitioning is performed, dividing the feature map into multiple non-overlapping window blocks, each containing information about a local region. Then, multi-scale structural feature extraction is performed; for each window block, multi-scale structural features are extracted using adaptive average pooling. These features reflect the structural information of the window block at different scales. Then, these multi-scale features are concatenated and compressed to obtain the structural statistics of each window block. Global attention is calculated based on the compressed structural statistics, calculating global attention weights, and using these weights to perform a weighted summation of the value matrix to obtain the global attention features. This process allows the model to efficiently capture global structural similarity in a low-dimensional space. Finally, the output is obtained: the calculated global attention features are residually connected to the input feature map, and then further processed by a multilayer perceptron (MLP) to obtain the module's output. The SG-GAB module efficiently computes global attention in a low-dimensional space through adaptive average pooling and multi-scale feature extraction, thereby significantly improving the model's ability to perceive the global structure of an image while maintaining computational efficiency.

[0046] Step 3.4: The frequency-driven channel-specific attention mechanism in the frequency-aware fusion module of this embodiment is as follows: Figure 8 As shown, based on the high-frequency feature map and low-frequency feature map obtained in steps 3.1 and 3.2, the high-frequency feature map is first... Low-frequency characteristics Addition generates initial features Then on Split along the channel into For each subgraph, a two-dimensional discrete cosine transform (2D-DCT) is performed to extract frequency components. The frequency components are then concatenated and processed through a fully connected layer and an activation function to generate channel weights. At the same time, for Perform global average pooling and max pooling on the channel, and then concatenate the results to generate spatial weights via convolutional layers and activation functions. .Will and Generate an initial spatial importance graph by multiplying channel by channel. Subsequently, on and Perform channel shuffling and grouped convolution to generate the final spatial weights. Finally, the modulation feature is calculated according to the formula = Weighted summation. The modulation characteristics are then compared with... , After addition, the features are output through convolution.

[0047] Step 4: Image Reconstruction: Shallow and deep feature maps are fused through convolution and subpixel convolution operations and upsampled to obtain the super-resolution reconstruction result. This process integrates the extracted features to generate the final high-resolution image.

[0048] This embodiment uses bicubic interpolation to obtain the corresponding low-resolution image from the original high-resolution image. The training set images are randomly rotated. , , Alternatively, the images can be horizontally flipped and then randomly cropped into 64x64 pixel blocks. Each training batch contains 64 randomly selected image blocks, with a total of 500K training iterations. The optimizer uses Adam (… , ), loss function used Loss. The initial learning rate is set to... The iteration count is halved when the number of iterations reaches the milestone of [250K, 400K, 450K, 475K]. The number of RLGCIMs is set to 4, with two LGCIMs in each RLGCIM. The first LGCIM consists of an EDHFB and a module composed of a cascaded HAB and GG-GAB, while the second LGCIM consists of an EDHFB and a module composed of a cascaded HAB and SG-GAB. This is used for shallow feature extraction. The number of filters is 48. The number of attention heads in HAB, GG-GAB, and SG-GAB are 6, 4, and 6, respectively. The sampling interval in GG-GAB is set to... The image patch size in SG-GAB is The method is implemented based on the PyTorch framework, and model training is performed using an NVIDIA GeForce RTX 4090 GPU.

[0049] like Figure 9 As shown, the super-resolution reconstruction results of the present invention are presented. The obtained super-resolution reconstruction results not only achieve performance improvement in terms of indicators, but also have very good visual effects.

[0050] In this invention, a low-resolution image is first acquired as input for super-resolution reconstruction. Then, convolutional operations are used to extract shallow features from the image, capturing basic texture and structural information. Subsequently, multiple cascaded residual local-global cross-integration modules are used to further mine deep features of the image. Each module includes three main components: Enhanced Detail High-Frequency Feature Extraction (EDHFB), Global Information Guided Mesh Attention (GG-GAB) or Structure Guided Global Attention (SG-GAB), and Frequency Aware Fusion (FAFM). EDHFB is used to capture detail information, GG-GAB and SG-GAB are used to explicitly model global information, and FAFM is used to fuse high-frequency and low-frequency features. Finally, shallow and deep features are fused and upsampled through convolutional and subpixel convolutional operations to obtain the super-resolution reconstruction result.

Claims

1. A lightweight single-image super-resolution reconstruction method based on local and global feature collaborative enhancement perception, characterized in that, include: Step 1: Acquire a low-resolution image; Step 2: Use convolution operations to extract shallow features from the low-resolution image to obtain a shallow feature map; Step 3: Use N cascaded residual local-global cross-integration modules and a convolution operation to extract deep features from the shallow feature map to obtain the deep feature map; Step 4: Fuse the shallow and deep feature maps through convolution and subpixel convolution operations and perform upsampling to obtain the super-resolution reconstruction result; In step 3, the residual local-global cross-integration module includes an enhanced detail high-frequency feature extraction (EDHFB) module for extracting local high-frequency features from the input feature map; a hybrid attention module (HAB) cascaded with a globally information-guided grid attention module (GG-GAB) or a structure-guided global attention module (SG-GAB) for extracting low-frequency global information from the input feature map; the results of the two feature maps are fused in the frequency-aware fusion module (FAFM); Output characteristics of each global cross-integration module Represented as: ,in, Indicates the first The frequency sensing fusion module in the global cross-integration module Indicates the first The enhanced detail high-frequency feature extraction module in the global cross-integration module Indicates the first Global information-guided grid attention modules or structure-guided global attention modules are included in the global cross-integration modules. Indicates the first Hybrid attention modules in a global cross-integration module; The enhanced detail high-frequency feature extraction module performs the following process: splitting the input feature map into a first sub-feature map along the channel dimension. Second sub-feature map ;right Simultaneously perform 3×3 convolution and multiple difference convolution collaborative computation, sum the results, and output them after passing them through an activation function; Max pooling, 1×1 convolution, and activation function processing are performed sequentially; the above outputs are concatenated, and after 1×1 convolution, they are added to the original input feature map; wherein, the multi-difference convolution collaborative computation includes parallel operations of central difference convolution, axial difference convolution, vertical difference convolution, and horizontal difference convolution; during the inference phase, 3×3 convolution and multi-difference convolution are combined into a single convolution computation; The frequency sensing fusion module performs the following process: It converts high-frequency features... Low-frequency characteristics Addition generates initial features ;right Split along the channel into For each subgraph, a two-dimensional discrete cosine transform is performed to extract frequency components; the frequency components are then concatenated and processed through a fully connected layer and an activation function to generate channel weights. ;right Perform global average pooling and max pooling on the channel, and then concatenate the results to generate spatial weights via convolutional layers and activation functions. ;Will and Generate an initial spatial importance graph by multiplying channel by channel. ;right and Perform channel shuffling and grouped convolution to generate the final spatial weights. Modulation characteristics according to formula = Weighted summation; combining modulation features with , After addition, the features are output through convolution.

2. The lightweight single-image super-resolution reconstruction method based on local and global feature collaborative enhancement perception as described in claim 1, characterized in that, The shallow feature extraction formula in step 2 is as follows: ,in, The input is a low-resolution image. This is a 3×3 convolution operation. This is a shallow feature map.

3. The hybrid attention module according to claim 1, characterized in that, Execute the following process: The input feature map is split into the first branch. Second branch ;right Execute window multi-head self-attention W-MSA, for Perform Shift-Window Multi-Head Self-Attention (SW-MSA); concatenate the output features from step 2, normalize them, and add them to the original input features to generate intermediate features; The intermediate features are computed using a multilayer perceptron (MLP) and then normalized by layers before being added to the intermediate features for output. W-MSA uses a fixed window partitioning, while SW-MSA uses window cyclic shifting to achieve cross-window interaction.

4. The globally information-guided grid attention module according to claim 1, characterized in that, Perform the following process: Perform layer normalization on the input feature map; according to grid intervals Divide the feature map into Grid blocks; for each grid block, perform maximum channel extraction and average channel extraction, and after concatenation, generate global interactive features through 1×1 convolution; For each grid tile: the mapping generates a query vector. Key vector Value vector Based on global interaction features and Calculate the first attention weight, and then combine it with Calculate the second attention weights; then combine the second attention weights with... Multiplication generates grid block attention results; all grid block results are restored to their original size, processed by layer normalization and MLP, and then concatenated with the original input residuals before output; the maximum / average value of each channel is extracted using global max pooling and average pooling, respectively; a uniform sampling strategy is used for grid partitioning, with a grid spacing of... These are configurable parameters.

5. The structure-guided global attention module according to claim 1, characterized in that, Perform the following process: Perform layer normalization on the input feature tensor; divide the feature map into... indivual Window blocks; generate a query matrix for each window block. Key matrix Value matrix ;right and Parallel execution of multi-scale adaptive average pooling extracts structural features at resolution scales of 1×1, 2×2, and 3×3. Flattening and stitching multi-scale structural features to generate compressed structural statistics; based on the compressed structural statistics and... Calculate the global attention weights, and Attention features are generated by multiplication; the attention features are added to the original input, and then processed by layer normalization and MLP to output the results; the multi-scale pooling results are flattened and then concatenated along the channel dimension. The fusion of attention features and the original input is achieved using residual connections.

Images