An image super-resolution reconstruction method and system based on double feature clustering

Abstract

The application discloses an image super-resolution reconstruction method and system based on double feature clustering, belongs to the technical field of image processing, and is used for solving the problems of low global receptive field acquisition efficiency of an existing CNN architecture, large calculation overhead of a Transformer architecture, and details and background interference and global information fragmentation of an existing lightweight Transformer architecture, and is mainly used for realizing efficient and high-quality reconstruction of a low-resolution image into a high-resolution image and suitable for relevant fields such as image processing and computer vision, and the method comprises the following steps: a shallow feature extraction module is used for mapping a low-resolution image from a three-dimensional color space to a high-dimensional space, and outputting initial serialized semantic features; a deep feature extraction module is composed of a plurality of backbone modules in cascade, and is internally provided with a double feature clustering unit, a local self-attention unit, a global guide cross-attention unit and a local region self-attention unit.

Description

Technical Field

[0001] This invention belongs to the field of image processing technology, specifically relating to an image super-resolution reconstruction method and system based on dual feature clustering. Background Technology

[0002] In the field of deep learning-based image super-resolution, the current mainstream technical solutions are mainly based on two major architectures: Convolutional Neural Networks (CNN) and Visual Transformers.

[0003] Convolutional neural networks (CNNs) rely on sliding window operations of local convolutional kernels to extract features. Their inherent inductive bias enables them to efficiently capture local texture and edge information. However, in order to obtain the global receptive field, CNNs must stack multiple layers of convolutions or use large kernel convolutions. This approach is inefficient and makes it difficult to directly establish dependencies between distant pixels. For high-frequency details (such as narrow edges and large-scale textures) that need to be recovered using long-range structural information, CNNs struggle to maintain their continuity and integrity, easily resulting in blurring or breakage.

[0004] Visual Transformers use a self-attention mechanism, which can calculate the association weight between any two image patches in a sequence. They inherently possess powerful global modeling capabilities. However, the computational complexity of their self-attention mechanism is proportional to the square of the length of the input sequence, resulting in huge computational and memory overhead when processing high-resolution images. This makes it difficult to deploy in real time on edge devices with limited computing resources. To balance global modeling capabilities with lightweight deployment requirements, lightweight Transformer architectures such as CATANet have emerged in recent years.

[0005] However, from a structural perspective, the token aggregation of the CATA module in this architecture relies solely on the semantic features implicitly learned by the network, without introducing explicit image structure priors such as gradient magnitude and direction. This leads to mutual interference between details and background content, causing subsequent intra-group self-attention to introduce background smoothing information into edge reconstruction, resulting in edge blurring and loss of texture details. On the other hand, the global center of the IRCA module is a static matrix, with each center isolated from the others and without information interaction, failing to form a global center that reflects the overall scene structure, leading to fragmentation of global information transmitted through cross-attention. Summary of the Invention

[0006] The purpose of this invention is to provide an image super-resolution reconstruction method and system based on dual feature clustering to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: an image super-resolution reconstruction method based on dual feature clustering, comprising the following steps: S1. Convert the original low-resolution image Input the shallow feature extraction module, map it from the three-dimensional color space to a high-dimensional space, and output the initial sequential semantic features of the image. ; S2. The initial serialized semantic features are fed into the deep feature interaction backbone consisting of multiple backbone modules. Before each feature routing, the macroscopic semantic identifier and microscopic HOG physical edge features of the current input features are calculated respectively. Based on the semantic information and HOG features, the full image features are jointly clustered. Then, global one-dimensional sorting and equidistant segmentation are performed to accurately divide the image into multiple local subgroups that maintain semantic consistency and edge smoothness. S3. Within the deep feature interaction backbone, based on the aforementioned segmented local subgroups and the current feature, a dual-path attention interaction is performed. On one hand, within the same group, linear projection is used to generate Query, Key, and Value, and local self-attention calculation is performed to output sharp and high-fidelity local aggregated features. ; On the other hand, global guided cross-attention is used to evolve the global center, and cross-attention is performed with the input features to output the result. Then, the two feature streams are fused in situ using residual fusion to output deep features. Then output through a local region self-attention module. ; S4. After processing by multiple deep feature extraction modules, the final features will be obtained. The input is fed to the tail-end dual-branch feature enhancement module. The dual-branch feature enhancement module will then input the final features... The traffic is split into two branches for separate processing. This module processes deep features in two ways: one way extracts and enhances the low-frequency skeleton features of the image. Another path is used to extract and enhance sharp high-frequency details. ; S5, the enhanced version and The residuals are added and fused, then fed into a sub-pixel convolutional layer for scaling, reconstructing and outputting the final high-resolution image. .

[0008] As a preferred implementation, the macro-semantic identifier in step S2 is obtained by: performing cosine similarity measurement on the features after layer normalization and the preset global center pool to obtain a similarity matrix; and extracting the semantic cluster index corresponding to each image token through the argmax operation as the macro-semantic identifier.

[0009] As a preferred implementation, the microscopic HOG physical edge features in step S2 are obtained by: using the Sobel operator to calculate the horizontal and vertical gradients of the input feature map, obtaining the gradient magnitude based on the horizontal and vertical gradients, and normalizing the gradient magnitudes to obtain the microscopic HOG physical edge features.

[0010] As a preferred implementation, the joint clustering in step S2 specifically involves: weighted fusion of macroscopic semantic identifiers and microscopic HOG physical edge features to construct a joint localization guidance value; global one-dimensional sorting of image tokens based on the joint localization guidance value; and then equal segmentation according to a fixed group size to obtain local subgroups with consistent internal semantics and smooth edges.

[0011] As a preferred implementation, the global guided cross-attention execution process in step S3 is as follows: first, perform center self-attention on the original global cluster centers to generate a high-order center matrix containing the global context; then, map the image feature sequence to a query matrix, and map the high-order center matrix to a key matrix and a value matrix respectively, and calculate the global aggregated features through cross-attention.

[0012] In a preferred implementation, during the local self-attention calculation in step S3, the current subgroup is concatenated with the features of the next adjacent group to expand the perception range. Then, the query, key, and value matrix is ​​calculated based on the concatenated features to complete the local self-attention operation within the group.

[0013] As a preferred implementation, the dual-branch feature enhancement module in step S4 includes a structural branch and a detail branch: the structural branch sequentially extracts and enhances low-frequency skeleton features through 3×3 convolution, pixel rearrangement downsampling, 1×1 convolution fusion, and pixel rearrangement upsampling; The detail branch extracts and enhances high-frequency sharp detail features through two layers of 3×3 convolution and the GCSA spatial channel attention module.

[0014] This invention also discloses an image super-resolution reconstruction system based on dual-feature clustering, comprising: The shallow feature extraction module is used to map low-resolution images from three-dimensional color space to high-dimensional space and output initial serialized semantic features; The deep feature extraction module is composed of multiple cascaded main modules and has built-in dual feature clustering units, local self-attention units, global guided cross-attention units, and local region self-attention units, which are used to output enhanced deep features. A dual-branch feature enhancement module is used to enhance low-frequency skeleton features and high-frequency detail features respectively; The upsampling reconstruction module is used to scale up the enhanced features and output a high-resolution image.

[0015] In a preferred embodiment, the dual-feature clustering unit includes a semantic identifier extraction subunit, a gradient feature extraction subunit, and a joint sorting and segmentation subunit, which are used to fuse semantic information and physical edge information to achieve accurate clustering.

[0016] In a preferred embodiment, the global guided cross-attention unit includes a center self-attention submodule and a cross-attention submodule. The center self-attention submodule is used to realize information interaction between global centers, and the cross-attention submodule is used to inject global context into local features.

[0017] Compared with the prior art, the beneficial effects of the present invention are: This invention innovatively introduces gradient magnitude prior (Sobel operator) in the semantic clustering stage. This multi-feature joint clustering effectively separates high-frequency edge regions from low-frequency smooth regions, avoiding feature aliasing. In the subsequent intra-group self-attention (IASA) process, edge region information can fully interact, thereby reconstructing sharper and more realistic edge and texture details. Moreover, the gradient operator requires no additional training parameters, thus achieving lightweight design.

[0018] The GGCA module of this invention adds a center self-attention mechanism before performing cross-attention. This structure enables the originally isolated cluster centers to communicate with each other and pre-aggregate to form a "high-order center semantic graph" that reflects the overall scene structure. Since the number of centers M is much smaller than the total number of pixels N, this structure solves the problem of global information fragmentation with almost no increase in computational burden, and greatly improves the reconstruction fidelity of large-scale geometric structures (such as building outlines).

[0019] This invention designs a separate dual-branch processing architecture. The structural branch utilizes PixelUnshuffle to achieve receptive field expansion without information loss, effectively preserving the low-frequency skeleton. The detail branch introduces the GCSA module, which forces information to flow across heterogeneous channels through serial channel attention, channel shuffling, and spatial attention. This dual-path parallel structure design breaks the information silos of single attention and greatly enhances the model's ability to sharpen high-frequency details at the end of reconstruction. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the overall architecture of the present invention; Figure 2 This is a schematic diagram of semantic-structural joint clustering in this invention; Figure 3 This is a schematic diagram of the globally guided cross-attention mechanism of the present invention; Figure 4 This is a schematic diagram of the dual-branch feature enhancement module of the present invention. Detailed Implementation

[0021] The present invention will be further described below with reference to embodiments.

[0022] The following embodiments are used to illustrate the present invention, but should not be used to limit the scope of protection of the present invention. The conditions in the embodiments can be further adjusted according to specific conditions, and simple improvements to the method of the present invention under the premise of the concept of the present invention are all within the scope of protection claimed by the present invention.

[0023] Please see Figures 1-4 This invention provides an image super-resolution reconstruction system based on dual feature clustering, which mainly includes: a shallow feature extraction module, a deep feature extraction module, and a tail dual-branch feature enhancement and reconstruction module. The deep feature extraction module includes a dual feature clustering mechanism, an intra-group self-attention mechanism, a global guided cross-attention mechanism, and a local self-attention module. The present invention also discloses the overall method flow as follows: Step 1: Extract the original low-resolution image Input the shallow feature extraction module, map it from the three-dimensional color space to a high-dimensional space, and output the initial sequential semantic features of the image. .

[0024] Step 2: The initial features are fed into the deep feature interaction backbone, which consists of multiple cascaded backbone modules.

[0025] Before each feature routing, the macroscopic semantic identifier and microscopic HOG physical edge features of the current input features are calculated respectively, and the full-image features are jointly clustered based on the semantic information and HOG features; then global one-dimensional sorting and equidistant segmentation are performed to accurately divide the image into multiple local subgroups that maintain semantic consistency and edge smoothness.

[0026] Step 3: Within the deep feature interaction backbone, based on the aforementioned segmented local subgroups and the current feature, perform dual-path attention interaction. On one hand, within the same group, use linear projection to generate Query, Key, and Value, perform local self-attention calculation, and output sharp and faithful local aggregated features. ; On the other hand, Global Guided Cross-Attention (GGCA) is used to evolve the global center and perform cross-attention with the input features to output the result. .

[0027] The two feature streams are then fused in situ using residual fusion to output deep features. Then, the output is processed through a Local Region Self-Attention Module (LRSA). .

[0028] Step 4: After processing by multiple deep feature extraction modules, the final features will be obtained. The input is fed into the tail-end dual-branch feature enhancement module. This module will then obtain the final features. The traffic is split into two branches for separate processing. This module processes deep features in two ways: one way extracts and enhances the low-frequency skeleton features of the image. Another path is used to extract and enhance sharp high-frequency details. .

[0029] Finally, the enhanced and The residuals are added and fused, then fed into a sub-pixel convolutional layer (PixelShuffle) for scaling, reconstructing and outputting the final high-resolution image. .

[0030] Specifically, the shallow feature extraction module is as follows: Let the input low-resolution image be The shallow feature extraction module uses a Convolutional layers map images from a three-dimensional color space to a high-dimensional latent space: in, For input low-resolution images ， For convolution operations with 3×3 kernels, These are high-dimensional features extracted from shallow layers. is the feature dimension, where C = number of channels, H = height, and W = width.

[0031] Deep feature extraction: The deep feature extraction backbone consists of multiple THAB modules (such as...) Figure 2 )constitute.

[0032] The THAB module is the core basic unit for achieving deep feature interaction in this invention. Its meaning lies in performing hybrid attention computation on clustered features through a dual-path parallel architecture. This module integrates a joint clustering module, an intra-group local self-attention module (IASA), and a globally guided cross-attention module (GGCA). This design breaks the limitations of traditional single attention mechanisms, enabling features within the same module to achieve high-frequency interaction and sharpening of edge details through local branches, while also capturing contextual information of macroscopic geometric structures through global branches. Finally, the two feature paths are fused in situ residuals, achieving deep feature extraction that balances detail fidelity and structural coherence.

[0033] Each THAB module first performs joint semantic and structural clustering on the input features.

[0034] Semantic-structural joint clustering Features that have undergone layer normalization (LayerNorm) , with the preset global central pool Perform similarity measurement.

[0035] Calculate the similarity matrix D between each image token and the M cluster centers using cosine similarity: in, For the first i The token and the first j Cosine similarity of cluster centers For the i-th image token feature, For the first j Cluster center features Transpose of cluster center features.

[0036] Subsequently, the semantic cluster index corresponding to each pixel is extracted through the argmax operation. This index serves as a macroscopic semantic identifier for that pixel: in, For the semantic cluster index corresponding to the pixel, This refers to the operation that retrieves the index corresponding to the maximum value.

[0037] in, This ensures that pixels with similar semantic features (such as different parts of the same object) can be identified and assigned the same integer label.

[0038] Secondly, to capture the edge and texture information of the image, a physical gradient prior is introduced. The input feature map is restored to a two-dimensional spatial structure, and the horizontal gradient at each pixel location is calculated using an edge detection operator (such as the Sobel operator). Gx with vertical gradient Gy Thus, the gradient magnitude is obtained. G : in, This is the minimum tolerance value, designed to prevent numerical overflow or infinity in subsequent calculations when the gradient magnitude is zero in smooth image regions, thus ensuring the numerical stability of the feature extraction process. The gradient magnitude is then flattened and normalized to obtain the normalized gradient features. : in, This represents the global minimum value of the gradient magnitude. This represents the global maximum value of the gradient magnitude. This is the minimum tolerance value.

[0039] in, This feature reflects the structural saliency of a pixel in physical space. The larger the gradient value, the higher the probability that the pixel is located at an edge or in a region with complex texture.

[0040] In obtaining semantic identifiers With physical priors Then, a joint positioning guidance value is constructed through weighted fusion. This is used to guide the final cluster ranking: in, and These are the learnable weight coefficients that are dynamically updated during network training.

[0041] Then perform global sorting and subgrouping. Perform a global sort on the tokens to obtain the sort index. : in, To return indices in ascending / descending order, the sorted sequence is divided into several local subgroups of a fixed group size G: For the k-th local subgroup, For the first The feature vectors of image tokens, where G is a fixed group size and N is the total number of image tokens. Round up.

[0042] Global Guided Cross-Attention Branch (GGCA) GGCA module (such as) Figure 3 Injecting global contextual information into local features can improve the global consistency and structural fidelity of the reconstructed image.

[0043] Using the center self-attention mechanism to analyze the original cluster centers Perform global interaction to generate a high-order central matrix containing global context information. This matrix participates in subsequent cross-attention calculations: in, It is a high-order central matrix. As the original cluster center, For layer normalization operation, For self-attention operations, This is the output of global features.

[0044] Input image feature sequence X Mapped to query matrix Q , will the higher-order center Mapped to key matrices respectively K Sum matrix V : Where N is the total number of pixels and M is the number of centers. , , Let B be the learnable projective weight matrix, C be the channel dimension of the input features, and d be the dimension of the projected features in the attention calculation. The calculated weights are... : In the formula, This is a scaling factor used to prevent the gradient of the Softmax function from vanishing due to excessively large dot product values. This represents the transpose of matrix K.

[0045] Local self-attention branch (IASA) Within each group, to break the boundaries between groups, the current group is concatenated with the next group to form a group of length [length missing]. 2L The range of perception.

[0046] Let the first The characteristics of each subgroup are: Then its corresponding query, key, and value matrix is: in, For the first i+1 Subgroup features For the first i A query matrix, For the first i A sliding window key matrix, For the first i A sliding window value matrix, for The transpose of , where d is the dimension of the attention projection feature and c is the dimension of the feature channels. Softmax To normalize the similarity, For feature splicing operations, L The length of a single subgroup. 2L The feature length after sliding splicing. Output local features.

[0047] Then merge the output. : Where Z is the intermediate quantity for the fusion of global and local features. For convolution operations ， It is a feedforward convolutional network. This is for deep output features.

[0048] Output via Local Self-Attention Module (LRSA) : in It is a shared weight matrix across patches. For deep output features, MSA stands for multi-head self-attention operation.

[0049] Dual-branch feature enhancement (DBFE), dual-branch feature enhancement module (such as...) Figure 4 This is executed after the last layer of the backbone network and contains two branches for structural preservation and detail sharpening of deep features.

[0050] Structural branching: Features of the input after cascading processing through multiple THAB modules. Through a Convolution extraction of preliminary features: :

[0051] Then, by halving the spatial size and quadrupling the number of channels through pixel unshuffle, the features are decomposed into four interconnected "subgroups". in For a pixel inverse rearrangement operation with a scaling factor of 2, then using a convolution( The features are obtained by fusing the four subbands. : right Pixel shuffle restores the features to their original spatial dimensions, while simultaneously restoring the channel count and fusing them with the original input features.

[0053] in This represents a pixel rearrangement operation with a scaling factor of 2. This represents the features after pixel rearrangement and restoration. Characteristics representing skip connection branches, This represents the final feature, which is the feature of the input dual-branch module.

[0054] Represents the activation function of the Gaussian error linear unit. This is a 3x3 convolution operation. This represents the low-frequency characteristics of the final output.

[0055] Detail branch, the input features first go through a ( Convolution extracts preliminary edge features :

[0056] Then through another Convolution generates edge features :

[0057] Then it is fed into the GCSA (Spatial Channel Attention) module for enhancement. GCSA internally contains a serial combination of channel attention and spatial attention, and introduces channel shuffling to enhance cross-channel information interaction. Let the mapping of GCSA be... The output characteristics of the GCSA module It is expressed as follows: The output is related to After addition, the output characteristics of the high-frequency branch are obtained by GELU activation. :

[0058] Finally, the outputs of the low-frequency and high-frequency branches are concatenated along the channel dimension and then combined with the original input. The residuals are summed to obtain the texture-enhanced features. : In the formula For feature splicing operations, , These represent low-frequency and high-frequency features, respectively, indicating the fused features after channel splicing. B is the batch size, C is the number of channels in the input feature map, H is the height of the feature map, and W is the width of the feature map.

[0059] Finally, image upsampling and reconstruction are performed. For super-resolution tasks, a two-stage upsampling approach is used. Each stage expands the number of channels to four times the original number through a convolutional layer, and then the spatial size is magnified by two times through pixel shaving. in, and It consists of two levels of upsampling features. Features after texture enhancement This is a 3x3 convolution operation; for or For this task, only one convolution and pixel rearrangement are performed to obtain the upsampled features. : All upsampling was performed using the ReLU activation function.

[0060] Then the upsampled features are passed through a... Convolution maps back to RGB space to obtain the residual image: in For 3x3 convolution operations

[0061] Final high-resolution image ( ) is the residual image ( ) and basic image ( Add them together:

[0062] Existing clustering mechanisms rely solely on implicit semantics, which can easily lead to the aliasing of smooth backgrounds and high-frequency edges. This invention innovatively introduces a gradient magnitude prior (Sobel operator) in the semantic clustering stage. This multi-feature joint clustering effectively separates high-frequency edge regions from low-frequency smooth regions, avoiding feature aliasing. In the subsequent intra-group self-attention (IASA) process, edge region information can fully interact, thereby reconstructing sharper, more realistic edges and texture details. Furthermore, the gradient operator requires no additional training parameters, thus maintaining lightweight design.

[0063] To address the issue of isolated global centers and lack of macroscopic connections in existing methods, which leads to insufficient fidelity in the reconstruction of large-scale geometric structures (such as building outlines) due to breaks and distortions, the GGCA module of this invention adds a center self-attention mechanism before performing cross-attention. This structure enables the previously isolated cluster centers to communicate with each other, pre-aggregating to form a "high-order center semantic graph" reflecting the overall scene structure. Since the number of centers M is much smaller than the total number of pixels N, this structure solves the problem of global information fragmentation with almost no increase in computational burden, significantly improving the reconstruction fidelity of large-scale geometric structures (such as building outlines).

[0064] In this invention, the Dual-Driven Joint Clustering Module (JATA) utilizes the Sobel operator to calculate the gradient magnitude of the input features and extracts the semantic information of the input features. It then constructs a joint localization guide value through weighted fusion to guide the cluster ranking. This avoids the erroneous spread of smoothing information to edge regions by intra-group self-attention, thus improving the sharpness and clarity of reconstructed narrow edges and complex textures. The Global Guided Cross-Attention (GGCA) module includes a center self-attention submodule: it performs multi-head self-attention on the global center pool, enabling information exchange between centers and generating a high-order center matrix. Then, global-local cross-attention is performed: the higher-order centers are projected as keys K and values ​​V respectively, and cross-attention is calculated with the query Q of the local token, injecting the global context into the local features. This effectively overcomes the drawbacks of global feature fragmentation while ensuring low computational overhead, enabling the network to exhibit excellent structural coherence and physical fidelity when reconstructing macroscopic geometric targets.

[0065] The Dual-Branch Feature Enhancement (DBFE) structure comprises a parallel structural branch and a detail branch. The structural branch consists of sequentially connected 3×3 convolutional layers, pixel rearrangement downsampling layers, 1×1 convolutional fusion layers, and pixel rearrangement upsampling layers, forming residual connections with the input features. The detail branch contains convolutional layers and a GCSA spatial channel attention module. The outputs of the two branches are concatenated along the channel dimension and reduced in dimensionality by 1×1 convolutions before being integrated into the main residual flow.

[0066] In the tail structure of the network, this invention designs a separate dual-branch processing architecture. The structural branch utilizes PixelUnshuffle to achieve receptive field expansion without information loss, effectively preserving the low-frequency skeleton. The detail branch introduces a GCSA module, which forces cross-group information flow between heterogeneous channels through serial channel attention, channel shuffling, and spatial attention. This dual-path parallel structure design breaks down the information silos of single attention, greatly enhancing the model's ability to sharpen high-frequency details at the end of reconstruction.

[0067] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An image super-resolution reconstruction method based on dual-feature clustering, characterized in that, Includes the following steps: S1. Convert the original low-resolution image Input the shallow feature extraction module, map it from the three-dimensional color space to a high-dimensional space, and output the initial sequential semantic features of the image. ; S2. The initial serialized semantic features are fed into the deep feature interaction backbone consisting of multiple backbone modules. Before each feature routing, the macroscopic semantic identifier and microscopic HOG physical edge features of the current input features are calculated respectively. Based on the semantic information and HOG features, the full image features are jointly clustered. Then, global one-dimensional sorting and equidistant segmentation are performed to accurately divide the image into multiple local subgroups that maintain semantic consistency and edge smoothness. S3. Within the deep feature interaction backbone, based on the aforementioned segmented local subgroups and the current feature, a dual-path attention interaction is performed. On one hand, within the same group, linear projection is used to generate Query, Key, and Value, and local self-attention calculation is performed to output sharp and high-fidelity local aggregated features. ; On the other hand, global guided cross-attention is used to evolve the global center, and cross-attention is performed with the input features to output the result. Then, the two feature streams are fused in situ using residual fusion to output deep features. Then output through a local region self-attention module. ; S4. After processing by multiple deep feature extraction modules, the final features will be obtained. The input is fed into the tail-end dual-branch feature enhancement module, where the dual-branch feature enhancement module will finalize the features. The traffic is split into two branches for separate processing. One path is used to extract and enhance low-frequency skeleton features of the image. Another path is used to extract and enhance sharp high-frequency details. ; S5, the enhanced version and The residuals are added and fused, then fed into a sub-pixel convolutional layer for scaling, reconstructing and outputting the final high-resolution image. .

2. The image super-resolution reconstruction method based on dual feature clustering according to claim 1, characterized in that: The macro-semantic identifier mentioned in step S2 is obtained by performing cosine similarity measurement on the features after layer normalization and the preset global center pool to obtain a similarity matrix, and extracting the semantic cluster index corresponding to each image token through the argmax operation as the macro-semantic identifier.

3. The image super-resolution reconstruction method based on dual feature clustering according to claim 1, characterized in that: The microscopic HOG physical edge features mentioned in step S2 are obtained by: using the Sobel operator to calculate the horizontal and vertical gradients of the input feature map, obtaining the gradient magnitude based on the horizontal and vertical gradients, and normalizing the gradient magnitudes to obtain the microscopic HOG physical edge features.

4. The image super-resolution reconstruction method based on dual feature clustering according to claim 1, characterized in that: The joint clustering in step S2 specifically involves: weighted fusion of macroscopic semantic identifiers and microscopic HOG physical edge features to construct a joint localization guide value; global one-dimensional sorting of image tokens based on the joint localization guide value; and then equal segmentation according to a fixed group size to obtain local subgroups with consistent internal semantics and smooth edges.

5. The image super-resolution reconstruction method based on dual feature clustering according to claim 1, characterized in that: The global guided cross-attention execution process described in step S3 is as follows: First, perform center self-attention on the original global cluster centers to generate a high-order center matrix containing the global context; then, map the image feature sequence to the query matrix, and map the high-order center matrix to the key matrix and the value matrix respectively, and calculate the global aggregated features through cross-attention.

6. The image super-resolution reconstruction method based on dual feature clustering according to claim 1, characterized in that: In step S3, when calculating local self-attention, the current subgroup is concatenated with the features of the next adjacent group to expand the perception range. Then, the query, key, and value matrix is ​​calculated based on the concatenated features to complete the local self-attention operation within the group.

7. The image super-resolution reconstruction method based on dual feature clustering according to claim 1, characterized in that: The dual-branch feature enhancement module in step S4 includes a structural branch and a detail branch: the structural branch sequentially extracts and enhances low-frequency skeleton features through 3×3 convolution, pixel rearrangement downsampling, 1×1 convolution fusion, and pixel rearrangement upsampling; The detail branch extracts and enhances high-frequency sharp detail features through two layers of 3×3 convolution and the GCSA spatial channel attention module.

8. An image super-resolution reconstruction system based on dual-feature clustering, characterized in that, The image super-resolution reconstruction method based on dual-feature clustering as described in any one of claims 1-7 includes: The shallow feature extraction module is used to map low-resolution images from three-dimensional color space to high-dimensional space and output initial serialized semantic features; The deep feature extraction module is composed of multiple cascaded main modules and has built-in dual feature clustering units, local self-attention units, global guided cross-attention units, and local region self-attention units, which are used to output enhanced deep features. A dual-branch feature enhancement module is used to enhance low-frequency skeleton features and high-frequency detail features respectively; The upsampling reconstruction module is used to scale up the enhanced features and output a high-resolution image.

9. The image super-resolution reconstruction system based on dual feature clustering according to claim 8, characterized in that: The dual-feature clustering unit includes a semantic identifier extraction subunit, a gradient feature extraction subunit, and a joint sorting and segmentation subunit, which are used to fuse semantic information and physical edge information to achieve accurate clustering.

10. The image super-resolution reconstruction system based on dual feature clustering according to claim 8, characterized in that: The global guided cross-attention unit includes a center self-attention submodule and a cross-attention submodule. The center self-attention submodule is used to realize information interaction between global centers, and the cross-attention submodule is used to inject global context into local features.

Images