Image Super-Resolution Reconstruction System and Method Based on Centralized Top-K Semantic Bucket Mamba Network
The image super-resolution reconstruction system based on the centrally refined Top-K semantic bucket Mamba network solves the problems of spatial partitioning incompatibility and high computational complexity in existing technologies, and achieves efficient recovery of high-frequency details and global structure of images.
Patent Information
- Application Number
- CN202611289009.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-24
- Publication Date
- 2026-11-17
AI Technical Summary
Existing image super-resolution reconstruction methods suffer from problems such as lack of content adaptability in fixed spatial partitioning, high computational complexity of global self-attention, and difficulty in balancing local details with global structure.
An image super-resolution reconstruction system based on a centrally refined Top-K semantic bucket Mamba network is adopted. The system describes the image structure and texture distribution through adaptive global prototyping, organizes similar tokens using Top-K sparse routing, and enhances edge and texture recovery by using dual sequential scanning and a global context-aware local self-attention module, thereby reducing computational complexity.
It achieves high-frequency details and global structure of images while reducing artifacts, maintaining clear edges and textures, and improving image reconstruction performance under the premise of efficient computation.
Smart Images

Figure CN123118366A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the fields of image processing, computer vision and deep learning technology, and specifically relates to an image super-resolution reconstruction system and method based on a center-refined Top-K semantic bucket Mamba network. Background Technology
[0002] Single-image super-resolution reconstruction aims to recover a high-resolution image with higher spatial resolution and richer texture details from a low-resolution image. Because degradation processes such as downsampling, blurring, and noise cause irreversible loss of high-frequency information, the same low-resolution image may correspond to multiple plausible high-resolution images. Therefore, single-image super-resolution reconstruction is a typical ill-posed inverse problem.
[0003] Existing convolutional neural networks primarily aggregate information within a fixed local neighborhood, enabling the extraction of local edges and textures. However, limited by the local receptive field, they struggle to directly establish dependencies between repetitive edges, lines, and textures that are far apart. Networks based on global self-attention can enhance long-distance feature interactions, but their computational and storage requirements typically increase quadratically with the number of image tokens, hindering high-resolution image processing and lightweight deployment.
[0004] To reduce computational complexity, some existing technologies use fixed windows, strips, or predefined scanning paths to segment image features. These segmentations are primarily determined by spatial location and are insensitive to input content. Therefore, they may assign tokens that are spatially distant but have similar textures or structures to different regions, while forcing weakly related tokens within the same region to interact. Although selective state-space models can propagate long-distance information with near-linear complexity, the propagation path of two-dimensional features, after being unfolded into a one-dimensional sequence in a fixed order, is still primarily determined by spatial location.
[0005] Furthermore, directly merging tokens or replacing the original tokens with a small number of fixed centers may result in the loss of high-frequency details required for image super-resolution reconstruction; hard allocation strategies are difficult to represent tokens at structural boundaries or with mixed texture properties; and using only local window self-attention is prone to lacking guidance on global structure and texture distribution.
[0006] Therefore, there is a need for an image super-resolution reconstruction method that can dynamically organize similar tokens based on the current input content, while taking into account cross-regional long-distance modeling, local high-frequency detail restoration, and computational efficiency. Summary of the Invention
[0007] This application addresses the problems of existing image super-resolution reconstruction methods, such as the lack of content adaptability in fixed spatial partitioning, high computational complexity of global self-attention, and difficulty in balancing local details and global structure. It provides an image super-resolution reconstruction system and method based on a centrally refined Top-K semantic bucket Mamba network.
[0008] A first aspect of this application provides an image super-resolution reconstruction system based on a centrally refined Top-K semantic bucket Mamba network, comprising:
[0009] Shallow feature extraction module: Performs convolution processing on the input low-resolution image to extract shallow features;
[0010] The deep feature extraction module includes multiple residual groups. The shallow features are fused with the outputs of the multiple residual groups through global residual connections to obtain the fused deep features. Each residual group includes a content-aware main branch and a local convolutional residual branch. The content-aware main branch includes a centrally refined Top-K semantic bucket Mamba module and a global context-aware local self-attention module.
[0011] Image reconstruction module: Based on the fused deep features, predict high-frequency residuals through convolution and pixel shuffle, and add the high-frequency residuals to the reference image obtained by interpolation and upsampling of the low-resolution image to obtain a super-resolution image.
[0012] A second aspect of this application provides an image super-resolution reconstruction method based on a centrally refined Top-K semantic bucket Mamba network, comprising the following steps:
[0013] Convolutional processing is performed on the input low-resolution image to extract shallow features;
[0014] The shallow features are fused with the outputs of multiple residual groups through global residual connections to obtain fused deep features; wherein, each residual group includes a content-aware main branch and a local convolutional residual branch, and the content-aware main branch includes a centrally refined Top-K semantic bucket Mamba module and a global context-aware local self-attention module.
[0015] Based on the fused deep features, high-frequency residuals are predicted through convolution and pixel shuffle, and the high-frequency residuals are added to the reference image obtained by interpolating and upsampling the low-resolution image to obtain a super-resolution image.
[0016] The beneficial effects of this application are as follows:
[0017] This application utilizes an input adaptive global prototype to describe the main structure and texture distribution of the current image, and propagates the global context on a compact prototype sequence through a central-level Mamba module. It employs Top-K sparse routing to organize spatially distant but content-similar image tokens into the same semantic bucket, enabling bucket-level Mamba to propagate information along content-related paths. A dual-order scanning approach, combining spatial order and similarity order, is used to balance geometric continuity and semantic proximity. The refined global prototype modulates the value branch of local window self-attention, enhancing the recovery of edges, corners, and repetitive textures. When the number of global prototypes M and the number of routes K are much smaller than the number of image tokens N, the overall computational cost is approximately linear with respect to N, avoiding the quadratic computational bottleneck of global self-attention. Attached Figure Description
[0018] Figure 1 The overall structure diagram of the centrally refined Top-K semantic bucket Mamba network provided in the embodiments of this application;
[0019] Figure 2 This is a flowchart illustrating the adaptive global prototype generation process provided in an embodiment of this application.
[0020] Figure 3 A flowchart illustrating the Top-K sparse routing and semantic bucket construction process provided in this application embodiment;
[0021] Figure 4 A flowchart of semantic bucket dual-sequential scanning provided for embodiments of this application;
[0022] Figure 5 A flowchart of value branch modulation in global context-aware local self-attention provided in an embodiment of this application. Detailed Implementation
[0023] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention. Other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are all within the scope of protection of this invention.
[0024] In this application, "image token" refers to a feature vector consisting of a spatial location in a two-dimensional feature map and its corresponding channel features; "global prototype" refers to a feature vector that compactly represents the main structure or texture pattern in the current input image; "semantic bucket" refers to a variable-length token set consisting of image tokens that select the same refined global prototype as one of the Top-K routing targets; and "Mamba module" refers to a network module that performs sequence modeling based on a selective state-space model.
[0025] like Figure 1 As shown, this application provides an image super-resolution reconstruction system based on a center-refined Top-K semantic bucket Mamba network, including a shallow feature extraction module, a deep feature extraction module, and an image reconstruction module.
[0026] Furthermore, embodiments of this application provide low-resolution images. The shallow feature extraction module extracts shallow features using a 3×3 convolutional layer. Subsequently, the deep feature extraction module models deep features using multiple residual groups and fuses the shallow features with the outputs of these residual groups using global residual connections. Finally, the image reconstruction module predicts high-frequency residuals through convolution and pixel shuffle, and adds them to the baseline image obtained through bilinear interpolation. The above process can be represented as follows:
[0027]
[0028]
[0029]
[0030] in, , and These represent shallow feature extraction, deep feature extraction, and image reconstruction operations, respectively. Indicates the magnification factor as Bilinear interpolation; and These represent shallow features and deep fusion features, respectively. This represents the final super-resolution image.
[0031] In this embodiment, the shallow feature extraction module performs convolution operations on the low-resolution image to obtain shallow features that retain color, edge, and low-frequency structural information.
[0032] In the deep feature extraction module, each residual group includes a content-aware main branch and a local convolutional residual branch.
[0033] Furthermore, the content-aware main branch sequentially performs center-refined Top-K semantic bucket Mamba processing and global context-aware local self-attention processing; the local convolutional residual branch preferably uses 3×3 convolution and ReLU activation function to provide stable local inductive bias. The outputs of the two branches are fused at the end of the residual group and added to the residual group input.
[0034] Furthermore, this embodiment, based on the centrally refined Top-K semantic bucket Mamba module, sequentially completes adaptive global prototype generation and central refinement, Top-K sparse routing and semantic bucket construction.
[0035] The specific implementation steps for adaptive global prototype generation and central refinement are as follows:
[0036] like Figure 2 As shown, suppose the input features of a certain residual group are... After flattening the spatial dimensions, we obtain the image token sequence:
[0037]
[0038] in, , For feature dimensions.
[0039] First, the two-dimensional input features are divided into... Divide the regions into non-overlapping regions and perform average pooling within each region to obtain... An initial global prototype with uniform spatial distribution is used. This initialization method ensures that the initial global prototype covers different spatial regions of the input features, reducing the possibility that multiple prototypes are concentrated in a small number of high-response regions during the early stages of training.
[0040] In this embodiment, the initial global prototype obtained by region average pooling, covering different spatial regions of the input features, mainly reflects local spatial statistics. Therefore, iterative soft allocation is further employed to update the initial global prototype based on the current content distribution of the input. In the... In the nth iteration, the 1st The image token pair The soft-assignment weights and prototype update formulas for each global prototype are as follows:
[0041]
[0042]
[0043] in, Indicates soft-assigned weights. Indicates the similarity between normalized features. Indicates feature normalization, To prevent positive numbers with a denominator of zero, This represents the j-th global prototype in the t-th iteration. This represents the m-th global prototype at the t-th iteration, where m is the index variable used for normalized summation of all M global prototypes. In this embodiment, soft allocation allows image tokens at texture boundaries or with mixed attributes to contribute to multiple related global prototypes simultaneously, thereby reducing information loss caused by hard allocation.
[0044] The training phase can also maintain the current adaptive global prototype. Exponential moving average statistics:
[0045]
[0046] in, This represents the statistics of historical prototypes. The momentum coefficient, This indicates that the gradient operation is stopped.
[0047] In this embodiment, the exponential moving average prototype serves only as a training stability reference and does not replace the current input adaptive global prototype during forward propagation. Subsequently, a prototype consistency regularization term is constructed based on the distance between the current adaptive global prototype and the exponential moving average prototype after gradient cessation. The prototype consistency regularization term is defined as:
[0048]
[0049] In this embodiment, the inference phase does not use the exponential moving average prototype branch, but instead directly generates an adaptive global prototype based on the current input features.
[0050] Furthermore, after completing the iterative update, The adaptive global prototypes are arranged into a compact prototype sequence and input into the central Mamba module:
[0051]
[0052] in, Representation layer normalization, This indicates a central-level Mamba operation. This represents the refined global prototype. The central Mamba module enables each global prototype to incorporate contextual information from other semantic regions. Because... Central-level modeling is performed only on shorter prototype sequences.
[0053] The specific implementation steps for Top-K sparse routing and semantic bucket construction are as follows:
[0054] like Figure 3 As shown, the normalized similarity between all image tokens and the refined global prototype is calculated, and the i-th token is denoted as . The image token and the first The similarity between the refined global prototypes is Regarding the first Image tokens, only retaining the most similar ones. A set of refined global prototypes is defined, and sparse route weights are determined within the selected set of refined global prototypes; wherein, the selected set of global prototypes is denoted as . And sparse route weight is defined as :
[0055]
[0056]
[0057]
[0058] in, Indicates sparse route weights. This refers to the routing temperature parameter. Indicates the i-th image token and the i-th image token. Normalized similarity between refined global prototypes The index variable is used in the selected Top-K global prototype set to perform normalized summation of the exponential similarity of each candidate prototype in the set.
[0059] In this embodiment, Top-K sparse soft routing allows the same image token to be associated with multiple related semantic patterns. For example, an image token located at a structural boundary can contain both contour and texture information, and routing it to multiple related semantic buckets helps preserve hybrid attributes; at the same time, each image token retains at most K valid routing paths, avoiding redundant interactions with the entire global prototype.
[0060] Furthermore, regarding the first A refined global prototype is used to collect image tokens that are selected as one of the Top-K targets, forming the... Each image token is organized into semantic buckets. Routing operations only copy and reorganize image tokens, without merging or discarding the original tokens. Each routed copy retains its original spatial index and corresponding sparse routing weights for subsequent recovery. Thus, image tokens from different spatial regions but with similar textures, edge orientations, or structural patterns can be organized within the same semantic bucket.
[0061] The specific implementation steps of semantic bucket double-order Mamba scan and weighted scattering are as follows:
[0062] like Figure 4 As shown, considering that information propagation in the Mamba module is related to sequence order, two complementary image token sequences are constructed within each semantic bucket. The first is a spatial order sequence, which is ordered according to the original spatial position of the image tokens in the two-dimensional feature map to maintain the geometric continuity of edges and textures; the second is a similarity order sequence, which is ordered according to the similarity between the image tokens and their corresponding refined global prototypes, so that image tokens with similar semantics are adjacent in the sequence.
[0063] Furthermore, in this embodiment, the two image token sequences are input into the bucket-level Mamba module to perform selective state space scanning, and then their respective inverse ordering indices are used to restore the common semantic bucket order. Element-by-element summation is then performed, and the sequences are fused using linear projection. This dual-order design organizes image tokens within the same semantic bucket from both spatial structure and semantic similarity perspectives.
[0064] Furthermore, after the bucket-level scan is completed, the outputs from different semantic buckets are aligned to the original image token positions and aggregated according to the sparse routing weights:
[0065]
[0066] in, Indicates the first The image token in the first The responses obtained from each semantic bucket This represents the aggregated characteristics after weighted scattering.
[0067] Furthermore, the aggregated features are linearly projected and restored to their original two-dimensional spatial layout, and then added to the input of the center-refined Top-K semantic bucket Mamba module through residual connections.
[0068] In this embodiment, the computational complexity of the similarity calculation between the image token and the global prototype is approximately Since each image token can be routed to at most [number] times. The semantic buckets, and the total number of image token copies after routing, do not exceed [number]. The computational complexity of a bucket-level Mamba scan is approximately Therefore, the overall computational complexity of the centrally refined Top-K semantic bucket Mamba module is approximately... ).when and much smaller At that time, this complexity is relative to the number of image tokens. Approximately linear and avoids global self-attention. Square complexity.
[0069] Furthermore, the specific implementation steps of the global context-aware local self-attention module are as follows:
[0070] like Figure 5 As shown, the center-refined Top-K semantic bucket Mamba module is mainly used for cross-region structural modeling. To enhance the recovery of local edges, corners, and repetitive textures, the refined global prototype is utilized. Provide global semantic guidance for local window self-attention.
[0071] First of all The refined global prototypes are averaged and aggregated to obtain the global context representation. :
[0072]
[0073] Standard local window self-attention based on query ,key Sum Perform the calculation:
[0074]
[0075] This embodiment maintains and The computation remains unchanged, ensuring that attention weights are still determined by local spatial neighborhood relations, utilizing only the global context representation. Modulation value branch To obtain the modulated value characteristics:
[0076]
[0077] in, and For learnable mapping matrix, This represents the Sigmoid activation function. For learnable scaling factor, This represents element-wise multiplication. The multiplicative modulation term adaptively recalibrates the feature channels based on global semantics, while the additive modulation term injects contextual biases relevant to the current image into the local features. The final output is:
[0078]
[0079] In this embodiment, due to and While maintaining consistency, local spatial correlations and edge and texture continuity are stably preserved; value branch modulation introduces global structure and texture priors without disrupting local correspondences. The local self-attention output is linearly projected and fused with parallel local convolutional residual branches.
[0080] Furthermore, in this embodiment, the image reconstruction module performs convolution and pixel rearrangement operations on the fused deep features to obtain the predicted high-frequency residual; performs interpolation upsampling on the low-resolution image to obtain the reference image; and adds the predicted high-frequency residual to the reference image to obtain the super-resolution image.
[0081] Furthermore, the specific model training and inference process involved in this application is as follows:
[0082] In a preferred embodiment, the network model described above is implemented using the PyTorch framework. Training data consists of pairs of low-resolution images and high-resolution real images. High-resolution training images are cropped into 256×256 image patches, and low-resolution input is obtained through appropriate downsampling operations. The total training loss is defined as:
[0083]
[0084] in, For the super-resolution image predicted by the model, For high-resolution real images, For prototype consistency regularization The weight.
[0085] In a preferred embodiment, Number of Top-K routes The Adam optimizer is used, with the following parameters: , No weight decay is used; batch size is set to 16, and initial learning rate is set to... The total number of training iterations was set to 250,000, and the learning rate was decayed to 0.5 times the original value at the 125,000, 200,000, 225,000 and 237,500 iterations.
[0086] During inference, the low-resolution image to be processed is input into the trained network model. The current input features are then processed by region average pooling and iterative soft allocation to generate an adaptive global prototype in real time. The exponential moving average prototype branch during the training phase is not enabled. The image reconstruction module outputs a super-resolution image corresponding to the magnification factor.
[0087] Based on this, the specific implementation effects of this application are as follows:
[0088] In one embodiment, the model was trained using the DIV2K dataset and tested on the Set5, Set14, B100, Urban100, and Manga109 datasets. Evaluation metrics included Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity (SSIM), calculated on the luminance channel of the YCbCr color space. Table 1 lists the test results of the model at different magnifications.
[0089] Table 1. PSNR / SSIM results of the model in this application on different datasets.
[0090]
[0091] As shown in Table 1, the model in this application achieves good image reconstruction performance on multiple test datasets and at magnifications of ×2, ×3, and ×4. When reconstructing regular architectural lines, dense textures, and repetitive structures, the model in this application can reduce structural deformation and artifacts while maintaining relatively clear edges.
[0092] Furthermore, component ablation experiments were conducted under the ×4 super-resolution setting. The PSNR of the baseline model on Set5 / B100 was 32.48 / 27.59. After sequentially adding the global context-aware local self-attention module, the central Mamba module, the Top-K sparse routing module, and the bucket-level Mamba module, the PSNR increased to 32.51 / 27.63, 32.54 / 27.68, 32.58 / 27.73, and 32.63 / 27.79, respectively, indicating that each component module can gradually improve the super-resolution reconstruction performance.
[0093] When the number of Top-K routes K is 1, 2, 4, and 8, the PSNR on Set5 / B100 is 32.50 / 27.66, 32.55 / 27.70, 32.63 / 27.79, and 32.57 / 27.72, respectively. In this embodiment, a good balance between performance and route sparsity is achieved when K=4. When only spatial order, only similarity order, and dual-order scanning are used, the PSNR on Set5 / B100 is 32.56 / 27.71, 32.53 / 27.69, and 32.63 / 27.79, respectively, indicating that spatial geometric continuity and semantic similarity have complementary effects.
[0094] This application also provides an image super-resolution reconstruction method based on a centrally refined Top-K semantic bucket Mamba network, including the following steps:
[0095] Convolutional processing is performed on the input low-resolution image to extract shallow features;
[0096] The shallow features are fused with the outputs of multiple residual groups through global residual connections to obtain fused deep features; wherein, each residual group includes a content-aware main branch and a local convolutional residual branch, and the content-aware main branch includes a centrally refined Top-K semantic bucket Mamba module and a global context-aware local self-attention module.
[0097] Based on the fused deep features, high-frequency residuals are predicted through convolution and pixel shuffle, and the high-frequency residuals are added to the reference image obtained by interpolating and upsampling the low-resolution image to obtain a super-resolution image.
[0098] The above description is merely a preferred embodiment of this application and is not intended to limit the scope of protection of this application. All equivalent substitutions, structural changes, parameter adjustments, or module combinations made within the spirit and principles of this application should be included within the scope of protection of this application.
Claims
1. An image super-resolution reconstruction system based on a centrally refined Top-K semantic bucket Mamba network, characterized in that, include: Shallow feature extraction module: Performs convolution processing on the input low-resolution image to extract shallow features; The deep feature extraction module includes multiple residual groups. The shallow features are fused with the outputs of the multiple residual groups through global residual connections to obtain the fused deep features. Each residual group includes a content-aware main branch and a local convolutional residual branch. The content-aware main branch includes a centrally refined Top-K semantic bucket Mamba module and a global context-aware local self-attention module. Image reconstruction module: Based on the fused deep features, predict high-frequency residuals through convolution and pixel shuffle, and add the high-frequency residuals to the reference image obtained by interpolation and upsampling of the low-resolution image to obtain a super-resolution image.
2. The image super-resolution reconstruction system based on a centrally refined Top-K semantic bucket Mamba network according to claim 1, characterized in that, The local convolutional residual branch uses 3×3 convolution and the ReLU activation function.
3. The image super-resolution reconstruction system based on a centrally refined Top-K semantic bucket Mamba network according to claim 1, characterized in that, Based on the centrally refined Top-K semantic bucket Mamba module, adaptive global prototype generation and central refinement, Top-K sparse routing and semantic bucket construction are completed sequentially.
4. The image super-resolution reconstruction system based on a centrally refined Top-K semantic bucket Mamba network according to claim 3, characterized in that, The specific steps for adaptive global prototype generation and center refinement are as follows: The input features are flattened in space to obtain an image token sequence, and the input features are divided into M non-overlapping regions. Average pooling is performed on each region to obtain M initial global prototypes with uniform spatial distribution; Based on the current input content distribution characteristics, iterative soft allocation is used to update the initial global prototype; After iterative updates, The adaptive global prototypes are arranged into a compact prototype sequence, which is then refined by inputting the central-level Mamba module.
5. The image super-resolution reconstruction system based on a centrally refined Top-K semantic bucket Mamba network according to claim 4, characterized in that, When updating the initial global prototype using iterative soft allocation: During the training phase, maintain the exponential moving average statistics of the current adaptive global prototype; Based on the distance between the current adaptive global prototype and the exponential moving average prototype of the stopping gradient, a prototype consistency regularization term is constructed. During the inference phase, the exponential moving average prototype branch is removed, and an adaptive global prototype is generated based on the current input features.
6. The image super-resolution reconstruction system based on a centrally refined Top-K semantic bucket Mamba network according to claim 3, characterized in that, The steps for constructing the Top-K sparse routes and semantic buckets are as follows: Calculate the normalized similarity between the image token and the refined global prototype; Take the image token with the highest similarity A refined global prototype is generated, and the similarity is normalized to obtain sparse route weights. Image tokens that are selected from the same refined global prototype are collected into semantic buckets, and each routed image token retains its original spatial index and sparse routing weight.
7. The image super-resolution reconstruction system based on a centrally refined Top-K semantic bucket Mamba network according to claim 6, characterized in that, Two types of image token sequences are constructed within the semantic bucket, including a spatial order sequence and a similarity order sequence, wherein: The spatial order sequence is ordered according to the original spatial position of the image token in the two-dimensional feature map; The similarity order sequence is sorted according to the similarity between the image token and the corresponding refined global prototype.
8. The image super-resolution reconstruction system based on a centrally refined Top-K semantic bucket Mamba network according to claim 7, characterized in that, The following steps are performed within the semantic bucket to perform a dual-sequential Mamba scan and weighted scattering operation: The two image token sequences are input into the bucket-level Mamba module to complete selective state space scanning. The sequence is restored to the common semantic bucket order through the corresponding inverse sorting index. The feature fusion is completed by summing the elements one by one and then performing linear projection. After the bucket-level scan is completed, the outputs of different semantic buckets are aligned to the token positions of the original image and aggregated according to the sparse route weights to obtain the aggregated features after weighted scattering. The aggregated features are restored to the original two-dimensional spatial layout through linear projection, and then added and fused with the input of the center-refined Top-K semantic bucket Mamba module through residual connections.
9. The image super-resolution reconstruction system based on a centrally refined Top-K semantic bucket Mamba network according to claim 1, characterized in that, The global context-aware local self-attention module uses a refined global prototype to provide global semantic guidance for local window self-attention, as follows: The refined global prototype is averaged and aggregated to obtain the global context representation; Keeping the query branch Q and key branch K in the local window self-attention unchanged, and using the global context representation to perform multiplicative modulation and additive modulation on the value branch V to obtain the modulated value features.
10. An image super-resolution reconstruction method based on a center-refined Top-K semantic bucket Mamba network, characterized in that, Includes the following steps: Convolutional processing is performed on the input low-resolution image to extract shallow features; The shallow features are fused with the outputs of multiple residual groups through global residual connections to obtain fused deep features; wherein, each residual group includes a content-aware main branch and a local convolutional residual branch, and the content-aware main branch includes a centrally refined Top-K semantic bucket Mamba module and a global context-aware local self-attention module. Based on the fused deep features, high-frequency residuals are predicted through convolution and pixel shuffle, and the high-frequency residuals are added to the reference image obtained by interpolating and upsampling the low-resolution image to obtain a super-resolution image.
