Anchor guidance and boundary refinement collaborative image super-resolution reconstruction method and device

Abstract

The invention relates to an anchor point guidance and boundary refinement collaborative image super-resolution reconstruction method and device, and belongs to the field of computer vision. The method comprises the following steps: performing shallow feature extraction on an input low-resolution image to obtain an initial feature; constructing a deep feature enhancement network comprising a plurality of series collaborative reconstruction units, inputting the initial features, and sequentially enhancing the initial features through the units; each unit respectively executes anchor point routing global interaction and boundary sensing area refinement on the input features to obtain global enhanced features and local refined features, the global enhanced features and the local refined features are fused with input feature residual errors after being subjected to adaptive gating fusion, current collaborative enhanced features are output and transmitted to the next unit, and finally deep enhanced features are obtained; and performing up-sampling and pixel reconstruction on the deep enhancement feature, and outputting a super-resolution image. According to the method, long-distance structure dependence modeling and local boundary detail recovery are considered, and collaborative optimization of global structure consistency and local detail fidelity is realized under lightweight deployment constraint.

Description

An image super-resolution reconstruction method and apparatus that combines anchor point guidance and boundary refinement Technical Field

[0001] This invention relates to the field of computer vision, and in particular to an image super-resolution reconstruction method and apparatus that combines anchor point guidance and boundary refinement. Background Technology

[0002] Single-Image Super-Resolution (SISR) aims to reconstruct high-resolution images from low-resolution images and is one of the most representative classic problems in low-level vision. Because low-resolution observations inevitably lose a significant amount of high-frequency texture and fine-grained structural information during the imaging process, this task is essentially a highly ill-posed inverse problem. Nevertheless, image super-resolution has significant application value in scenarios such as intelligent surveillance, medical diagnosis, mobile photography, and edge vision terminals. These applications typically require high reconstruction quality while imposing strict constraints on model complexity, inference speed, and deployment costs. Therefore, achieving high-fidelity reconstruction within a limited computational budget remains a core challenge in image super-resolution research.

[0003] In recent years, deep learning-based methods have significantly advanced the development of single-image super-resolution. From the early establishment of nonlinear mappings from low to high resolution using convolutional neural networks, to the successive introduction of residual learning, attention mechanisms, and Transformer structures, the model's ability to enhance local texture and model cross-regional structural relationships has been continuously improved. There is also a consensus in the industry that high-quality super-resolution reconstruction requires both effective local detail restoration and global contextual dependency characterization.

[0004] However, existing technologies still face the core challenge of balancing reconstruction performance with lightweight deployment requirements. On one hand, many high-quality super-resolution models often rely on increasing network layers, widening feature channels, or designing dense global attention mechanisms to improve performance. This directly leads to a significant increase in the model's parameter size and computational cost, making it difficult to adapt to resource-constrained lightweight deployment scenarios such as edge vision terminals. On the other hand, while existing lightweight super-resolution methods reduce computational complexity by simplifying network structures, most methods rely solely on local convolutions or limited-range attention interactions to complete feature modeling, lacking sufficient ability to characterize the long-range structural dependencies of images. Furthermore, the few lightweight methods that introduce attention mechanisms to enhance global modeling lack explicit constraints on local boundaries, texture transitions, and structural continuity, making it impossible to achieve coordinated optimization of global structural consistency and local detail fidelity within a limited computational budget. This makes it difficult to meet the dual requirements of reconstruction quality and deployment efficiency in practical applications. Summary of the Invention

[0005] Therefore, it is necessary to provide an image super-resolution reconstruction method and apparatus that can balance reconstruction quality and deployment efficiency by combining anchor point guidance and boundary refinement.

[0006] An image super-resolution reconstruction method that combines anchor point guidance and boundary refinement, the method comprising:

[0007] A shallow feature extraction process is performed on the input low-resolution image to obtain initial features. A deep feature enhancement network is constructed, comprising multiple cascaded collaborative reconstruction units. The initial features are input into the deep feature enhancement network and sequentially processed by each of the collaborative reconstruction units. In the current collaborative reconstruction unit, the input features are processed by global interaction of anchor routing and boundary-aware region refinement, resulting in global enhanced features and local refined features. The global enhanced features and the local refined features are then subjected to adaptive structural fusion and residual fusion with the input features to output the current collaborative enhanced features. The current collaborative enhanced features are then input into the next collaborative reconstruction unit for processing, ultimately outputting deep enhanced features. Upsampling and pixel reconstruction processes are sequentially performed on the deep enhanced features to output the final super-resolution image.

[0008] On the other hand, an image super-resolution reconstruction device combining anchor-guided and boundary-refinement synergy is also provided, comprising: a shallow feature extraction module for performing shallow feature extraction processing on an input low-resolution image to obtain initial features; a network construction module for constructing a deep feature enhancement network, the deep feature enhancement network including multiple cascaded collaborative reconstruction units; a deep feature enhancement module for inputting the initial features into the deep feature enhancement network, which are then sequentially processed by each of the collaborative reconstruction units for feature enhancement processing, wherein, in the current collaborative reconstruction unit, the input features are subjected to global interactive processing of anchor-guided routing and boundary-aware region refinement processing respectively, resulting in global enhanced features and local refined features, which are then subjected to adaptive structural fusion processing and residual fusion with the input features to output the current collaborative enhanced features; the current collaborative enhanced features are input into the next collaborative reconstruction unit for processing, and finally, deep enhanced features are output; and an image reconstruction module for sequentially performing upsampling and pixel reconstruction processing on the deep enhanced features to output the final super-resolution image.

[0009] Compared with existing technologies, the image super-resolution reconstruction method and apparatus provided by this invention, which combines anchor point guidance and boundary refinement, has the following beneficial effects: 1. The deep feature enhancement network adopts multiple cascaded collaborative reconstruction units, combined with adaptive structure fusion and residual fusion mechanisms, effectively retains the underlying visual information while optimizing features in stages, avoids feature decay and excessive smoothing, improves the model training stability and reconstruction robustness, and makes the restoration effect of complex structures and high-frequency texture regions better.

[0010] 2. By performing global interaction of anchor point routing and boundary-aware region refinement in parallel through collaborative reconstruction units, it not only efficiently portrays long-distance structural dependencies by leveraging anchor point compact representation and sparse routing mechanism, avoiding the high computational cost brought by dense global attention, but also enhances the fine restoration of local boundaries and texture transitions through boundary bias and alternating window shifting. This fundamentally solves the core contradiction that existing lightweight methods cannot balance global structural consistency and local detail fidelity, and significantly improves the fidelity of super-resolution reconstruction.

[0011] 3. The overall architecture does not rely on deep networks, wide channels, or dense global interactions. Through modular design with functional decoupling, it achieves efficient feature modeling with a limited computing budget. The model complexity is controllable, the inference speed is fast, and it can be adapted to lightweight deployment scenarios such as intelligent monitoring and edge vision terminals, truly achieving dual optimization of reconstruction quality and deployment efficiency. Attached Figure Description

[0012] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of the present invention, and those skilled in the art can obtain other related drawings based on these drawings without creative effort.

[0013] Figure 1 is a flowchart illustrating the image super-resolution reconstruction method that combines anchor point guidance and boundary refinement in Example 1; Figure 2 is a flowchart of the image super-resolution reconstruction method that combines anchor point guidance and boundary refinement in Example 1; Figure 3 is a flowchart of the global interaction module for anchor point routing in Example 1; Figure 4 is a flowchart of the boundary-aware region refinement module in Example 1; Figure 5 is a structural block diagram of the image super-resolution reconstruction device that combines anchor point guidance and boundary refinement in Example 2; Figure 6 is an internal structural diagram of the computer device in Example 3.

[0014] The objectives, features, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0015] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0016] It should be noted that in this invention, the use of terms such as "first," "second," etc., is for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0017] It is understood that the technical solutions of the various embodiments of the present invention can be combined with each other, but only if they are based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0018] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0019] Example 1, as shown in Figure 1, provides an image super-resolution reconstruction method that combines anchor point guidance and boundary refinement, including the following steps: Step 201, perform shallow feature extraction processing on the input low-resolution image to obtain initial features.

[0020] In this step, the low-resolution image is a single input image that has lost high-frequency texture and fine-grained structural information through downsampling operations, and it serves as the basic input for super-resolution reconstruction. The initial features are high-channel dimensionality feature representations obtained by extracting the low-resolution image through shallow convolution, which can characterize the low-level visual information of the image.

[0021] Step 202: Construct a deep feature enhancement network, which includes multiple cascaded collaborative reconstruction units.

[0022] In this step, the deep feature enhancement network is the core network that realizes the progressive enhancement of image features from coarse to fine. The collaborative reconstruction unit (CRU) is the basic feature enhancement unit of this network. Each unit has three functions: global structure modeling, local detail refinement and feature adaptive gating fusion.

[0023] Step 203: Input the initial features into the deep feature enhancement network, and perform feature enhancement processing in each collaborative reconstruction unit in sequence. In the current collaborative reconstruction unit, the input features are subjected to global interactive processing of anchor point routing and boundary-aware region refinement processing, respectively, to obtain global enhanced features and local refined features. After adaptive structural fusion processing of global enhanced features and local refined features, residual fusion is performed with the input features to output the current collaborative enhanced features. The current collaborative enhanced features are then input into the next collaborative reconstruction unit for processing, and finally, deep enhanced features are output.

[0024] This step is the core process of deep feature enhancement. The deep enhancement features are high-dimensional features that, after iterative optimization by multiple collaborative reconstruction units, possess both global structural consistency and local detail fidelity, and form the core feature foundation for subsequent image reconstruction. The entire process starts with the initial features. As the input to the first collaborative reconstruction unit, after the... Collaborative enhancement features output after processing each unit As the first The input of the nth unit is iterated sequentially until the nth unit. After processing each unit, the final deep augmentation feature is output. .

[0025] Step 204: Perform upsampling and pixel reconstruction processing on the deep enhancement features in sequence to output the final super-resolution image.

[0026] In this step, upsampling is the process of mapping high-dimensional deep enhancement features from a low-resolution feature space to a target high-resolution feature space, pixel reconstruction is the process of restoring the feature representation of the high-resolution feature space to a pixel domain image, and the super-resolution image is the final output high-resolution result that restores high-frequency texture and fine-grained structure.

[0027] In the specific implementation of step 201, as shown in Figure 2, a shallow feature extraction is first performed on the input low-resolution image using a convolutional head to obtain initial features, represented as: In the formula, Indicates initial features; Indicates batch size; Indicates the number of channels; and These represent the spatial height and width of the initial feature, respectively. This indicates a low-resolution image.

[0028] This step uses lightweight convolution operations to quickly extract key information such as basic texture, edge contours, and brightness distribution of the image, avoiding the inefficiency caused by starting feature modeling from the original pixels. At the same time, it establishes a unified feature representation foundation for global structure modeling and local detail restoration.

[0029] In the specific implementation of step 202, as shown in Figure 2, a system was constructed using... A deep feature enhancement network composed of interconnected collaborative reconstruction units. The number can be flexibly adjusted according to the actual reconstruction accuracy requirements and deployment resource constraints.

[0030] Each collaborative reconstruction unit integrates three parts: Anchor Point Routing Global Interaction Module (ARGI), Boundary Aware Region Refinement Module (BRR), and Adaptive Structure Fusion Mechanism (ASF), forming a feature processing architecture that is functionally decoupled and works in concert.

[0031] Among them, the anchor point routing global interaction module is specifically responsible for extracting long-distance structural dependencies from a global perspective. Through compact anchor point representation and sparse routing mechanism, it achieves efficient aggregation and propagation of global information with low computational cost.

[0032] The boundary-aware region refinement module focuses on local structure restoration, enhancing the fine restoration of edge contours and texture details through boundary-aware constraints and cross-regional continuity modeling.

[0033] The adaptive structural fusion mechanism is responsible for dynamically integrating global enhanced features and local refined features in an adaptive manner, adjusting the contribution ratio of the two information streams according to the feature attributes of different spatial locations.

[0034] This step abandons the unreasonable architecture of traditional methods where a single mechanism simultaneously undertakes global inference and local recovery. Through a serialized modular design, it achieves decoupling optimization between global structure modeling and local detail refinement, which not only ensures the interpretability of the network structure, but also provides flexible and efficient architectural support for subsequent stage-by-stage feature enhancement.

[0035] In the specific implementation of step 203, the initial features are used as the basis. As the input to the first collaborative reconstruction unit, for the current collaborative reconstruction unit (i.e., the first...) (the nth collaborative reconstruction unit), given the input features as the nth Cooperative enhancement features output by each cooperative reconstruction unit Then, through parallel global interaction processing of anchor point routing and boundary-aware region refinement processing, and then through adaptive structural fusion and residual fusion, the current collaborative enhancement feature (i.e., the first) is obtained. Cooperative enhancement features output by each cooperative reconstruction unit ), ultimately passed The iterative output of deep enhancement features from each collaborative reconstruction unit .

[0036] To more efficiently model long-range dependencies in deep feature enhancement networks, the anchor point routing global interaction module designed in this invention differs from traditional self-attention, which directly establishes pairwise associations between all spatial locations, and from simply relying on local convolutions to expand the receptive field layer by layer. ARGI first extracts a small number of representative structural anchor points from the compressed feature representation, and then selectively propagates global structural information back to dense spatial locations through a sparse routing mechanism, thereby achieving global context aggregation at a lower cost. Simultaneously, to avoid excessive smoothing of high-frequency textures and edge details during spatial compression, ARGI further introduces a local compensation branch to enhance local structural responses. Through a hierarchical modeling process of "structural anchor point extraction, anchor point-level global interaction, sparse routing backpropagation, and local detail compensation," it is possible to balance global structural consistency and local detail fidelity under conditions of limited model complexity.

[0037] Specifically, in the first In the collaborative reconstruction unit, the first The anchor point routing global interaction module targets input features. The extraction of global enhanced features is completed, and the overall processing expression is as follows: In the formula, Indicates the first One global enhancement feature; Indicates the first The mapping function for each anchor point route in the global interaction module; Indicates the first The collaborative enhancement features output by each collaborative reconstruction unit.

[0038] More specifically, as shown in Figure 3, in the global interaction module for anchor point routing, the input features are first processed by a depthwise convolution with stride. The compression ratio is The downsampling mapping is used to reduce the computational burden of subsequent global modeling, resulting in compact features, represented as: In the formula, Indicates compactness; Indicates the compression ratio. downsampling mapping, The compressed height represents the compact feature. This represents the compressed width of the compact feature.

[0039] This learnable spatial compression method differs from fixed average pooling. It can selectively preserve the structural responses that are critical to the reconstruction task, so that the compressed features still have strong discriminativeness and recoverability, and avoid excessive smoothing of high-frequency information.

[0040] In obtaining compact features Subsequently, instead of directly performing global attention on all spatial tokens, this invention first constructs a set of structure-aware anchors. Specifically, it first predicts the importance map of each spatial location through a lightweight scoring branch: In the formula, This indicates the importance score for each position; This indicates that a 1×1 pointwise convolution operation is performed.

[0041] Subsequently, compact features Divided into Non-overlapping local regions Within each local region, a region anchor point is generated through normalized weighted aggregation. The expression for a single region anchor point is: In the formula, Indicates the area anchor point; Indicates position Importance score Indicates position eigenvectors; Represents the first compact feature partition. A local area; This indicates the total number of anchor points in the region.

[0042] All regional anchor points constitute the anchor point set. Due to the number of anchor points Much smaller than the total number of spatial locations in the compressed feature Therefore, subsequent global relation modeling can be shifted from the dense spatial domain to the compact anchor domain, which significantly reduces complexity while preserving the main structural semantics.

[0043] Next, to characterize long-range structural dependencies across regions, multi-head self-attention processing is performed on the set of regional anchors to obtain a global enhanced anchor set. The process expression is as follows: In the formula, Represents the set of global enhancement anchor points; , and These represent the sets of regional anchor points respectively. The query vector, key vector, and value vector obtained by linear projection; Indicates the single-head channel dimension; This represents the normalized exponential function; This represents the matrix transpose operation.

[0044] This step enables representative structural patterns from different regions to interact efficiently in a low-dimensional anchor space, thereby obtaining more complete global layout information. Compared to directly building global attention on all space tokens, its complexity can be reduced by... Reduced to It is more suitable for the actual deployment needs of lightweight image super-resolution networks.

[0045] After completing the anchor-level global interaction, it is also necessary to effectively propagate the enhanced global structural information back to each spatial location. To this end, the compact features are first... Flatten into a token sequence ,in Indicates compact features The total number of tokens in the flattened space. Then, through the sparse routing mechanism of MaskTop-k masking, the information of the global enhanced anchor set is propagated back to the compact feature to obtain the global propagation feature. The process expression is: In the formula, A token sequence representing global propagation characteristics; This means only retaining each space token and the top-ranked response. Masking operations for connecting anchor points in each region; Represents a sequence of tokens in space. The query vector obtained by linear mapping; and These represent the sets of globally enhanced anchor points. The key vector and value vector obtained by linear projection; This indicates the number of anchor points connected to each space token.

[0046] Unlike dense attention, which requires each token to be connected to all anchors, sparse routing allows ARGI to retain only the connections between each token and its most relevant anchors. The interaction relationships between anchor points are masked as invalid connections in other locations, allowing the global interaction module of anchor point routing to inject global structural priors into different spatial locations in a more selective manner. This transforms information propagation from undifferentiated dense associations to content-adaptive, structure-guided interactions, with a computational cost of approximately [missing information]. ,in .

[0047] However, while anchor-based sparse global propagation effectively enhances structural consistency, spatial compression and anchor aggregation may still weaken the expressive power of local high-frequency textures and edge details. Therefore, a 3×3 depthwise convolution is introduced to perform local detail compensation processing on compact features, refining local detail features: In the formula, Indicates local detailed features; This represents a 3×3 depthwise convolution operation. Through local detail compensation processing, it can explicitly preserve texture variations, edge transitions, and fine-grained structural responses within the neighborhood, thereby compensating for the loss of detail that may occur during global compression modeling.

[0048] Subsequently, the obtained local detail features With global propagation characteristics Perform adaptive gating fusion, record To recover a two-dimensional feature map from the token sequence of globally propagated features, the fusion process is defined as follows: ; In the formula, Represents the adaptive gating fusion weight graph; This represents the sigmoid activation function. This indicates concatenation along the channel dimension. This represents element-wise multiplication; This indicates a compact fusion feature. Through this gated fusion mechanism, the deep feature enhancement network can adaptively balance the contributions of global structural information and local detail information in different spatial regions based on the input content. In smooth regions, the anchor point routing global interaction module can rely more on the global priors obtained from anchor point propagation to maintain overall structural consistency; while in edge or texture-rich regions, the response of the local compensation branch can be appropriately enhanced to improve the detail clarity and spatial realism of the reconstruction result.

[0049] Finally, regarding the fusion compact features Learnable spatial restoration processing is performed, mapping back to the original resolution using a learnable spatial restoration operator. Residual fusion is then performed with the input features using residual connections, preserving low-frequency consistency information in the input to obtain globally enhanced features. The expression is as follows: In the formula, Indicates the first A global enhancement feature, This indicates the upsampling recovery operator. This indicates a point-by-point projection layer.

[0050] Residual connections help improve training stability and avoid destroying the basic structural information in the original representation during feature reconstruction. Therefore, the entire anchor routing global interaction module can be summarized as follows: .

[0051] It can be seen that the anchor point routing global interaction module proposed in this invention has a significantly different design approach from conventional lightweight attention units. Firstly, global modeling no longer directly applies to all spatial tokens, but instead uses structural anchor points for compact representation, thereby significantly reducing the computational cost of long-distance dependency modeling; secondly, it leverages top- The sparse routing mechanism enables modules to propagate global semantics back to dense spatial locations in a more targeted manner, improving information interaction efficiency. Thirdly, by introducing local compensation branches, modules effectively preserve sensitivity to high-frequency textures and edge details while strengthening global structural consistency. Therefore, ARGI provides a more structured global representation learning approach for deep feature enhancement networks, better meeting the needs of efficient deployment.

[0052] To complement the global interaction module of anchor point routing, a boundary-aware region refinement module (BRR) is also designed in the collaborative reconstruction unit for more refined modeling of local structure, texture details, and edge transitions. Unlike the global interaction module of anchor point routing, which is mainly responsible for long-distance structural consistency modeling, BRR focuses more on the recovery of high-frequency information and the preservation of spatial continuity within local regions. Although traditional local window attention can effectively reduce computational complexity, it usually adopts a uniform interaction mode for all tokens within the region, lacking explicit constraints on edges, texture abrupt changes, and structural break locations, which can easily introduce over-smoothing or boundary aliasing problems during local reconstruction. To address this shortcoming, BRR introduces boundary-aware priors during feature interaction within the region and combines an alternating window mechanism to achieve cross-region information flow, thereby enhancing local detail fidelity and spatial consistency while maintaining low complexity.

[0053] Specifically, in the first In the collaborative reconstruction unit, the first Each boundary-aware region refinement module targets input features Extracting local refinement features, the overall processing expression is as follows: In the formula, Indicates the first A localized refinement feature; Indicates the first Mapping functions for each boundary-aware region refinement module.

[0054] More specifically, as shown in Figure 4, in the boundary-aware region refinement module, given input features... First, based on the preset window size The input features are divided into several non-overlapping local regions, and each region is expanded into a token sequence. The token sequence after region division is denoted as: In the formula, Indicates the number of regions in each image; This indicates the number of tokens in each region; The window height indicating the region division; The width of the window indicating the region division; This represents the feature dimension. Through this regional modeling approach, the module can perform fine-grained feature interactions within a controlled local scope, thereby avoiding the huge computational overhead of global self-attention under high-resolution input.

[0055] To make the interaction within the region more focused on structural boundaries and texture changes, boundary-aware prior extraction is performed on the input features to generate a single-channel structural response map, expressed as: In the formula, This represents the local structural response intensity at each spatial location. The response map is synchronously divided into various regions and used to characterize the boundary differences between different tokens within each region.

[0056] Based on the structural response graph, a boundary bias is constructed for any two tokens within the region, represented as follows: In the formula, Indicates the first in the region The token and the first Boundary offset between tokens; Indicates the scaling factor; The first element in the structural response diagram is represented by the second element. The structural response value corresponding to each token position; The first element in the structural response diagram is represented by the second element. The structural response value corresponding to each token's location. The intuitive meaning of this bias term is: when the structural responses of two locations differ significantly, they are more likely to be located on opposite sides of an edge or in different texture regions, in which case feature mixing between them should be appropriately suppressed; while when their responses are similar, more sufficient aggregation of local information is allowed. Thus, BRR can establish an adaptive interaction mechanism within the region, constrained by boundary priors.

[0057] Based on boundary bias, the region token sequence Linear mappings are performed separately to obtain query, key, and value representations, and a boundary-modulated region attention is constructed. , and Let each represent the corresponding projection result. Then, the feature interaction within the region can be represented as: In the formula, The attention weights represent the boundary modulation. This represents the normalized exponential function; This represents the matrix transpose operation; , and These represent the query vector, key vector, and value vector obtained by linear projection of the region token sequence, respectively. Indicates the single-head channel dimension; Represents a learnable relative position offset, used to encode spatial relationships within a region; Indicates boundary offset; This indicates optional region mask items. When the boundary-aware region refinement module uses a window-shifting strategy... This is used to mask invalid connections, ensuring that attention is only performed within legitimate regions. Compared to standard region attention, the above formula explicitly introduces structural boundary information during local interactions, making feature aggregation not only dependent on semantic similarity but also modulated by local geometric and texture changes. Therefore, it is more conducive to the stable recovery of edge contours and texture transitions in super-resolution tasks.

[0058] After boundary-modulated region attention processing, region enhancement features are obtained: In the formula, A token sequence representing region enhancement features; This represents a linear projection operation. Region enhancement features emphasize responses with consistent structural properties within a region while suppressing irrational mixing across boundaries, thereby improving the structural clarity of local reconstruction results.

[0059] While boundary modulation attention (BRR) can effectively improve the quality of detail modeling within a region, if a fixed region division is always used, information interaction may still be limited to within a static window. Therefore, this invention further introduces an alternating windowing mechanism to enhance the continuous modeling between adjacent regions. In specific implementation, some BRR modules first perform cyclic translation on the input features before region division, then perform region division and boundary modulation attention calculation, and finally perform inverse translation recovery in the output stage to output the spatial feature map. By alternating between "non-moving window" and "moving window" configurations between adjacent layers, the model can progressively achieve cross-regional information exchange without resorting to more costly global intensive interactions. This design not only retains the efficiency advantages of local region modeling but also effectively alleviates the feature fragmentation problem caused by fixed window boundaries.

[0060] To further preserve texture and edge responses within the neighborhood, BRR also introduces a lightweight local compensation branch that performs convolutional local detail compensation on the input features to obtain local detail features: In the formula, This represents local detail features. The local detail compensation branch does not rely on explicit token interactions, but instead enhances the representation of high-frequency details within the local neighborhood through convolutional operations.

[0061] Subsequently, the spatial feature map with local details Adaptive gating fusion is performed, expressed as follows: ; In the formula, Indicates the first This local refinement feature is achieved through a fusion mechanism. The deep feature enhancement network can dynamically balance the contribution ratio of boundary-constrained region refinement features and convolution-compensated local detail features based on the input content. For regions with smooth textures but clear structural boundaries, the boundary-aware region refinement module can rely more on boundary-aware region interactions to maintain contour continuity. For regions rich in high-frequency details, the response of the local compensation branch can be appropriately enhanced to obtain more realistic and sharp reconstruction results.

[0062] Therefore, the entire boundary sensing area refinement module can be summarized as follows: .

[0063] In summary, BRR differs significantly from conventional local window attention modules in several aspects: First, it does not rely solely on the semantic similarity of tokens within a region for feature aggregation, but rather enhances local structural transitions and edge preservation capabilities by explicitly introducing boundary biases. Second, it combines an alternating windowing mechanism to improve cross-regional continuity modeling capabilities without significantly increasing computational burden. Third, by introducing local compensation branches and gating fusion strategies, the module achieves a more robust balance between local texture enhancement and structural boundary preservation. Therefore, BRR provides deep feature enhancement networks with a more structure-aware and locally faithful modeling unit, and forms a clear global-to-local functional division with ARGI.

[0064] After obtaining global enhancement features With local refinement features Then, the two are integrated through an adaptive structural fusion mechanism. The fusion weight map, which predicts content adaptively using pointwise convolution and the sigmoid activation function, is expressed as: In the formula, Indicates the first Fusion weight diagram of collaborative reconstruction units; This represents the Sigmoid activation function; This represents a 1×1 pointwise convolution operation; Indicates the first One global enhancement feature; Indicates the first Local refinement features.

[0065] Then, based on this fusion weight map, the two features are subjected to gated weighted fusion and combined with the input features of the current collaborative reconstruction unit. Forming residual connections, we obtain the first... The collaborative enhancement feature of each unit is expressed as follows: In the formula, Indicates the first The collaborative enhancement features output by each collaborative reconstruction unit; Indicates the first The collaborative enhancement features output by each collaborative reconstruction unit; Indicates the total number of collaborative reconstruction units; This indicates an element-wise multiplication operation.

[0066] Through this adaptive gating fusion method, the network can dynamically adjust the contribution ratio of global structural information and local detail information according to the content features of different spatial locations. In smooth regions, the model can rely more on the global structural prior provided by ARGI to maintain the consistency of the overall layout; while in edge and texture-rich regions, the local refinement response of BRR can be appropriately enhanced to improve the quality of detail recovery.

[0067] From an overall structural perspective, ABCR is not a simple stacking of several independent modules, but a staged collaborative reconstruction framework with a clear division of functions. ARGI is responsible for efficiently aggregating long-range structural information in a compact representation space, BRR is responsible for performing boundary-aware fine-grained refinement within local regions, and ASF undertakes the content adaptive coordination of global and local information. As multiple CRUs are stacked layer by layer, the network can gradually complete the multi-stage reconstruction process from coarse-grained structural organization to fine-grained texture restoration, thereby achieving more stable and interpretable feature evolution.

[0068] The first Cooperative enhancement features output by each cooperative reconstruction unit As the first The input features of the first collaborative reconstruction unit are processed iteratively until the second collaborative reconstruction unit is reached. After processing by each collaborative reconstruction unit, the final deep augmentation features are output. .

[0069] In the specific implementation of step 204, the mapping function is reconstructed. Deep enhancement features Upsampling and pixel reconstruction are performed to generate the final super-resolution image. The expression is: In the formula, Represents a super-resolution image. Reconstruction mapping function. This system integrates a complete workflow of upsampling and pixel reconstruction. The upsampling stage uses a learnable upsampling recovery operator to map deep enhancement features from a low-resolution feature space to the target high-resolution feature space. The pixel reconstruction stage uses lightweight convolution operations to restore the feature representation from the high-resolution feature space to a pixel-domain image, ultimately outputting a super-resolution image that restores high-frequency textures and fine-grained structures. The reconstruction head employs a lightweight design, matching the computational complexity of the preceding anchor point routing global interaction module and boundary-aware region refinement module. This ensures the overall model's lightweight characteristics, enabling efficient deployment in resource-constrained scenarios such as intelligent monitoring and edge vision terminals. Simultaneously, it accurately restores the structural details and visual fidelity of the image, achieving the ultimate goal of super-resolution reconstruction.

[0070] In one embodiment, the reconstruction performance and generalization ability of the proposed method are verified. Systematic experiments were conducted on five commonly used single-image super-resolution benchmark datasets: Set5, Set14, BSD100, Urban100, and Manga109. These datasets cover various types, including smooth textures, complex urban scenes, and high-detail line drawings, enabling a comprehensive evaluation of the model's reconstruction capabilities under different structural complexities and texture distributions. In the experiments, low-resolution images were generated from high-resolution images through bicubic downsampling. Quantitative evaluation metrics included Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity (SSIM). All metrics were calculated on the luminance channel (Y channel) of the YCbCr color space. Before evaluation, the image was cropped by four pixels around the edges to reduce the impact of convolution boundary effects.

[0071] The experiments were implemented using the PyTorch framework and the BasicSR code library, and were trained and tested on a single NVIDIA RTX 4090 GPU. The model was trained from scratch, without pre-trained weights and without relying on a distributed training environment, ensuring the reproducibility of the experimental setup. The DIV2K dataset was used as the supervised training set. Low-resolution inputs were generated using bicubic degradation. During training, 256×256 image patches were randomly cropped from high-resolution images as training samples. Data augmentation strategies including random horizontal flipping and 90° rotation were employed. The batch size was set to 16, and CUDA prefetching and persistent workers were enabled to improve data loading efficiency. For optimization, the AdamW optimizer was used. , (weight decay to 0), initial learning rate is The total number of training iterations was 500k. A warm-up strategy was used for the first 20k iterations, with multi-step learning rate decay (multiplying by 0.5 each time) performed at iterations of 250k, 400k, 450k, 475k, and 490k. The loss function used was... The reconstruction loss balances training stability with high-fidelity image restoration. To further stabilize and validate performance, the model parameters are maintained using an exponential moving average (EMA) with a decay coefficient of 0.999 during training.

[0072] During the validation and testing phase, the model was validated on the five benchmark datasets mentioned above every 5000 iterations. During testing, inference was performed directly on the entire image without any block operations. Experiments showed that the PSNR of the validation set continued to improve steadily during training and gradually approached saturation in the later stages of training, without any obvious optimization oscillations, proving that the training configuration used has good stability.

[0073] The ABCR method of this invention was quantitatively compared with several representative lightweight super-resolution methods such as IMDN, RFDN-L, SwinIR-L, Omni-SR, and SRFormer-L at a magnification of ×4. The results are shown in Table 1.

[0074] Table 1. Quantitative comparison with existing lightweight super-resolution methods at ×4 magnification.

[0075] The comparison results show that ABCR has 777K parameters and 62.7G FLOPs, which is within a reasonable range for current lightweight super-resolution models. It does not sacrifice performance for a significant increase in model size. In terms of reconstruction performance, ABCR achieves PSNRs of 32.62dB, 28.97dB, 27.80dB, 26.97dB, and 31.52dB on Set5, Set14, BSD100, Urban100, and Manga109, respectively, and SSIM values ​​of 0.9001, 0.7900, 0.7443, 0.8110, and 0.9202, respectively. Overall, its performance is better than most existing lightweight methods. Especially on the more structurally complex Urban100 and more detailed Manga109 datasets, ABCR significantly outperforms representative methods such as IMDN, RFDN-L, and SwinIR-L. Even compared to some methods that perform well on specific datasets (such as EFATSR), ABCR demonstrates superior overall performance on datasets like Set14 and BSD100, which better reflect structural recovery capabilities. These experimental results fully validate the significant advantages of the global-local collaborative modeling mechanism proposed in this invention for complex structural recovery and high-frequency detail reconstruction. They demonstrate that the method can more effectively utilize feature representation capabilities under limited computational budgets, achieving an optimal balance between reconstruction quality and model complexity. Furthermore, they verify that under lightweight deployment constraints, this method possesses superior reconstruction performance and generalization capabilities compared to existing technologies, meeting the practical needs of lightweight deployment scenarios such as intelligent monitoring, edge vision terminals, and mobile photography.

[0076] It should be understood that although the steps in Figures 1 to 4 of this embodiment are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some of the steps in Figures 1 to 4 may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least a portion of the sub-steps or stages of other steps.

[0077] Example 2 is based on the image super-resolution reconstruction method of anchor point guidance and boundary refinement coordination in Example 1. This example discloses an image super-resolution reconstruction device of anchor point guidance and boundary refinement coordination, as shown in Figure 5. The image super-resolution reconstruction device of anchor point guidance and boundary refinement coordination includes: a shallow feature extraction module 401, a network construction module 402, a deep feature enhancement module 403, and an image reconstruction module 404. The shallow feature extraction module 401 is used to perform shallow feature extraction processing on the input low-resolution image to obtain initial features.

[0078] The network building module 402 is used to build a deep feature enhancement network, which includes multiple cascaded collaborative reconstruction units.

[0079] The deep feature enhancement module 403 is used to input the initial features into the deep feature enhancement network, and then perform feature enhancement processing through each collaborative reconstruction unit in sequence. In the current collaborative reconstruction unit, the input features are subjected to global interactive processing of anchor point routing and boundary-aware region refinement processing, respectively, to obtain global enhanced features and local refined features. The global enhanced features and local refined features are then subjected to adaptive structural fusion processing, and then residual fusion is performed with the input features to output the current collaborative enhanced features. The current collaborative enhanced features are then input into the next collaborative reconstruction unit for processing, and finally the deep enhanced features are output.

[0080] The image reconstruction module 404 is used to perform upsampling and pixel reconstruction processing on the deep enhancement features in sequence, and output the final super-resolution image.

[0081] In this embodiment, the specific working process and working principle of the shallow feature extraction module 401, network construction module 402, deep feature enhancement module 403, and image reconstruction module 404 are the same as those in Embodiment 1, and therefore will not be described again in this embodiment. Each unit module can be implemented entirely or partially through software, hardware, or a combination thereof. Each unit module can be embedded in or independent of the processor in a computer device in hardware form, or it can be stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each of the above unit modules.

[0082] Example 3, as shown in Figure 6, discloses a computer device including a transmitter, a receiver, a memory, and a processor. The transmitter sends instructions and data, the receiver receives instructions and data, the memory stores computer execution instructions, and the processor executes the computer execution instructions stored in the memory to implement the method in Example 1.

[0083] It is important to note that the aforementioned memory can be either standalone or integrated with the processor. When the memory is set up independently, the terminal device also includes a bus for connecting the memory and the processor.

[0084] Example 4 This example discloses a computer-readable storage medium storing computer-executable instructions. When a processor executes the computer-executable instructions, it implements the method in Example 1 above.

[0085] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0086] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0087] The embodiments described above are merely examples of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention.

Claims

1. A method for image super-resolution reconstruction that combines anchor point guidance and boundary refinement, characterized in that, The method includes: performing shallow feature extraction processing on the input low-resolution image to obtain initial features; constructing a deep feature enhancement network, the deep feature enhancement network including multiple cascaded collaborative reconstruction units; inputting the initial features into the deep feature enhancement network, and sequentially performing feature enhancement processing through each of the collaborative reconstruction units, wherein, in the current collaborative reconstruction unit, the input features are subjected to global interactive processing of anchor point routing and boundary-aware region refinement processing respectively, resulting in global enhanced features and local refined features; the global enhanced features and the local refined features are subjected to adaptive structural fusion processing, and then residual fusion is performed with the input features to output the current collaborative enhanced features; the current collaborative enhanced features are input into the next collaborative reconstruction unit for processing, and finally output deep enhanced features; the deep enhanced features are sequentially subjected to upsampling and pixel reconstruction processing to output the final super-resolution image.

2. The image super-resolution reconstruction method based on anchor point guidance and boundary refinement as described in claim 1, characterized in that, The process involves performing global interactive anchor routing on the input features, including: performing learnable spatial compression on the input features to obtain compact features; extracting structural anchors from the compact features to generate a set of regional anchors; performing multi-head self-attention processing on the set of regional anchors to obtain a set of global enhancement anchors; propagating the information from the set of global enhancement anchors back to the compact features based on a sparse routing mechanism to obtain global propagation features; performing local detail compensation processing on the compact features, and adaptively gating fusion of the obtained local detail features with the global propagation features to obtain fused compact features; performing learnable spatial restoration processing on the fused compact features, and then performing residual fusion with the input features to output the global enhancement features.

3. The image super-resolution reconstruction method based on anchor point guidance and boundary refinement as described in claim 2, characterized in that, The region anchor point generation expression in the set of region anchor points is: In the formula, Indicates the area anchor point; Indicates position Importance score Indicates position eigenvectors; The first part represents the compact feature partitioning. A local area; This indicates the total number of anchor points in the region.

4. The image super-resolution reconstruction method based on anchor point guidance and boundary refinement as described in claim 2, characterized in that, Perform multi-head self-attention processing on the set of regional anchor points to obtain a global enhanced anchor point set. The process expression is as follows: In the formula, Represents the set of global enhancement anchor points; 、 and These represent the sets of regional anchor points respectively. The query vector, key vector, and value vector obtained by linear projection; Indicates the single-head channel dimension; This represents the normalized exponential function; This represents the matrix transpose operation.

5. The image super-resolution reconstruction method based on anchor point guidance and boundary refinement as described in claim 4, characterized in that, Based on the sparse routing mechanism, the information of the global enhanced anchor set is propagated back to the compact feature to obtain the global propagation feature. The process expression is as follows: In the formula, A token sequence representing global propagation characteristics; This means only retaining each space token and the top-ranked response. Masking operations for connecting anchor points in each region; Represents a sequence of tokens in space. The query vector obtained by linear mapping; and These represent the sets of globally enhanced anchor points. The key vector and value vector obtained by linear projection; This indicates the number of anchor points connected to each space token.

6. The image super-resolution reconstruction method according to any one of claims 1 to 5, characterized in that, The input features are subjected to boundary-aware region refinement processing, including: dividing the input features into multiple non-overlapping local regions according to a preset window size to obtain a region token sequence; performing boundary-aware prior extraction on the input features to generate a single-channel structure response map; constructing a boundary bias based on the structure response map; performing boundary-modulated region attention processing on the region token sequence based on the boundary bias to obtain region enhancement features; performing cross-region information interaction processing with alternating window shifting on the region enhancement features to obtain a spatial feature map; performing convolutional local detail compensation processing on the input features to obtain local detail features; performing adaptive gating fusion of the spatial feature map and the local detail features, and then performing residual fusion with the input features to output the local refinement features.

7. The image super-resolution reconstruction method based on anchor point guidance and boundary refinement as described in claim 6, characterized in that, The boundary bias is constructed based on the structural response map and is expressed as follows: In the formula, Indicates the first in the region The token and the first Boundary offset between tokens; Indicates the scaling factor; The first element in the structural response diagram is represented by the second element. The structural response value corresponding to each token position; The first element in the structural response diagram is represented by the second element. The structure response value corresponding to each token position.

8. The image super-resolution reconstruction method based on anchor point guidance and boundary refinement as described in claim 6, characterized in that, Based on the boundary bias, region attention processing with boundary modulation is performed on the region token sequence. The process expression is as follows: ； In the formula, The attention weights represent the boundary modulation. This represents the normalized exponential function; This represents the matrix transpose operation; 、 and These represent the query vector, key vector, and value vector obtained by linear projection of the region token sequence, respectively. Indicates the single-head channel dimension; Indicates the learnable relative position offset; Indicates boundary offset; Indicates an optional region mask item; A token sequence representing region enhancement features; This indicates a linear projection operation.

9. The image super-resolution reconstruction method according to any one of claims 1 to 5, characterized in that, After adaptive structural fusion of the global enhancement features and the local refinement features, residual fusion is performed with the input features to output the current collaborative enhancement features. The process expression is as follows: ； In the formula, Indicates the first Fusion weight diagram of collaborative reconstruction units; This represents the Sigmoid activation function; This represents a 1×1 pointwise convolution operation; Indicates the first One global enhancement feature; Indicates the first A localized refinement feature; Indicates the first The collaborative enhancement features output by each collaborative reconstruction unit; Indicates the first The collaborative enhancement features output by each collaborative reconstruction unit; Indicates the total number of collaborative reconstruction units; This indicates an element-wise multiplication operation.

10. An image super-resolution reconstruction device that combines anchor point guidance and boundary refinement, characterized in that, The device includes: a shallow feature extraction module for performing shallow feature extraction processing on an input low-resolution image to obtain initial features; a network construction module for constructing a deep feature enhancement network, the deep feature enhancement network including multiple cascaded collaborative reconstruction units; a deep feature enhancement module for inputting the initial features into the deep feature enhancement network, which are then sequentially processed by each of the collaborative reconstruction units for feature enhancement processing. Specifically, in the current collaborative reconstruction unit, the input features are subjected to global interactive processing of anchor point routing and boundary-aware region refinement processing, resulting in global enhanced features and local refined features. The global enhanced features and the local refined features are then subjected to adaptive structural fusion processing, followed by residual fusion with the input features to output the current collaborative enhanced features. The current collaborative enhanced features are then input into the next collaborative reconstruction unit for processing, ultimately outputting deep enhanced features; and an image reconstruction module for sequentially performing upsampling and pixel reconstruction processing on the deep enhanced features to output the final super-resolution image.

Images