Remote sensing image building extraction method fusing double-space attention features

Abstract

The invention discloses a remote sensing image building extraction method fusing double-space attention features, and belongs to the technical field of image processing. The method comprises the following steps: firstly, inputting a training set of an Inria aviation data set into a ConvNeXt backbone network to extract shallow feature information; then, carrying out feature information enhancement processing on the feature information through a double-space attention feature fusion module (DAFF); and finally, a multi-modal dynamic enhancement module (FAM) is used in the last time of feature information up-sampling operation to carry out channel alignment operation of different levels of features, and feature information fusion errors are reduced. According to the method, the accuracy of building extraction in the remote sensing image is remarkably improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of image processing, and particularly relates to a method for extracting buildings from remote sensing images by fusing dual-spatial attention features. Background Art

[0002] Image segmentation, as one of the core technologies in computer vision research, aims to finely divide an image into several regions or independent objects, laying a solid foundation for subsequent in-depth analysis and processing. This technology is of inestimable value for understanding the connotation of images and extracting key information, and its application scope spans multiple key fields such as medical image diagnosis, remote sensing detection, security monitoring, and autonomous driving. Traditional image segmentation techniques focus on the fine analysis at the pixel level and use means such as edge recognition, threshold definition, and region expansion to effectively distinguish regions. For example, edge recognition technology outlines the object contour by capturing the drastic fluctuations in pixel intensity; while the threshold segmentation strategy simply and efficiently divides the image into foreground and background parts based on the gray-scale attributes of pixels.

[0003] In recent years, the booming development of deep learning technology has brought revolutionary breakthroughs to image segmentation. The segmentation methods based on convolutional neural networks (CNNs) have achieved a significant improvement in segmentation accuracy due to their powerful feature learning ability. Among them, classic models such as the Fully Convolutional Network (FCN), U-Net, and Mask R-CNN, by constructing an encoder-decoder framework, deeply mine the image context information and achieve accurate positioning of the target region.

[0004] It is worth mentioning that the introduction of the Transformer architecture has injected new vitality into the field of image segmentation. Its unique self-attention mechanism can efficiently process the sequential representation of images, thus significantly improving the segmentation performance. This innovation enables the model to better capture long-range dependencies and comprehensively grasp the global context information, providing strong support for processing large-scale and complex image segmentation tasks.

[0005] However, despite many advancements, the field of image segmentation still faces many challenges. Problems such as incomplete segmentation and insufficient fusion of feature information still exist and urgently need more advanced technical means and algorithm innovations to solve. Summary of the Invention

[0006] The purpose of the present invention is to provide a method for extracting buildings from remote sensing images by fusing dual-spatial attention features, which uses the designed dual-spatial attention module DAFF to perform feature enhancement of global information on the shallow features extracted by the ConvNeXt network, then fuses them with the upsampled features through skip connections, and uses the feature alignment module FAM to efficiently fuse the channel-level features of the feature information and finally outputs a segmentation map with a higher precision rate.

[0007] The present invention is implemented through the following technical solutions:

[0008] Step 1: Apply random local deformation field perturbations (deformation amplitude δ ≤ 15% of the image size) to the training images in the Inria aerial dataset to generate adversarial samples containing occluded buildings and broken contours. Use an improved ConvNeXt backbone network (integrating a gradient-sensitive deformable convolution module) to extract features through a two-stage training strategy;

[0009] Step 2: Perform axial window partitioning on the shallow features output by the encoder (horizontal window 1×H, vertical window W×1). Calculate sparse self-attention weights within the window, fuse with the original features through gated residual connections, and adopt a four-way rotatable convolution kernel group (0°, 45°, 90°, 135°). After generating the direction response map, screen the response branches aligned with the main axis of the building through a dynamic routing algorithm, calculate the gradient magnitude difference map of the dual-path output features, input them into a lightweight LSTM network to generate pixel-level fusion weights, and achieve non-linear interpolation;

[0010] Step 3: Dynamically select the interpolation multiple (1×, 2×, 4×) according to the building scale distribution, compensate for the deformation error through a deformable convolution kernel, embed a frequency-domain attention module in the encoder-decoder skip path (perform DCT transformation on the feature map to generate a frequency-domain mask to suppress high-frequency noise), and finally filter out non-closed artifacts through a graph structure connectivity detection module (based on the adjacency matrix of the building contour vertices) for the output features to generate a topologically complete building mask.

[0011] Furthermore, the specific content of Step 1 is as follows: Dynamically adapt the task for data partitioning: Based on the Inria aerial dataset, aiming at the statistical distribution of the geometric complexity (aspect ratio, edge curvature, topological connectivity) of the building contours, adopt a gradient-sensitive hierarchical sampling strategy to ensure that the training set and the test set cover all building form categories (such as rectangles, L-shapes, polygons); adjust the partitioning ratio to 75% training set + 25% test set, where 25% of the test set contains 20% of regular samples and 5% of adversarial samples (such as occluded buildings, shadow interference images) for verifying the robustness of the model; use an improved ConvNeXt backbone network for feature decoupling pre-training.

[0012] Furthermore, the specific steps of Step 2 are as follows: Input the extracted shallow features into the DAFF module, and generate enhanced feature maps through the following steps: Adaptive noise suppression and alignment: Use wavelet domain gradient guided filtering (WGF) to suppress high-frequency noise in the shallow features, correct the spatial offset of the feature map through optical flow-guided spatial alignment, use deformable group convolution (DGC) and gated residual learning (GRL) to fuse multi-scale contexts, design spatio-temporal-spectral collaborative attention (TSFCA), calculate attention weights by jointly considering spatial, temporal, and frequency domain features, and gradually fuse shallow details and deep semantics through a multi-stage feature pyramid (MFP). In each stage, an adaptive interpolation kernel is used for resolution reconstruction.

[0013] Furthermore, the specific DAFF module is as follows: Feature dynamic decomposition: The multi-level feature maps (F1 - F4) output by the ConvNeXt backbone network are dimensionally reduced by 1×1 convolution and then input into a feature decomposition unit driven by a deformable convolution kernel. By dynamically adjusting the deformation parameters (offsets Δx, Δy) of the convolution kernel, high-frequency edge features and low-frequency semantic features are separated; Heterogeneous attention paths: Path A (axial sparse attention): Perform axial shift window partitioning (horizontal window size 1×11, vertical window size 11×1) on the high-frequency features in sequence, calculate long-range dependencies through sparse self-attention within the window, and fuse with the original features through gated residual connections; Path B (multi-directional convolution enhancement): Use a direction-sensitive convolution group (deploy a group of 3×3 rotatable convolution kernels in each of the 0°, 45°, 90°, and 135° directions) for the low-frequency features. After generating the direction response map, screen the response branches strongly related to the target building contour through a dynamic routing algorithm; Gradient-sensitive fusion: Calculate the gradient magnitude difference map of the output features of Path A and Path B, and input it into a lightweight LSTM network to predict pixel-level fusion weights; Nonlinear interpolation fusion of the dual-path features based on the weights (formula: Fout = α⊙FA+(1 - α)⊙FB, where α is dynamically generated by the LSTM output); Feature recalibration: The fused features are processed by a channel-spatial decoupled attention module and finally integrated and output through a cross-branch gating mechanism to form topologically optimized features of the building contour.

[0014] Furthermore, the specific steps of Step 3 are as follows: In the feature enhancement stage, the system performs multi-scale invertible upsampling (MIRV-Up) operations, which specifically include:

[0015] DAFF Processing: Dynamically calibrate the decoder input features through a deformable channel-spatial gating module (DCSGM). Adopt multi-branch residual learning to fuse shallow details and deep semantic information. Generate a dynamic interpolation kernel through a lightweight inverse network and gradually restore the resolution. After each upsampling, perform edge-guided residual connection, extract edge features using the Canny operator and fuse them with weights; at the end of the encoder, adopt a multi-modal dynamic enhancement module (FAM) to perform adaptive domain alignment on multi-scale features, specifically including: dynamically adjust the local receptive field through a deformable convolution kernel (DeformConv), capture features in irregular regions, introduce an inter-domain similarity metric (DSM), calculate the cross-domain correlation weights between feature channels using the Mahalanobis distance, and input the aligned features into a learnable upsampling kernel to generate dynamic interpolation weights through a lightweight MLP.

[0016] Furthermore, the multi-modal dynamic enhancement module (FAM) is specifically as follows: Feature dynamic encoding stage: Perform dynamic convolution operations on multi-level feature maps using an adaptive receptive field convolutional kernel, adjust the convolution kernel size (from 3×3 to 7×7) according to local texture complexity, and generate multi-scale context-aware features; perform dynamic channel weighting on the feature maps through a channel recalibration network (CRN), and the CRN is driven by a gated recurrent unit (GRU). Graph attention feature distillation stage: Map the encoded feature maps into graph-structured data, with nodes representing local feature blocks and edge weights calculated by the cosine similarity between nodes; perform feature distillation based on a graph attention network (GAT), and model the long-range dependencies between nodes through a multi-head graph attention mechanism. Dual-path feature recombination stage: Path 1: Perform asymmetric spatial pooling (ASP) on the key-value feature matrices K / V, compress the feature maps along the height and width directions respectively to generate spatially compact context features; Path 2: Generate a query vector Q through a dynamic channel interaction module, and the DCIM captures features in irregular regions using deformable convolution (Deformable Conv), constructs a Q-K association matrix using the Manhattan distance instead of the cosine similarity, and performs sparse normalization through Gumbel-Softmax. Resolution reconstruction and fusion stage: Replace bilinear interpolation with a learnable upsampling kernel (LUK), and the LUK generates dynamic interpolation weights through a lightweight MLP; introduce an edge-aware residual connection (EARC), and perform weighted fusion on the upsampled feature maps and the contour masks extracted from the initial feature maps by an edge detection operator (such as Sobel).

[0017] The beneficial effects of the present invention are as follows:

[0018] Through the DAFF module and the FAM module, the present invention enables the model to pay more attention to multi-dimensional scale feature information, enhances the feature information, and enriches the diversity of information features, providing an effective method for processing image feature information. Brief Description of the Drawings

[0019] To more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the accompanying drawings required for the description of the embodiments. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other accompanying drawings can also be obtained based on these drawings.

[0020] Figure 1 It is a schematic diagram of the model structure of the present invention;

[0021] Figure 2 It is a schematic diagram of the structure of the DAFF module;

[0022] Figure 3 It is a schematic diagram of the structure of the FAM module;

[0023] Figure 4 It is a visual comparison of the model of the present invention and other advanced models on the Inria aerial dataset; Specific implementation manners

[0024] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the protection scope of the present invention.

[0025] Please refer to Figures 1-4 As shown, the present invention is a method for extracting buildings from remote sensing images that integrates dual - spatial attention features, including the following steps:

[0026] Step 1: Apply a random local deformation field perturbation (deformation amplitude δ ≤ 15% of the image size) to the training images in the Inria aerial dataset to generate adversarial samples containing occluded buildings and broken contours. Use an improved ConvNeXt backbone network (integrating gradient - sensitive deformable convolution modules) to extract features through a two - stage training strategy;

[0027] Step 2: Perform axial window partitioning on the shallow features output by the encoder (horizontal window 1×H, vertical window W×1). Calculate sparse self - attention weights within the window, fuse with the original features through gated residual connections, use a four - way rotatable convolution kernel group (0°, 45°, 90°, 135°). After generating the direction response map, screen the response branches aligned with the main axis of the building through a dynamic routing algorithm, calculate the gradient magnitude difference map of the dual - path output features, and input it into a lightweight LSTM network to generate pixel - level fusion weights to achieve non - linear interpolation;

[0028] Step 3: Dynamically select the interpolation factor (1×, 2×, 4×) according to the building scale distribution, compensate for the deformation error through a deformable convolution kernel, embed a frequency-domain attention module in the encoder-decoder skip path (perform DCT transformation on the feature map to generate a frequency-domain mask to suppress high-frequency noise), and finally filter out non-closed artifacts through a graph structure connectivity detection module (based on the building contour vertex adjacency matrix) to generate a topologically complete building mask.

[0029] As an optimized solution of the above embodiment, the specific content of Step 1 is as follows: Dynamically adapt the task for data partitioning: Based on the Inria aerial dataset, for the statistical distribution of the geometric complexity (aspect ratio, edge curvature, topological connectivity) of building contours, adopt a gradient-sensitive hierarchical sampling strategy to ensure that the training set and the test set cover all building form categories (such as rectangles, L-shaped, polygons); adjust the partitioning ratio to 75% training set + 25% test set, where 25% of the test set contains 20% regular samples and 5% adversarial samples (such as occluded buildings, shadow-interfered images) for verifying the robustness of the model; use an improved ConvNeXt backbone network for feature decoupling pre-training.

[0030] As an optimized solution of the above embodiment, the specific content of Step 2 is as follows: Input the extracted shallow features into the DAFF module, and generate an enhanced feature map through the following steps: Adaptive noise suppression and alignment: Use wavelet-domain gradient-guided filtering (WGF) to suppress high-frequency noise in the shallow features, correct the spatial offset of the feature map through optical flow-guided spatial alignment, use deformable group convolution (DGC) and gated residual learning (GRL) to fuse multi-scale context, design spatio-temporal-spectrum collaborative attention (TSFCA), jointly calculate the attention weights of spatial, temporal, and frequency-domain features, and gradually fuse shallow details and deep semantics through a multi-stage feature pyramid (MFP), and use an adaptive interpolation kernel for resolution reconstruction at each stage.

[0031] As an optimized solution for the above embodiments, the DAFF module specifically includes: Feature Dynamic Decomposition: The multi-level feature maps (F1-F4) output by the ConvNeXt backbone network are reduced in dimension through 1×1 convolution and then input into a feature decomposition unit driven by deformable convolution kernels. By dynamically adjusting the convolution kernel deformation parameters (offsets Δx, Δy), high-frequency edge features and low-frequency semantic features are separated; Heterogeneous Attention Path: Path A (Axial Sparse Attention): Axial shift window partitioning (horizontal window size 1×11, vertical window size 11×1) is sequentially performed on the high-frequency features. Long-range dependencies are calculated through sparse self-attention within the window, and the original features are fused through gated residual connections; Path B (Multi-directional Convolution Enhancement): For low-frequency features, a direction-sensitive convolution group (a set of 3×3 rotatable convolution kernels are deployed in each of the 0°, 45°, 90°, and 135° directions) is used. After generating the direction response map, the response branches strongly related to the target building contour are screened through a dynamic routing algorithm; Gradient-Sensitive Fusion: Calculate the gradient magnitude difference map of the output features of Path A and Path B, and input it into a lightweight LSTM network to predict the pixel-level fusion weight; Nonlinear interpolation fusion is performed on the dual-path features based on the weight (formula: Fout = α⊙FA+(1-α)⊙FB, where α is dynamically generated by the LSTM output); Feature Re-calibration: The fused features are processed by a channel-spatial decoupled attention module and finally integrated and output through a cross-branch gating mechanism to form the topologically optimized features of the building contour.

[0032] +(1-α)⊙FB, where α is dynamically generated by the LSTM output); Feature Re-calibration: The fused features are processed by a channel-spatial decoupled attention module and finally integrated and output through a cross-branch gating mechanism to form the topologically optimized features of the building contour.

[0033] As an optimized solution for the above embodiments, Step 3 specifically includes: In the feature enhancement stage, the system performs multi-scale invertible upsampling (MIRV-Up) operations, which specifically include: DAFF processing: The decoder input features are dynamically calibrated through a deformable channel-spatial gating module (DCSGM). Multi-branch residual learning is used to fuse shallow details and deep semantic information, and a dynamic interpolation kernel is generated through a lightweight inverse network to gradually restore the resolution. After each upsampling, edge-guided residual connections are performed, and edge features are extracted using the Canny operator and weighted and fused; At the end of the encoder, a multi-modal dynamic enhancement module (FAM) is used to perform adaptive domain alignment on multi-scale features, which specifically includes: Dynamically adjusting the local receptive field through a deformable convolution kernel (DeformConv) to capture features in irregular regions, introducing an inter-domain similarity measure (DSM), calculating the cross-domain correlation weights between feature channels using the Mahalanobis distance, and inputting the aligned features into a learnable upsampling kernel to generate dynamic interpolation weights through a lightweight MLP.

[0034] As an optimized solution of the above embodiments, the multi-modal dynamic enhancement module (FAM) is specifically as follows: Feature dynamic encoding stage: An adaptive receptive field convolutional kernel is used to perform dynamic convolution operations on multi-level feature maps, and the convolutional kernel size (3×3 to 7×7) is adjusted according to the local texture complexity to generate multi-scale context-aware features; the feature maps are dynamically weighted in the channel domain through a channel recalibration network (CRN), and the CRN is driven by a gated recurrent unit (GRU). Graph attention feature distillation stage: The encoded feature maps are mapped into graph-structured data, the nodes represent local feature blocks, and the edge weights are calculated by the cosine similarity between nodes; feature distillation is performed based on a graph attention network (GAT), and the long-range dependencies between nodes are modeled through a multi-head graph attention mechanism. Dual-path feature recombination stage: Path one: Asymmetric spatial pooling (ASP) is performed on the key-value feature matrices K / V to compress the feature maps along the height and width directions respectively, generating spatially compact context features; Path two: A query vector Q is generated through a dynamic channel interaction module, the DCIM uses deformable convolution (Deformable Conv) to capture features of irregular regions, uses the Manhattan distance to replace the cosine similarity to construct a Q-K association matrix, and performs sparse normalization through Gumbel-Softmax. Resolution reconstruction and fusion stage: A learnable upsampling kernel (LUK) is used to replace bilinear interpolation, and the LUK generates dynamic interpolation weights through a lightweight MLP; an edge-aware residual connection (EARC) is introduced, and the upsampled feature maps are weighted and fused with the contour masks extracted from the initial feature maps through an edge detection operator (such as Sobel).

[0035] Although the specific implementation manners of the present invention are described above in conjunction with the accompanying drawings, it is not a limitation on the protection scope of the present invention. Those skilled in the art should understand that various modifications or deformations that can be made without creative efforts on the basis of the technical solutions of the present invention are still within the protection scope of the present invention.

Claims

1. A remote sensing image building extraction method integrating dual spatial attention features, characterized in that: The remote sensing image building extraction method integrating dual spatial attention features comprises the following steps: Step 1: Apply random local deformation field perturbations (deformation amplitude δ≤15% image size) to the training images in the Inria aerial dataset to generate adversarial samples containing occluded buildings and broken contours. Use an improved ConvNeXt backbone network (with an integrated gradient-sensitive deformable convolution module) to extract features through a two-stage training strategy. Step 2: Divide the shallow features output by the encoder into axial windows (horizontal window 1×H, vertical window W×1), calculate the sparse self-attention weights in the window, fuse them with the original features through gated residual connection, use a four-way rotatable convolution kernel group (0°, 45°, 90°, 135°), generate a directional response map, and then use a dynamic routing algorithm to select the response branches aligned with the main axis of the building, calculate the gradient amplitude difference map of the dual-path output features, and input the lightweight LSTM network to generate pixel-level fusion weights to achieve nonlinear interpolation; Step 3: Dynamically select the interpolation multiple (1×, 2×, 4×) according to the building scale distribution, compensate the deformation error through the deformable convolution kernel, embed the frequency domain attention module in the encoder-decoder skip path (perform DCT transformation on the feature map, generate frequency domain mask to suppress high-frequency noise), and finally output the feature through the graph structure connectivity detection module (based on the building outline vertex adjacency matrix) to filter out the non-closure artifacts and generate a topologically complete building mask.

2. The method according to claim 1, characterized in that The step one is specifically as follows: dynamic task adaptation data partitioning: based on the Inria aerial dataset, a gradient-sensitive stratified sampling strategy is adopted for the statistical distribution of the geometric complexity of building outlines (aspect ratio, edge curvature, topological connectivity) to ensure that the training set and the test set cover all building form categories (such as rectangles, L-shapes, polygons); the partition ratio is adjusted to 75% training set + 25% test set, of which the 25% test set contains 20% regular samples and 5% adversarial samples (such as occluded buildings, shadow-interfered images) to verify the robustness of the model; and an improved ConvNeXt backbone network is used for feature decoupling pre-training.

3. The method according to claim 1, characterized in that The step 2 is specifically as follows: the extracted shallow features are input into the DAFF module, and an enhanced feature map is generated through the following steps: Adaptive noise suppression and alignment: high-frequency noise of shallow features is suppressed by using wavelet-domain gradient-guided filtering (WGF), the spatial offset of the feature map is corrected by optical flow-guided spatial alignment, multi-scale context is fused by using deformable grouped convolution (DGC) and gated residual learning (GRL), spatiotemporal-spectral collaborative attention (TSFCA) is designed, and attention weights are calculated by combining spatial, temporal and frequency domain features. Shallow details and deep semantics are gradually fused through a multi-stage feature pyramid (MFP), and resolution is reconstructed by using an adaptive interpolation kernel at each stage.

4. The method according to claim 3, characterized in that The DAFF module is specifically as follows: dynamic feature decomposition: the multi-level feature map (F1-F4) output by the ConvNeXt backbone network is reduced in dimension by 1×1 convolution, and then input into the feature decomposition unit driven by the deformable convolution kernel, and the high-frequency edge features and low-frequency semantic features are separated by dynamically adjusting the convolution kernel deformation parameters (offset Δx, Δy); heterogeneous attention path: Path A (axial sparse attention): axial shift window division is performed on high-frequency features in sequence (horizontal window size 1×11, vertical window size 11×1), long-range dependencies are calculated in the window through sparse self-attention, and fused with the original features through gated residual connection; Path B ( Multi-directional convolution enhancement): A direction-sensitive convolution group is used for low-frequency features (a group of 3×3 rotatable convolution kernels are deployed in each direction of 0°, 45°, 90°, and 135°). After generating a directional response map, a dynamic routing algorithm is used to select response branches that are strongly related to the target building outline; Gradient-sensitive fusion: The gradient amplitude difference map of the output features of path A and path B is calculated, and it is input into a lightweight LSTM network to predict pixel-level fusion weights; Nonlinear interpolation fusion of dual-path features is performed based on weights (formula: Fout=α⊙FA+(1-α)⊙FBFout=α⊙FA+(1-α)⊙FB, where α is dynamically generated by LSTM output); feature Recalibration: The fused features are processed by the channel-space decoupled attention module and finally integrated through the cross-branch gating mechanism to form the topological optimization features of the building outline.

5. The method according to claim 1, characterized in that The step three is specifically as follows: in the feature enhancement stage, the system performs a multi-scale reversible upsampling (MIRV-Up) operation, specifically including: DAFF processing: dynamically calibrating the decoder input features through a deformable channel-spatial gating module (DCSGM), using multi-branch residual learning to fuse shallow details with deep semantic information, generating a dynamic interpolation kernel through a lightweight inverse network, and gradually restoring the resolution. After each upsampling, edge-guided residual connections are performed, and edge features are extracted and weightedly fused using the Canny operator; at the end of the encoder, a multi-modal dynamic enhancement module (FAM) is used to adaptively align multi-scale features to the domain, specifically including: dynamically adjusting the local receptive field through a deformable convolution kernel (DeformConv), capturing irregular region features, introducing an inter-domain similarity measure (DSM), using the Mahalanobis distance to calculate the cross-domain correlation weights between feature channels, inputting the aligned features into the learnable upsampling kernel, and generating dynamic interpolation weights through a lightweight MLP.

6. The method according to claim 5, characterized in that The multimodal dynamic enhancement module (FAM) is specifically as follows: feature dynamic encoding stage: using adaptive receptive field convolution kernel to perform dynamic convolution operation on multi-level feature maps, adjusting the convolution kernel size (3×3 to 7×7) according to local texture complexity, and generating multi-scale context-aware features; The feature map is dynamically weighted in the channel domain through the channel recalibration network (CRN), and the CRN is driven by the gated recurrent unit (GRU). Graph attention feature distillation stage: the encoded feature map is mapped to graph structure data, the nodes represent local feature blocks, and the edge weights are calculated by the cosine similarity between nodes; feature distillation is performed based on the graph attention network (GAT), and the long-range dependency relationship between nodes is modeled through the multi-head graph attention mechanism. Dual-path feature recombination stage: Path 1: Asymmetric spatial pooling (ASP) is implemented on the key-value feature matrix K / V, and the feature map is compressed along the height and width directions respectively to generate spatially compact context features; Path 2: Generate query vector Q through dynamic channel interaction module. DCIM uses deformable convolution (Deformable Conv) to capture irregular area features, Manhattan distance is used instead of cosine similarity to construct QK association matrix, and sparse normalization is performed through Gumbel-Softmax. Resolution reconstruction and fusion stage: Use learnable upsampling kernel (LUK) to replace bilinear interpolation. The LUK generates dynamic interpolation weights through lightweight MLP; introduce edge-aware residual connection (EARC), and perform weighted fusion of the upsampled feature map and the contour mask extracted by the initial feature map through edge detection operator (such as Sobel).

Images