A building edge perception method based on dual coding feature interaction

Through dual-coded feature interaction and deep supervision technology, the problems of low extraction accuracy and poor robustness in building edge perception are solved, and the fine extraction and robustness of building edges in high-resolution remote sensing images are achieved.

CN118644785BActive Publication Date: 2024-11-26TAIYUAN UNIVERSITY OF TECHNOLOGY
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411125192.9
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-08-16
Publication Date
2024-11-26
Estimated Expiration
2044-08-16

AI Technical Summary

Technical Problem

The prior art has problems in building edge perception with low extraction accuracy, poor extraction effect on irregular edges, and poor robustness. Especially in high-resolution remote sensing images, traditional semantic segmentation methods are time-consuming and generalized. Poor, deep learning algorithms ignore global context information, resulting in poor building edge recognition effect.

Method used

Using a dual-coded feature interaction method, the feature interaction module and edge perception module are designed through the dual-path coding structure of Swin Transformer and convolutional neural network, and combined with deep supervision technology, the network's recognition accuracy and adaptability to the edge of the building is improved.

Benefits of technology

High-precision extraction of building edges is achieved, the extraction effect of irregular edges is improved, the model's robustness to noise and outliers is enhanced, and the training speed and segmentation performance are improved.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118644785B_ABST
    Figure CN118644785B_ABST
Patent Text Reader

Abstract

The invention provides a building edge perception method based on dual-coding feature interaction, belonging to the technical field of building edge perception; the technical problem to be solved is: to provide an improvement of the building edge perception method based on dual-coding feature interaction; the technical scheme adopted to solve the technical problem is: to establish a dual-path coding structure in a network model, which is composed of an encoder based on a Swin Transformer and an encoder based on a residual module; to design a feature interaction module in the network model, which includes two parallel convolution branches and an interaction branch, and to design an edge perception module in the network model based on deformable convolution and ordinary convolution, and to dynamically assign different weights to different types of convolution branches, so that the network model allocates different attention levels to regular buildings and irregular buildings, thereby improving the extraction effect of the irregular building boundary; the invention is applied to building edge perception.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention provides a building edge perception method based on dual-coding feature interaction, belonging to the technical field of building edge perception. Background Art

[0002] The rapid and accurate extraction of building information is an important means for urban planning, disaster prevention and mitigation, and the construction and updating of geographic information databases. The main task is to process the collected building image information. At present, in order to apply it to different fields, the semantic segmentation method of efficient interpretation of high-resolution remote sensing images is used to analyze the building images, mainly using traditional semantic segmentation methods and deep learning semantic segmentation methods.

[0003] The traditional semantic segmentation method actually obtains geometric features manually and then uses classic machine learning algorithms to classify the features. This method is time-consuming, labor-intensive, costly, and requires a lot of prior knowledge to support it, resulting in the interpretation results being not objective and stable enough. In addition, the remote sensing images obtained using this method will also be affected by objective conditions, resulting in poor generalization ability and unable to meet the application requirements of high-resolution remote sensing images under complex backgrounds.

[0004] The deep learning algorithms currently used have excellent performance in the field of semantic segmentation. Deep learning is a learning method based on deep neural networks. It automatically learns deep abstract features through a network structure of multi-layer neurons. It does not need to consider features such as spectrum, geometry, texture and shadow, and has good interpretation ability in complex backgrounds. The semantic segmentation method based on convolutional neural networks can achieve automatic interpretation, but it still has defects in the application field of edge extraction of high-resolution remote sensing building images. For example, convolutional neural networks pay more attention to local information of the image and ignore the equally important global context information, which can help the network better understand the important role of a certain pixel or area in the entire image. Convolutional neural networks usually improve the accuracy of building extraction by deepening the network depth, but in this process it will cause the learning ability of shallow networks to decrease, which greatly affects the model's prediction effect on building edges. In addition, remote sensing images contain rich semantic information and there is a certain degree of redundancy, which still has obstacles to the fine extraction of building edge information. Summary of the invention

[0005] In order to overcome the deficiencies in the prior art, the present invention aims to solve the technical problem of providing an improved building edge perception method based on dual-coding feature interaction.

[0006] In order to solve the above technical problems, the technical solution adopted by the present invention is: a building edge perception method based on dual coding feature interaction, comprising the following steps:

[0007] Step 1: Establish a dual-path encoding structure in the network model, which consists of an encoder based on Swin Transformer and an encoder based on residual module;

[0008] Step 2: Design a feature interaction module in the network model, including two parallel convolution branches and one interaction branch. The feature interaction module further fuses and refines the feature information of the Swin Transformer and the convolutional neural network to perform effective information interaction.

[0009] Step 3: Based on deformable convolution and ordinary convolution, design the edge perception module in the network model. By dynamically assigning different weights to different types of convolution branches, the network model allocates different levels of attention to regular buildings and irregular buildings, thereby improving the extraction effect of irregular building boundaries.

[0010] Step 4: Use deep supervision to fully train the network model and improve the segmentation performance of the network;

[0011] Step 5: Process the building image information based on the trained network model and extract the building edge contours in detail.

[0012] The specific design method of the two parallel convolution branches in the feature interaction module of step 2 is:

[0013] First, the feature map from the Swin Transformer is adjusted using a 1×1 convolution to adjust the number of channels. Then, two independent convolution branches are used to adjust the number of channels of the Swin Transformer feature map F swin conv And the convolutional neural network feature map F cnn For feature extraction, the calculation formula is:

[0014] ;

[0015] Where: BN(.) represents batch normalization, σ represents the rectified linear unit activation function, Con 3×3 represents a convolution operation using a 3×3 convolution kernel, F a conv and F b conv Represents different convolution branches to F swin conv and F cnn Feature map after feature extraction;

[0016] Finally, we use cross-branch connections to promote the fusion of different semantic information, where: a concat Yes Fa conv and F cnn Through the connection operation, the connection feature F b concat It is through F b conv and F swin conv The connection operation is formed, and the calculation formula is:

[0017] ;

[0018] Where: Concat(.) represents the concatenation operation.

[0019] The specific design method of the interaction branch in the feature interaction module of step 2 is:

[0020] First, the feature maps from different encoders are concatenated and fused, and then global average pooling and global maximum pooling operations are performed on the concatenated feature maps, focusing on overall information and significant information. Then, a fully connected layer with shared weights is used to reduce the feature dimension and refine the extracted features. Then, a fully connected layer with non-shared weights and a Softmax activation function are used to weight the features respectively. Finally, based on the obtained weight vector, the weights of the two convolution branches are adjusted and recalibrated to achieve adaptive capture of interactive information. The calculation formula used in this operation process is:

[0021] ;

[0022] Where: F ab concat Yes F swin conv and F cnn It is formed by the connection operation, GAP is the global average pooling operation, GMP is the global maximum pooling operation, FC share is a fully connected layer with shared weights, and δ is the common representation of cross attention;

[0023] ;

[0024] Where: FC α and FC β represents two fully connected layers with non-shared weights, α and β represent weight vectors for adjusting different convolution branches;

[0025] ;

[0026] Where: represents element-wise matrix multiplication operation, and Represents the feature map after the convolution branch weights are adjusted;

[0027] This module finally concatenates the Swin Transformer, the convolutional neural network, and the original feature map, and then uses a 1×1 convolution operation to obtain the final output result Out. The calculation formula is:

[0028] .

[0029] The specific design method of the edge perception module in step 3 is:

[0030] The branch using standard 3×3 convolution is used for feature extraction of regular buildings;

[0031] The deformable 3×3 convolution branch is used to extract features of irregularly shaped buildings, and the extracted features are weighted to increase the focus on the target object.

[0032] Design dynamic weight branch structure:

[0033] First, the global maximum pooling operation is used to compress the input feature map, and then two fully connected layers are used. The rectified linear unit activation function is used after the first fully connected layer, and the Softmax activation function is used after the second fully connected layer to generate the attention weights of the remaining two branches. The weights are calculated by the network based on the input features.

[0034] Finally, the module assigns the two calculated attention weights to different branches through dynamic weight update, and then adds them to the original input features to improve the adaptability of the network model to targets with unstable shapes. The calculation formula used in this operation process is:

[0035] ;

[0036] Where: F in and F out Represent the input feature map and output feature map of the module respectively, F standard out and F deformable out They represent the output feature maps using the standard convolution branch and the output feature maps using the deformable convolution branch, respectively. a and b They represent the attention weights assigned to different convolution branches.

[0037] The method of using deep supervision in step 4 is specifically as follows:

[0038] The prediction graph and the true label graph generated at each stage of the decoding process are optimized using the binary cross entropy loss function. By providing label signals at multiple levels, the network can reduce the noise interference of irrelevant factors and improve the training speed and segmentation performance of the network.

[0039] The beneficial effects of the present invention over the prior art are as follows: the building edge perception method provided by the present invention is implemented based on a designed dual-coding network model structure, and by setting a feature interaction module in the network model, the global context information extracted from the Swin Transformer and the local information extracted by the convolutional neural network are complemented by advantages, and the feature maps extracted by the Swin Transformer and the convolutional neural network branches are sent to the designed feature interaction module, and the features extracted by different convolution branches in the feature interaction module are recalibrated, so as to achieve the purpose of adaptively capturing interactive information, improve the recognition accuracy of the network model for the building edge, and ensure the integrity of edge extraction; in addition, in order to solve the problem of poor extraction effect of irregular edges, an edge perception module is designed, which introduces deformable convolution and further adjusts the importance of conventional factors and deformable factors in building extraction by updating dynamic weights, effectively improves the adaptability of the network model to targets with non-fixed shapes, strengthens the extraction of building edge contours, and achieves the goal of fine extraction; in addition, deep supervision technology is introduced in the model design process, so as to enhance the robustness of the network model to noise and outliers, and improve the training speed and effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] The present invention will be further described below in conjunction with the accompanying drawings:

[0041] Figure 1 A flowchart of the steps of using a network model to perceive the edge of a building in the present invention;

[0042] Figure 2 This is a flow chart of the steps of data processing using the feature interaction module of the present invention;

[0043] Figure 3 This is a flow chart of the steps of using the edge perception module to process data in the present invention;

[0044] Figure 4 A comparison diagram of the visualized results of the ablation experiment in an embodiment of the present invention;

[0045] Figure 5 Comparison of different network visualization results in the Inria Aerial Labeling dataset in an embodiment of the present invention Figure 1 ;

[0046] Figure 6Comparison of different network visualization results in the Inria Aerial Labeling dataset in an embodiment of the present invention Figure 2 . DETAILED DESCRIPTION

[0047] The present invention provides a building edge perception method based on dual-coding feature interaction, which specifically relates to high-resolution remote sensing image semantic segmentation technology. The high-resolution remote sensing image semantic segmentation technology specifically focuses on how to use advanced image processing technologies such as deep learning to classify and understand high-resolution remote sensing images at the pixel level, thereby realizing automatic recognition and analysis of ground objects and surface coverage.

[0048] The present invention applies the Transformer model used in the field of natural language processing to the field of computer vision in a pioneering and cross-domain manner. The advantages of Transformer are that it can retain more global context information and has excellent parallel processing capabilities. It has achieved good results in image classification, image detection and image segmentation tasks. SwinTransformer introduces a hierarchical construction method similar to a convolutional neural network to construct a hierarchical Transformer, which enables it to have the ability to serve as a general backbone network, but it ignores the equally important local information. In order to improve the edge missing and edge blur caused by information loss in the feature extraction process of the convolutional neural network and improve the edge recognition accuracy of buildings, especially irregular buildings, the present invention proposes a building edge perception method based on dual-coding feature interaction, and the dual-path coding structure adopted in the method is composed of SwinTransformer and a convolutional neural network.

[0049] The currently used method for extracting buildings from high-resolution remote sensing images based on convolutional neural networks focuses on obtaining local information and ignores the equally important global context information in the extraction process; the attention mechanism in Swin Transformer can model long-distance dependencies and better capture global context information, however, it does not take into account the importance of obtaining local information to some extent; due to the different focuses of the above two network models, the feature extraction of the network model is not sufficient, which further affects the prediction effect of the model. Therefore, the present invention aims to design a network model to complement the global context information from Swin Transformer and the local information from the convolutional neural network to solve the problem of low accuracy in extracting buildings from high-resolution remote sensing images; and since the buildings in high-resolution remote sensing images are usually irregular, and the traditional convolution operation is through a convolution kernel of fixed size and fixed shape Implementing the convolution operation will cause the receptive field corresponding to a certain position in the feature map to be relatively fixed, which is not conducive to the model's good modeling of geometric transformations; therefore, the present invention aims to improve the network's adaptability to targets with non-fixed shapes, enhance the network's ability to extract building edge contours, and solve the problem of imprecise extraction of irregular building edges in high-resolution remote sensing image building extraction tasks; in addition, deep learning models, especially complex models, are prone to overfitting, that is, they perform well on training data but poorly on unseen data, because the model learns the noise and details in the training data but does not capture the potential distribution of the data; therefore, the present invention aims to improve the training speed and effect of the network and solve the problem of poor robustness of the network model in the face of noise and outliers.

[0050] In order to solve the above technical problems, the present invention proposes a building edge perception method based on dual coding feature interaction. In order to achieve the corresponding perception effect, it is necessary to establish a corresponding network model; Figure 1As shown in the figure, the dual-path encoding structure in the network model consists of an encoder based on Swin Transformer and an encoder based on Residual Block. On this basis, the Feature Interaction Module (FIM) is first introduced to further fuse and refine the feature information of Swin Transformer and convolutional neural network for effective information interaction. Then, an Edge Awareness Module (EAM) is designed using deformable convolution and ordinary convolution. It dynamically assigns different weights to different types of convolution branches, so that the network model allocates different levels of attention to regular buildings and irregular buildings, thereby improving the extraction effect of irregular building boundaries. Finally, the network uses deep supervision (DS) to enable the network to be more fully trained and improve the segmentation performance of the network.

[0051] The specific structure of the feature interaction module is as follows: Figure 2 As shown in Figure 1, the module contains two parallel convolution branches and one interactive branch. It first uses 1×1 convolution to adjust the number of channels of the feature map from Swin Transformer, and then uses two independent convolution branches to adjust the number of channels of the Swin Transformer feature map (F swin conv ) and convolutional neural network feature map (F cnn ) for feature extraction. The specific calculation process is shown in formulas (1-1) to (1-3).

[0052] ;

[0053] Where BN(.) represents batch normalization, σ represents the rectified linear unit activation function, and Con 3×3 Represents a convolution operation using a 3×3 convolution kernel. a conv and F b conv Represents different convolution branches to F swin conv and F cnn Feature map after feature extraction.

[0054] Finally, we use cross-branch connections to promote the fusion of different semantic information. a concat Yes F a conv and F cnn Through the connection operation, the connection feature F bconcat It is through F b conv and F swin conv The connection operation is formed, and the calculation formula is shown in formulas (1-4) and (1-5).

[0055] ;

[0056] In the formula, Concat(.) represents the concatenation operation.

[0057] The interactive branch first concatenates and fuses the feature maps from different encoders, and then recalibrates the convolution branch features through a series of operations to achieve the purpose of adaptively capturing interactive information. See formulas (1-6) to (1-9). The interactive branch first performs global average pooling (GAP) and global maximum pooling (GMP) on the concatenated feature maps to focus on overall information and significant information. Then, a fully connected layer with shared weights is used to reduce the feature dimension and refine the extracted features. Then, a fully connected layer with non-shared weights and a Softmax operation are used to weight the features respectively. Finally, the weight vector obtained in the previous step is used to adjust the weights of the two convolution branches respectively.

[0058] ;

[0059] Where: F ab concat Yes F swin conv and F cnn It is formed by the connection operation, GAP is the global average pooling operation, GMP is the global maximum pooling operation, FC share is a fully connected layer with shared weights, and δ is the common representation of cross attention.

[0060] (1-8);

[0061] Where FC α and FC β represents two fully connected layers with non-shared weights, and α and β represent the weight vectors for adjusting different convolution branches.

[0062] (1-9);

[0063] In the formula represents element-wise matrix multiplication operation, and Represents the feature map after the convolution branch weights are adjusted.

[0064] This module finally concatenates the Swin Transformer branch, the convolutional neural network branch, and the original feature map, and then uses a 1×1 convolution operation to obtain the final output result Out. The specific calculation process is shown in formula (1-10).

[0065] (1-10).

[0066] The specific structure of the edge perception module is as follows Figure 3 As shown in the figure, the module is divided into three parts. The branch using standard 3×3 convolution is used for feature extraction of regular buildings, and the branch using deformable 3×3 convolution is used for feature extraction of irregularly shaped buildings. While using deformable 3×3 convolution, the extracted features are also weighted to increase the attention to the target object, thereby enhancing the network's feature extraction capability for deformable targets.

[0067] Dynamic weight branch structure see Figure 3 In the top part, it first uses the global maximum pooling operation to compress the input feature map, and then uses two fully connected layers. The rectified linear unit activation function is used after the first fully connected layer, and the Softmax function is used after the second fully connected layer. This function is used to generate the attention weights of the remaining two branches. This weight is calculated by the network model based on the input features, and is not two fixed values ​​set artificially.

[0068] Finally, the module assigns the two calculated attention weights to different branches through dynamic weight update, and then adds them to the original input features to further adjust the importance of conventional factors and deformable factors in building extraction, thereby improving the adaptability of the network model to objects with variable shapes. The calculation process is shown in formula (1-11).

[0069] (1-11);

[0070] Where F in and F out Represent the input feature map and output feature map of the module respectively, F standard out and F deformable out They represent the output feature maps using the standard convolution branch and the output feature maps using the deformable convolution branch, respectively. a and b They represent the attention weights assigned to different convolution branches.

[0071] The present invention adds deep supervision technology to the decoding process, and optimizes the prediction graph and the real label graph generated at each stage of the decoding process using the binary cross entropy loss function. By providing label signals at multiple levels, the network can reduce the noise interference of other irrelevant factors, improve the network training speed and segmentation performance. Example

[0072] (1) Experimental environment and parameter settings:

[0073] The server system used in the experiment is Ubuntu (20.04), the GPU is RTX 3090, the video memory is 24GB, the PyTorch deep learning framework 1.10.0, cuda 11.3, and python3.8. The AdamW optimizer and cosine strategy are used for model training in the experiment. Random horizontal and vertical flipping of data is used as data enhancement methods. For the Inria Aerial Labeling dataset, this experiment trained the model for 150 epochs, and for the Massachusetts Buildings dataset, this experiment trained the model for 100 epochs, with the basic learning rate set to 5e-4 and the batch size set to 8.

[0074] (2) Experimental strategy:

[0075] The experiment used two datasets, namely the Inria Aerial Labeling dataset and the Massachusetts Buildings dataset.

[0076] A total of four groups of ablation experiments were set up to verify whether the three improvements proposed in the present invention can improve the extraction accuracy of buildings on the Inria Aerial Labeling dataset by removing different modules based on the overall network model.

[0077] Group 1: Overall network model (All).

[0078] The second group: remove the feature interaction module (All-FIM) in the overall network model.

[0079] Group 3: Remove the edge-aware module (All-EAM) from the overall network model.

[0080] Group 4: Remove deep supervision from the overall network model (All-DS).

[0081] (3) Evaluation indicators:

[0082] This experiment used intersection over union (IoU), mean intersection over union (mIoU), recall rate, precision rate and F1 score to evaluate the extraction results of buildings.

[0083] (4) Ablation experiment:

[0084] The ablation experiment results are shown in Table 1 below.

[0085]

[0086] Table 1 Ablation experiment results

[0087] As shown in Table 1, the network model (All) proposed in the present invention has the highest building recognition accuracy, with an average intersection-over-union ratio of 88.02%, and precision and recall rates of 90.22% and 89.45% respectively. Compared with All-FIM without the introduction of FIM, F1 and mIoU of All increased by 1.16% and 1.29% respectively, compared with All-EAM without the introduction of EAM, F1 and mIoU of All increased by 1.6% and 1.69% respectively, and compared with All-DS without the use of deep supervision technology, F1 and mIoU of All increased by 0.63% and 0.65% respectively. It can be seen that the improvements proposed in the present invention can improve the extraction accuracy of buildings. The introduction of FIM and EAM has a significant effect on improving the accuracy, while DS is second.

[0088] Comparison of ablation experiment visualization results Figure 4 As shown in the figure, it can be seen that All has the best detail extraction ability, especially the most refined edge contour extraction of irregular buildings. Compared with All-FIM, All introduces a feature interaction module, which significantly improves the extraction integrity of building edges and reduces the boundary loss caused by insufficient feature extraction capabilities. Compared with All-EAM, All adds a boundary perception module, which improves the refinement of complex building boundary extraction and makes the boundary more regular. Compared with All-DS, All introduces deep supervision technology, which further improves the extraction effect of buildings and reduces the error between the real label and the real label.

[0089] (5) Comparative experiment:

[0090] The comparison results of different methods on the Inria Aerial Labeling dataset are shown in Table 2 below.

[0091]

[0092] Table 2 Comparison results of different methods on the Inria Aerial Labeling dataset

[0093] As can be seen from Table 2, the building edge perception network based on dual-coding feature interaction proposed in this invention has the best extraction effect, and all evaluation indicators have reached the optimal level. The intersection-over-union ratio of a single category of buildings reached 81.54%, exceeding Unet, Deeplabv3+, TransUnet, SegFormer, Attention U-Net and CE-Net by 1.23%, 4.54%, 2.35%, 4.95%, 3.35% and 1.4% respectively.

[0094] The present invention Figure 5 and Figure 6 The experimental results of several classic semantic segmentation models and the method proposed in the present invention are compared by visualization. Through analysis, it can be seen that the network model of the present invention has the best segmentation effect in edge detail extraction.

[0095] like Figure 5 and Figure 6 As shown in the second, third and fourth rows, for the situation where the environment is complex, the buildings are arranged in a disorderly manner and the shapes are irregular, the network model proposed in the present invention not only realizes the precise positioning of the buildings, but also has better processing ability for edge details than other network models; while the other network models cannot eliminate interference very well, and there are some errors, omissions and voids. At the same time, they cannot segment out finer building boundaries, and are prone to boundary missing and roughness problems caused by insufficient extraction; the network model adaptively and dynamically adjusts the model extraction process through feature interaction, improves the network's feature extraction ability for buildings in complex environments, reduces the phenomenon of errors and omissions, and the edge perception module enhances the ability to learn and model irregular buildings, improves the network's adaptability to targets with irregular shapes, and makes the edge extraction of irregular buildings more refined; in addition, by adding deep supervision, additional constraints are further added to the network model, thereby improving the accuracy of building extraction.

[0096] like Figure 5 and Figure 6 As shown in the first row, the network model has the smallest error with the true label when extracting the boundaries of buildings with various shapes and tortuous shapes. As shown in the fifth row, the network model also shows the best effect in extracting buildings with straight boundaries.

[0097] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein by equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A building edge perception method based on dual coding feature interaction, Features: The steps include: Step 1: Establish a dual-path encoding structure in the network model, which consists of an encoder based on Swin Transformer and an encoder based on residual module; Step 2: Design a feature interaction module in the network model, including two parallel convolution branches and one interaction branch. The feature interaction module further fuses and refines the feature information of the Swin Transformer and the convolutional neural network to perform effective information interaction. Step 3: Based on deformable convolution and ordinary convolution, design the edge perception module in the network model. By dynamically assigning different weights to different types of convolution branches, the network model allocates different levels of attention to regular buildings and irregular buildings, thereby improving the extraction effect of irregular building boundaries. Step 4: Use deep supervision to fully train the network model and improve the segmentation performance of the network; Step 5: Process the building image information based on the trained network model and extract the building edge contours in detail.

2. A building edge perception method based on dual coding feature interaction according to claim 1, Features: The specific design method of the two parallel convolution branches in the feature interaction module of step 2 is: First, the feature map from the Swin Transformer is adjusted using a 1×1 convolution to adjust the number of channels. Then, two independent convolution branches are used to adjust the number of channels of the Swin Transformer feature map F swin conv And the convolutional neural network feature map F cnn For feature extraction, the calculation formula is: ; Where: BN(.) represents batch normalization, σ represents the rectified linear unit activation function, Con 3×3 represents a convolution operation using a 3×3 convolution kernel, F a conv and F b conv Represents different convolution branches to F swin conv and F cnn Feature map after feature extraction; Finally, we use cross-branch connections to promote the fusion of different semantic information, where: a concat Yes F a conv and F cnn Through the connection operation, the connection feature F b concat It is through F b conv and F swin conv The connection operation is formed, and the calculation formula is: ; Where: Concat(.) represents the concatenation operation.

3. A building edge perception method based on dual coding feature interaction according to claim 2, Features: The specific design method of the interaction branch in the feature interaction module of step 2 is: First, the feature maps from different encoders are concatenated and fused, and then global average pooling and global maximum pooling operations are performed on the concatenated feature maps, focusing on overall information and significant information. Then, a fully connected layer with shared weights is used to reduce the feature dimension and refine the extracted features. Then, a fully connected layer with non-shared weights and a Softmax activation function are used to weight the features respectively. Finally, based on the obtained weight vector, the weights of the two convolution branches are adjusted and recalibrated to achieve adaptive capture of interactive information. The calculation formula used in this operation process is: ; Where: F ab concat Yes F swin conv and F cnn It is formed by the connection operation, GAP is the global average pooling operation, GMP is the global maximum pooling operation, FC share is a fully connected layer with shared weights, and δ is the common representation of cross attention; ; Where: FC α and FC β represents two fully connected layers with non-shared weights, α and β represent weight vectors for adjusting different convolution branches; ; Where: represents element-wise matrix multiplication operation, and Represents the feature map after the convolution branch weights are adjusted; This module finally concatenates the Swin Transformer, the convolutional neural network, and the original feature map, and then uses a 1×1 convolution operation to obtain the final output result Out. The calculation formula is: 。 4. A building edge perception method based on dual coding feature interaction according to claim 3, Features: The specific design method of the edge perception module in step 3 is: The branch using standard 3×3 convolution is used for feature extraction of regular buildings; The deformable 3×3 convolution branch is used to extract features of irregularly shaped buildings, and the extracted features are weighted to increase the focus on the target object. Design dynamic weight branch structure: First, the global maximum pooling operation is used to compress the input feature map, and then two fully connected layers are used. The rectified linear unit activation function is used after the first fully connected layer, and the Softmax activation function is used after the second fully connected layer to generate the attention weights of the remaining two branches. The weights are calculated by the network based on the input features. Finally, the module assigns the two calculated attention weights to different branches through dynamic weight update, and then adds them to the original input features to improve the adaptability of the network model to targets with unstable shapes. The calculation formula used in this operation process is: ; Where: F in and F out Represent the input feature map and output feature map of the module respectively, F standard out and F deformable out They represent the output feature maps using the standard convolution branch and the output feature maps using the deformable convolution branch, respectively. a and b They represent the attention weights assigned to different convolution branches.

5. A building edge perception method based on dual coding feature interaction according to claim 4, Features: The method of using deep supervision in step 4 is specifically as follows: The prediction graph and the true label graph generated at each stage of the decoding process are optimized using the binary cross entropy loss function. By providing label signals at multiple levels, the network can reduce the noise interference of irrelevant factors and improve the network training speed and segmentation performance.