High-resolution remote sensing image double-branch segmentation method

Abstract

The invention discloses a high-resolution remote sensing image double-branch segmentation method, which comprises the following steps: acquiring a high-resolution remote sensing image data set, preprocessing the high-resolution remote sensing image data set, and dividing the high-resolution remote sensing image data set into a training set, a verification set and a test set; constructing a semantic segmentation network model of the high-resolution remote sensing image, wherein the semantic segmentation network model comprises a detail branch, a semantic branch and an up-sampling part; inputting the training set into a high-resolution remote sensing image semantic segmentation network model for training, calculating a loss function, performing back propagation, updating network model parameters, obtaining an optimal parameter network model, and verifying the optimal parameter network model by using the verification set; and inputting the test set into the trained optimal parameter network model, and outputting an accurate segmentation image of the high-resolution remote sensing image. According to the method, feature extraction and fusion are carried out in a multi-level and multi-scale manner by capturing target information of different types and scales in the high-resolution remote sensing image, so that the segmentation precision of the high-resolution remote sensing image is improved.

Description

Technical Field

[0001] The invention relates to a high-resolution remote sensing image dual-branch segmentation method, belonging to the technical field of remote sensing image semantic segmentation. Background Art

[0002] Semantic segmentation of remote sensing images is one of the most important advances in remote sensing image technology. Its goal is to semantically classify each pixel in a remote sensing image, thereby providing a thorough understanding of the spatial layout and range of different elements in the remote sensing image and promoting the development of remote sensing applications such as land surveying, environmental monitoring, and urban planning.

[0003] With the development of deep learning, convolutional neural networks have been widely used in the field of image processing due to their powerful feature extraction capabilities. However, with the development of remote sensing technology, the multi-scale characteristics of remote sensing images have become increasingly prominent. The inherent limitations of convolution make convolutional neural networks limited in their perception of long-distance contextual information and lack a global understanding of images. The Transformer architecture, which can better capture global contextual information, provides a new idea for the semantic segmentation of remote sensing images. Wang et al. combined UNet and Transformer and proposed a new U-shaped network structure (Wang LB, Li R, Zhang C, et al. (2022). UNetFormer: A UNet-like transformer for efficient semantic segmentation of remote sensing urban scene imagery. ISPRS Journal of Photogrammetry and Remote Sensing, 190, 196-214), which has a CNN-based encoder and a Transformer-based decoder for effective semantic segmentation of remote sensing urban scene images. Chen et al. proposed the TransUNet network (Chen JN, Lu YY, Yu QH, et al. (2021). Transunet: Transformers make strong encoders for medical image segmentation. arXivpreprint arXiv:2102.04306), integrating the Transformer into the UNet encoder to enhance global information. However, the encoder-decoder structure will lead to the problem of small target information loss. The network may not be able to retain detailed information with high segmentation accuracy. In addition, the degree of fusion between CNN and Transformer is shallow. Some models simply splice the two and fail to give full play to the advantages of the network. In view of the high resolution and multi-scale features of remote sensing images, the segmentation effect of the model on targets of different types and scales in remote sensing images still needs to be improved. Therefore, how to further optimize these strategies to achieve more effective feature integration remains a challenge. Summary of the invention

[0004] The technical problem to be solved by the present invention is to provide a dual-branch segmentation method for high-resolution remote sensing images, which captures target information of different types and scales in high-resolution remote sensing images, performs feature extraction and fusion at multiple levels and scales, thereby improving the segmentation accuracy of high-resolution remote sensing images.

[0005] The present invention adopts the following technical solutions to solve the above technical problems: A high-resolution remote sensing image dual-branch segmentation method comprises the following steps: Step 1, obtain a high-resolution remote sensing image dataset, including the original remote sensing image and the corresponding label map, pre-process the high-resolution remote sensing image dataset, and divide it into a training set, a validation set, and a test set according to a set ratio; Step 2, constructing a high-resolution remote sensing image semantic segmentation network model, including a detail branch, a semantic branch and an upsampling part, the preprocessed remote sensing image is respectively passed through the detail branch and the semantic branch to obtain the output of the detail branch and the semantic branch respectively, and the output of the detail branch and the semantic branch respectively is passed through the upsampling part to obtain the output of the high-resolution remote sensing image semantic segmentation network model; The detail branch includes five layers, wherein the first layer includes a convolution layer and a pooling layer connected in sequence, the second layer includes a first residual module, a second residual module and a first bilateral feature guided fusion module connected in sequence, the third layer includes a third residual module, a fourth residual module and a second bilateral feature guided fusion module connected in sequence, the fourth layer includes a fifth residual module, a sixth residual module and a third bilateral feature guided fusion module connected in sequence, and the fifth layer includes a seventh residual module and an eighth residual module connected in sequence; the structures of the first to eighth residual modules are the same; The semantic branch includes six layers, wherein the first layer includes a bilateral feature aggregation module, the second layer includes a first improved depthwise separable convolution module and a first multi-layer perceptual axial attention module connected in sequence, the third layer includes a second improved depthwise separable convolution module and a second multi-layer perceptual axial attention module connected in sequence, the fourth layer includes a third improved depthwise separable convolution module, a first MobileVit module and a third multi-layer perceptual axial attention module connected in sequence, the fifth layer includes a fourth improved depthwise separable convolution module, a second MobileVit module and a fourth multi-layer perceptual axial attention module connected in sequence, and the sixth layer includes a fifth improved depthwise separable convolution module, a third MobileVit module and a fifth multi-layer perceptual axial attention module connected in sequence; the structures of the first to third MobileVit modules are the same; The preprocessed remote sensing image, i.e., the original input image, is used as the input of the first convolution layer of the detail branch; the detail feature map outputted from the first layer of the detail branch and the original input image are used as the input of the first bilateral feature aggregation module of the semantic branch; the output of the second residual module of the second layer of the detail branch and the output of the third multi-layer perceptual axial attention module of the fourth layer of the semantic branch are used as the input of the first bilateral feature guided fusion module of the second layer of the detail branch; the output of the fourth residual module of the third layer of the detail branch and the output of the fourth multi-layer perceptual axial attention module of the fifth layer of the semantic branch are used as the input of the second bilateral feature guided fusion module of the third layer of the detail branch; the output of the sixth residual module of the fourth layer of the detail branch and the output of the fifth multi-layer perceptual axial attention module of the sixth layer of the semantic branch are used as the input of the third bilateral feature guided fusion module of the fourth layer of the detail branch; The upsampling part includes a bilateral guided aggregation module and a segmentation head module connected in sequence. The detail feature map output by the fifth layer of the detail branch and the global feature map output by the sixth layer of the semantic branch are used as the input of the bilateral guided aggregation module. The segmentation head module outputs the final remote sensing image segmentation map. Step 3, input the training set of step 1 into the high-resolution remote sensing image semantic segmentation network model constructed in step 2 for training, calculate the loss function and perform back propagation, update the network model parameters, obtain the optimal parameter network model, and use the verification set of step 1 to verify the optimal parameter network model; Step 4: Input the test set of step 1 into the optimal parameter network model trained in step 3, and output the accurate segmentation map of the high-resolution remote sensing image.

[0006] As a preferred embodiment of the present invention, the specific process of step 1 is as follows: Get the Vaihingen and Potsdam remote sensing image datasets. Both the Vaihingen and Potsdam remote sensing image datasets include original remote sensing images of six categories: low vegetation, trees, impervious surfaces, cars, buildings, and clutter, as well as corresponding label images. The original remote sensing images and corresponding label images in the Vaihingen remote sensing image dataset are cropped using the farthest point random cropping data augmentation method so that the cropped image size is Pixels, get the preprocessed Vaihingen remote sensing image dataset; The original remote sensing images and corresponding label images in the Potsdam remote sensing image dataset are cropped into Pixel image, cropped The pixel images are enhanced using image enhancement technology to obtain the preprocessed Potsdam remote sensing image dataset; The preprocessed Vaihingen and Potsdam remote sensing image datasets were The training set, validation set, and test set are obtained by dividing the set into two parts.

[0007] As a preferred solution of the present invention, the bilateral feature aggregation module includes two branches, left and right. The left branch aggregates the detail feature map output by the first layer of the detail branch. Perform convolution with a kernel size of 3 and a step size of 1, BN normalization and ReLU activation operations to obtain the original input image The feature map of resolution a1; the right branch is for the original input image Perform convolution with a kernel size of 3 and a stride of 2, BN normalization and ReLU activation operations to obtain the original input image The feature map a2 of resolution is obtained by performing maximum pooling and average pooling with a pooling kernel size of 3 and a step size of 2 on the feature map a2 to obtain two original input images. After adding the feature maps a1, a3, and a4, the added result is convolved with a kernel size of 3 and a step size of 1, and then BN normalized and ReLU activated to obtain the original input image. The feature map of resolution is the first-layer global feature map of the semantic branch output by the bilateral feature aggregation module.

[0008] As a preferred solution of the present invention, the first to fifth improved depthwise separable convolution modules have the same structure, and all include a first standard convolution module, a variable channel spatial pyramid module, and a second standard convolution module connected in sequence, wherein the first standard convolution module performs a first step on the input of the improved depthwise separable convolution module. Convolution, BN normalization and SiLU activation operations are performed to obtain the preprocessed feature map; The variable channel spatial pyramid module includes three branches, wherein the first branch performs convolution on the preprocessed feature map with a kernel size of 3, a hole rate of 1, and a grouping number of The separation of the hole convolution operation and the BN normalization operation is used to obtain the first feature map. represents the number of channels of the preprocessed feature map; the second branch performs convolution on the preprocessed feature map with a kernel size of 3, a void rate of 2, and a number of groups. The separation of the dilated convolution operation and the BN normalization operation is performed to obtain the second feature map; the third branch performs convolution on the preprocessed feature map with a kernel size of 3, a dilation rate of 5, and a number of groups of Separate hole convolution operation and BN normalization operation to obtain the third feature map; fuse the first, second and third feature maps, and then perform Convolution, BN normalization and ReLU activation operations are performed to obtain the multi-scale feature map output by the variable channel spatial pyramid module; The second standard convolution module performs multi-scale feature maps Convolution, BN normalization and SiLU activation operations are performed to obtain the output of the improved depthwise separable convolution module.

[0009] As a preferred solution of the present invention, the first to fifth multi-layer perception axial attention modules have the same structure, and the two-dimensional input feature map of the multi-layer perception axial attention module is reshaped into horizontal one-dimensional features and vertical one-dimensional features along the horizontal and vertical directions; the horizontal one-dimensional features are subjected to maximum pooling and average pooling with a pooling kernel size of 3, and the two pooling results are output and added to obtain a first-layer pooling feature map; the first-layer pooling feature map is subjected to maximum pooling and average pooling with a pooling kernel size of 5, and the two pooling results are output and added to obtain a second-layer pooling feature map; the second-layer pooling feature map is subjected to maximum pooling and average pooling with a pooling kernel size of 7, and the two pooling results are output and added to obtain a third-layer pooling feature map; the third-layer pooling feature map is subjected to Convolution, BN normalization, ReLU activation operation and feature reshaping are performed to output a horizontal two-dimensional feature map; the horizontal one-dimensional feature operation steps are repeated for the vertical one-dimensional features to output a vertical two-dimensional feature map; A Sigmoid activation function is applied to the horizontal two-dimensional feature map to obtain a horizontal weight feature map, which is dot-multiplied by the two-dimensional input feature map of the multi-layer perceptual axial attention module to obtain a horizontal attention weighted feature map. A Sigmoid activation function is applied to the vertical two-dimensional feature map to obtain a vertical weight feature map, which is dot-multiplied by the two-dimensional input feature map of the multi-layer perceptual axial attention module to obtain a vertical attention weighted feature map. The horizontal attention weighted feature map, the vertical attention weighted feature map, and the two-dimensional input feature map of the multi-layer perceptual axial attention module are added together to obtain the output of the multi-layer perceptual axial attention module.

[0010] As a preferred solution of the present invention, the first to third bilateral feature guided fusion modules have the same structure, and the two inputs of the bilateral feature guided fusion module are and Splicing is performed, and the splicing results are subjected to maximum average pooling and adaptive average pooling with a pooling kernel size of 1. The two pooling results are added and two layers of channel-by-channel convolution with a convolution kernel size of 3 and a step size of 1 are performed to output the channel feature map; the Sigmoid activation function is used for the channel feature map and the two inputs of the fusion module are guided according to the bilateral features and Dimension splitting generates two channel weights and ,Will and After multiplication, Add together to get the feature map ,Will and After multiplication, Add together to get the feature map ; The feature map and Splicing, after performing ordinary convolution with a convolution kernel size of 3 and a step size of 1 on the splicing result, it is re-split into two feature maps from the channel dimension, and a convolution operation is performed on one of the feature maps to obtain a feature map , perform adaptive average pooling with a pooling kernel size of 1, channel-by-channel convolution with a convolution kernel size of 3 and a step size of 1, and channel-by-channel convolution with a convolution kernel size of 3 and a step size of 1 on another feature map to obtain a feature map , the feature map and Multiply and use the Sigmoid activation function to generate two weights and , and The result of multiplication is and The multiplication results are added together to obtain the output of the bilateral feature guided fusion module.

[0011] A computer device comprises a memory, a processor, and a computer program stored in the memory and capable of running on the processor. When the processor executes the computer program, the steps of the high-resolution remote sensing image dual-branch segmentation method are implemented.

[0012] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the high-resolution remote sensing image dual-branch segmentation method.

[0013] Compared with the prior art, the present invention adopts the above technical solution and has the following technical effects: 1. The present invention adds a BFGF module in the detail branch, which promotes the information interaction and fusion between the CNN and Transformer branches, and realizes the complementarity and enhancement of features.

[0014] 2. The present invention adds the MPRA module to the semantic branch, which significantly improves the model's ability to capture long-distance spatial information and special-scale targets.

[0015] 3. The present invention can adaptively process spatial and channel feature information of different depths by introducing the SCSP module into the depth-separable convolution module of the semantic branch. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 It is a schematic diagram of the process of the dual-branch segmentation method of high-resolution remote sensing images of the present invention; Figure 2 It is a structural diagram of a high-resolution remote sensing image semantic segmentation network model constructed by the present invention; Figure 3 is a structural diagram of a bilateral feature aggregation module (BFA) provided by the present invention; Figure 4 It is a structural diagram of a variable channel spatial pyramid module (SCSP) provided by the present invention; Figure 5 is a structural diagram of a multi-layer perceptual axial attention module (MPRA) provided by the present invention; Figure 6 is a structural diagram of a bilateral feature guided fusion module (BFGF) provided by the present invention; Figure 7 It is a structural diagram of a bilateral guide aggregation module (BGA) provided by the present invention; Figure 8 is a structural diagram of a segmentation head module (SegHead) provided by the present invention; Fig. 9 Schematic diagram of the segmentation results of the high-resolution remote sensing image dual-branch segmentation method proposed in the present invention on the Potsdam dataset; Fig.10 It is a schematic diagram of the segmentation results of the high-resolution remote sensing image dual-branch segmentation method proposed in the present invention on the Potsdam dataset. DETAILED DESCRIPTION

[0017] The embodiments of the present invention are described in detail below, and examples of the embodiments are shown in the accompanying drawings. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be interpreted as limiting the present invention.

[0018] like Figure 1 As shown, the present invention proposes a high-resolution remote sensing image dual-branch segmentation method, and the specific steps are as follows: Step 1: Obtain the Vaihingen and Potsdam remote sensing datasets, including the original images and the corresponding label images. After preprocessing the dataset, divide it into training set, validation set and test set according to the set ratio.

[0019] The specific data set is: Get the Vaihingen remote sensing data set, which contains five foreground categories (low vegetation, trees, impervious surfaces, cars, buildings) and one background category (clutter) and corresponding label maps, with a total of 33 average High-resolution remote sensing images of pixels; obtain the Potsdam remote sensing dataset, which contains six categories (low vegetation, trees, impervious surfaces, cars, buildings, clutter) and corresponding label maps, a total of 38 images Pixel high-resolution remote sensing imagery.

[0020] The data set preprocessing is as follows: the original images and the corresponding label images in the Vaihingen data set are cropped into sizes of In order to prevent overfitting and improve the spatial information content of the dataset, the farthest point random cropping data enhancement method was used, and finally 19335 remote sensing original images and corresponding label images were obtained; the original images and corresponding label images in the Potsdam dataset were cropped in order to a size of Pixel images were processed using a variety of image enhancement techniques, such as flipping, translation, scaling, and cropping, and ultimately 20,102 original remote sensing images and corresponding label images were obtained.

[0021] The specific division of the data set is: The size of the data set is randomly The set is divided into training set, validation set and test set according to the ratio of

[0022] Step 2: Build a high-resolution remote sensing image segmentation network that integrates BiSeNetV2 and Transformer, such as Figure 2 As shown, it includes ResNet-18 network, MobileVit module (MViT), BFGF module, BFA module, improved depth separable convolution module (DSBlock), MPRA module, BGA module and SegHead module, among which ResNet-18 network and BFGF module constitute the detail branch of the high-resolution remote sensing image semantic segmentation network, BFA module, improved depth separable convolution module, MPRA module and MobileVit module constitute the semantic branch of the high-resolution remote sensing image semantic segmentation network, and BGA module and SegHead module constitute the upsampling part of the high-resolution remote sensing image semantic segmentation network.

[0023] Step 2.1: Input the remote sensing image JPG original image into the detail branch and semantic branch respectively, and obtain the shallow feature maps of the detail branch and semantic branch respectively. The detail branch is mainly composed of 5 layers. Through operations such as convolution, pooling, and feature aggregation, efficient extraction of remote sensing images from low-level detail features to high-level detail features is achieved. The semantic branch is mainly composed of 6 layers. Through operations such as convolution, pooling, and feature weighting, efficient extraction of remote sensing images from low-level global features to high-level global features is achieved. The specific steps are: perform convolution and pooling operations on the first layer of the detail branch on the JPG original image to obtain the original input image The shallow detail feature map with a resolution of 1000 s is then input into the BFA module of the first layer of the semantic branch to obtain a shallow global feature map containing more detail information with a resolution of 1000 s. The BFA module consists of five parts, which are implemented using standard convolution, maximum pooling and average pooling. Specifically, the original JPG image is Convolution, BN normalization and ReLU activation operations to get the original input image The feature map of resolution is then subjected to maximum pooling and average pooling with a pooling kernel size of 3 and a step size of 2 to obtain two original input images. The feature map of the detail branch shallow feature map Convolution, BN normalization and ReLU activation operations to get the original input image The above three feature maps are added together and then Convolution, BN normalization and ReLU activation operations are performed to obtain a shallow global feature map.

[0024] Step 2.2: The shallow detail feature map and shallow global feature map obtained in step 2.1 are further propagated to the deep layer of the branch. The second layer detail feature map is obtained from the detail branch, and the shallow global feature map is successively passed to the second, third, and fourth layers of the semantic branch to obtain the fourth layer global feature map. The second layer of the detail branch stacks 2 ResBlock modules and 1 BFGF module to output the original input image. The second and third layers of the semantic branch are improved depth-separable convolution modules and MPRA modules, which output the original input images respectively. , The fourth layer is the improved depth separable convolution module, MobileVit module and MPRA module, which outputs the original input image. The BFGF module consists of two blocks, which are implemented using standard convolution, pooling operations, and channel-by-channel convolution; the improved depthwise separable convolution module consists of three blocks, which are implemented using standard convolution and SCSP modules; the SCSP module consists of four blocks, which are implemented using standard convolution and dilated convolution; the MPRA module consists of three blocks, which weight features using feature reshaping, standard convolution, and pooling operations; the MobileVit module consists of three blocks, which are implemented using standard convolution, feature reshaping, and fully connected operations. The specific steps are: In the second layer of the semantic branch, the shallow global feature map is reshaped. Convolution, BN normalization and SiLU activation operations are performed to obtain a preprocessed feature map; the preprocessed feature map is input into the SCSP module, and a separate hole convolution operation and a BN normalization operation with a convolution kernel size of 3, a hole rate of 1, and a group number of 128 are performed on it, and the first feature map is output. The preprocessed feature map is subjected to a separate hole convolution operation and a BN normalization operation with a convolution kernel size of 3, a hole rate of 2, and a group number of 64, and the second feature map is output. The preprocessed feature map is subjected to a separate hole convolution operation and a BN normalization operation with a convolution kernel size of 3, a hole rate of 5, and a group number of 32, and the third feature map is output. The three feature maps are feature fused, and then the feature map is obtained by Convolution, BN normalization and ReLU activation operations are performed to output the multi-scale feature map of the SCSP module; the multi-scale feature map of the SCSP module is Convolution, BN normalization and SiLU activation operations are performed to output the multi-scale fusion feature map of the improved depth-separable convolution module. The multi-scale fusion feature map is input into the MPRA attention module, and the two-dimensional input features are flattened into horizontal one-dimensional features and vertical one-dimensional features in the horizontal and vertical directions respectively by feature reshaping. The horizontal one-dimensional features are subjected to maximum pooling and average pooling with a pooling kernel size of 3, and the two pooling results are output and added to obtain the first layer of pooling feature maps. The first layer of pooling feature maps are subjected to maximum pooling and average pooling with a pooling kernel size of 5, and the two pooling results are output and added to obtain the second layer of pooling feature maps. The second layer of pooling feature maps are subjected to maximum pooling and average pooling with a pooling kernel size of 7, and the two pooling results are output and added to obtain the third layer of pooling feature maps. The third layer of pooling feature maps are subjected to Convolution, BN normalization, ReLU activation operation and feature reshaping are performed to output a horizontal two-dimensional feature map; the horizontal one-dimensional feature operation steps are repeated for the vertical one-dimensional features to output a vertical two-dimensional feature map. The Sigmoid activation function is used for the horizontal two-dimensional feature map and the vertical two-dimensional feature map to output two weighted feature maps and multiply them with the multi-scale fusion feature map to obtain the attention weighted feature map, which is then added to the multi-scale fusion feature map to output the second-layer global feature map of the semantic branch. Repeat the above second-layer operation steps of the semantic branch to output the third-layer global feature map of the semantic branch. Repeat the above improved depth-separable convolution module operation steps for the third-layer global feature map of the semantic branch to output the fourth-layer multi-scale fusion feature map of the semantic branch. The multi-scale fusion feature map is input into the MobileVit module, and standard convolution with kernel sizes of 3 and 1, BN normalization and SiLU activation are performed on it. The preprocessed feature map is output and divided into patches for rearrangement. The Transformer module maps the features into Q, K, and V vectors and processes them through multi-head self-attention mechanism, full connection, SiLU activation and Dropout. The processed features are residually connected and layer normalized with the preprocessed feature map, and the Transformer module feature map is output. After convolution, BN normalization and SiLU activation operations, it is concatenated with the fourth-layer multi-scale fusion feature map of the semantic branch, and then Convolution, BN normalization and SiLU activation operations are performed to output the global feature map of the MobileVit module. Repeat the above MPRA module operation steps for the global feature map of the MobileVit module to output the fourth-layer global feature map of the semantic branch. The shallow detail feature map obtained in step 2.1 is input into the ResBlock module for two layers. The convolution, BN normalization and ReLU activation operations are repeated twice to obtain the ResBlock module feature map. The ResBlock module feature map and the fourth-layer global feature map of the semantic branch are used as the input of the BFGF module. After splicing, the maximum average pooling and adaptive average pooling are performed respectively. Then the two pooling results are added and two layers of channel-by-channel convolution are performed to output the channel feature map. The Sigmoid activation function is used on it and two channel weights are generated according to the dimension of the original input feature map. They are multiplied with the two original inputs for weighting, and then added with the other original input to output the two output feature maps of the first part of the BFGF module. The two output feature maps are spliced, and a layer of channel-by-channel convolution operation is performed. After that, they are re-split into two feature maps from the channel dimension. One feature map is convolved, and the other feature map is subjected to maximum pooling and two layers of channel-by-channel convolution operations. The processed feature maps are multiplied and weights are generated using the Sigmoid activation function. Then they are multiplied with the output feature maps of the first part of the BFGF module, and finally added to output the detail feature map of the second layer of the detail branch.

[0025] Step 2.3: The second-layer detail feature map and the fourth-layer global feature map obtained in step 2.2 are further propagated to the deep layer of the branch. The third-layer detail feature map is obtained from the detail branch, and the fifth-layer global feature map is obtained from the semantic branch. The third layer of the detail branch stacks 2 ResBlock modules and 1 BFGF module to output the original input image. The fifth layer of the semantic branch is an improved deep separable convolution module, MobileVit module and MPRA module, which outputs the original input image. The specific steps are as follows: for the fourth-layer global feature map, perform the improved depth-separable convolution module operation steps, the MobileVit module operation steps, and the MPRA attention module operation steps in sequence, and output the fifth-layer global feature map. Repeat the step 2.2 ResBlock module operation steps twice for the second-layer detail feature map, and output the ResBlock module feature map. Use the ResBlock module feature map and the fifth-layer global feature map as the BFGF module input, perform the step 2.2 BFGF module operation steps, and output the third-layer detail feature map.

[0026] Step 2.4: The third-layer detail feature map and the fifth-layer global feature map obtained in step 2.3 are further propagated to the deep layer of the branch. The fourth-layer detail feature map is obtained from the detail branch, and the sixth-layer global feature map is obtained from the semantic branch. The fourth layer of the detail branch stacks 2 ResBlock modules and 1 BFGF module to output the original input image. The sixth layer of the semantic branch is an improved deep separable convolution module, MobileVit module and MPRA module, which outputs a feature map with a resolution of 1 / 32 of the original input image. The specific steps are as follows: perform step 2.2 of the improved deep separable convolution module operation steps, the MobileVit module operation steps and the MPRA attention module operation steps on the fifth layer global feature map, and output the sixth layer global feature map. Repeat step 2.2 of the ResBlock module operation steps twice for the third layer detail feature map, and output the ResBlock module feature map. Use the ResBlock module feature map and the sixth layer global feature map as the BFGF module input, perform step 2.2 of the BFGF module operation steps, and output the fourth layer detail feature map.

[0027] Step 2.5: Pass the fourth-layer detail feature map obtained in step 2.4 to the fifth layer of the detail branch, repeat the ResBlock module operation steps in step 2.2 twice, output the fifth-layer detail feature map, use the fifth-layer detail feature map and the sixth-layer global feature map obtained in step 2.4 as the input of the BGA module, and output the original input image. The SegHead module is used to upsample and predict the feature map of the high-resolution remote sensing image, and the final accurate high-resolution remote sensing image segmentation map is obtained. The BGA module consists of three parts, which are implemented by deep convolution, standard convolution, pooling operation and upsampling operation. The SegHead module consists of three parts, which are implemented by standard convolution and upsampling operation. The specific steps are: In the BGA module, the fifth layer detail feature map is sequentially subjected to deep convolution with a convolution kernel size of 3 and a grouping number of 256 and Convolution to obtain detailed features Figure 1 ; The fifth layer detail feature map is subjected to ordinary convolution with a convolution kernel size of 3 and a step size of 2 and maximum average pooling with a pooling kernel size of 3 to obtain the detail feature Figure 2 ; Perform ordinary convolution with a kernel size of 3 and 2 times upsampling operations on the sixth layer global feature map to obtain the global feature Figure 1 ; Perform deep convolution with kernel size of 3 and group number of 256 on the sixth layer global feature map. Convolution to obtain global features Figure 2 . Figure 1 and global features Figure 1 Multiply to get fusion features Figure 1 , the detail features Figure 2 and global features Figure 2 Multiply and upsample by 2 times to get the fused features Figure 2 , the fusion features Figure 1 and fusion features Figure 2 Add and proceed Convolution, output BGA module fusion feature map. In SegHead module, the fusion feature map is processed. Convolution, BN normalization and ReLU activation operations, followed by two 4-fold upsampling, and finally Convolution outputs the final accurate high-resolution remote sensing image segmentation map.

[0028] Step 3: Input the preprocessed training set in step 1 into the high-resolution remote sensing image semantic segmentation network in step 2 for training, calculate the loss function and perform back propagation, update the network parameters, and obtain the optimal parameter model. After each training for a certain number of times, input the preprocessed validation set in step 1 into the high-resolution remote sensing image semantic segmentation network in step 2 for verification, observe the training effect and record it.

[0029] Step 3.1: Randomly initialize the parameters of the high-resolution remote sensing image semantic segmentation network, input the preprocessed training set and validation set data in step 1 into the high-resolution remote sensing image segmentation network in step 2, generate a semantic segmentation probability map, and calculate the cross entropy loss and Dice loss; The calculation formula of the cross entropy loss function is: , in, N is the number of pixels, C is the number of categories, For the i Pixels k The predicted probability of each category, It represents the i pixels belong to the class k The value in the one-hot encoding (if the pixel i Belongs to category k , then it is 1, otherwise it is 0).

[0030] The calculation formula of Dice loss function is: , in, TP is the number of true positives, i.e., the number of pixels for which both the predicted and true labels are positive; FP is the number of false positives, i.e., the number of pixels predicted to be positive but actually negative; FN is the number of false negatives, i.e., the number of pixels predicted as negative but actually positive.

[0031] Step 3.2: Loss back propagation, update network parameters, minimize the loss function as the optimization goal, obtain the optimal parameter model and save it.

[0032] Step 4: Input the preprocessed test set in step 1 into the optimal parameter model trained in step 3, and output an accurate segmentation map of the high-resolution remote sensing image.

[0033] The structure of the BFA module provided by the present invention is as follows Figure 3 As shown in the figure: the BFA module adopts two branches, the left branch is the convolution layer, the right branch is the convolution layer and the pooling layer in sequence, and the last layer is the convolution layer; the left branch convolution layer performs convolution with a kernel size of 3 and a step size of 1, BN normalization and ReLU activation operations; the right branch convolution layer performs convolution with a kernel size of 3 and a step size of 2, BN normalization and ReLU activation operations; the pooling layer includes parallel pooling kernel size of 3 and a step size of 2 maximum pooling and average pooling; the last convolution layer performs convolution with a kernel size of 3 and a step size of 1, BN normalization and ReLU activation operations.

[0034] The JPG original image and shallow detail feature map are used as the input of the BFA module. , convolution operation with a step size of 2, BN normalization and ReLU activation operation, get the original input image The feature map of resolution is then subjected to maximum pooling and average pooling with a pooling kernel size of 3 and a step size of 2 to obtain two original input images. The feature map of the detail branch shallow feature map , convolution operation with a step size of 1, BN normalization and ReLU activation operation to obtain the original input image The above three feature maps are added together and then , convolution operation with a step size of 1, BN normalization and ReLU activation operation, outputting the original input image The feature map of resolution is the shallow global feature map output by the BFA module.

[0035] The structure of the SCSP module provided by the present invention is as follows Figure 4 As shown in the figure, the first, second and third layers of the SCSP module are separated dilated convolutions, and the fourth layer is a normal convolution layer. The convolution kernel size of the separated dilated convolution layer is 3, the step size is 1, and the number of groups is , convolution operations and BN normalization operations with dilation coefficients of 1, 2, and 5 respectively, where is the number of channels of the current input feature map; the ordinary convolution layer performs , convolution operation with a stride of 1, BN normalization, and ReLU activation operation.

[0036] The SCSP module takes the preprocessed feature map of step 2.2 as input, performs a separate dilated convolution operation and a BN normalization operation with a convolution kernel size of 3, a dilation rate of 1, and a grouping number of 128, and outputs the first feature map. The preprocessed feature map is subjected to a separate dilated convolution operation and a BN normalization operation with a convolution kernel size of 3, a dilation rate of 2, and a grouping number of 64, and outputs the second feature map. The preprocessed feature map is subjected to a separate dilated convolution operation and a BN normalization operation with a convolution kernel size of 3, a dilation rate of 5, and a grouping number of 32, and outputs the third feature map. The three feature maps are subjected to feature fusion, and then the feature map is merged by Convolution, BN normalization and ReLU activation operations, outputting the original input image The feature map of the resolution is the multi-scale feature map output by the SCSP module.

[0037] The structure of the MPRA module provided by the present invention is as follows Figure 5 As shown: the first layer of the MPRA module is the feature reshaping layer, which flattens the two-dimensional input features into horizontal one-dimensional features and vertical one-dimensional features in the horizontal and vertical directions; the second, third and fourth layers are maximum pooling and average pooling with pooling kernel sizes of 3, 5, and 7 respectively and a step size of 1; the fifth layer is an ordinary convolution layer with feature reshaping, which performs one-dimensional convolution operations with a convolution kernel size of 1 and a step size of 1, BN normalization, ReLU activation operations and feature reshaping operations; the sixth layer is feature weighting, which uses the Sigmoid function for activation operations.

[0038] The SCSP module will Figure 5 The output multi-scale fusion feature map is used as input, and feature reshaping is used to flatten the two-dimensional input features into horizontal one-dimensional features and vertical one-dimensional features in the horizontal and vertical directions respectively. The horizontal one-dimensional features are subjected to maximum pooling and average pooling with a pooling kernel size of 3, and the two pooling results are output and added to obtain the first layer pooling feature map. The first layer pooling feature map is subjected to maximum pooling and average pooling with a pooling kernel size of 5, and the two pooling results are output and added to obtain the second layer pooling feature map. The second layer pooling feature map is subjected to maximum pooling and average pooling with a pooling kernel size of 7, and the two pooling results are output and added to obtain the third layer pooling feature map. The third layer pooling feature map is subjected to Convolution, BN normalization, ReLU activation operation and feature reshaping are performed to output a horizontal two-dimensional feature map. Repeat the horizontal one-dimensional feature operation steps for the vertical one-dimensional features to output a vertical two-dimensional feature map. Use the Sigmoid activation function for the horizontal two-dimensional feature map and the vertical two-dimensional feature map to output two weighted feature maps and perform dot multiplication with the multi-scale fusion feature map to obtain an attention weighted feature map, which is then added to the multi-scale fusion feature map to output the original input image. The feature map of the resolution is the feature map output by the MPRA module.

[0039] The structure of the BFGF module provided by the present invention is as follows Figure 6 As shown in the figure: the first layer of the BFGF module is the parallel maximum average pooling and adaptive average pooling with a pooling kernel size of 1; the second and fourth layers are channel-by-channel convolution with a convolution kernel size of 3 and a step size of 1; the third layer is a channel-by-channel convolution with a convolution kernel size of 3 and a step size of 1 with a Sigmoid activation function; the fifth layer is a normal convolution with a convolution kernel size of 3 and a step size of 1 with a mean operation; the sixth layer consists of an adaptive average pooling part and a channel-by-channel convolution part. The adaptive average pooling part performs a pooling operation with a pooling kernel size of 1, and the channel-by-channel convolution part performs two channel-by-channel convolutions with a convolution kernel size of 3 and a step size of 1; the seventh layer is a feature weighting layer, which uses the Sigmoid function for activation.

[0040] The BFGF module takes the feature map of the ResBlock module in step 2.2 and the fourth-layer global feature map of the semantic branch as input, and performs maximum average pooling and adaptive average pooling respectively after splicing. Then, the two pooling results are added and two layers of channel-by-channel convolution are performed to output the channel feature map. The Sigmoid activation function is used on it and two channel weights are generated according to the dimension of the original input feature map. The weights are multiplied with the two original inputs, and then added with the other original input to output the two output feature maps of the first part of the BFGF module. The two output feature maps are spliced, and a layer of channel-by-channel convolution operation is performed. Then, they are re-split into two feature maps from the channel dimension. One feature map is convolved, and the other feature map is subjected to maximum pooling and two layers of channel-by-channel convolution operations. The processed feature maps are multiplied and weights are generated using the Sigmoid activation function. Then, they are multiplied with the output feature maps of the first part of the BFGF module, and finally added to output the original input image. The feature map of the resolution is the feature map output by the BFGF module.

[0041] The structure of the BGA module provided by the present invention is as follows Figure 7 As shown: the first layer of the BGA module consists of depth convolution with a kernel size of 3, a stride of 1, and a group number of 256, and ordinary convolution with a kernel size of 1 and a stride of 1; the second layer consists of ordinary convolution with a kernel size of 3 and a stride of 2, and flat pooling with a pooling kernel size of 3 and a stride of 1; the third layer consists of ordinary convolution with a kernel size of 3 and a stride of 1 and 2x linear upsampling; the fourth layer consists of depth convolution with a kernel size of 3, a stride of 1, and a group number of 256, and ordinary convolution with a kernel size of 1 and a stride of 1; the fifth layer consists of ordinary convolution with a kernel size of 3 and a stride of 1.

[0042] The BGA module takes the fifth-layer detail feature map in step 2.5 and the sixth-layer global feature map in step 2.4 as input, and performs deep convolution with a convolution kernel size of 3 and a grouping number of 256 on the fifth-layer detail feature map. Convolution to obtain detailed features Figure 1 ; The fifth layer detail feature map is subjected to ordinary convolution with a convolution kernel size of 3 and a step size of 2 and maximum average pooling with a pooling kernel size of 3 to obtain the detail feature Figure 2 ; Perform ordinary convolution with a kernel size of 3 and 2 times upsampling operations on the sixth layer global feature map to obtain the global feature Figure 1 ; Perform deep convolution with kernel size of 3 and group number of 256 on the sixth layer global feature map. Convolution to obtain global features Figure 2 . Figure 1 and global features Figure 1 Multiply to get fusion features Figure 1 , the detail features Figure 2 and global features Figure 2 Multiply and upsample by 2 times to get the fused features Figure 2 , the fusion features Figure 1 and fusion features Figure 2 Add and proceed Convolution, outputting the original input image The characteristic map of the resolution is the output characteristic map of the BGA module.

[0043] The structure of the SegHead module provided by the present invention is as follows Figure 8 As shown: the first layer of the SegHead module performs convolution operations with a kernel size of 3 and a stride of 1, BN normalization, and ReLU activation operations; the second and third layers perform fourfold linear interpolation upsampling; the fourth layer performs convolution operations with a kernel size of 1 and a stride of 1.

[0044] The SegHead module will Figure 7 The output feature map is used as input and the feature map is processed Convolution, BN normalization and ReLU activation operations, followed by two 4-fold upsampling, and finally Convolution outputs the final accurate high-resolution remote sensing image segmentation map.

[0045] In order to demonstrate the effectiveness of the dual-branch segmentation method for high-resolution remote sensing images provided in this embodiment, the Vaihingen and Potsdam remote sensing datasets are used to train, verify and test the model.

[0046] The index comparison and segmentation results of the dual-branch segmentation method for high-resolution remote sensing images proposed in this invention on the Vaihingen dataset are shown in Tables 1 and Fig. 9 shown.

[0047] Table 1

[0048] The index comparison and segmentation results of the dual-branch segmentation method for high-resolution remote sensing images proposed in this invention on the Potsdam dataset are shown in Table 2 and Table 3. Fig.10 shown.

[0049] Table 2

[0050] It can be seen from the table and figure that each evaluation index of the method proposed in the present invention is higher than that of the existing segmentation network, and the segmentation effect is closest to the original image.

[0051] Based on the same inventive concept, an embodiment of the present application provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of the aforementioned high-resolution remote sensing image dual-branch segmentation method are implemented.

[0052] Based on the same inventive concept, an embodiment of the present application provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, the steps of the aforementioned high-resolution remote sensing image dual-branch segmentation method are implemented.

[0053] It will be appreciated by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0054] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0055] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0056] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0057] The above embodiments are only for illustrating the technical idea of ​​the present invention, and cannot be used to limit the protection scope of the present invention. Any changes made on the basis of the technical solution in accordance with the technical idea proposed by the present invention shall fall within the protection scope of the present invention.

Claims

1. A dual-branch segmentation method for high-resolution remote sensing images, characterized in that: The steps include: Step 1, obtain a high-resolution remote sensing image dataset, including the original remote sensing image and the corresponding label map, pre-process the high-resolution remote sensing image dataset, and divide it into a training set, a validation set, and a test set according to a set ratio; Step 2, constructing a high-resolution remote sensing image semantic segmentation network model, including a detail branch, a semantic branch and an upsampling part, the preprocessed remote sensing image is respectively passed through the detail branch and the semantic branch to obtain the output of the detail branch and the semantic branch respectively, and the output of the detail branch and the semantic branch respectively is passed through the upsampling part to obtain the output of the high-resolution remote sensing image semantic segmentation network model; The detail branch includes five layers, wherein the first layer includes a convolution layer and a pooling layer connected in sequence, the second layer includes a first residual module, a second residual module and a first bilateral feature guided fusion module connected in sequence, the third layer includes a third residual module, a fourth residual module and a second bilateral feature guided fusion module connected in sequence, the fourth layer includes a fifth residual module, a sixth residual module and a third bilateral feature guided fusion module connected in sequence, and the fifth layer includes a seventh residual module and an eighth residual module connected in sequence; the structures of the first to eighth residual modules are the same; The semantic branch includes six layers, wherein the first layer includes a bilateral feature aggregation module, the second layer includes a first improved depthwise separable convolution module and a first multi-layer perceptual axial attention module connected in sequence, the third layer includes a second improved depthwise separable convolution module and a second multi-layer perceptual axial attention module connected in sequence, the fourth layer includes a third improved depthwise separable convolution module, a first MobileVit module and a third multi-layer perceptual axial attention module connected in sequence, the fifth layer includes a fourth improved depthwise separable convolution module, a second MobileVit module and a fourth multi-layer perceptual axial attention module connected in sequence, and the sixth layer includes a fifth improved depthwise separable convolution module, a third MobileVit module and a fifth multi-layer perceptual axial attention module connected in sequence; the structures of the first to third MobileVit modules are the same; The preprocessed remote sensing image, i.e., the original input image, is used as the input of the first convolution layer of the detail branch; the detail feature map outputted from the first layer of the detail branch and the original input image are used as the input of the first bilateral feature aggregation module of the semantic branch; the output of the second residual module of the second layer of the detail branch and the output of the third multi-layer perceptual axial attention module of the fourth layer of the semantic branch are used as the input of the first bilateral feature guided fusion module of the second layer of the detail branch; the output of the fourth residual module of the third layer of the detail branch and the output of the fourth multi-layer perceptual axial attention module of the fifth layer of the semantic branch are used as the input of the second bilateral feature guided fusion module of the third layer of the detail branch; the output of the sixth residual module of the fourth layer of the detail branch and the output of the fifth multi-layer perceptual axial attention module of the sixth layer of the semantic branch are used as the input of the third bilateral feature guided fusion module of the fourth layer of the detail branch; The upsampling part includes a bilateral guided aggregation module and a segmentation head module connected in sequence. The detail feature map output by the fifth layer of the detail branch and the global feature map output by the sixth layer of the semantic branch are used as the input of the bilateral guided aggregation module. The segmentation head module outputs the final remote sensing image segmentation map. Step 3, input the training set of step 1 into the high-resolution remote sensing image semantic segmentation network model constructed in step 2 for training, calculate the loss function and perform back propagation, update the network model parameters, obtain the optimal parameter network model, and use the verification set of step 1 to verify the optimal parameter network model; Step 4: Input the test set of step 1 into the optimal parameter network model trained in step 3, and output the accurate segmentation map of the high-resolution remote sensing image.

2. The high-resolution remote sensing image dual-branch segmentation method according to claim 1, characterized in that: The specific process of step 1 is as follows: Get the Vaihingen and Potsdam remote sensing image datasets. Both the Vaihingen and Potsdam remote sensing image datasets include original remote sensing images of six categories: low vegetation, trees, impervious surfaces, cars, buildings, and clutter, as well as corresponding label images. The original remote sensing images and corresponding label images in the Vaihingen remote sensing image dataset are cropped using the farthest point random cropping data augmentation method so that the cropped image size is Pixels, get the preprocessed Vaihingen remote sensing image dataset; The original remote sensing images and corresponding label images in the Potsdam remote sensing image dataset are cropped into Pixel image, cropped The pixel images are enhanced using image enhancement technology to obtain the preprocessed Potsdam remote sensing image dataset; The preprocessed Vaihingen and Potsdam remote sensing image datasets were The training set, validation set, and test set are obtained by dividing the set into two parts.

3. The high-resolution remote sensing image dual-branch segmentation method according to claim 1, characterized in that: The bilateral feature aggregation module includes two branches, left and right. The left branch aggregates the detail feature map output by the first layer of the detail branch. Perform convolution with a kernel size of 3 and a step size of 1, BN normalization and ReLU activation operations to obtain the original input image Feature map of resolution a1; The right branch is for the original input image Perform convolution with a kernel size of 3 and a stride of 2, BN normalization and ReLU activation operations to obtain the original input image The feature map a2 of resolution is obtained by performing maximum pooling and average pooling with a pooling kernel size of 3 and a step size of 2 on the feature map a2 to obtain two original input images. After adding the feature maps a1, a3, and a4, the added result is convolved with a kernel size of 3 and a step size of 1, and then BN normalized and ReLU activated to obtain the original input image. The feature map of resolution is the first-layer global feature map of the semantic branch output by the bilateral feature aggregation module.

4. The high-resolution remote sensing image dual-branch segmentation method according to claim 1, characterized in that: The first to fifth improved depthwise separable convolution modules have the same structure, and all include a first standard convolution module, a variable channel spatial pyramid module, and a second standard convolution module connected in sequence, wherein the first standard convolution module performs a transformation on the input of the improved depthwise separable convolution module. Convolution, BN normalization and SiLU activation operations are performed to obtain the preprocessed feature map; The variable channel spatial pyramid module includes three branches, wherein the first branch performs convolution on the preprocessed feature map with a kernel size of 3, a hole rate of 1, and a grouping number of The separation of the hole convolution operation and the BN normalization operation is used to obtain the first feature map. represents the number of channels of the preprocessed feature map; the second branch performs convolution on the preprocessed feature map with a kernel size of 3, a void rate of 2, and a number of groups. The separation of the dilated convolution operation and the BN normalization operation is performed to obtain the second feature map; the third branch performs convolution on the preprocessed feature map with a kernel size of 3, a dilation rate of 5, and a number of groups of Separate hole convolution operation and BN normalization operation to obtain the third feature map; fuse the first, second and third feature maps, and then perform Convolution, BN normalization and ReLU activation operations are performed to obtain the multi-scale feature map output by the variable channel spatial pyramid module; The second standard convolution module performs multi-scale feature maps Convolution, BN normalization and SiLU activation operations are performed to obtain the output of the improved depthwise separable convolution module.

5. The high-resolution remote sensing image dual-branch segmentation method according to claim 1, characterized in that: The structures of the first to fifth multi-layer perception axial attention modules are the same. The two-dimensional input feature map of the multi-layer perception axial attention module is reshaped into horizontal one-dimensional features and vertical one-dimensional features in the horizontal and vertical directions; the horizontal one-dimensional features are subjected to maximum pooling and average pooling with a pooling kernel size of 3, and the two pooling results are output and added to obtain the first layer pooling feature map; the first layer pooling feature map is subjected to maximum pooling and average pooling with a pooling kernel size of 5, and the two pooling results are output and added to obtain the second layer pooling feature map; the second layer pooling feature map is subjected to maximum pooling and average pooling with a pooling kernel size of 7, and the two pooling results are output and added to obtain the third layer pooling feature map; the third layer pooling feature map is subjected to Convolution, BN normalization, ReLU activation operation and feature reshaping are performed to output a horizontal two-dimensional feature map; the horizontal one-dimensional feature operation steps are repeated for the vertical one-dimensional features to output a vertical two-dimensional feature map; A Sigmoid activation function is applied to the horizontal two-dimensional feature map to obtain a horizontal weight feature map, which is dot-multiplied by the two-dimensional input feature map of the multi-layer perceptual axial attention module to obtain a horizontal attention weighted feature map. A Sigmoid activation function is applied to the vertical two-dimensional feature map to obtain a vertical weight feature map, which is dot-multiplied by the two-dimensional input feature map of the multi-layer perceptual axial attention module to obtain a vertical attention weighted feature map. The horizontal attention weighted feature map, the vertical attention weighted feature map, and the two-dimensional input feature map of the multi-layer perceptual axial attention module are added together to obtain the output of the multi-layer perceptual axial attention module.

6. The high-resolution remote sensing image dual-branch segmentation method according to claim 1, characterized in that: The structures of the first to third bilateral feature guided fusion modules are the same, and the two inputs of the bilateral feature guided fusion module are and Splicing is performed, and the splicing results are subjected to maximum average pooling and adaptive average pooling with a pooling kernel size of 1. The two pooling results are added and two layers of channel-by-channel convolution with a convolution kernel size of 3 and a step size of 1 are performed to output the channel feature map; the Sigmoid activation function is used for the channel feature map and the two inputs of the fusion module are guided according to the bilateral features and Dimension splitting generates two channel weights and ,Will and After multiplication, Add together to get the feature map ,Will and After multiplication, Add together to get the feature map ; The feature map and Splicing, after performing ordinary convolution with a convolution kernel size of 3 and a step size of 1 on the splicing result, it is re-split into two feature maps from the channel dimension, and a convolution operation is performed on one of the feature maps to obtain a feature map , perform adaptive average pooling with a pooling kernel size of 1, channel-by-channel convolution with a convolution kernel size of 3 and a step size of 1, and channel-by-channel convolution with a convolution kernel size of 3 and a step size of 1 on another feature map to obtain a feature map , the feature map and Multiply and use the Sigmoid activation function to generate two weights and , and The result of multiplication is and The multiplication results are added together to obtain the output of the bilateral feature guided fusion module.

7. A computer device comprising a memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that: When the processor executes the computer program, the steps of the high-resolution remote sensing image dual-branch segmentation method as described in any one of claims 1 to 6 are implemented.

8. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the high-resolution remote sensing image dual-branch segmentation method as described in any one of claims 1 to 6 are implemented.

Images