Remote sensing image change detection method based on PAP-UNet + + network

By adopting the PAP-UNet++ network in remote sensing image change detection, using the hybrid cavity convolution and parallel cavity pyramid module and CBAM attention mechanism, the problem of lack of semantic information on shallow features and inaccurate spatial positioning of deep features is solved, and the more Accurate remote sensing change detection and better generalization performance.

CN119068352APending Publication Date: 2024-12-03WUXI UNIV
View PDF 0 Cites 5 Cited by

Patent Information

Application Number
CN202411105189.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-12
Publication Date
2024-12-03

AI Technical Summary

Technical Problem

In the existing remote sensing image change detection methods, the shallow features lack semantic information and the deep features are inaccurately positioned, resulting in the model being unable to accurately identify the changing scenes in the remote sensing image.

Method used

Using the PAP-UNet++ network method, the receptive field is expanded by mixing the hollow convolution module, combining the parallel hollow pyramid module and the CBAM attention mechanism, and integrating multi-scale features and global context information to improve the recognition accuracy and generalization of the model's change detection ability.

Benefits of technology

Effectively avoiding the grid effect, improving the multi-scale information recognition accuracy and segmentation stability of remote sensing change detection, enhancing the ability to express edges and tiny targets, and improving the generalization performance of the model.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119068352A_ABST
    Figure CN119068352A_ABST
Patent Text Reader

Abstract

The invention discloses a remote sensing image change detection method based on a PAP-UNet + + network, which relates to the technical field of remote sensing image processing, and comprises the following steps: firstly, preprocessing an original data set, uniformly scaling images in a training set into image blocks with the size of 512 * 512, and taking the image blocks as input; in an encoder stage, a mixed cavity convolution module is used for realizing step-by-step extraction of multi-scale features and performing down-sampling operation, and in a decoder stage, a parallel cavity pyramid module is inserted so as to improve the capability of the model for capturing global context features. According to the method, a CBAM attention mechanism is introduced in the jump connection process of the same level, so that the recognition and distinguishing capability on a change area is improved, a pixel-level classification prediction result consistent with the input size is finally generated, the problem that multi-scale feature information is insufficiently utilized by a network can be effectively solved through the method, and change detail information is accurately detected.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of remote sensing image processing, and in particular to a remote sensing image change detection method based on a PAP-UNet++ network. Background Art

[0002] Change detection refers to identifying the differences between images taken at two different times within the same geographic area. Remote sensing image change detection is a basic and challenging task in the field of remote sensing. It has promoted the development of related research and applications in the field of remote sensing and has attracted widespread attention in recent years. The rich remote sensing data provides a basic guarantee for the change detection task, so that changes on the earth's surface can be dynamically detected in real time.

[0003] With the continuous development of deep learning technology, it has achieved outstanding performance in recent years. Many scholars have introduced deep learning technology into remote sensing processing tasks and achieved excellent performance. Compared with traditional change detection methods, deep learning-based methods can automatically extract features from original images without the need to manually design specific feature descriptors, which can not only reduce manual intervention and effectively avoid errors caused by preprocessing, but also reduce the workload of post-processing. However, most of the current remote sensing change detection methods face the problem of lack of semantic information in shallow features and inaccurate spatial positioning of deep features, which makes the model unable to accurately identify and detect the changed scenes in remote sensing images. Summary of the invention

[0004] In order to solve the above technical problems, the present invention provides a remote sensing image change detection method based on a PAP-UNet++ network, comprising the following steps:

[0005] S1, data set division, dividing the image data in the original data set into a training set and a test set;

[0006] S2, image preprocessing;

[0007] S3, establish a network model, input the dual-phase training set images to iteratively train the model, and select the optimal model based on the classification detection evaluation index, and finally obtain the trained PAP-UNet++ network; specifically, it includes the following sub-steps:

[0008] S3.1, inputting dual-phase remote sensing images at two different times, T1 and T2, into the encoder structure in the network simultaneously;

[0009] S3.2, while the encoder processes the feature map layer by layer, the current layer x i,0 With the next layer x i+1,0 The calculation results are added through the concat operation to obtain x i,1 Feature map;

[0010] S3.3. Merge x through concat operation i,0 、x i,1 and x i+1,1 Get x i,2 The feature map of the layer; merge x through the concat operation i,0 、x i,1 、x i,2 and x i+1,2 Get x i,3 Feature map of the layer;

[0011] S3.4. In the decoder process, a parallel atrous pyramid PAP module is added during upsampling, and a CBAM attention module is added during the jump connection between the encoder and decoder. Finally, the outputs of the above two modules and x are fused. i,1 、x i,2 and x i,3 Get x i,4 Layer results;

[0012] S4. Input the test set into the trained remote sensing image change detection network to obtain the image change detection result.

[0013] The technical solution further defined in the present invention is:

[0014] Furthermore, in step S1, the image data in the original data set is divided into a training set and a test set in a ratio of 4:1.

[0015] As described above, in the remote sensing image change detection method based on the PAP-UNet++ network, in step S2, the image preprocessing is specifically as follows: the image is uniformly scaled to a resolution of 512×512 as the input image, and then horizontally flipped with a probability of 0.5 to expand the data set.

[0016] As described above, in the remote sensing image change detection method based on the PAP-UNet++ network, in step S3, each convolution layer in the PAP-UNet++ network includes a series of nonlinear transformations, which include convolution operations, batch normalization, and ReLU activation functions; wherein the convolution operation formula is as follows:

[0017]

[0018] Among them, f represents the input feature map, g represents the convolution kernel, N represents the size of the original image, S represents the stride size, and M represents the convolution kernel size;

[0019] Batch normalization is used to transform the input image of each layer into a normal distribution with a mean of 0 and a variance of 1 to obtain a standardized result. The formula is as follows:

[0020]

[0021] Among them, x i represents a certain characteristic value of the i-th sample, μ B represents the feature average, represents the variance of the feature, B represents the sample set in the current batch, γ and β represent the introduced parameters, and ∈ represents the constant added to the minimum batch variance.

[0022] As mentioned above, the remote sensing image change detection method based on the PAP-UNet++ network includes an encoder structure and a decoder structure. The encoder structure is used for feature extraction and multi-level downsampling, and the decoder structure is used for multi-level upsampling using deconvolution operations. The PAP-UNet++ network merges two 3-channel images into a 6-channel image, that is, the PAP-UNet++ network processes the information of the two images at the same time, and uses the pixel difference between them to complete the change detection task. The final output results are X 0,1 , X 0,2 , X 0,3 and X 0,4 , and finally take the average of the four output results.

[0023] As mentioned above, in the remote sensing image change detection method based on the PAP-UNet++ network, the encoder structure replaces the convolution module of the original UNet++ network with a hybrid dilated convolution module. The hybrid dilated convolution module consists of two dilated convolutions, a BN layer, and a ReLU activation function layer, and the expansion coefficients of the two layers of dilated convolutions are 2 and 3, respectively.

[0024] As described above, in the remote sensing image change detection method based on the PAP-UNet++ network, the decoder structure includes a parallel atrous pyramid PAP module, which includes four parallel convolutional layer branches and a global average pooling branch. The four parallel convolutional layer branches use a 3×3 atrous convolution structure to capture contextual information at different levels. Each convolutional layer branch is equipped with a BN layer and a ReLU activation function to perform feature normalization and introduce nonlinear expression capabilities respectively.

[0025] The global average pooling branch is used to capture global context features. The branch consists of a global average pooling layer, two 1×1 convolutional layers, and a Sigmoid activation function. A learnable scaling factor mechanism is introduced to dynamically adjust the importance of global context features.

[0026] As mentioned above, in the remote sensing image change detection method based on the PAP-UNet++ network, in the parallel atrous pyramid PAP module, the first convolution branch of the four parallel convolutional layer branches consists of a 1×1 convolutional layer and a Sigmoid activation function; the second convolution branch consists of two 3×3 atrous convolutional layers with atrous rates of 6 and 8 respectively; the third convolution branch consists of a 3×3 atrous convolutional layer with a atrous rate of 12; and the fourth convolution branch consists of a 3×3 atrous convolutional layer with a atrous rate of 18.

[0027] As described above, in the remote sensing image change detection method based on the PAP-UNet++ network, in step S3, the PAP-UNet++ network is embedded with a CBAM attention mechanism module in the same-level skip connection process. The CBAM attention mechanism module includes two submodules, a channel attention CAM module and a spatial attention SAM module. The CBAM attention mechanism module receives the feature map from the previous layer as input, calculates the importance of each channel through the channel attention CAM module, and passes the result to the spatial attention SAM module to further refine the attention on the spatial position. The feature map processed by the two attention modules is used as output for use by subsequent layers;

[0028] The channel attention CAM module applies global average pooling and global maximum pooling to the input features to obtain two one-dimensional vectors, and then passes these two one-dimensional vectors through two shared fully connected layers to obtain two channel attention weight vectors, and uses the activation function to activate the output of the fully connected layer, and then passes through another fully connected layer, and then uses the Sigmoid activation function to obtain the final channel attention weight, and finally performs channel-by-channel multiplication with the input feature map to obtain the weighted feature map;

[0029] The spatial attention SAM module applies maximum pooling and average pooling to the output features of the channel attention module, performs pooling operations along the channel dimension to obtain two feature maps, which are then concatenated and convolved to obtain a single-channel feature map. The convolution output is then processed using the Sigmoid activation function to obtain the spatial attention weight, and the obtained spatial attention weight is element-wise multiplied with the output feature map of the channel attention module to obtain the final weighted feature map.

[0030] As described above, in the remote sensing image change detection method based on the PAP-UNet++ network, in step S3, the following L1 loss function is set in the trained PAP-UNet++ network:

[0031]

[0032] Among them, f(x i ) represents the predicted value of the model, y iRepresents the true value of the model.

[0033] The beneficial effects of the present invention are:

[0034] (1) In the present invention, a hybrid dilated convolution module is designed by introducing dilated convolution technology. This design effectively expands the receptive field of each layer of the network without significantly increasing the computational complexity, which is conducive to processing multi-scale features and global-local information fusion in remote sensing change detection, effectively avoiding gridding effect and improving overall performance;

[0035] (2) In the present invention, a parallel dilated pyramid module is proposed and applied to the constructed network, which includes four parallel dilated convolution branches and one global average pooling branch to integrate global context features. After the five branches are processed independently, the results of each branch are fused through feature mapping splicing. The output of the global average pooling branch is adjusted to match the dimension of the convolution branch, thereby effectively improving the recognition accuracy and segmentation stability of the model for multi-scale information.

[0036] (3) In the present invention, an attention mechanism CBAM module is embedded between the mixed void convolution modules at each level and the output of the last layer in the constructed network, aiming to effectively integrate multi-scale shallow and deep features, especially the accurate expression of edges and small targets. The CBAM module strengthens the model's cognition and extraction of remote sensing image features, and improves the model's generalization ability when dealing with unseen changing scenes. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 A schematic diagram of the structure of the PAP-UNet++ network in an embodiment of the present invention;

[0038] Figure 2 Schematic diagram of the structure of a hybrid dilated convolution module in an embodiment of the present invention;

[0039] Figure 3 Schematic diagram of the structure of a parallel hollow pyramid module in an embodiment of the present invention;

[0040] Figure 4 It is a schematic diagram of the structure of the CBAM attention mechanism module in an embodiment of the present invention;

[0041] Figure 5 Schematic diagram of the structure of the channel attention CAM module in an embodiment of the present invention;

[0042] Figure 6 Schematic diagram of the structure of the spatial attention SAM module in an embodiment of the present invention. DETAILED DESCRIPTION

[0043] This embodiment provides a remote sensing image change detection method based on the PAP-UNet++ network. Figures 1 to 6As shown, the following steps are included

[0044] S1. Dataset division: The image data in the original dataset is divided into a training set and a test set in a ratio of 4:1.

[0045] S2. Image preprocessing: The images are uniformly scaled to a resolution of 512×512 as input images, and then horizontally flipped with a probability of 0.5 to expand the dataset.

[0046] S3. Establish a network model, input the dual-phase training set images to iteratively train the model, and select the optimal model based on the classification detection evaluation index, and finally obtain the trained PAP-UNet++ (Parallel AtrousPyramid-UNet++) network.

[0047] The classification detection evaluation indicators include Mean Intersection over Union (MIoU), Overall Accuracy (OA), Kappa coefficient and Dice coefficient. The model will calculate the above four evaluation indicators at a certain number of steps during the training process, and then automatically save the model parameters with the best comprehensive evaluation indicators based on the calculation results.

[0048] The overall process of the PAP-UNet++ network includes the following steps:

[0049] S3.1. The dual-phase remote sensing images at two different times, T1 and T2, are simultaneously input into the encoder structure in the network.

[0050] S3.2, while the encoder processes the feature map layer by layer, the current layer x i,0 With the next layer x i+1,0 The calculation results are added through the concat operation to obtain x i,1 Feature map.

[0051] S3.3. Merge x through concat operation i,0 、x i,1 and x i+1,1 Get x i,2 The feature map of the layer; merge x through the concat operation i,0 、x i,1 、x i,2 and x i+1,2 Get x i,3 The feature map of the layer.

[0052] S3.4. In the decoder process, a parallel atrous pyramid PAP module is added during upsampling, and a CBAM attention module is added during the jump connection between the encoder and decoder. Finally, the outputs of the above two modules and x are fused. i,1 、x i,2 and x i,3 Get x i,4 The result of the layer.

[0053] In step S3, each convolutional layer in the PAP-UNet++ network contains a series of nonlinear transformations, including convolution operations, batch normalization, and ReLU activation functions; the convolution operation formula is as follows:

[0054]

[0055] Among them, f represents the input feature map, g represents the convolution kernel, N represents the size of the original image, S represents the stride size, and M represents the convolution kernel size.

[0056] Batch normalization is used to transform the input image of each layer into a normal distribution with a mean of 0 and a variance of 1 to obtain a standardized result. The formula is as follows:

[0057]

[0058] Among them, x i represents a certain characteristic value of the i-th sample, μ B represents the feature average, represents the variance of the feature, B represents the sample set in the current batch, γ and β represent the introduced parameters, and ∈ represents the constant added to the minimum batch variance.

[0059] The ReLU activation function is used to add nonlinear factors, which can enhance the learning ability of the model without the problem of gradient disappearance; the batch normalization operation can speed up the convergence of the network and greatly speed up the training.

[0060] like Figure 1 As shown in Figure 1, the PAP-UNet++ network includes an encoder structure and a decoder structure. The encoder structure is used for feature extraction and multi-level downsampling, and the decoder structure is used for multi-level upsampling using deconvolution operations. The PAP-UNet++ network merges two 3-channel images into a 6-channel image, that is, the PAP-UNet++ network processes the information of the two images at the same time and uses the pixel differences between them to complete the change detection task. The final output results are X 0,1 , X 0,2 , X 0,3 and X 0,4 , and finally take the average of the four output results.

[0061] The encoder structure replaces the convolution module of the original UNet++ network with a hybrid dilated convolution module, and selects dilated convolutions of different resolutions to broaden the effective receptive field of the convolution kernels at each level of the network; Figure 2 As shown in the figure, the hybrid atrous convolution module consists of two atrous convolutions, a BN layer, and a ReLU activation function layer. In the hybrid atrous convolution module, two layers of atrous convolutions with expansion coefficients of 2 and 3 are selected respectively to learn rich multi-scale features, which can enhance the capture of details of edges and tiny targets while maintaining the size of the feature map, which is crucial to improving segmentation accuracy.

[0062] The decoder structure includes a Parallel Atrous Pyramid (PAP) module, such as Figure 3 As shown in the figure, the parallel hole pyramid module includes four parallel convolutional layer branches and a global average pooling branch, which aims to improve the accuracy of the model in identifying and segmenting multi-scale information.

[0063] In the PAP module, four parallel convolutional layer branches use a 3×3 dilated convolution structure to capture contextual information at different levels. Each convolutional layer branch is equipped with a BN layer and a ReLU activation function for feature normalization and the introduction of nonlinear expression capabilities, respectively. The first convolutional branch consists of a 1×1 convolutional layer and a Sigmoid activation function. The second convolutional branch consists of two 3×3 dilated convolutional layers with dilation rates of 6 and 8 respectively. The third convolutional branch consists of a 3×3 dilated convolutional layer with a dilation rate of 12. The fourth convolutional branch consists of a 3×3 dilated convolutional layer with a dilation rate of 18.

[0064] The global average pooling branch is used to capture global context features. It consists of a global average pooling layer, two 1×1 convolutional layers, and a Sigmoid activation function. A learnable scaling factor mechanism is introduced to dynamically adjust the importance of global context features. The Sigmoid activation function constrains it to the range of (0, 1) to form a scalar feature map that reflects the importance of global context information. The pixel value at each position is used as the adaptive weight of the corresponding global feature. Subsequently, these dynamically calculated scaling factors are multiplied by the global average pooling output to flexibly adjust the importance of global context features.

[0065] After the five branches are processed independently, the results of each branch are fused through feature mapping splicing, in which the output of the global average pooling branch is adjusted to match the dimension of the convolution branch; finally, the spliced ​​features are compressed and interacted through a 1×1 convolutional layer to generate a comprehensive feature map that integrates multi-scale and global information, which is used to optimize the subsequent image segmentation performance.

[0066] The PAP-UNet++ network embeds the CBAM attention mechanism module in the same-level skip connection process. This design aims to enable the network to effectively fuse shallow and deep feature information at the multi-scale feature level. The CBAM attention mechanism module helps the model focus on significant features related to change detection through adaptively weighted channels and spatial features, while suppressing irrelevant redundant information, thereby improving the ability to identify and distinguish changed areas, and is expected to achieve more accurate change detection effects.

[0067] like Figure 4 As shown in the figure, the CBAM attention mechanism module includes two sub-modules: the channel attention CAM module and the spatial attention SAM module. The CBAM attention mechanism module receives the feature map from the previous layer as input, calculates the importance of each channel through the channel attention CAM module, and passes its result to the spatial attention SAM module to further refine the attention on the spatial position. The feature map processed by the two attention modules is used as output for use by subsequent layers.

[0068] like Figure 5 As shown in the figure, the channel attention CAM module applies global average pooling and global maximum pooling to the input features respectively to obtain two one-dimensional vectors, and then passes these two one-dimensional vectors through two shared fully connected layers to obtain two channel attention weight vectors, and uses the activation function to activate the output of the fully connected layer, and then passes through another fully connected layer, and then uses the Sigmoid activation function to obtain the final channel attention weight, and finally performs channel-by-channel multiplication with the input feature map to obtain the weighted feature map.

[0069] like Figure 6 As shown in the figure, the spatial attention SAM module applies maximum pooling and average pooling to the output features of the channel attention module, performs pooling operations along the channel dimension, obtains two feature maps, and then concatenates them and obtains a single-channel feature map through convolution operation. The convolution output is then processed with the Sigmoid activation function to obtain the spatial attention weight, and the obtained spatial attention weight is multiplied element-by-element with the output feature map of the channel attention module to obtain the final weighted feature map.

[0070] In the trained PAP-UNet++ network, set the following L1 loss function:

[0071]

[0072] Among them, f(x i ) represents the predicted value of the model, y i Represents the true value of the model.

[0073] S4. Input the test set into the trained remote sensing image change detection network to obtain the image change detection result.

[0074] In this embodiment, the datasets used are open source remote sensing image change detection datasets, namely the WHU dataset, the SYSU-CD dataset, and the LEVIR-CD dataset.

[0075] During the implementation process, the number of training times was set to 100 times. After reaching the specified number of training times, the result graph of the change detection image was output; the development environments of the experiments were: development environment Python 3.8, deep learning framework PaddlePaddle, integrated development tool Pycharm 2021.1.3, and operating system Windows 10 (x64).

[0076] In view of the inherent limitations of the currently widely used remote sensing change detection models, namely the incomplete extraction of multi-scale information of remote sensing images, this embodiment designs a remote sensing image change detection method based on the PAP-UNet++ network. Compared with the traditional image segmentation method, the PAP-UNet++ model proposed in the present invention shows excellent performance in remote sensing change detection tasks that is superior to traditional convolutional neural networks, and has good generalization performance.

[0077] In order to enhance the model's ability to extract multi-scale features, this embodiment replaces the standard convolution in the downsampling stage with a hybrid dilated convolution module to expand the receptive field, and captures multi-scale features through dilated convolution layers with different expansion coefficients; at the same time, the CBAM attention mechanism is embedded to optimize the fusion of shallow and deep features and enhance the recognition accuracy of changed areas; in addition, the parallel dilated pyramid module proposed in this embodiment includes four convolution branches with increasing dilated rates to extract contextual information of each scale, and a global average pooling branch to integrate the global context, and dynamically adjusts the importance by learning the scaling factor.

[0078] In addition to the above embodiments, the present invention may also have other implementation modes. Any technical solution formed by equivalent replacement or equivalent transformation falls within the protection scope required by the present invention.

Claims

1. A remote sensing image change detection method based on PAP-UNet++ network, Features: The following steps are involved: S1, data set division, dividing the image data in the original data set into a training set and a test set; S2, image preprocessing; S3, establish a network model, input the dual-phase training set images to iteratively train the model, and select the optimal model based on the classification detection evaluation index, and finally obtain the trained PAP-UNet++ network; The following steps are included: S3.1, inputting dual-phase remote sensing images at two different times, T1 and T2, into the encoder structure in the network simultaneously; S3.2, while the encoder processes the feature map layer by layer, the current layer x i,0 With the next layer x i+1,0 The calculation results are added through the concat operation to obtain x i,1 Feature map; S3.

3. Merge x through concat operation i,0 、x i,1 and x i+1,1 Get x i,2 The feature map of the layer; merge x through the concat operation i,0 、x i,1 、x i,2 and x i+1,2 Get x i,3 Feature map of the layer; S3.

4. In the decoder process, a parallel atrous pyramid PAP module is added during upsampling, and a CBAM attention module is added during the jump connection between the encoder and decoder. Finally, the outputs of the above two modules and x are fused. i,1 、x i,2 and x i,3 Get x i,4 Layer results; S4. Input the test set into the trained remote sensing image change detection network to obtain the image change detection result.

2. According to claim 1, the remote sensing image change detection method based on the PAP-UNet++ network, Features: In the step S1, the image data in the original data set is divided into a training set and a test set in a ratio of 4:

1.

3. According to claim 1, the remote sensing image change detection method based on PAP-UNet++ network, Features: In step S2, the image preprocessing is specifically as follows: the image is uniformly scaled to a resolution of 512×512 as the input image, and then horizontally flipped with a probability of 0.5 to expand the data set.

4. The remote sensing image change detection method based on PAP-UNet++ network according to claim 1, Features: In step S3, each convolution layer in the PAP-UNet++ network includes a series of nonlinear transformations, which include convolution operations, batch normalization, and ReLU activation functions; wherein the convolution operation formula is as follows: Among them, f represents the input feature map, g represents the convolution kernel, N represents the size of the original image, S represents the stride size, and M represents the convolution kernel size; Batch normalization is used to transform the input image of each layer into a normal distribution with a mean of 0 and a variance of 1 to obtain a standardized result. The formula is as follows: Among them, x i represents a certain characteristic value of the i-th sample, μ B represents the feature average, represents the variance of the feature, B represents the sample set in the current batch, γ and β represent the introduced parameters, and ∈ represents the constant added to the minimum batch variance.

5. According to claim 4, the remote sensing image change detection method based on PAP-UNet++ network, Features: The PAP-UNet++ network includes an encoder structure and a decoder structure, the encoder structure is used for feature extraction and multi-level downsampling, and the decoder structure is used for multi-level upsampling using deconvolution operations; The PAP-UNet++ network merges two 3-channel images into a 6-channel image, that is, the PAP-UNet++ network processes the information of the two images at the same time and uses the pixel differences between them to complete the change detection task. The final output results are X 0,1 , X 0,2 , X 0,3 and X 0,4 , and finally take the average of the four output results.

6. According to claim 5, the remote sensing image change detection method based on PAP-UNet++ network, Features: The encoder structure replaces the convolution module of the original UNet++ network with a hybrid dilated convolution module, which consists of two dilated convolutions, a BN layer, and a ReLU activation function layer, and the expansion coefficients of the two layers of dilated convolutions are 2 and 3, respectively.

7. The remote sensing image change detection method based on PAP-UNet++ network according to claim 5, Features: The decoder structure includes a parallel atrous pyramid PAP module, which includes four parallel convolutional layer branches and a global average pooling branch. The four parallel convolutional layer branches use a 3×3 atrous convolution structure to capture context information at different levels. Each convolutional layer branch is equipped with a BN layer and a ReLU activation function to respectively perform feature normalization and introduce nonlinear expression capabilities; The global average pooling branch is used to capture global context features. The branch consists of a global average pooling layer, two 1×1 convolutional layers, and a Sigmoid activation function. A learnable scaling factor mechanism is introduced to dynamically adjust the importance of global context features.

8. The remote sensing image change detection method based on PAP-UNet++ network according to claim 7, Features: In the parallel atrous pyramid PAP module, the first convolution branch of the four parallel convolutional layer branches consists of a 1×1 convolutional layer and a Sigmoid activation function; the second convolution branch consists of two 3×3 atrous convolutional layers with atrous rates of 6 and 8 respectively; the third convolution branch consists of a 3×3 atrous convolutional layer with a atrous rate of 12; and the fourth convolution branch consists of a 3×3 atrous convolutional layer with a atrous rate of 18.

9. The remote sensing image change detection method based on PAP-UNet++ network according to claim 1, Features: In step S3, the PAP-UNet++ network embeds a CBAM attention mechanism module in the same-level skip connection process. The CBAM attention mechanism module includes two submodules: a channel attention CAM module and a spatial attention SAM module. The CBAM attention mechanism module receives a feature map from the previous layer as input, calculates the importance of each channel through the channel attention CAM module, and passes the result to the spatial attention SAM module to further refine the attention on the spatial position. The feature map processed by the two attention modules is used as output for use by subsequent layers; The channel attention CAM module applies global average pooling and global maximum pooling to the input features to obtain two one-dimensional vectors, and then passes these two one-dimensional vectors through two shared fully connected layers to obtain two channel attention weight vectors, and uses the activation function to activate the output of the fully connected layer, and then passes through another fully connected layer, and then uses the Sigmoid activation function to obtain the final channel attention weight, and finally performs channel-by-channel multiplication with the input feature map to obtain the weighted feature map; The spatial attention SAM module applies maximum pooling and average pooling to the output features of the channel attention module, performs pooling operations along the channel dimension to obtain two feature maps, which are then concatenated and convolved to obtain a single-channel feature map. The convolution output is then processed using the Sigmoid activation function to obtain the spatial attention weight, and the obtained spatial attention weight is element-wise multiplied with the output feature map of the channel attention module to obtain the final weighted feature map.

10. The remote sensing image change detection method based on PAP-UNet++ network according to claim 1, Features: In step S3, in the trained PAP-UNet++ network, the following L1 loss function is set: Among them, f(x i ) represents the predicted value of the model, y i Represents the true value of the model.

Citation Information

Cited By

  • Underground pipeline detection and identification method and system based on ground penetrating radar and neural network

    CN119805401A

  • Positive sample filtering method and device combining image comparison and sensitive defect detection

    CN120339220A

  • Road recognition and extraction method, device, equipment and medium

    CN120472302A

  • Road weak information extraction method and system based on ASU-Net

    CN120808157A

  • Mangrove forest identification method and system based on AI deep learning

    CN121121487A