SAR image change detection method and system based on dynamic bilinear fusion network

Abstract

The invention discloses an SAR image change detection method and system based on a dynamic bilinear fusion network, and the method comprises the steps: collecting dual-temporal SAR images of the same region in different periods, forming a dual-temporal data set, and selecting a part of image blocks and labels thereof as training samples for the training and optimization of a model; constructing an SAR image change detection model; double-time input in the training sample is sent into the dynamic bilinear fusion network for training, and model parameters are updated; inputting the dual-time SAR image into the trained model to obtain a change detection prediction result; according to the method, dynamic shift convolution and self-attention are combined in parallel, a self-attention mechanism is used for capturing global information, and the dynamic shift convolution dynamically gathers a plurality of convolution kernels and is combined with the shift convolution to extract more detailed and rich feature representation, so that the representation capability of the model is enhanced; the bilinear convergence module interactively fuses parallel features from among images and in the images, and captures spatial correlation among pixels in the images.

Description

Technical Field

[0001] The present invention belongs to remote sensing image change detection, and in particular relates to a SAR image change detection method and system based on a dynamic bilinear fusion network. Background Art

[0002] Synthetic Aperture Radar (SAR) imagery, with its unique imaging capabilities unrestricted by weather and lighting conditions, offers irreplaceable advantages in numerous applications. Change detection using SAR imagery is particularly crucial in scenarios such as land cover monitoring, disaster assessment, and urban planning. By analyzing image changes over time over the same area, change detection can provide powerful support for decision-making in these areas and has become a hot topic in remote sensing research.

[0003] With the rapid growth of remote sensing data and the increasing complexity of application requirements, traditional change detection methods face numerous challenges in terms of accuracy and robustness. Deep learning technology, due to its powerful feature learning capabilities, has become the mainstream choice in the current change detection field. Deep learning models can automatically extract effective features through multi-level learning, eliminating the need for manual feature extraction, thereby improving detection performance. One current research hotspot is how to fully leverage the feature representation capabilities of deep learning models to effectively process complex change information in dual-temporal SAR imagery. In "Change detection in synthetic aperture radar images using a dual-domain network," Qu et al. proposed a dual-domain network that utilizes multi-region convolution to adaptively model contextual information and central region features. In "Deep Shearlet Network for Change Detection in SAR Images," Dong et al. proposed ShearNet, which enhances the feature representation capabilities of CNNs and effectively reduces the impact of noise by introducing shearing transformations and denoising layers. In order to make up for the limitations of CNN in capturing global dependencies and processing long-distance features, the Transformer structure has gradually been introduced into remote sensing image change detection; in the paper "Convolution and attention mixer for synthetic aperture radar image change detection", Zhang et al. combined the self-attention mechanism and shifted convolution to capture global semantic information and local features in a parallel design.

[0004] In addition, feature fusion is a key strategy to further improve the accuracy of change detection. By integrating feature information from different images, networks or layers, the model can capture subtle changes while avoiding over-sensitivity to background noise. For example, by fusing low-level spatial features and high-level semantic information, the model can simultaneously obtain the fine edges and overall structure of the object, thereby improving the detection accuracy of complex change areas. In the paper "Change Detection in SAR Images Based on A Multi-Scale Attention Convolution Network", Li et al. introduced multi-scale feature extraction and linear attention weight modules, and simultaneously fused contextual information of different scales, effectively improving the detection accuracy. In the patent "SAR Image Change Detection Method Based on Multi-Directional Attention Representation Enhancement", a method of multi-scale convolution and multi-directional feature extraction is proposed, which can effectively enhance the center point representation information in SAR image change detection, solve the noise interference problem, and improve the change detection effect. To suppress speckle noise, Zhao et al. proposed a multi-domain fusion module in their paper "SAR Image Change Detection in Spatial-Frequency Domain Based on Attention Mechanism and Gated Linear Unit" that combines spatial and frequency domain features to construct complementary feature representations. Similarly, the patent "SAR Image Change Detection Method Based on Unsupervised Spatial-Frequency Representation Learning Fusion" effectively extracts key information in unsupervised scenarios by fusing spatial and frequency domain features, improving detection accuracy and reliability.

[0005] Although significant progress has been made in SAR change detection through the use of CNNs and Transformers for feature representation, as well as joint feature fusion methods, the field still faces several challenges. First, CNNs' feature representation capabilities are limited by the network's depth and channel width, making it difficult to capture more complex spatiotemporal dependencies. Furthermore, existing models primarily focus on fusing features at different scales or in different domains, while paying insufficient attention to the complex interactions between SAR image features. This limits the models' ability to capture and identify subtle image changes. Summary of the Invention

[0006] In order to solve the problems existing in the prior art, the present invention proposes a SAR image change detection method based on a dynamic bilinear fusion network, which effectively establishes the interaction between images and within images, thereby significantly improving the representation ability of the model.

[0007] In order to achieve the above objectives, in a first aspect, the present invention provides a SAR image change detection method based on a dynamic bilinear fusion network, comprising the following steps: Acquire a dual-time SAR image, and input the dual-time SAR image into a trained SAR image change detection model to obtain a change detection prediction result; The SAR image change detection model includes two branches with the same structure, a bilinear fusion module, and a change detection result prediction module. The branches with the same structure are equipped with a dynamic shift convolution and attention module, a gated feedforward network, and a bilinear fusion module within the image. The change detection result prediction module includes a fully connected neural network layer for performing linear transformation on the input features.

[0008] Furthermore, training the SAR image change detection model includes the following steps: S1, collects dual-temporal SAR images of the same area at different times to form a dual-temporal dataset; selects some image blocks and their labels as training samples for model training and optimization; S2, feeds the dual-time input in the training sample into the dynamic bilinear fusion network for training and updates the model parameters; In S3, the dual-time SAR image is input into the trained model to obtain the change detection prediction results.

[0009] Furthermore, the dynamic shift convolution module and the attention module process the input data in parallel. Specifically: First, adaptive modeling is performed through a dynamic convolution layer. Dynamic convolution uses a dynamic weight calculation mechanism to integrate multiple convolution kernels and dynamically adjust the weight coefficients of the convolution kernels. It is defined as follows:

[0010] in, and are the weight matrix and bias vector respectively, sharing the same attention, is the activation function, is the number of convolution kernels, Indicates the The attention weight of the convolution kernel; The feature map processed by dynamic convolution is sent to the shift convolution layer for shift operations in four directions: up, down, left, and right; The shifted feature map is input again into another dynamic convolution layer to extract features; The self-attention module divides the image into non-overlapping sub-tiles and encodes each sub-tile into a token embedding. The query (Q), key (K), and value (V) are calculated through linear transformation of the token embedding. The output of the self-attention module is calculated by the following formula:

[0011] in is the scaling factor; The outputs of the dynamic shift convolution module and the self-attention module are fused through element-wise summation to obtain a fusion feature that combines local spatial information and global dependencies, thereby enhancing the model's ability to express complex patterns and potential features.

[0012] Furthermore, squeeze-and-excitation is used to calculate the attention weights. First, the global spatial information is compressed by global average pooling, and then two fully connected layers and softmax are used to obtain the kernel attention weights. With input , pay attention to the weight Represents the linear model under a given input The optimal set, the set model Through nonlinear aggregation of attention convolution kernels.

[0013] Furthermore, the feature map processed by dynamic convolution is sent to the shift convolution layer for shifting in four directions: up, down, left, and right. The number of channels is expanded to ,in is the expansion ratio, is the number of channels. Secondly, the input features are divided into the shift convolution part along the channel direction, and then divided into four equal parts, each feature is shifted along the left, right, up, and down directions.

[0014] Furthermore, for the bilinear confluence module: The outputs of the dynamic shift convolution and attention modules of the two branches and the output of the gated feedforward network will enter the bilinear confluence module as the input of the bilinear confluence module. Perform a convolution operation of a set size, by matrix outer product at position Integrate these two feature maps and calculate as:

[0015] in, 、 They are the feature maps obtained by convolution, Then it is the bilinear combination feature at position p; For the bilinear fusion module within the image, the two parallel feature extraction modules are the dynamic shift convolution and attention module and the gated feedforward network. The hybrid module that combines the dynamic shift convolution and attention module with the gated feedforward network extracts deep features through stacking, and inputs the results of the two-branch bilinear fusion module into a top-level bilinear fusion module to generate the final detection result through the fully connected layer and softmax function.

[0016] Furthermore, the Adam optimizer is used to train the SAR image change detection model, and the cross entropy loss function is adopted.

[0017] In a second aspect, the present invention provides a SAR image change detection system based on a dynamic bilinear fusion network, comprising a data acquisition module, two branches with the same structure, and a detection result prediction module; the two branches with the same structure are provided with a dynamic shift convolution and attention module, a gated feedforward network, and a bilinear fusion module within the image; The data acquisition module is used to acquire dual-time SAR images and serve as the input of the SAR image change detection model; Dynamic shift convolution and attention modules are used to process input data in parallel to obtain fused features that combine local spatial information and global dependencies; The gated feedforward network is used to modulate the input features, strengthen or suppress some features, and form gated features; The bilinear confluence module in the image performs convolution and integrates feature maps based on the outputs of the dynamic shift convolution and attention modules of the two branches and the output of the gated feedforward network. It uses the second-order information between features to model the pairwise interaction of features at each position and extracts deep features through stacking. The detection result prediction module uses a bilinear confluence module to generate the final detection result through a fully connected layer and a softmax function.

[0018] In a third aspect, the present invention may also provide a computer device comprising a processor and a memory, wherein the memory is used to store a computer executable program, and the processor reads the computer executable program from the memory and executes it. When the processor executes the computer executable program, the SAR image change detection method based on the dynamic bilinear fusion network described in the present invention can be implemented.

[0019] A computer-readable storage medium is also provided. The computer-readable storage medium stores a computer program. When the computer program is executed by a processor, the SAR image change detection method based on a dynamic bilinear fusion network according to the present invention can be implemented.

[0020] Compared with the existing technology, the present invention has at least the following beneficial effects: the present invention proposes a SAR image change detection method based on a dynamic bilinear fusion network, adopts multiple parallel feature extraction modules and their fusion, effectively establishes the interaction between images and within images, thereby significantly improving the representation ability of the model. This network architecture can not only comprehensively utilize the advantages of different modules, but also improve the recognition ability of complex change patterns; in addition, the present invention designs a dynamic shift convolution module, which combines the advantages of dynamic convolution and shift convolution, significantly enriches the effective extraction of contextual information, and thus improves the sensitivity to change areas; at the same time, the introduction of a bilinear fusion module generates bilinear features through matrix outer product operations, which can effectively obtain complementary information and capture the comprehensive interactive correlation between multiple features, thereby enhancing the depth and breadth of information integration; extensive experimental results on three actual SAR data sets verify the effectiveness of the method of the present invention, providing an innovative and efficient solution for SAR image change detection, which has important theoretical significance and application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 This is the overall framework diagram of the dynamic bilinear fusion network in an embodiment of the present invention.

[0022] Figure 2 This is a structural diagram of the dynamic shift convolution and attention module in an embodiment of the present invention.

[0023] Figure 3 4 is a structural diagram of a bilinear merging module in an embodiment of the present invention.

[0024] Figure 4 2 is a structural diagram of dynamic convolution in an embodiment of the present invention.

[0025] Figure 5 2 is a comparison chart of the detection results of different methods in the embodiments of the present invention. DETAILED DESCRIPTION

[0026] The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0027] The structure of the dynamic bilinear fusion network of the present invention is as follows Figure 1 As shown in Figure 3, the network architecture consists of a dynamic shifted convolution and attention module, a gated feedforward network, and a bilinear fusion module. The network first uses an unsupervised pre-classification method to obtain training samples. Specifically, a log-ratio operator is used to generate a difference map, which is then divided into change, invariant, and uncertain classes using a hierarchical clustering algorithm. Then, samples from the change and invariant classes are randomly selected as training data. Next, the training data is fed into a dynamic bilinear fusion network for feature extraction. Finally, the output features are passed through a fully connected layer to obtain the final change map.

[0028] like Figure 2 and Figure 4 As shown in Figure 1, the dynamic shift convolution module and the self-attention module process the input in parallel. Specifically, the input data x is first adaptively modeled through a dynamic convolution layer. Dynamic convolution uses a dynamic weight calculation mechanism. By integrating multiple convolution kernels and dynamically adjusting the weight coefficients of the convolution kernels, it can better capture the complex patterns and potential characteristics of the input data. It is defined as follows:

[0029] in, and are the weight matrix and bias vector respectively, sharing the same attention, is the activation function, is the number of convolution kernels, Indicates the The attention weights of the convolution kernels are calculated using squeeze-and-excitation. First, the global spatial information is compressed by global average pooling:

[0030] Then two fully connected layers and softmax are used to get kernel attention. Attention weights It is not fixed, but changes with the input , pay attention to the weight Represents the linear model under a given input The optimal set of. Set model By nonlinearly aggregating 𝐾 convolution kernels through attention, it has stronger representation ability than static convolution.

[0031] refer to Figure 4 Next, the dynamic convolution feature map is sent to the shift convolution layer for shift operation, including shift in four directions: up, down, left, and right. First, the input data The number of channels is expanded to ,in is the expansion ratio, is the number of channels; secondly, the input feature map is divided into shifted convolution parts along the channel direction, and then the divided shifted convolution parts are evenly divided into four parts along the channel. Each feature is shifted pixel by pixel along the left, right, top, and bottom directions respectively. Finally, the feature map after the shift operation is input again to another dynamic convolution layer for further feature extraction. Therefore, the dynamic shift convolution can be expressed as:

[0032] The module that integrates dynamic convolution and shifted convolution can simultaneously exploit the adaptability and more structural information of the input features, thereby better capturing the complex local patterns and potential characteristics of the input data while maintaining high computational efficiency.

[0033] At the same time, the self-attention module divides the image into multiple non-overlapping sub-tiles and encodes each sub-tile into a token embedding. Next, the query (Q), key (K), and value (V) are calculated through linear transformation of the token embedding. The output of the self-attention module is calculated by the following formula:

[0034] in is the scaling factor. Finally, the outputs of the dynamic shift convolution module and the self-attention module are fused together by element-wise summation.

[0035] At the same time, the gated feedforward network processes the input data x in parallel. First, the input is processed by a nonlinear activation function. The GELU or ReLU function can be used to activate the input feature map element by element. The activation process is calculated as follows:

[0036] Next, the activated feature map will be element-wise multiplied with another input feature map to implement the gating mechanism. This process can be seen as modulating the input features, strengthening important features and suppressing unimportant features, thereby forming gated features:

[0037] in, It is another input feature map combined with the activation feature map, ⊙ represents element-by-element multiplication. The core of this step is the gating mechanism, which can dynamically adjust the transmission of features.

[0038] Finally, the feature map processed by the gating mechanism is output as input for subsequent processing. The structure of the gated feedforward network enables the model to flexibly capture important features, improving the model's learning ability and expressiveness of input features.

[0039] refer to Figure 3 , the output of the dynamic shift convolution and attention modules of the two branches and the output of the gated feedforward network are used as the input of the bilinear confluence module, and a 3×3 convolution operation is first performed; then, the matrix outer product is performed at position Integration and These two feature maps are calculated as:

[0040] in, 、 They are the feature maps obtained by convolution, is the bilinear combination feature at position p. Unlike conventional fusion operations (such as summation or concatenation), the bilinear fusion module uses the second-order information between features to model the pairwise interactions of the features at each position, thereby capturing more feature relationships. In the present invention, a bilinear fusion module is used to obtain the joint feature representation of different feature extraction modules, which is not limited to features in the same image, but also includes feature interactions from top-level networks of different images. In this way, feature interactions within and between images are fully modeled, making feature extraction more discriminative and more suitable for fine-grained change detection tasks. For the bilinear fusion module within the image, the two parallel feature extraction modules are dynamic shift convolution and attention module with gated feedforward network. The two parallel feature extraction modules extract deep features by stacking. Finally, the outputs of the bilinear fusion modules from the two branches are input into a bilinear fusion module again, and the final detection result is generated through a fully connected layer and softmax function.

[0041] Specifically, this paper employs a dynamic bilinear fusion network for change detection in SAR images. The image is divided into non-overlapping 9×9 pixel sub-blocks. Training samples are obtained using an unsupervised pre-classification method. The network is trained for 20 epochs, using cross-entropy as the loss function to optimize model performance. During each training epoch, the model output and loss are calculated using forward propagation, followed by backpropagation to update the model parameters. During the inference phase, the model runs in evaluation mode, processing test data pixel by pixel, classifying it according to pre-classified labels, and ultimately generating a change map.

[0042] In order to verify the effects of the present invention, the following verification examples were carried out.

[0043] In this paper, experiments were conducted on three datasets: a river, a lake, and San Francisco. The river dataset covers its estuary area and consists of images acquired with the Radarset-2 SAR sensor in June 2008 and June 2009, with a size of 291×444 pixels. The lake dataset covers a region of the lake, with images acquired with the Sentinel-1 sensor in May and July 2020, with a size of 384×384 pixels. During this period, the lake experienced its highest water level on record. The San Francisco dataset consists of images acquired with the ERS-2 satellite in August 2003 and May 2004, with a resolution of 25m and a size of 256×256 pixels. The performance of the proposed model was measured using five common evaluation metrics in change detection: false positives (FP), false negatives (FN), overall error (OE), percentage correctly classified (PCC), and kappa coefficient (KC).

[0044] Figure 5 Tables 1, 2, and 3 show the detection results of different methods on the three datasets. The comparison algorithms include Convolutional-Wavelet Neural Networks (CWNN), Dual-domain network (DDNet), LayerAttention-Based Noise-Tolerant Network (LANTNet), Convolution and attention mixer (CAM), and Wavelet-Based Bi-Dimensional Aggregation Network (WBANet).

[0045] Table 1 Evaluation indicators of detection results of a river dataset

[0046] Table 2 Evaluation indicators of detection results of a lake dataset

[0047] Table 3 Evaluation indicators of San Francisco dataset detection results

[0048] In a river dataset, PCC improved by 0.09%, 0.16%, and 0.44% compared to DDNet, LANTNet, and CAM, respectively, while KC improved by 0.18%, 1.45%, and 4.86%, respectively. In a lake dataset, PCC improved by 1.13%, 1.49%, and 0.19% compared to DDNet, LANTNet, and CAM, respectively, while KC improved by 5.53%, 7.62%, and 1.04%, respectively. This improvement is primarily due to the powerful feature fusion and interaction capabilities of the bilinear confluence module, which fully utilizes the diverse information extracted by convolution and attention mechanisms, enhancing sensitivity to complex water scenes. Furthermore, the dynamic shift convolution module further enhances the capture of local features, especially when dealing with changing region boundaries, effectively reducing false positives and improving change detection accuracy. In the San Francisco dataset, the dynamic bilinear fusion network achieved improvements in PCC by 0.87%, 0.74%, and 0.08% over DDNet, LANTNet, and CAM, respectively, and improvements in KC by 5.91%, 5.12%, and 0.41%, respectively. This dataset has relatively few regions with small variations, so accurately capturing these subtle changes is crucial for improving performance. The combination of dynamic convolution and shifted convolution captures more detailed information, while the introduction of the bilinear fusion module fully exploits the complex interactions between features. This enables this method to accurately detect regions of change while maintaining low false positives.

[0049] The present invention also provides a SAR image change detection system based on a dynamic bilinear fusion network, comprising a data acquisition module, two branches with the same structure, and a detection result prediction module; the two branches with the same structure are provided with a dynamic shift convolution and attention module, a gated feedforward network, and a bilinear fusion module within the image; The data acquisition module is used to acquire dual-time SAR images and serve as the input of the SAR image change detection model; Dynamic shift convolution and attention modules are used to process input data in parallel to obtain fused features that combine local spatial information and global dependencies; The gated feedforward network is used to modulate the input features, strengthen or suppress some features, and form gated features; The bilinear confluence module in the image performs convolution and integrates feature maps based on the outputs of the dynamic shift convolution and attention modules of the two branches and the output of the gated feedforward network. It uses the second-order information between features to model the pairwise interaction of features at each position and extracts deep features through stacking. The detection result prediction module uses a bilinear confluence module to generate the final detection result through a fully connected layer and a softmax function.

[0050] On the other hand, the present invention provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, it can implement the SAR image change detection method based on the dynamic bilinear fusion network described in the present invention.

[0051] The computer device may be a laptop computer, a desktop computer or a workstation.

[0052] The present invention can also provide a computer device, including a processor and a memory, the memory is used to store a computer executable program, the processor reads the computer executable program from the memory and executes it, and when the processor executes the computer executable program, it can implement the SAR image change detection method based on the dynamic bilinear fusion network described in the present invention.

[0053] The processor can be a central processing unit (CPU), a graphics processing unit (GPU), a digital signal processor (DSP), an application-specific integrated circuit (ASIC), or an off-the-shelf field-programmable gate array (FPGA).

[0054] The memory of the present invention may be an internal storage unit of a laptop computer, desktop computer or workstation, such as a memory or a hard disk; or an external storage unit, such as a mobile hard disk or a flash memory card.

[0055] Computer-readable storage media may include computer storage media and communication media. Computer storage media include volatile and non-volatile, removable and non-removable media implemented by any method or technology for storing information such as computer-readable instructions, data structures, program modules or other data. Computer-readable storage media may include: read-only memory (ROM), random access memory (RAM), solid-state drives (SSD) or optical disks, etc. Among them, random access memory may include resistance random access memory (ReRAM) and dynamic random access memory (DRAM).

[0056] The above content is only for explaining 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 claims of the present invention.

Claims

1. A SAR image change detection method based on a dynamic bilinear fusion network is characterized by: The following steps are involved: Acquire a dual-time SAR image, and input the dual-time SAR image into a trained SAR image change detection model to obtain a change detection prediction result; The SAR image change detection model includes two branches with the same structure, a bilinear fusion module and a change detection result prediction module. The branches with the same structure are provided with a dynamic shift convolution and attention module, a gated feedforward network and a bilinear fusion module within the image; the change detection result prediction module includes a fully connected neural network layer for performing linear transformation on input features.

2. The SAR image change detection method based on dynamic bilinear fusion network according to claim 1 is characterized in that: Training a SAR image change detection model involves the following steps: S1, collect dual-temporal SAR images of the same area at different periods to form a dual-temporal data set; select some image blocks and their labels as training samples for model training and optimization; S2, sending the dual-time input in the training sample into the dynamic bilinear fusion network for training and updating the model parameters; S3, input the dual-time SAR image into the trained model to obtain the change detection prediction result.

3. The SAR image change detection method based on dynamic bilinear fusion network according to claim 1 is characterized in that: The dynamic shift convolution module and the attention module process the input data in parallel. Specifically: First, adaptive modeling is performed through a dynamic convolution layer. Dynamic convolution uses a dynamic weight calculation mechanism to integrate multiple convolution kernels and dynamically adjust the weight coefficients of the convolution kernels. Its definition is as follows: in, and are the weight matrix and bias vector, respectively, sharing the same attention, is the activation function, is the number of convolution kernels, Indicates The attention weights of the convolution kernels; The feature map processed by dynamic convolution is sent to the shift convolution layer for shift operations in four directions: up, down, left, and right; The feature map after the shift operation is input again into another dynamic convolution layer to extract features; The self-attention module divides the image into non-overlapping sub-tiles and encodes each sub-tile into a token embedding. The query (Q), key (K), and value (V) are calculated through linear transformation of the token embedding. The output of the self-attention module is calculated by the following formula: in is the scaling factor; The outputs of the dynamic shift convolution module and the self-attention module are fused by element-wise summation to obtain a fused feature that combines local spatial information and global dependencies.

4. The SAR image change detection method based on dynamic bilinear fusion network according to claim 3 is characterized in that: The attention weights are calculated using squeeze-and-excitation. First, the global spatial information is compressed by global average pooling, and then two fully connected layers and softmax are used to obtain the core attention weights. With input , pay attention to the weight Represents the linear model under a given input The optimal set of ensemble models Non-linear aggregation via attention convolution kernel.

5. The SAR image change detection method based on dynamic bilinear fusion network according to claim 3 is characterized in that: The feature map processed by dynamic convolution is sent to the shift convolution layer for shifting in four directions: up, down, left, and right. The operation includes: The number of channels is expanded to ,in is the expansion ratio, is the number of channels. Secondly, the input feature is divided into the shifted convolution part along the channel direction, and then divided into four parts on average, and each feature moves along the left, right, up and down directions.

6. The SAR image change detection method based on dynamic bilinear fusion network according to claim 1 is characterized in that: For the bilinear confluence module: The outputs of the dynamic shift convolution and attention modules of the two branches and the output of the gated feedforward network will enter the bilinear confluence module as the input of the bilinear confluence module. Perform a convolution operation of a set size by performing a matrix outer product at position Integrate these two feature maps and calculate as: in, , They are the feature maps obtained after convolution, Then it is the bilinear combination feature at position p; For the bilinear fusion module within the image, the two parallel feature extraction modules are the dynamic shift convolution and attention module and the gated feedforward network. The hybrid module of the dynamic shift convolution and attention module and the gated feedforward network is stacked to extract deep features. The results of the two-branch bilinear fusion module are input into a top-level bilinear fusion module, and the final detection result is generated through the fully connected layer and softmax function.

7. The SAR image change detection method based on dynamic bilinear fusion network according to claim 1 is characterized in that: The Adam optimizer is used to train the SAR image change detection model, and the cross entropy loss function is adopted.

8. A SAR image change detection system based on a dynamic bilinear fusion network, characterized in that: It includes a data acquisition module, two branches with the same structure, and a detection result prediction module; the two branches with the same structure are provided with a dynamic shift convolution and attention module, a gated feedforward network, and a bilinear fusion module within the image; The data acquisition module is used to acquire dual-time SAR images and use them as input for the SAR image change detection model; Dynamic shift convolution and attention modules are used to process input data in parallel to obtain fused features that combine local spatial information and global dependencies; The gated feedforward network is used to modulate the input features, strengthen or suppress some features, and form gated features; The bilinear confluence module in the image performs convolution and integration of feature maps based on the output of the dynamic shift convolution and attention modules of the two branches and the output of the gated feedforward network. It uses the second-order information between features to model the pairwise interaction of features at each position and extracts deep features through stacking. The detection result prediction module uses a bilinear confluence module to generate the final detection result through a fully connected layer and a softmax function.

9. A computer device, characterized in that: The invention comprises a processor and a memory, wherein the memory is used to store a computer executable program, the processor reads the computer executable program from the memory and executes it, and when the processor executes the computer executable program, the SAR image change detection method based on a dynamic bilinear fusion network as described in any one of claims 1 to 7 can be implemented.

10. A computer-readable storage medium, characterized in that: A computer program is stored in a computer-readable storage medium. When the computer program is executed by a processor, the SAR image change detection method based on a dynamic bilinear fusion network as described in any one of claims 1 to 7 can be implemented.

Images