A multi-scale attention space-frequency dual-domain collaborative SAR image change detection method

Abstract

The application relates to the technical field of change detection, in particular to a multi-scale attention space-frequency dual-domain collaborative SAR image change detection method.The application constructs a space-frequency dual-domain collaborative fusion module, generates a dynamic frequency weight mask by using space features, performs frequency spectrum modulation on frequency domain features under the guidance of a space structure, and suppresses noise frequency components irrelevant to real changes; meanwhile, the modulated frequency domain features are reconstructed back to the space dimension, a space attention map is generated to perform secondary calibration on the space features, and the frequency domain global information can reversely enhance the space local change features; therefore, the application does not simply splice the space features and the frequency domain features, but realizes bidirectional modulation and collaborative fusion between the space domain and the frequency domain, can reduce the interference of noise spectrum response and invalid space response on the detection result, improves the inhibition capacity for SAR speckle noise, a complex texture background and a pseudo change disturbance while keeping the edge and detail information of the change region.

Description

Technical Field

[0001] This application relates to the field of change detection technology, and in particular to a multi-scale attention-based spatial-frequency dual-domain collaborative SAR image change detection method. Background Technology

[0002] Change detection is the process of identifying changes by analyzing images of the same scene observed at different time intervals. In the past, many change detection methods have been applied to various scenarios, such as Earth resource monitoring, forest resource assessment, desertification analysis, urban or natural environmental change detection, and natural disaster impact assessment. Today, the application of change detection in remote sensing imagery has become an important research direction. SAR, as an important remote sensing imaging technology, possesses all-weather, all-day imaging capabilities, and can stably acquire high-quality surface information even in complex environments. Therefore, SAR exhibits unique advantages in change detection.

[0003] Existing technology (patent application number: 202310790988.5) discloses an unsupervised deep multi-scale SAR image change detection method based on spatial frequency domain. This method uses hierarchical fuzzy C-means clustering to extract pseudo-labels from difference images generated from SAR images of the detection area, solving the problem of insufficient labels. This makes the SAR image change detection method applicable to more scenarios. This invention utilizes the spatial and frequency domain information of the input SAR image and proposes a spatial multi-regional, multi-scale deep feature extraction method. The image features captured in this way are more conducive to detection. This invention introduces attention mechanisms and gated linear units in the spatial and frequency domains respectively to improve the sensitivity to changes in details and reduce the influence of inherent speckle noise in SAR images, thereby improving detection accuracy.

[0004] The existing technology has the following drawbacks:

[0005] While existing technologies can utilize both spatial and frequency domain features, they often employ splicing or simple fusion methods, making it difficult to achieve bidirectional modulation between spatial structural information and frequency domain noise response. This results in the inability to adaptively filter frequency domain noise components and spatial detail features, leading to false detections, missed detections, and loss of boundary details in complex textures and strong speckle noise scenarios. Summary of the Invention

[0006] To overcome the aforementioned problems in the existing technology, this application provides a multi-scale attention-based spatial-frequency dual-domain cooperative SAR image change detection method and system, which adopts the following technical solution:

[0007] In the first aspect, this application provides a multi-scale attention-based spatial-frequency dual-domain collaborative SAR image change detection method, the steps of which include: pre-classifying the SAR input image using a clustering method to generate pseudo-label samples;

[0008] Spatial features are obtained by extracting spatial features from pseudo-labeled samples in multiple regions and at multiple scales;

[0009] Frequency domain features are extracted from the pseudo-labeled samples to obtain frequency domain features;

[0010] Based on the dual-domain collaborative fusion module, spatial and frequency domain features are subjected to spatial-frequency dual-domain collaborative operation to obtain dual-domain fused features;

[0011] SAR image change detection based on dual-domain fusion features.

[0012] Furthermore, the extraction of multi-region and multi-scale spatial domain features from the pseudo-label samples specifically involves:

[0013] Multi-region selection is performed on image patches in pseudo-label samples, selecting horizontal regions, vertical regions, and complete regions, and then convolutional operations are performed to obtain feature maps;

[0014] Multi-scale spatial feature extraction is performed on the feature map to obtain three types of region features;

[0015] A multi-scale attention mechanism is used to obtain key features from the complete region features among the three types of region features;

[0016] The horizontal region features, vertical region features, and key features are added together to obtain the extracted features.

[0017] The extracted features are subjected to size matching and convolution operations to obtain the final spatial features.

[0018] Furthermore, the frequency domain feature extraction of the pseudo-label samples specifically involves: using the frequency components of the pseudo-label samples to generate a first weight matrix and a second weight matrix using Conv; activating the second weight matrix with the Swish activation function and then performing a weighted operation with the first weight matrix to obtain the frequency domain features, which can be expressed as: ,in, and These represent the first and second weight matrices, respectively. This represents the frequency components of the pseudo-labeled samples. and This is a bias term.

[0019] Furthermore, the spatial-frequency dual-domain collaborative operation based on the dual-domain collaborative fusion module for spatial and frequency domain features specifically involves:

[0020] Spatial domain modulation is applied to the frequency domain features to obtain the modulated frequency domain features;

[0021] The spatial features are frequency-domain modulated to obtain the modulated spatial features;

[0022] The modulated spatial and frequency domain features are reconstructed and fused to obtain dual-domain fused features.

[0023] Furthermore, the step of spatially modulating the frequency domain features to obtain modulated frequency domain features specifically involves:

[0024] Using the DCF module to extract spatial features Channel descriptors are compressed using global average pooling. Then, a bottlenecked MLP is used to model the channel dependencies between frequency components through nonlinear mapping. Finally, a dynamic frequency weight mask is generated using the sigmoid function. The dynamic frequency weight mask It can be represented as ,in, This represents a multilayer perceptron. This indicates a global average pooling operation. This represents the Sigmoid activation function;

[0025] Dynamic frequency weight mask Frequency domain characteristics Performing element-wise dot multiplication enables dynamic modulation of the spectral response, yielding the modulated frequency domain characteristics, which can be expressed as: ,in, This represents the frequency domain characteristics after modulation. Represents frequency domain characteristics, This represents a dynamic frequency weighted mask.

[0026] Furthermore, the spatial features are frequency-domain modulated to obtain modulated spatial features. Specifically, the modulated frequency-domain features are... use The spatial dimension is reconstructed, then passed through convolutional layers and a sigmoid function to generate a spatial attention map; spatial features are then analyzed based on the spatial attention map. Secondary calibration and enhancement are performed to obtain the modulated spatial features: ,in, Frequency domain features are reconstructed back to the spatial dimension, and convolutional layers are used to perceive the mapping weights of global spectral information in spatial location.

[0027] Furthermore, the reconstruction and fusion of the modulated spatial and frequency domain features to obtain dual-domain fused features specifically involves: reconstructing and fusing the modulated frequency domain features... The reconstructed frequency domain features are obtained by using inverse Fourier transform; the modulated spatial features and the reconstructed frequency domain features are then fused by addition to obtain dual-domain fused features, which can be expressed as: ;in, This indicates the dual-domain fusion feature.

[0028] Secondly, this application also provides a multi-scale attention-based spatial-frequency dual-domain collaborative SAR image change detection system, including: a pseudo-label sample generation module, a spatial feature extraction module, a frequency domain feature extraction module, a dual-domain feature fusion module, and an image change detection module;

[0029] The pseudo-label sample generation module is used to pre-classify SAR input images using clustering methods and generate pseudo-label samples.

[0030] The spatial feature extraction module is used to extract spatial features from pseudo-label samples in multiple regions and at multiple scales to obtain spatial features.

[0031] The frequency domain feature extraction module is used to extract frequency domain features from pseudo-label samples to obtain frequency domain features;

[0032] The dual-domain feature fusion module is used to perform spatial-frequency dual-domain collaborative operations on spatial features and frequency domain features based on the dual-domain collaborative fusion module to obtain dual-domain fused features.

[0033] The image change detection module is used to detect changes in SAR images based on dual-domain fusion features.

[0034] Thirdly, this application provides an electronic device, comprising:

[0035] One or more processors; a memory; and one or more computer programs, wherein the one or more computer programs are stored in the memory, and the one or more computer programs include instructions that, when executed by the device, cause the device to perform the method as described in the first aspect.

[0036] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when run on a computer, causes the computer to perform the method described in the first aspect.

[0037] This application has the following technical effects:

[0038] This application constructs a space-frequency dual-domain collaborative fusion module, utilizes spatial features to generate a dynamic frequency weight mask, and performs spectral modulation of frequency domain features under spatial structure guidance to suppress noise frequency components unrelated to real changes. Simultaneously, the modulated frequency domain features are reconstructed back to the spatial dimension, and a spatial attention map is generated to perform secondary calibration of the spatial features, enabling global frequency domain information to inversely enhance local spatial change features. Therefore, this application does not simply splice spatial and frequency domain features, but achieves bidirectional modulation and collaborative fusion between the spatial and frequency domains, which can reduce the interference of noise spectral response and invalid spatial response on the detection results, and improve the suppression of SAR speckle noise, complex texture background, and spurious change disturbances while maintaining the edge and detail information of the change region. Attached Figure Description

[0039] To more clearly illustrate the solutions in this application, the accompanying drawings used in the description of the embodiments of this application will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0040] Figure 1 This is an exemplary system architecture diagram to which embodiments of this application can be applied;

[0041] Figure 2 This is a detailed operational framework diagram of an embodiment of this application;

[0042] Figure 3 This is a detailed process diagram of multi-region selection in an embodiment of this application;

[0043] Figure 4 This is a comparison diagram of traditional methods and multi-scale attention modules in multi-region feature extraction in embodiments of this application;

[0044] Figure 5 This is a flowchart of frequency domain feature extraction according to an embodiment of this application;

[0045] Figure 6 This is a detailed process of dual-domain feature fusion in the embodiments of this application;

[0046] Figure 7 These are detection maps of Sendai City, Japan, and Xiong'an New Area, China, the study areas of this application's embodiments;

[0047] Figure 8-11 The figures show various experimental data from embodiments of this application. Figure 12 This is a schematic diagram of a computer device according to an embodiment of this application. Detailed Implementation

[0048] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains; the terminology used herein in the specification of the application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application; the terms "comprising" and "having," and any variations thereof, in the specification, claims, and foregoing drawings of this application, are intended to cover non-exclusive inclusion. The terms "first," "second," etc., in the specification, claims, or foregoing drawings of this application are used to distinguish different objects, not to describe a particular order.

[0049] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0050] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings.

[0051] like Figure 1 As shown, system architecture 100 may include terminal devices 101, 102, and 103, a network 104, and a server 105. Network 104 serves as the medium for providing communication links between terminal devices 101, 102, and 103 and server 105. Network 104 may include various connection types, such as wired or wireless communication links, or fiber optic cables, etc.

[0052] Users can use terminal devices 101, 102, and 103 to interact with server 105 via network 104 to receive or send messages, etc. Various communication client applications can be installed on terminal devices 101, 102, and 103, such as web browser applications, shopping applications, search applications, instant messaging tools, email clients, social media platform software, etc.

[0053] Terminal devices 101, 102, and 103 can be various electronic devices with displays and support web browsing, including but not limited to smartphones, tablets, e-book readers, MP3 players (Moving Picture Experts Group Audio Layer III), MP4 players (Moving Picture Experts Group Audio Layer IV), laptops, and desktop computers, etc.

[0054] Server 105 can be a server that provides various services, such as a backend server that supports the pages displayed on terminal devices 101, 102, and 103.

[0055] It should be noted that the multi-scale attention-based spatial-frequency dual-domain cooperative SAR image change detection method provided in this application embodiment is generally executed by a server / terminal device, and correspondingly, the multi-scale attention-based spatial-frequency dual-domain cooperative SAR image change detection system is generally set in the server / terminal device.

[0056] It should be understood that Figure 1 The number of terminal devices, networks, and servers shown is merely illustrative. Depending on implementation needs, any number of terminal devices, networks, and servers can be included.

[0057] Continue to refer to Figure 2 A multi-scale attention-based spatial-frequency dual-domain collaborative SAR image change detection method, comprising the following steps:

[0058] S1, use clustering method to pre-classify SAR input images and generate pseudo-label samples;

[0059] It should be noted that the clustering method is fuzzy C-means clustering based on spider wasps (SWOFCM). This method transforms the search process for cluster centers into a population optimization process by simulating the predation, nest building and mating behavior of spider wasps, thereby improving the algorithm's ability to escape local optima.

[0060] In this embodiment, the pre-classification of SAR input images using clustering methods specifically involves:

[0061] S101, by processing the SAR input image and Perform a logarithmic ratio operation to obtain the first difference plot;

[0062] S102, the first difference map is processed by the first SWOFCM algorithm to obtain classification data; specifically, in this process, the spider-wasp population in the first SWOFCM algorithm iterates continuously in the solution space to minimize the objective function of FCM and obtain the initial cluster centers; based on the initial cluster centers, the difference map pixels are divided into two basic clusters. and The number of pixels are respectively and .

[0063] S103, regarding the basic clusters in categorical data A second SWOFCM algorithm is performed to obtain a second difference map; specifically, two prior thresholds are set. The second SWOFCM algorithm is used to divide the basic clusters. Divided into five sub-clusters and Calculate the average pixel value within each sub-cluster and sort them in ascending order, satisfying the following conditions: Based on heuristic principles, Predefine it as a core change class and associate it with The pixel ratio is incremented step by step for judgment. If the incremented result is less than 1, the pixel ratio is incremented step by step for judgment. If the accumulated result is between 1 / 2 and 2 / 3, then the pixels in the current cluster are classified as changed pixels; if ... and If the sum exceeds a certain value, the corresponding pixel is classified as an uncertain pixel; if the sum exceeds a certain value... These pixels are then classified as unchanged pixels; the classified pixels are then mapped onto the SAR input image as pseudo-label pixels.

[0064] S104, centered on the pseudo-label pixel, select a size of... Image blocks are stitched together to form pseudo-label samples.

[0065] S2, perform multi-region and multi-scale spatial feature extraction on pseudo-label samples to obtain spatial features;

[0066] In this embodiment of the application, the extraction of multi-region and multi-scale spatial domain features from the pseudo-label samples specifically includes:

[0067] like Figure 3 As shown in the figure, a multi-region selection method according to an embodiment of this application is illustrated. Specifically,

[0068] S201, perform multi-region selection on the image patch in the pseudo-label sample, select the horizontal region of the first region, the vertical region of the second region, and the complete region of the third region, and perform convolution operation to obtain feature map;

[0069] It should be noted that all information of the complete region image block is preserved; while the horizontal and vertical regions emphasize the central region by ignoring certain edge information; the complete region ensures the integrity of the information; the multi-region selection method effectively utilizes contextual information while highlighting the central region.

[0070] Furthermore, convolution operations are performed on the image patches after multi-region selection to obtain feature maps. Specifically, multi-region selection divides the image patch equally along the channel direction to obtain a feature map, which can be represented as: ,in, , , Representing feature maps, Indicates transpose. Indicates the number of channels;

[0071] It should be noted that, according to the principle of convolution, the... Each channel contains complete information about the pseudo-labeled sample.

[0072] S202, performs multi-scale spatial feature extraction on the feature map, obtaining three types of region features. Specifically...

[0073] In the spatial feature extraction process, each layer uses a convolutional layer with bias and different kernel sizes, as shown in the following expression: ,in, Indicates the first The output of a convolutional operation, Indicates the first The convolution kernels of the layers are different in size from one layer to another. This represents the convolution operation. Indicates the first In the layer area, This indicates the bias term; This represents the feature map corresponding to the i-th channel;

[0074] The features after convolution are normalized, and the horizontal and vertical region features are padded to match the size of the complete region features. Specifically, the padded horizontal and vertical region features are zero-padding in their corresponding directions to fill the parts of the horizontal and vertical region features that were ignored during multi-region selection, ensuring their size matches the complete region features. After padding, the... The horizontal region features, vertical region features, and complete region features of a layer are represented as follows: , and ;

[0075] like Figure 4 As shown in the figure, a comparison diagram of the multi-scale attention mechanism and the traditional feature extraction module of the present application embodiment is presented. (a) represents the traditional feature extraction module, and (b) represents the multi-scale attention mechanism.

[0076] S203, employs a multi-scale attention mechanism to obtain key features from the complete region features among the three types of region features; specifically, a multi-scale attention mechanism is used for the complete region features. Refinement is carried out in both channel and spatial dimensions to obtain key features. Key features It can be represented as: ,in, This represents channel attention mapping. Represents spatial attention mapping, This indicates element-wise multiplication. This represents the channel weight characteristics.

[0077] S2031, the channel attention mapping The specific steps for obtaining the data are as follows: targeting global features. Global average pooling and global max pooling are used to extract smooth features respectively. and salient features For smooth features and salient features The channel attention map is obtained by element-wise addition using a multilayer perceptron (MLP), and the calculation formula is as follows: ,in, and For the weights of the MLP, Represents the Sigmoid activation function. This indicates that global average pooling is performed on the features of the entire region. This indicates that global max pooling is performed on the features of the complete region.

[0078] S2032, the spatial attention mapping The specific steps for obtaining the data are as follows: For channel weight features... Perform global average pooling and global max pooling to obtain ;

[0079] In two sets of feature vectors After splicing, the scale used is The convolutional layer extracts spatial features, represented as:

[0080] ;

[0081] use The convolutional layer extracts spatial features, represented as:

[0082] ;

[0083] use Convolutional layers extract spatial features, represented as:

[0084] ;

[0085] right and Spatial features extracted from three different sizes are fused to obtain fused features; , This represents the splicing operation; the spatial attention mapping is represented as: ,in This represents the convolution operation. This indicates the splicing of channel dimensions.

[0086] S204, the horizontal region features, vertical region features, and key features are added together to obtain the extracted features. Layer feature extraction Represented as: .

[0087] S205, size matching and convolution operations are performed on the extracted features to obtain the final spatial features; specifically, to ensure that the final spatial features contain information at different scales, the extracted features obtained after the second layer are upsampled to match the size of the extracted features from the first layer; features of different scales with the same size are concatenated and convolution operations are performed to obtain the final spatial features. : in, This represents the first layer of feature extraction. to Indicates from the second layer to the first layer. Features are extracted after the layer is upsampled.

[0088] S3, extract frequency domain features from pseudo-labeled samples to obtain frequency domain features;

[0089] In this embodiment of the application, the frequency domain feature extraction of the pseudo-label samples specifically involves: obtaining frequency features of the pseudo-label samples based on a deep SWiGLU architecture, such as... Figure 5 As shown in the figure, the detailed architecture of the frequency domain branch integrating the dynamic weight module and the deep SWiGLU layer is illustrated. Specifically, the frequency components of the pseudo-label samples are used to generate a first weight matrix and a second weight matrix using Conv. After activating the second weight matrix with the Swish activation function, a weighted operation is performed with the first weight matrix to obtain the frequency domain features. The frequency domain features can be expressed as: ,in, and These represent the first and second weight matrices, respectively. This represents the frequency components of the pseudo-labeled samples. and This is a bias term.

[0090] S4, Based on the dual-domain collaborative fusion module, spatial features and frequency domain features are subjected to spatial-frequency dual-domain collaborative operation to obtain dual-domain fused features;

[0091] It should be noted that in SAR image change detection tasks, single feature domain extraction methods often struggle to simultaneously suppress noise and preserve details of changed regions. For example, spatial features can capture local edges, textures, and structural information, and are particularly sensitive to the representation of change boundaries, but their anti-interference ability is limited when faced with speckle noise, which is prevalent in SAR images, and they are prone to false detections. Similarly, frequency domain features are mainly used to characterize global structural information and periodic pattern changes in images, and are advantageous in suppressing noise and maintaining overall consistency, but they are insufficient in capturing minute changes and local details, resulting in insufficient perception of subtle changes. To address this issue, this application proposes a dual-domain collaborative fusion (DCF) module.

[0092] like Figure 6 As shown in the figure, the detailed process of dual-domain collaborative feature fusion performed by the dual-domain collaborative fusion module includes:

[0093] S401 performs spatial domain modulation on the frequency domain features to obtain the modulated frequency domain features. Specifically, the DCF module modulates the spatial features... Channel descriptors are compressed using global average pooling. Then, a bottlenecked MLP is used to model the channel dependencies between frequency components through nonlinear mapping. Finally, a dynamic frequency weight mask is generated using the sigmoid function. The dynamic frequency weight mask It can be represented as ,in, This represents a multilayer perceptron. This indicates a global average pooling operation. This represents the Sigmoid activation function;

[0094] Dynamic frequency weight mask Frequency domain characteristics Performing element-wise dot multiplication enables dynamic modulation of the spectral response, yielding the modulated frequency domain characteristics, which can be expressed as: ,in, This represents the frequency domain characteristics after modulation. Represents frequency domain characteristics, This represents a dynamic frequency weighted mask.

[0095] S402, frequency domain modulation of spatial features, specifically, the modulated frequency domain features... use The spatial dimension is reconstructed, then passed through convolutional layers and a sigmoid function to generate a spatial attention map; spatial features are then analyzed based on the spatial attention map. Secondary calibration and enhancement are performed to obtain the modulated spatial features: ,in, Frequency domain features are reconstructed back to the spatial dimension, and convolutional layers are used to perceive the mapping weights of global spectral information in spatial location.

[0096] S403 reconstructs and fuses the modulated spatial and frequency domain features to obtain dual-domain fused features. Specifically, the modulated frequency domain features... The reconstructed frequency domain features are obtained by using inverse Fourier transform; the modulated spatial features and the reconstructed frequency domain features are then fused by addition to obtain dual-domain fused features, which can be expressed as: ;in, This indicates the dual-domain fusion feature.

[0097] It should be noted that while preserving local edge details in the spatial domain, the dual-domain fusion features achieve complementary representation of features in different dimensions by coupling global structural information of frequency domain modulation. With the help of the spatial-frequency domain collaborative mechanism, it can suppress speckle noise interference in the background and enhance the distinguishability and stability of the features.

[0098] S5 performs SAR image change detection based on dual-domain fusion features. Specifically, the fusion features are input into the decoder, and the change image is gradually recovered through the upsampling layer structure.

[0099] To better demonstrate the effect of this application compared with the prior art, the following uses 6 real datasets to verify the effect of the multi-scale attention spatial-frequency dual-domain cooperative SAR image change detection method;

[0100] As shown in Table 1, the following data are presented: The study area covers Sendai City, Japan, and Xiong'an New Area, China. The former mainly presents abrupt changes in land cover caused by a sudden tsunami disaster, while the latter mainly includes seasonal land cover disturbances. Table 1 details the characteristics of these datasets, revealing their distinct features: including different spatial resolutions of 9 meters and 20 meters, and image sizes ranging from 200 to 1000 pixels; furthermore, these datasets cover various types of change patterns. The scenes detected in these datasets are also diverse (e.g., water areas and densely built-up areas); this diversity of scenes further validates the technical effects produced by this application.

[0101]

[0102] Table 1

[0103] like Figure 8-11As shown, to more intuitively evaluate the detection performance of this application, the following experiments were conducted, and visualizations of the detection results for each dataset are provided. In these images, black represents unchanged pixels, white represents changed pixels, blue represents false positives, and red represents missed detections. To objectively evaluate the detection performance of the proposed model, five evaluation metrics were used for quantitative assessment: false positives (FP), missed detections (FN), total error (OE), correct classification rate (PCC), and Kappa coefficient (KC).

[0104] Experimental setup:

[0105] First, the image patch size is set to 15, the number of spatial feature extraction layers is set to 3, and the kernel sizes of each layer are set to 3, 5, and 7 respectively; the frequency domain part uses a three-layer SWiGLU.

[0106] In the process of spatial domain feature extraction, targeting , and Image patches centered on selected pixels are formed and stitched together to obtain an input of size (3, 15, 15). First, multi-region feature extraction with a convolution kernel size of 3 is performed to generate a first-layer feature map of size (5, 15, 15). Then, multi-region feature extraction with a convolution kernel size of 5 is applied to the first-layer features to obtain a second-layer feature map of size (5, 13, 13). Next, multi-feature extraction with a convolution kernel size of 7 is performed on the second-layer features to obtain a third-layer feature map of size (5, 9, 9). Then, the second and third-layer features are upsampled to (5, 15, 15). Subsequently, the upsampled second and third-layer features are stitched together with the first-layer features, and a convolution operation is performed to obtain a spatial feature of size (5, 15, 15). Finally, this spatial feature is reshaped to (1, 1125).

[0107] In the frequency domain feature extraction process, the input is first processed by Fast Fourier Transform (FFT), followed by bilinear interpolation to obtain features of size (3, 16, 16). The interpolated features are reshaped to (3, 256) and fed into the first SWiGLU layer, generating output features of size (3, 128). These features are then fed into the second SWiGLU layer, producing features of size (3, 64). The reshaped interpolated features are concatenated with the output of the second SWiGLU layer and used as the input to the third SWiGLU layer, thus generating frequency domain features of size (3, 160). Finally, the frequency domain features are reshaped to (1, 480) and concatenated with the spatial features, then fed into a fully connected layer to obtain the final mapping result.

[0108] In addition to the parameters mentioned above, this application also compared the detection results of the Ottawa dataset under different learning rates, as shown in Table 2. When the learning rate is set to 0.001, the PCC and KC values ​​reach their highest levels. Therefore, this application uses 0.001 as the learning rate. The Adam optimizer is used during training, and the batch size is set to 128.

[0109]

[0110] Table 2

[0111] In FCM pre-classification, the choice of threshold affects model performance. To better determine the optimal threshold, this application designed two sets of experiments to obtain the optimal threshold. ,like Figure 8 and Figure 9 As shown, Figure 8 It shows Middle threshold Impact on algorithm performance Figure 9 It shows Middle threshold Impact on algorithm performance; Figure 8 and Figure 9 In the table, (a) represents the results on dataset III, and (b) represents the results on dataset IV;

[0112] In order to select the optimal The range was set between 0.75 and 1.75; experiments were conducted on two sets of real SAR datasets, and PCC and KC were used as evaluation criteria. Figure 6 The graph shows the PCC and KC values ​​as the parameters vary between 0.75 and 1.75. As can be seen, the algorithm achieves its highest accuracy when the parameter is set to 1.25. Simultaneously, the algorithm's Kappa coefficient also reaches its maximum value, indicating that when... When the threshold is set to 1.25, the final image shows a high degree of consistency with the reference image. Therefore, the threshold for MR is... Set it to 1.25.

[0113] In order to select the optimal The value is set to range from 0.6 to 1.4. Experiments were conducted on two sets of real SAR datasets, and PCC and KC were used as evaluation criteria. Figure 7 The graph shows the PCC and Kappa coefficients when the parameter varies between 0.6 and 1.4. As can be seen from the graph, for dataset IV, the values ​​are highest when the parameter is 0.6; however, this dataset was discarded to avoid overfitting. Considering all factors, the algorithm achieves the highest accuracy when the parameter is set to 1.2. Simultaneously, the Kappa coefficient for dataset IV also reaches its maximum value. Therefore, the threshold for ML is... Set to 1.2.

[0114] Furthermore, to demonstrate the effectiveness of the multi-scale attention spatial-frequency dual-domain collaborative SAR image change detection method (MADCNet) of this application, it was compared with several change detection algorithms, including RUSACD, SAFNet, DDNet, LANTNet, WBANet and DBFNet.

[0115] like Figure 10 As shown in the figure, the change detection results of (a) RUSACD, (b) SAFNet, (c) DDNet, (d) LANTNet, (e) WBANet, (f) DBFNet, and (g) MADCNet on six real SAR datasets are presented. Overall, MADCNet accurately identifies changed regions in all scenarios while effectively suppressing false positives and false negatives, demonstrating excellent robustness and sensitivity. Compared to other methods, MADCNet performs stably in areas with small changes and complex backgrounds, balancing global consistency with local detail features, especially in detecting subtle targets and complex textures. SAFNet and DDNet have advantages in detail recovery in some areas with small changes, but still exhibit some false positives and false negatives in scenarios with strong noise interference. LANTNet and WBANet are relatively robust in detecting large-area changes, but their response to small targets is slightly insufficient. RUSACD has good overall detection capabilities, but MADCNet is more outstanding in detail and robustness against complex backgrounds. In summary, the experimental results show that MADCNet exhibits superior detection accuracy and robustness in SAR change detection tasks, and can take into account both local details and global information.

[0116] like Figure 11 As shown, to further quantify and evaluate the detection performance of this application, Figure 11The quantitative analysis results of PCC and KC coefficients of MADCNet and various comparison methods on different datasets are presented. (a) to (f) correspond to the results of datasets I–VI, respectively. Based on the provided multi-method experimental results, the following analysis can be obtained: Overall, MADCNet exhibits the best performance in both PCC and KC metrics, with PCC reaching approximately 98.47%–99.42% and KC reaching approximately 85.36%–94.24%, indicating that it accurately captures changing regions while significantly reducing false positives and false negatives. In contrast, SAFNet and RUSACD have relatively lower KC metrics, at approximately 73%–78% and 78%–85%, respectively, indicating certain limitations in suppressing false positives. DDNet, LANTNet, WBANet, and DBFNet all maintain PCC above 97%–98%, but their KC performance is slightly lower than MADCNet, especially in the identification of small or complex changing regions. In summary, MADCNet achieves a balance between local details and global structure through feature extraction via spatial-frequency dual-domain collaboration and multi-scale depth adaptive convolution. This makes its discriminative and robust change detection significantly superior to other comparative methods, enabling it to more stably and accurately identify changing targets in SAR images.

[0117] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. This computer program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. The aforementioned storage medium can be a non-volatile storage medium such as a magnetic disk, optical disk, or read-only memory (ROM), or random access memory (RAM).

[0118] It should be understood that although the steps in the flowcharts of the accompanying figures are shown sequentially as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the accompanying figures may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times, and their execution order is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.

[0119] This application also discloses a multi-scale attention-based spatial-frequency dual-domain collaborative SAR image change detection system, including: a pseudo-label sample generation module, a spatial feature extraction module, a frequency domain feature extraction module, a dual-domain feature fusion module, and an image change detection module;

[0120] The pseudo-label sample generation module is used to pre-classify SAR input images using clustering methods and generate pseudo-label samples.

[0121] The spatial feature extraction module is used to extract spatial features from pseudo-label samples in multiple regions and at multiple scales to obtain spatial features.

[0122] The frequency domain feature extraction module is used to extract frequency domain features from pseudo-label samples to obtain frequency domain features;

[0123] The dual-domain feature fusion module is used to perform spatial-frequency dual-domain collaborative operations on spatial features and frequency domain features based on the dual-domain collaborative fusion module to obtain dual-domain fused features.

[0124] The image change detection module is used to detect changes in SAR images based on dual-domain fusion features.

[0125] To address the aforementioned technical problems, embodiments of this application also provide a computer device. Please refer to [link / reference needed]. Figure 12 , Figure 12 This is a basic structural block diagram of the computer device in this embodiment.

[0126] The computer device 12 includes a memory 12a, a processor 12b, and a network interface 12c that are interconnected via a system bus. It should be noted that only the computer device 12 with components 12a-12c is shown in the figure; however, it should be understood that it is not required to implement all the shown components, and more or fewer components can be implemented alternatively. Those skilled in the art will understand that the computer device described here is a device capable of automatically performing numerical calculations and / or information processing according to pre-set or stored instructions, and its hardware includes, but is not limited to, microprocessors, application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), digital signal processors (DSPs), embedded devices, etc.

[0127] The computer device can be a desktop computer, laptop, handheld computer, or cloud server, etc. The computer device can interact with the user via a keyboard, mouse, remote control, touchpad, or voice control.

[0128] The memory 12a includes at least one type of readable storage medium, including flash memory, hard disk, multimedia card, card-type memory (e.g., SD or DX memory), random access memory (RAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic memory, magnetic disk, optical disk, etc. In some embodiments, the memory 12a may be an internal storage unit of the computer device 12, such as the hard disk or memory of the computer device 12. In other embodiments, the memory 12a may also be an external storage device of the computer device 12, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc. Of course, the memory 12a may also include both internal storage units and external storage devices of the computer device 12. In this embodiment, the memory 12a is typically used to store the operating system and various application software installed on the computer device 12, such as the program code for a multi-scale attention-based spatial-frequency dual-domain cooperative SAR image change detection method. Furthermore, the memory 12a can also be used to temporarily store various types of data that have already been output or will be output.

[0129] In some embodiments, the processor 12b may be a central processing unit (CPU), controller, microcontroller, microprocessor, or other data processing chip. The processor 12b is typically used to control the overall operation of the computer device 12. In this embodiment, the processor 12b is used to run program code stored in the memory 12a or process data, for example, to run the program code of the multi-scale attention-based spatial-frequency dual-domain cooperative SAR image change detection method.

[0130] The network interface 12c may include a wireless network interface or a wired network interface, which is typically used to establish communication connections between the computer device 12 and other electronic devices.

[0131] This application also provides another embodiment, namely, a non-volatile computer-readable storage medium storing a program for a multi-scale attention-space-frequency dual-domain cooperative SAR image change detection method, wherein the multi-scale attention-space-frequency dual-domain cooperative SAR image change detection can be executed by at least one processor to enable the at least one processor to perform the steps of the multi-scale attention-space-frequency dual-domain cooperative SAR image change detection method as described above.

[0132] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0133] Obviously, the embodiments described above are only some embodiments of this application, not all embodiments. The accompanying drawings show preferred embodiments of this application, but do not limit the patent scope of this application. This application can be implemented in many different forms; rather, the purpose of providing these embodiments is to provide a more thorough and comprehensive understanding of the disclosure of this application. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing specific embodiments, or make equivalent substitutions for some of the technical features. Any equivalent structures made using the content of this application's specification and drawings, directly or indirectly applied to other related technical fields, are similarly within the scope of patent protection of this application.

Claims

1. A method for detecting changes in multi-scale attention-based spatial-frequency dual-domain collaborative SAR images, characterized by the following steps: include: Clustering methods are used to pre-classify SAR input images and generate pseudo-label samples; Spatial features are obtained by extracting spatial features from pseudo-labeled samples in multiple regions and at multiple scales; Frequency domain features are extracted from the pseudo-labeled samples to obtain frequency domain features; Based on the dual-domain collaborative fusion module, spatial and frequency domain features are subjected to spatial-frequency dual-domain collaborative operation to obtain dual-domain fused features; SAR image change detection based on dual-domain fusion features.

2. The multi-scale attention-based spatial-frequency dual-domain cooperative SAR image change detection method according to claim 1, characterized in that, The extraction of multi-region and multi-scale spatial domain features from pseudo-labeled samples specifically involves: Multi-region selection is performed on image patches in pseudo-label samples, selecting horizontal regions, vertical regions, and complete regions, and then convolutional operations are performed to obtain feature maps; Multi-scale spatial feature extraction is performed on the feature map to obtain three types of region features; A multi-scale attention mechanism is used to obtain key features from the complete region features among the three types of region features; The horizontal region features, vertical region features, and key features are added together to obtain the extracted features. The extracted features are subjected to size matching and convolution operations to obtain the final spatial features.

3. The multi-scale attention-based spatial-frequency dual-domain collaborative SAR image change detection method according to claim 2, characterized in that, The frequency domain feature extraction of the pseudo-label samples specifically involves: using the frequency components of the pseudo-label samples to generate a first weight matrix and a second weight matrix respectively using Conv; activating the second weight matrix with the Swish activation function and then performing a weighted operation with the first weight matrix to obtain the frequency domain features, which can be expressed as follows: ,in, and These represent the first and second weight matrices, respectively. This represents the frequency components of the pseudo-labeled samples. and This is a bias term.

4. The multi-scale attention-based spatial-frequency dual-domain cooperative SAR image change detection method according to claim 3, characterized in that, The spatial-frequency dual-domain collaborative operation based on the dual-domain collaborative fusion module specifically involves: Spatial domain modulation is applied to the frequency domain features to obtain the modulated frequency domain features; The spatial features are frequency-domain modulated to obtain the modulated spatial features; The modulated spatial and frequency domain features are reconstructed and fused to obtain dual-domain fused features.

5. The multi-scale attention-based spatial-frequency dual-domain cooperative SAR image change detection method according to claim 4, characterized in that, The process of spatially modulating the frequency domain features to obtain the modulated frequency domain features is as follows: Using the DCF module to extract spatial features Channel descriptors are compressed using global average pooling. Then, a bottlenecked MLP is used to model the channel dependencies between frequency components through nonlinear mapping. Finally, a dynamic frequency weight mask is generated using the sigmoid function. The dynamic frequency weight mask It can be represented as ,in, This represents a multilayer perceptron. This indicates a global average pooling operation. This represents the Sigmoid activation function; Dynamic frequency weight mask Frequency domain characteristics Performing element-wise dot multiplication enables dynamic modulation of the spectral response, yielding the modulated frequency domain characteristics, which can be expressed as: ,in, This represents the frequency domain characteristics after modulation. Represents frequency domain characteristics, This represents a dynamic frequency weighted mask.

6. The multi-scale attention-based spatial-frequency dual-domain cooperative SAR image change detection method according to claim 5, characterized in that, The spatial features are frequency-domain modulated to obtain modulated spatial features. Specifically, the modulated frequency-domain features are... use The spatial dimension is reconstructed, and then passed through convolutional layers and the Sigmoid function to generate a spatial attention map; Spatial features based on spatial attention maps Secondary calibration and enhancement are performed to obtain the modulated spatial features: ,in, Frequency domain features are reconstructed back to the spatial dimension, and convolutional layers are used to perceive the mapping weights of global spectral information in spatial location.

7. The multi-scale attention-based spatial-frequency dual-domain cooperative SAR image change detection method according to claim 6, characterized in that, The reconstruction and fusion of the modulated spatial and frequency domain features to obtain dual-domain fused features specifically involves: reconstructing and fusing the modulated frequency domain features... The reconstructed frequency domain features are obtained by using inverse Fourier transform; The modulated spatial features and the reconstructed frequency domain features are additively fused to obtain dual-domain fused features, which can be expressed as follows: ;in, This indicates the dual-domain fusion feature.

8. A multi-scale attention-based spatial-frequency dual-domain cooperative SAR image change detection system, comprising: The module includes a pseudo-label sample generation module, a spatial feature extraction module, a frequency domain feature extraction module, a dual-domain feature fusion module, and an image change detection module. The pseudo-label sample generation module is used to pre-classify SAR input images using clustering methods and generate pseudo-label samples. The spatial feature extraction module is used to extract spatial features from pseudo-label samples in multiple regions and at multiple scales to obtain spatial features. The frequency domain feature extraction module is used to extract frequency domain features from pseudo-label samples to obtain frequency domain features; The dual-domain feature fusion module is used to perform spatial-frequency dual-domain collaborative operations on spatial features and frequency domain features based on the dual-domain collaborative fusion module to obtain dual-domain fused features. The image change detection module is used to detect changes in SAR images based on dual-domain fusion features.

9. An electronic device, characterized in that, include: One or more processors; Memory; And one or more computer programs, wherein the one or more computer programs are stored in the memory, the one or more computer programs including instructions that, when executed by the device, cause the device to perform the steps of the multi-scale attention spatial-frequency dual-domain cooperative SAR image change detection method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when run on a computer, causes the computer to perform the steps of the multi-scale attention-based spatial-frequency dual-domain cooperative SAR image change detection method as described in any one of claims 1 to 7.

Images