SAR image change detection method, system, device and storage medium based on interactive attention and median filtering

Abstract

The present invention discloses a SAR image change detection method, system, device and storage medium based on interactive attention and median filtering, belonging to the field of SAR image processing technology. The method includes obtaining a target data set, which includes a first time phase and a second time phase; inputting the first time phase and the second time phase into a pre-trained SAR change detection network model to obtain a final change detection map; wherein the SAR change detection network model includes a filter residual convolution module, a spatial interactive attention module, a decoding module and a connected domain judgment module. The present invention can reduce the interference of salt and pepper noise and improve the quality of SAR images through the filter residual convolution module; through the spatial interactive attention module, the first time phase and the second time phase are allowed to interact with each other in global spatial information, accurately focusing on the real change area; and through the connected domain judgment module, the accuracy and consistency of the detection results are further guaranteed.

Description

Technical Field

[0001] The present invention belongs to the technical field of SAR image processing, and in particular relates to a SAR image change detection method, system, device and storage medium based on interactive attention and median filtering. Background Art

[0002] Synthetic Aperture Radar (SAR) is an important active remote sensing technology widely used in fields such as Earth observation, environmental monitoring, military reconnaissance, and disaster assessment. Compared to optical remote sensing, SAR imaging relies on the emission and reception of microwave echoes. It can acquire high-resolution surface information in all weather conditions (including complex environments such as clouds and haze), day or night. This all-weather, all-day imaging capability gives SAR an irreplaceable advantage in specific application scenarios, making change detection in SAR imagery crucial.

[0003] Existing change detection methods include traditional methods and deep learning-based methods. Traditional methods cannot comprehensively extract and evaluate features from multiple images and rely on image quality, making them difficult to apply to real-world scenarios. Deep learning-based methods, on the other hand, learn the transformation features of multiple image data and ultimately generate a binary transformation detection map, enabling more complex transformation feature analysis.

[0004] However, in SAR image processing, noise suppression is key to improving image quality. Salt and pepper noise, a common type of noise in SAR images, significantly negatively impacts both the visual quality of SAR images and subsequent analysis tasks. However, most current deep learning methods are ineffective in processing areas contaminated by salt and pepper noise, easily leading to false and missed detections. Furthermore, due to the uncertainty of change distribution in dual-temporal images and the presence of task-irrelevant interference such as seasonal changes and building renovations, feature interaction plays a crucial role. However, most current SAR image change detection methods lack a feature interaction mechanism, which results in reduced detection accuracy. Summary of the Invention

[0005] The purpose of the present invention is to overcome the deficiencies in the prior art and provide a SAR image change detection method, system, device and storage medium based on interactive attention and median filtering, which effectively solves the problems of noise suppression and insufficient change detection accuracy in SAR image change detection.

[0006] The present invention provides the following technical solutions:

[0007] In a first aspect, a SAR image change detection method based on interactive attention and median filtering is provided, comprising: obtaining a target data set, which includes a first time phase and a second time phase; inputting the first time phase and the second time phase into a pre-trained SAR change detection network model to obtain a final change detection map; wherein the SAR change detection network model includes a filter residual convolution module, a spatial interactive attention module, a decoding module and a connected domain judgment module; the filter residual convolution module is used to hierarchically extract features in the first time phase and the second time phase to obtain output feature map pairs of each layer; multiple spatial interactive attention modules are used to respectively fuse the output feature map pairs of each layer to obtain corresponding interactive fused feature maps of each layer; the decoding module is used to fuse the interactive fused feature maps of each layer, and decode the fused feature maps to obtain a coarse change detection map; the connected domain judgment module is used to remove noise in the coarse change detection map to obtain a final change detection map.

[0008] As an optional technical solution of the present invention, the filter residual convolution module includes multiple filter residual convolution layers, and the output features of each filter residual convolution layer serve as input features of the next filter residual convolution layer;

[0009] No. The output feature map pair of the filter residual convolution layer is expressed as ,in, represents the output feature map of the first phase, Represents the output feature map of the second phase.

[0010] As an optional technical solution of the present invention, the filtered residual convolution layer includes a residual convolution block, a double downsampling and a median filter block;

[0011] The residual convolution block is used to extract the features of the input features through a convolution operation to obtain the first feature; the convolution operation is expressed as:

[0012] ;

[0013] in, Indicates that the output feature is at pixel location , No. The pixel value of each channel, Indicates that the input feature is at pixel location , No. The pixel value of each channel, Indicates that the convolution kernel is in the output channel For input channels The weight of Indicates the number of channels of the input feature map, represents the convolution kernel OK, represents the convolution kernel List, ;

[0014] The double downsampling is used to reduce the spatial dimension of the first feature to obtain the second feature;

[0015] The median filter block is used to process the noise levels of different regions of the second feature to obtain the output feature map pair of the current filter residual convolution layer, including:

[0016] For the pixel position of the second feature , define its local window as , expressed as:

[0017] ;

[0018] in, represents the minimum local window size, Indicates the maximum local window size, represents the adjustment factor, represents the noise level threshold, The noise level of the second feature local area is expressed as:

[0019] ;

[0020] in, Indicates that the second feature is at pixel position The pixel value of Represents the pixel mean within the local window, expressed as:

[0021] .

[0022] As an optional technical solution of the present invention, The output feature map of the filter residual convolution layer is Input the convolutional layer of the spatial interaction attention module respectively, and obtain the query matrix 、 , key matrix 、 Sum Matrix 、 , expressed as:

[0023] , ;

[0024] , ;

[0025] , ;

[0026] Among them, Conv represents the convolution operation, Indicates the total number of pixels, Indicates the number of channels of the input feature map;

[0027] Will and Superposition is performed to generate a fused value matrix , expressed as:

[0028] ;

[0029] Among them, concat represents the superposition operation;

[0030] For the bond matrix 、 Perform softmax operation in the row direction to obtain weight distribution 、 , expressed as:

[0031] , ;

[0032] Among them, softmax represents the softmax operation, dim represents the dimension of the softmax operation;

[0033] Generate attention score matrix 、 , expressed as:

[0034] , ;

[0035] Where T represents transpose;

[0036] For the query matrix 、 Perform softmax operation in the column direction and then compare it with the corresponding attention score matrix and Perform matrix multiplication to obtain Interactive fusion feature map of the layer , expressed as:

[0037] , ;

[0038] ;

[0039] in, Represents the adaptive adjustment parameters, and concat represents the superposition operation.

[0040] As an optional technical solution of the present invention, the decoding module includes an upsampling block and a residual convolution block; the decoding module is used to fuse the interactive fusion feature maps of each layer, including:

[0041] The last layer of interactive fusion feature map Upsampling matches the previous layer interactive fusion feature map The size of the last layer of interactive fusion feature map after matching Input the residual convolution block for feature extraction, and superimpose the extracted features on the previous layer of interactive fusion feature map , get the first fusion decoding feature ;

[0042] The first fusion decoding feature Upsampling matches the previous layer interactive fusion feature map The size of the first fusion decoded features will be matched Input the residual convolution block for feature extraction, and superimpose the extracted features on the previous layer of interactive fusion feature map , get the second fusion decoding feature ;

[0043] Iterate until the fused decoding features are obtained ;

[0044] The fused decoding features After upsampling, the residual convolution block is input to obtain the predicted probability map;

[0045] An argmax operation is performed on the predicted probability map to obtain a coarse change detection map.

[0046] As an optional technical solution of the present invention, the connected domain determination module removes noise in the coarse change detection graph, including:

[0047] Pixel The surrounding 8 adjacent pixels are defined as the neighborhood , expressed as:

[0048] ;

[0049] For each pixel at the edge of the connected domain in the coarse change detection map , if its neighborhood If there is only one adjacent pixel in the connected domain, and the pixels above, below, left, right, or diagonally adjacent to the adjacent pixel are not included in the connected domain, the adjacent pixel is determined to be noise and is set to 0 to eliminate the noise;

[0050] For the binary image after noise removal Mark the connected domains and assign a unique label to each connected domain , connected domain Area Expressed as:

[0051] ;

[0052] in, Represents pixels The connected domain to which it belongs, represents the label matrix, Represents an indicator function, which takes the value 1 when the condition is true and 0 when it is false;

[0053] Get the binary change detection map , expressed as:

[0054] ;

[0055] in, Indicates the threshold value set for the area of ​​the connected domain;

[0056] The binary change detection map The final change detection map is obtained by splicing.

[0057] In a second aspect, a SAR image change detection system based on interactive attention and median filtering is provided, comprising: a data acquisition module for acquiring a target data set, which includes a first time phase and a second time phase;

[0058] A detection module, configured to input the first time phase and the second time phase into a pre-trained SAR change detection network model to obtain a final change detection map;

[0059] The SAR change detection network model includes a filter residual convolution module, a spatial interaction attention module, a decoding module and a connected domain judgment module;

[0060] The filter residual convolution module is used to extract features in the first time phase and the second time phase in layers to obtain output feature map pairs of each layer;

[0061] The plurality of spatial interaction attention modules are used to fuse the output feature map pairs of each layer respectively to obtain the corresponding interactive fusion feature maps of each layer;

[0062] The decoding module is used to fuse the interactive fusion feature maps of each layer and decode the fused feature maps to obtain a coarse change detection map;

[0063] The connected domain judgment module is used to remove noise in the rough change detection map to obtain the final change detection map.

[0064] In a third aspect, a SAR image change detection device based on interactive attention and median filtering is provided, comprising a processor and a storage medium; the storage medium is used to store instructions; and the processor is used to operate according to the instructions to execute the steps of the SAR image change detection method based on interactive attention and median filtering described in the first aspect.

[0065] In a fourth aspect, a computer-readable storage medium is provided, on which a computer program is stored. When the program is executed by a processor, the steps of the SAR image change detection method based on interactive attention and median filtering described in the first aspect are implemented.

[0066] Compared with the prior art, the present invention has the following beneficial effects:

[0067] The present invention provides a SAR image change detection method based on interactive attention and median filtering, which can reduce the interference of salt and pepper noise and improve the quality of SAR images through the filter residual convolution module; the spatial interactive attention module enables the first phase and the second phase to interact with global spatial information, accurately focusing on the real change area and improving the detection accuracy; the connected domain judgment module further ensures the accuracy and consistency of the detection results. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1 is a flow chart of a SAR image change detection method based on interactive attention and median filtering in an embodiment of the present invention;

[0069] Figure 2 is a schematic diagram of a filter residual convolution module in an embodiment of the present invention;

[0070] Figure 3 is a schematic diagram of a residual convolution block in an embodiment of the present invention;

[0071] Figure 4 is a schematic diagram of a spatial interactive attention module in an embodiment of the present invention;

[0072] Figure 5 1 is a schematic diagram of a decision of a connected domain decision module in an embodiment of the present invention;

[0073] Figure 6 It is the final change detection graph of the SAR dual-phase graph in the embodiment of the present invention. DETAILED DESCRIPTION

[0074] The present invention will be further described below in conjunction with the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solutions of the present invention and are not intended to limit the scope of protection of the present invention.

[0075] Example 1

[0076] This embodiment provides a SAR image change detection method based on interactive attention and median filtering. Specifically, it includes the following steps:

[0077] Step 1: Obtain a target data set, which includes a first time phase and a second time phase.

[0078] Step 2: Input the first time phase and the second time phase into a pre-trained SAR change detection network model to obtain a final change detection map.

[0079] like Figure 1 As shown, the first phase img1 and the second phase img2 form an image pair As the input of the SAR change detection network model.

[0080] The SAR change detection network model includes a filter residual convolution module, a spatial interaction attention module, a decoding module and a connected domain judgment module. After inputting the SAR change detection network model, the final change detection map is obtained after being processed by each module in turn. The specific processing process is as follows:

[0081] Step (1), image pair The filter residual convolution module is input, and the filter residual convolution module is used to extract features in the first time phase and the second time phase in layers to obtain output feature map pairs of each layer.

[0082] Specifically, if Figure 2 As shown in FIG, the filter residual convolution module includes multiple filter residual convolution layers, and the output features of each filter residual convolution layer are used as the input features of the next filter residual convolution layer. Input two weight-shared filter residual convolution modules, and the first time phase img1 and the second time phase img2 are respectively used as input features of the first filter residual convolution layer. The filter residual convolution module includes four filter residual convolution layers.

[0083] like Figure 2 As shown in FIG, the filter residual convolution layer includes a residual convolution block, a double downsampling block, and a median filter block. The specific operations are as follows:

[0084] 1.1, the residual convolution block is used to extract the features of the input feature map through convolution operation to obtain the first feature. Figure 3 As shown in the figure, the residual convolution block first extracts features through a convolution Conv, then further extracts features through normalization-nonlinear activation-convolution-normalization operations, and adds the features. The sizes of both convolution kernels are set to 3*3, and the first convolution operation expands the number of channels to twice the original number. For a certain output pixel position , the convolution operation is expressed as:

[0085] ;

[0086] in, Indicates that the output feature is at pixel location , No. The pixel value of each channel, Indicates that the input feature is at pixel location , No. The pixel value of each channel, Indicates that the convolution kernel is in the output channel For input channels The weight of Indicates the number of channels of the input feature map, represents the convolution kernel OK, represents the convolution kernel List, .

[0087] 1.2. The double downsampling is used to reduce the spatial dimension of the first feature to obtain the second feature. Specifically, after passing through the residual convolution block, the first feature is subjected to a double downsampling operation to reduce the spatial resolution while retaining important feature information.

[0088] 1.3. The median filter block is used to process the noise level of different regions of the second feature, obtaining the output feature map pair of the current filter residual convolution layer. Specifically, the second feature is input into the median filter block, and the noise level of different regions is processed by dynamically adjusting the window size. The median filter block of each layer effectively eliminates noise in the SAR image. Specifically, it includes:

[0089] For the pixel position of the second feature , define its local window as , expressed as:

[0090] ;

[0091] in, Indicates the minimum local window size, set to 3×3. Indicates the maximum local window size, which is set according to the specific SAR image change detection requirements, such as 7×7 or 9×9. Represents an adjustment factor that controls the sensitivity of the window size to changes in the noise level. represents the noise level threshold, Indicates the noise level of the second feature local area. First, select a 3×3 size and then iterate step by step. It is used to measure the noise level and is expressed as:

[0092] ;

[0093] in, Indicates that the second feature is at pixel position The pixel value of Represents the pixel mean within the local window, expressed as:

[0094] .

[0095] Furthermore, after each filter residual convolution layer undergoes operations 1.1-1.3, an output feature pair is obtained. The output feature map pair of the filter residual convolution layer is expressed as ,in, represents the output feature map of the first phase, In this embodiment, the output feature map of the last three layers is extracted. , , .

[0096] Step (2) is to extract the last three layers of feature maps from step (1) , , , respectively input into the spatial interactive attention module, strengthen the attention weight of each SAR image on another SAR image, improve the attention level of the change area, and generate interactive fusion feature maps respectively , , .

[0097] Among them, the multiple spatial interaction attention modules are used to fuse the output feature map pairs of each layer respectively to obtain the corresponding interactive fusion feature maps of each layer. In this embodiment, there are 3 spatial interaction attention modules. Figure 4 As shown, the specific process is as follows:

[0098] 2.1, Chapter The output feature map of the filter residual convolution layer is Input the convolutional layer of the spatial interaction attention module respectively, and obtain the query matrix 、 , key matrix 、 Sum Matrix 、 , expressed as:

[0099] , ;

[0100] , ;

[0101] , ;

[0102] Among them, Conv represents the convolution operation, Indicates the total number of pixels, Indicates the number of channels of the input feature map.

[0103] 2.2、 and Superposition is performed to generate a fused value matrix , expressed as:

[0104] ;

[0105] Among them, concat represents the superposition operation.

[0106] 2.3. For the bond matrix 、 Perform softmax operation in the row direction to obtain weight distribution 、 , expressed as:

[0107] , ;

[0108] Among them, softmax represents the softmax operation, dim represents the dimension of the softmax operation,

[0109] 2.4. Generate attention score matrix 、 , expressed as:

[0110] , ;

[0111] Where T represents transpose.

[0112] 2.5. For the query matrix 、 Perform softmax operation in the column direction and then compare it with the corresponding attention score matrix and Perform matrix multiplication to obtain Interactive fusion feature map of the layer , expressed as:

[0113] , ;

[0114] ;

[0115] in, Represents the adaptive adjustment parameters, and concat represents the superposition operation.

[0116] Step (3) Interactive fusion feature maps of different sizes 、 、 After being upsampled and stacked step by step by the decoding module, the final change detection map is obtained through the connected domain judgment module after decoding.

[0117] The decoding module is used to fuse the interactive fusion feature maps of each layer and decode the fused feature maps to obtain a coarse change detection map. The connected domain judgment module is used to remove noise from the coarse change detection map to obtain the final change detection map. The specific process is as follows:

[0118] 3.1. The decoding module includes an upsampling block and a residual convolution block.

[0119] The last layer of interactive fusion feature map Upsampling matches the previous layer interactive fusion feature map The size of the last layer of interactive fusion feature map after matching Input the residual convolution block for feature extraction, and superimpose the extracted features on the previous layer of interactive fusion feature map , get the first fusion decoding feature ;

[0120] The first fusion decoding feature Upsampling matches the previous layer interactive fusion feature map The size of the first fusion decoded features will be matched Input the residual convolution block for feature extraction, and superimpose the extracted features on the previous layer of interactive fusion feature map , get the fusion decoding features .

[0121] 3.2. Fusion decoding features After upsampling, the residual convolution block is input to obtain a prediction probability map. The prediction probability map is subjected to an argmax operation to obtain a coarse change detection map.

[0122] Furthermore, due to The size of the original dual-phase input SAR image is , so a four-fold upsampling operation is required to generate the predicted probability map.

[0123] 3.3. The coarse change detection map is input into the connected domain judgment module, which removes the false positives caused by noise and obtains the final change detection map.

[0124] Pixel The surrounding 8 adjacent pixels are defined as the neighborhood , expressed as:

[0125] ;

[0126] For each pixel at the edge of the connected domain in the coarse change detection map , if its neighborhood If there is only one adjacent pixel in the connected domain, and the pixels above, below, left, right, or diagonally adjacent to the adjacent pixel are not included in the connected domain, the adjacent pixel is determined to be noise and is set to 0 to eliminate the noise. Figure 5 As shown in the figure, the left side shows the two connected domains caused by noise, and the right side shows the connected domain segmentation map after removing the noise. It can be seen that the left and right positions of the noise in the two connected domains are not included in the connected domain, so it is determined to be noise. This step helps to eliminate the false connected areas caused by single-pixel connections, thereby improving the accuracy of connected domain analysis.

[0127] Furthermore, the binary image after noise removal Mark the connected domains and assign a unique label to each connected domain , connected domain Area Expressed as:

[0128] ;

[0129] in, Represents pixels The connected domain to which it belongs, represents the label matrix, Represents an indicator function, which takes a value of 1 when the condition is true and a value of 0 when it is false. Figure 5 As shown, the two connected domains after segmentation are marked as 1 and 2 respectively.

[0130] Get the binary change detection map , expressed as:

[0131] ;

[0132] in, Indicates setting threshold for connected domain area.

[0133] Furthermore, the binary change detection map The final change detection map is obtained by splicing. Figure 6 As shown in the figure, (a) is the final change detection map of this method on the public dataset Mexico City, and (b) is the final change detection map of this method on the public dataset Muragia.

[0134] Example 2

[0135] This example provides an experimental application of the method described in Example 1.

[0136] This experiment used Python 3.9, an NVIDIA RTX 3090 GPU with 24GB of video memory, and implemented the model in PyTorch. Comparative experiments included FC-EF, FC-SiamDi, FC-SiamConc, and SNUNet. Change detection results on the test set were evaluated using five metrics: precision, recall, F1 score, overall accuracy, and Kappa overall accuracy. The experimental results are shown in Table 1.

[0137] Table 1 Experimental results

[0138] Accuracy Recall F1 score Overall accuracy Kappa overall accuracy FC-EF 0.86890 0.86880 0.86890 0.98432 0.86311 FC-SiamDi 0.86980 0.90730 0.88820 0.99080 0.8810 FC-SiamConc 0.86780 0.90150 0.88430 0.99050 0.87710 SNUNet 0.92700 0.97150 0.94870 0.98800 0.94070 This method 0.97851 0.97681 0.97766 0.99706 0.97611

[0139] It can be seen from the experimental results that various indicators of the method of the present invention are higher than those of other methods.

[0140] Compared to FC-EF, whose precision, recall, and F1 scores were 0.86890, 0.86880, and 0.86890, respectively, the proposed method achieved 0.97851, 0.97681, and 0.97766, respectively, demonstrating that the proposed method can not only more accurately identify the changed areas but also effectively reduce false detections. Furthermore, compared to the relatively well-performing FC-SiamDi and FC-SiamConc, the proposed method also achieved significant improvements in recall and F1 scores. The F1 score for FC-SiamDi was 0.88820, and the F1 score for FC-SiamConc was 0.88430, while the proposed method achieved 0.97766, demonstrating its overall performance superior to these methods. Compared to SNUNet, the proposed method performed even better in precision and F1 score. Although the recall rate of SNUNet reaches 0.97150, which is close to 0.97681 of the proposed method, its precision and F1 score are 0.92700 and 0.94870 respectively, which are both lower than the proposed method. This shows that the proposed method can more comprehensively improve the overall accuracy of the change detection task.

[0141] The above results show that, under the same conditions, this method uses median filtering to suppress salt-and-pepper noise interference and spatial interaction attention to improve the capture of true change areas, and further enhances change detection results through connected component judgment. Compared with other methods, this method demonstrates its advantages in SAR change detection.

[0142] Example 3

[0143] This embodiment provides a SAR image change detection system based on interactive attention and median filtering, including:

[0144] The data acquisition module is used to acquire a target data set, which includes a first time phase and a second time phase.

[0145] The detection module is configured to input the first time phase and the second time phase into a pre-trained SAR change detection network model to obtain a final change detection map.

[0146] The SAR change detection network model includes a filter residual convolution module, a spatial interaction attention module, a decoding module and a connected domain judgment module.

[0147] The filter residual convolution module is used to extract features in the first time phase and the second time phase in layers to obtain output feature map pairs of each layer.

[0148] The plurality of spatial interaction attention modules are used to fuse the output feature map pairs of each layer respectively to obtain the corresponding interactive fusion feature maps of each layer.

[0149] The decoding module is used to fuse the interactive fusion feature maps of each layer and decode the fused feature maps to obtain a coarse change detection map.

[0150] The connected domain judgment module is used to remove noise in the rough change detection map to obtain the final change detection map.

[0151] Example 4

[0152] This embodiment provides a SAR image change detection device based on interactive attention and median filtering, including a processor and a storage medium; the storage medium is used to store instructions; the processor is used to operate according to the instructions to execute the steps of the SAR image change detection method based on interactive attention and median filtering described in Example 1.

[0153] Example 5

[0154] This embodiment provides a computer-readable storage medium having a computer program stored thereon. When the program is executed by a processor, the steps of the SAR image change detection method based on interactive attention and median filtering described in Example 1 are implemented.

[0155] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware. Furthermore, the present application may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0156] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0157] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0158] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0159] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A SAR image change detection method based on interactive attention and median filtering, characterized in that: include: Acquire a target data set, which includes a first time phase and a second time phase; Inputting the first time phase and the second time phase into a pre-trained SAR change detection network model to obtain a final change detection map; The SAR change detection network model includes a filter residual convolution module, a spatial interaction attention module, a decoding module and a connected domain judgment module; The filter residual convolution module is used to extract features in the first time phase and the second time phase in layers to obtain output feature map pairs of each layer; The plurality of spatial interaction attention modules are used to fuse the output feature map pairs of each layer respectively to obtain the corresponding interactive fusion feature maps of each layer; The decoding module is used to fuse the interactive fusion feature maps of each layer and decode the fused feature maps to obtain a coarse change detection map; The connected domain judgment module is used to remove noise in the rough change detection map to obtain the final change detection map; The filter residual convolution module includes multiple filter residual convolution layers, and the output features of each filter residual convolution layer serve as the input features of the next filter residual convolution layer; No. The output feature map pair of the filter residual convolution layer is expressed as ,in, represents the output feature map of the first phase, The output feature map representing the second phase; The filtering residual convolution layer includes a residual convolution block, a double downsampling and a median filter block; The residual convolution block is used to extract the features of the input features through a convolution operation to obtain the first feature; the convolution operation is expressed as: ； in, Indicates that the output feature is at pixel location , No. The pixel value of each channel, Indicates that the input feature is at pixel location , No. The pixel value of each channel, Indicates that the convolution kernel is in the output channel For input channels The weight of Indicates the number of channels of the input feature map, represents the convolution kernel OK, represents the convolution kernel List, ; The double downsampling is used to reduce the spatial dimension of the first feature to obtain the second feature; The median filter block is used to process the noise levels of different regions of the second feature to obtain the output feature map pair of the current filter residual convolution layer, including: For the pixel position of the second feature , define its local window as , expressed as: ； in, represents the minimum local window size, Indicates the maximum local window size, represents the adjustment factor, represents the noise level threshold, The noise level of the second feature local area is expressed as: ； in, Indicates that the second feature is at pixel position The pixel value of Represents the pixel mean within the local window, expressed as: 。 2. The SAR image change detection method based on interactive attention and median filtering according to claim 1, characterized in that: No. The output feature map of the filter residual convolution layer is Input the convolutional layer of the spatial interaction attention module respectively, and obtain the query matrix 、 , key matrix 、 Sum Matrix 、 , expressed as: ， ； ， ； ， ； Among them, Conv represents the convolution operation, Indicates the total number of pixels, Indicates the number of channels of the input feature map; Will and Superposition is performed to generate a fused value matrix , expressed as: ； Among them, concat represents the superposition operation; For the bond matrix 、 Perform softmax operation in the row direction to obtain weight distribution 、 , expressed as: ， ； Among them, softmax represents the softmax operation, dim represents the dimension of the softmax operation; Generate attention score matrix 、 , expressed as: ， ； Where T represents transpose; For the query matrix 、 Perform softmax operation in the column direction and then compare it with the corresponding attention score matrix and Perform matrix multiplication to obtain Interactive fusion feature map of the layer , expressed as: ， ； ； in, Represents the adaptive adjustment parameters, and concat represents the superposition operation.

3. The SAR image change detection method based on interactive attention and median filtering according to claim 1, characterized in that: The decoding module includes an upsampling block and a residual convolution block; the decoding module is used to fuse the interactive fusion feature maps of each layer, including: The last layer of interactive fusion feature map Upsampling matches the previous layer interactive fusion feature map The size of the last layer after matching is interactively fused with the feature map Input the residual convolution block for feature extraction, and superimpose the extracted features on the previous layer of interactive fusion feature map , get the first fusion decoding feature ; The first fusion decoding feature Upsampling matches the previous layer interactive fusion feature map The size of the first fusion decoded features will be matched Input the residual convolution block for feature extraction, and superimpose the extracted features on the previous layer of interactive fusion feature map , get the second fusion decoding feature ; Iterate until the fused decoding features are obtained ; The fused decoding features After upsampling, the residual convolution block is input to obtain the predicted probability map; An argmax operation is performed on the predicted probability map to obtain a coarse change detection map.

4. The SAR image change detection method based on interactive attention and median filtering according to claim 1, characterized in that: The connected domain judgment module removes noise in the coarse change detection graph, including: Pixel The surrounding 8 adjacent pixels are defined as the neighborhood , expressed as: ； For each pixel at the edge of the connected domain in the coarse change detection map , if its neighborhood If there is only one adjacent pixel in the connected domain, and the pixels above, below, left, right, or diagonally adjacent to the adjacent pixel are not included in the connected domain, the adjacent pixel is determined to be noise and is set to 0 to eliminate the noise; For the binary image after noise removal Mark the connected domains and assign a unique label to each connected domain , connected domain Area Expressed as: ； in, Represents pixels The connected domain to which it belongs, represents the label matrix, Represents an indicator function, which takes the value 1 when the condition is true and 0 when it is false; Get the binary change detection map , expressed as: ； in, Indicates the threshold value set for the area of ​​the connected domain; The binary change detection map The final change detection map is obtained by splicing.

5. A SAR image change detection system based on interactive attention and median filtering, characterized in that: The method for detecting changes in SAR images based on interactive attention and median filtering according to any one of claims 1 to 4 comprises: A data acquisition module, configured to acquire a target data set, which includes a first time phase and a second time phase; A detection module, configured to input the first time phase and the second time phase into a pre-trained SAR change detection network model to obtain a final change detection map; The SAR change detection network model includes a filter residual convolution module, a spatial interaction attention module, a decoding module and a connected domain judgment module; The filter residual convolution module is used to extract features in the first time phase and the second time phase in layers to obtain output feature map pairs of each layer; The plurality of spatial interaction attention modules are used to fuse the output feature map pairs of each layer respectively to obtain the corresponding interactive fusion feature maps of each layer; The decoding module is used to fuse the interactive fusion feature maps of each layer and decode the fused feature maps to obtain a coarse change detection map; The connected domain judgment module is used to remove noise in the rough change detection map to obtain the final change detection map.

6. A SAR image change detection device based on interactive attention and median filtering, characterized in that: including processors and storage media; The storage medium is used to store instructions; The processor is configured to operate according to the instructions to execute the steps of the SAR image change detection method based on interactive attention and median filtering according to any one of claims 1 to 4.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the SAR image change detection method based on interactive attention and median filtering described in any one of claims 1 to 4 are implemented.

Images