SAR image change detection method based on multi-directional attention representation enhancement

Abstract

The present invention provides a SAR image change detection method based on multi-directional attention representation enhancement. By acquiring a dual-phase SAR image and performing multi-scale convolution and multi-directional feature extraction respectively, the multi-scale convolution can obtain multi-scale feature information of the image, and the multi-directional feature extraction can screen the regional features to obtain features close to the center point, and fuse these features. The fused features pass through a self-attention module that matches the design area of the directional window, so that the regional importance of the pixel points in different directions can be obtained through model training, the representation information of the center point is enhanced, and some interference information is also eliminated. Compared with the change detection method of the prior art, the present invention enhances the representation information of the center point, eliminates invalid information, solves the noise problem of remote sensing SAR image change detection, and improves its change detection effect.

Description

Technical Field

[0001] The present invention belongs to the field of radar technology, and in particular relates to a SAR image change detection method based on multi-directional attention representation enhancement. Background Art

[0002] SAR (Synthetic Aperture Radar) is an active Earth observation system that can be installed on aircraft, satellites, spacecraft, and other platforms. It provides all-weather, 24 / 7 Earth observation and possesses a certain degree of surface penetration. Therefore, SAR systems offer unique advantages in disaster monitoring, environmental monitoring, oceanographic surveillance, resource exploration, crop yield estimation, surveying and mapping, and military applications. They can perform functions that other remote sensing methods struggle to achieve, and are therefore gaining increasing attention worldwide.

[0003] Polarimetric synthetic aperture radar (SAR) is a new imaging system that emerged with the development of synthetic aperture radar (SAR) technology. It acquires multiple polarization information about a target by analyzing the scattering characteristics of the same ground resolution unit in different polarization modes. This allows for richer ground object information than traditional single-polarization SAR, significantly enhancing the radar's ability to acquire target information. Therefore, fully or semi-automated research on these complex, large-scale polarimetric SAR data is urgently needed.

[0004] Change detection is the process of identifying differences in the state of an object or phenomenon by observing it at different times. Polarimetric SAR image change detection is an important branch of polarimetric SAR image interpretation. However, unlike natural images, SAR images are acquired differently, and most change detection methods face the challenge of speckle noise. Speckle noise degrades image quality, destroys target structure and texture information, and hinders the detailed interpretation of SAR data. Therefore, effectively addressing speckle noise in SAR images and effectively performing change detection tasks are currently a hot and challenging issue.

[0005] A new method for SAR image change detection based on convolutional neural networks (CNNs) is described in the prior art. It extracts and captures the spatial features of the original image through automatic learning, resulting in more robust results. Celik proposed a new unsupervised change detection technique for multi-temporal satellite imagery based on principal component analysis (PCA) and k-means clustering. PCA extracts orthogonal feature vectors, and each pixel is assigned to one of two clusters using the minimum Euclidean distance between the pixel's feature vector and the cluster's average feature vector, achieving change detection. Qu proposed utilizing features from the discrete cosine transform domain and integrating the reshaped DCT coefficients into the proposed model as frequency domain branches. Feature representations in both the frequency and spatial domains are used to suppress speckle noise. Liu proposed a stacked Fisher autoencoder (SFAE) for SAR change detection. The stacked autoencoder is extended to accommodate the multiplicative noise environment in SAR change detection. The features extracted by the SFAE are more discriminative than those of the original stacked autoencoder.

[0006] However, the prior art has the following defects:

[0007] (1) Due to the way SAR images are acquired, the image noise is significantly higher than that of natural images. The noise will destroy the structure and texture information of the image target, and its noise will greatly affect the effect of change detection. Although traditional filtering can reduce the overall noise level, it will cause the edge information of the object to diffuse in the normal direction, resulting in a certain degree of loss of edge information.

[0008] (2) For the change detection task, whether the category of a pixel changes is largely affected by the category change of its surrounding pixels. Especially for the target texture and edge information, the existing technology does not make good use of the information of the surrounding pixels.

[0009] Therefore, it is extremely important to overcome the influence of SAR image noise itself and retain sufficient object edge information and make effective use of it to obtain better change detection results. Summary of the Invention

[0010] To address the above-mentioned problems in the prior art, the present invention provides a SAR image change detection method based on multi-directional attention representation enhancement. The technical problem to be solved by the present invention is achieved through the following technical solutions:

[0011] The present invention provides a SAR image change detection method based on multi-directional attention representation enhancement, comprising:

[0012] Step 101: Acquire two SAR images taken at the same location but at different times from radar satellite data to form a dual-phase SAR image;

[0013] Step 102: Using a multi-directional characterization enhancement processing method, a directional window closest to the central pixel point is determined in the difference map of the dual-phase SAR image in the form of windows in different directions, and the image of the directional window is processed in the difference map to obtain an image after multi-directional characterization enhancement processing;

[0014] Step 103: Construct a change detection network based on self-attention and Senet;

[0015] The change detection network includes an attention module and a multi-scale feature fusion layer;

[0016] Step 104: using the dual-phase SAR image as first training data and the image after multi-directional characterization enhancement processing as second training data;

[0017] Step 105: Input the first training data and the second training data into different input ports of the change detection network respectively to perform multi-scale feature extraction, scale fusion, self-attention module learning, and deep feature processing, and train the change detection network in combination with the cost function to obtain a trained change detection network;

[0018] Step 106: Use the trained change detection network to identify the SAR image to achieve ground object change detection in different SAR images.

[0019] Optionally, step 102 includes:

[0020] Step 201: input a dual-phase SAR image;

[0021] Step 202: Slice the dual-phase SAR image pixel by pixel according to the corresponding position area to obtain each SAR slice image;

[0022] Step 203: Calculate the difference between the features of the slice images corresponding to the dual-phase regions to obtain a difference map;

[0023] Step 204: setting a plurality of direction windows with different directions and radii and the number of iterations, and selecting the target direction window closest to the center pixel in the difference map;

[0024] Step 205: keep the difference value in the target direction window unchanged, set the difference values in other areas to 0, and use the difference map after setting to 0 as the image after multi-directional representation enhancement processing.

[0025] Optionally, the difference graph in step 203 is:

[0026]

[0027] Among them, img1 and img2 represent the features of slice images in different phases, and Dif represents the difference map generated by the two phases.

[0028] Optionally, step 204 includes:

[0029] Step 204a: Set eight directional windows of different sizes, where the directional window set is: S = {L, R, U, D, NW, NE, SW, SE}, and input the difference map;

[0030] Step 204b: Calculate the pixel averages of the difference map in the eight direction windows;

[0031] Step 204c: Calculate the similarity between the pixel average value of each direction window and the central pixel, and select the most similar direction window as the target direction window based on the difference value.

[0032] Optional,

[0033] The pixel average value of the n-th direction window is:

[0034]

[0035] Among them, ω ij is the weight of pixel j in the neighborhood of center pixel i, q j is the pixel value of pixel j in the neighborhood, is the nth neighborhood of the center pixel i, and j is the jth pixel in the neighborhood;

[0036] The similarity values are:

[0037]

[0038] Among them, Im represents the nearest direction window index, q i is the pixel value of the center pixel i, n is the direction window number, and S is the direction window set.

[0039] Optionally, step 103 includes:

[0040] 301: Obtain sequentially connected convolutional layers of different scales and a Senet backbone network;

[0041] 302: Add a feature fusion layer and a self-attention module between the convolutional layer and the Senet backbone network in sequence to obtain a change detection network based on self-attention and Senet.

[0042] Optionally, step 104 includes:

[0043] Step 401: normalize the dual-phase SAR image and the image after multi-directional characterization enhancement processing;

[0044] Step 402: using the normalized dual-phase SAR image as the first training data, and using the normalized image after multi-directional characterization enhancement processing as the second training data.

[0045] Optionally, step 105 includes:

[0046] Step 501: input the first training data into the first convolutional layer of the change detection network, and convolve through multiple convolutional layers to obtain multi-scale information with different receptive fields;

[0047] Step 502: Input the second training data into the feature fusion layer of the change detection network to perform alignment and addition fusion with the multi-scale information;

[0048] Step 503: Send the fused features to the self-attention module, so that the self-attention module sends the fused features to the matching blocks for self-attention learning;

[0049] Among them, the size of the block corresponds to the radius of the direction window;

[0050] Step 504: Send the self-attention learning results to the backbone network to learn deep features, and obtain the classification probability from the linear layer, and select the one with the highest classification probability as the prediction result;

[0051] Step 505: Calculate the cost function based on the original SAR image and the prediction result;

[0052] Step 506: Using the gradient descent method in batches, solve the cost function minimization problem to obtain a trained change detection network.

[0053] Optionally, step 503 includes:

[0054] Step 601: Align each feature layer in the self-attention module with the feature size position of the four direction windows with smaller radius, so as to divide the feature layer of the self-attention module into 4 patches;

[0055] Step 602: Flatten each patch and pass it through a linear layer, and feed the fused features into the patch to triple the number of original features fed into the patch, and divide it into three equal parts, namely q, k, and v.

[0056] Step 603: Measure the similarity of q, k, v of each patch with the q, k, v of other patches, and then perform scaling, normalization, and weighting to effectively learn the importance of each patch and its relationship with other patches;

[0057] Step 604: Restore the mapping output obtained by self-attention learning to the original patch size to complete the self-attention learning.

[0058] Beneficial effects of the present invention:

[0059] 1. The present invention provides a SAR image change detection method based on multi-directional attention representation enhancement. By acquiring dual-phase SAR images and performing multi-scale convolution and directional window feature extraction respectively, the multi-scale convolution can obtain multi-scale feature information of the image, and the directional window feature extraction can screen the regional features to obtain features close to the center point, and fuse these features. The fused features pass through a self-attention module that matches the directional window design area, so that the regional importance of the pixel points in different directions can be obtained through model training, and the representation information of the center point is enhanced and some interference information is eliminated. Compared with the change detection method of the prior art, the present invention enhances the representation information of the center point, eliminates invalid information, solves the noise problem of remote sensing SAR image change detection, and improves its change detection effect.

[0060] 2. The present invention provides a SAR image change detection method based on multi-directional attention representation enhancement. By constructing a self-attention mechanism module of corresponding patch size corresponding to the window area set for multi-directional enhancement, the model learns the attention of the corresponding area and avoids other irrelevant pixels affecting the change detection effect.

[0061] 3. The present invention provides a SAR image change detection method based on multi-directional attention representation enhancement. The detection model is based on Senet, adding a directional window module and using the self-attention mechanism corresponding to the directional window to realize a pixel-by-pixel change detection network. Compared with the full convolutional network, it pays more attention to feature details and has a better detection effect on object edge features. Although it will cause more noise, this problem is well solved through the directional window and self-attention.

[0062] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 This is a flow chart of a SAR image change detection method based on multi-directional attention representation enhancement provided by an embodiment of the present invention;

[0064] Figure 2 1 is a schematic diagram of the process of the multi-directional characterization enhancement processing method provided by an embodiment of the present invention;

[0065] Figure 3 This is a multi-directional characterization enhancement flow chart provided by an embodiment of the present invention;

[0066] Figure 4 This is a diagram of a multi-directional characterization enhancement framework provided by an embodiment of the present invention;

[0067] Figure 5 is a framework diagram of a change detection network provided by an embodiment of the present invention;

[0068] Figure 6 is a flow chart of change detection network training provided by an embodiment of the present invention;

[0069] Figure 7 This is a framework diagram of the self-attention mechanism provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0070] The present invention will be further described in detail below with reference to specific examples, but the embodiments of the present invention are not limited thereto.

[0071] refer to Figure 1 The present invention provides a SAR image change detection method based on multi-directional attention representation enhancement, which includes:

[0072] Step 101: Acquire two SAR images taken at the same location but at different times from radar satellite data to form a dual-phase SAR image;

[0073] Step 102: Using a multi-directional representation enhancement (MDRE) method, a directional window closest to the central pixel is determined in the difference map of the dual-phase SAR image using windows of different directions, and the image of the directional window is processed in the difference map to obtain an image after multi-directional representation enhancement.

[0074] It is worth noting that: the present invention adopts a multi-directional characterization enhancement mechanism, and selects adjacent similar areas as feature enhancement areas in change detection. This invention can effectively overcome the characteristic of filtering noise reduction in SAR image change detection that does not preserve edges. While reducing the overall noise of SAR image features, it can effectively retain the edge information of objects, so that the edge texture information is well preserved, solving Problem 1 of the prior art.

[0075] Step 103: Construct a change detection network based on self-attention and Senet;

[0076] The change detection network includes a self-attention module and a multi-scale feature fusion layer;

[0077] Step 104: using the dual-phase SAR image as first training data and the image after multi-directional characterization enhancement processing as second training data;

[0078] Step 105: Input the first training data and the second training data into different input ports of the change detection network respectively to perform multi-scale feature extraction, scale fusion, self-attention module learning, and deep feature processing, and train the change detection network in combination with the cost function to obtain a trained change detection network;

[0079] Step 106: Use the trained change detection network to identify the SAR image to achieve ground object change detection in different SAR images.

[0080] It is worth noting that the present invention constructs a change detection network based on multi-directional attention representation enhancement. The network is a pixel-by-pixel multi-scale change detection network that focuses on the information of each pixel and prevents the loss of SAR image features due to up- and down-sampling. Multi-scale convolution is used for feature extraction, and a self-attention mechanism is added to the change detection network, so that the network pays more attention to the area obtained by directional window enhancement, so that the change detection results can better match the real objects.

[0081] The present invention provides a SAR image change detection method based on multi-directional attention representation enhancement. By acquiring a dual-phase SAR image and performing multi-scale convolution and directional window feature extraction respectively, the multi-scale convolution can obtain multi-scale feature information of the image, and the directional window feature extraction can screen the regional features to obtain features close to the center point, and fuse these features. The fused features pass through a self-attention module that matches the directional window design area, so that the regional importance of the pixel points in different directions can be obtained through model training, the representation information of the center point is enhanced, and some interference information is also eliminated. Compared with the change detection method of the prior art, the present invention enhances the representation information of the center point, eliminates invalid information, solves the noise problem of remote sensing SAR image change detection, and improves its change detection effect.

[0082] As an optional embodiment of the present invention, refer to Figure 2 as well as Figure 3 , step S102 includes:

[0083] Step 201: input a dual-phase SAR image;

[0084] Step 202: Slice the dual-phase SAR image pixel by pixel according to the corresponding position area to obtain each SAR slice image;

[0085] Step 203: Calculate the difference between the features of the slice images corresponding to the dual-phase regions to obtain a difference map;

[0086] Among them, the difference map is:

[0087]

[0088] Among them, img1 and img2 represent the features of slice images in different phases, and Dif represents the difference map generated by the two phases.

[0089] Step 204: setting a plurality of direction windows with different directions and radii and the number of iterations, and selecting the target direction window closest to the center pixel in the difference map;

[0090] refer to Figure 4 As an optional embodiment of the present invention, step 204 includes:

[0091] Step 204a: Set eight directional windows of different sizes, where the window set is: S = {L, R, U, D, NW, NE, SW, SE}, and input the difference map;

[0092] Step 204b: Calculate the pixel averages of the difference map in the eight direction windows;

[0093] Step 204c: Calculate the similarity between the pixel average value of each direction window and the central pixel, and select the most similar direction window as the target direction window based on the difference value.

[0094] The pixel average value of the n-th direction window is:

[0095]

[0096] Among them, ω ij is the weight of pixel j in the neighborhood of center pixel i, q j is the pixel value of pixel j in the neighborhood, is the nth neighborhood of the center pixel i, and j is the jth pixel in the neighborhood;

[0097] The similarity values are:

[0098]

[0099] Among them, I m Indicates the nearest direction window index, q i is the pixel value of the center pixel i, n is the direction window number, and S is the direction window set.

[0100] It is worth noting that the absolute value obtained by subtracting the result from the original pixel value represents the degree of similarity, and the area in S that is most similar to the original pixel value is selected.

[0101] Step 205: keep the difference value in the target direction window unchanged, set the difference values in other areas to 0, and use the difference map after setting to 0 as the image after multi-directional representation enhancement processing.

[0102] As an optional embodiment of the present invention, refer to Figure 2 , step 104 includes:

[0103] Step 401: normalize the dual-phase SAR image and the image after multi-directional characterization enhancement processing;

[0104] Step 402: using the normalized dual-phase SAR image as first training data, and using the enhanced and normalized image as second training data.

[0105] refer to Figure 5 As an optional embodiment of the present invention, step 103 includes:

[0106] 301: Obtain sequentially connected convolutional layers of different scales and a Senet backbone network;

[0107] 302: Add a feature fusion layer and a self-attention module between the convolutional layer and the Senet backbone network in sequence to obtain a change detection network based on self-attention and Senet.

[0108] The present invention uses the Senet backbone network as part of the detection network, inputs the feature layer after the self-attention mechanism into the Senet backbone network, performs deep feature processing, and combines the label with the cost function to perform model training, and finally obtains the change detection result.

[0109] refer to Figure 6 As an optional embodiment of the present invention, step 105 includes:

[0110] Step 501: Input the first training data into the first convolution layer of the change detection network, and convolve it through multiple convolution layers to obtain multi-scale information with different receptive fields;

[0111] Step 502: Input the second training data into the feature fusion layer of the change detection network to perform alignment and addition fusion with the multi-scale information;

[0112] Step 503: Send the fused features to the attention module, so that the attention module sends the fused features to the matching blocks for self-attention learning;

[0113] Among them, the size of the block corresponds to the radius of the direction window;

[0114] Step 504: Send the self-attention learning results to the backbone network to learn deep features, and obtain the classification probability from the linear layer, and select the one with the highest classification probability as the prediction result;

[0115] Step 505: Calculate the cost function based on the original SAR image and the prediction result;

[0116] Step 506: Using the gradient descent method in batches, solve the cost function minimization problem to obtain a trained change detection network.

[0117] It is worth noting that: after inputting preprocessed data, the data is divided into two parts, namely data 1 and data 2 processed by multi-directional representation enhancement; data 1 is subjected to 1*1, 3*3, and 5*5 multi-scale convolution to obtain multi-scale information of different receptive fields, and data 2 is aligned and added with the features after multi-scale convolution, enhancing the area similar to the center point and suppressing other areas; the fused features correspond to the radius size blocks set in the multi-directional representation enhancement, and then self-attention calculation (4 patches) is performed; the Senet backbone network, as part of this network, contains a residual structure and a unique SE (Squeeze-and-Excitation) structure, which can learn deeper features. The feature layer obtained in step 403 is input to the Senet backbone network and the classification probability is obtained through a linear layer; the data label is a binary label (0, 1) of a part of the SAR image area, representing the changed area and the unchanged area, respectively. The final prediction result and the label are subjected to SoftmaxCrossEntropy and the label cost function.

[0118] Softmax:

[0119]

[0120] Where |C| represents the number of categories, z j Represents the output value corresponding to the j-th category, y i Represents the probability of the i-th category.

[0121] Cross Entropy Loss:

[0122]

[0123] in, Represents the predicted probability of category i, y i Represents the label value, L CE represents the cross entropy loss.

[0124] The cost function is minimized using the gradient descent method in batches to obtain the trained change detection network.

[0125] refer to Figure 7 As an optional implementation of the present invention, step 503 includes:

[0126] Step 601: each feature layer in the attention module is matched to the feature size position of the four direction windows with smaller radius, so as to divide the feature layer of the attention module into 4 patches;

[0127] Step 602: Flatten each patch and pass it through a linear layer, and feed the fused features into the patch to triple the number of original features fed into the patch, and divide it into three equal parts, namely q, k, and v.

[0128] Step 603: Measure the similarity of q, k, v of each patch with the q, k, v of other patches, and then perform scaling, normalization, and weighting to effectively learn the importance of each patch and its relationship with other patches;

[0129] Step 604: Restore the mapping output obtained by self-attention learning to the original patch size to complete the self-attention learning.

[0130] It is worth noting that the current feature layer is aligned with the window feature size and position of the multi-directional representation enhancement. This method divides the feature layer into four patches, each corresponding to the minimum window size set by the directional window enhancement. Each patch is flattened and passed through a linear layer, tripling the original number of features. The patches are then divided equally into three parts, designated query, key, and value (abbreviated as q, k, and v). The q, k, and v of each patch are then measured for similarity with the others, scaled and normalized, and finally weighted to effectively learn the importance of each patch and its relationship to other patches. The self-attention mapping output is then restored to the original patch size.

Claims

1. A SAR image change detection method based on multi-directional attention representation enhancement, characterized in that: include: Step 101: Acquire two SAR images taken at the same location but at different times from radar satellite data to form a dual-phase SAR image; Step 102: Using a multi-directional characterization enhancement processing method, a directional window closest to the central pixel point is determined in the difference map of the dual-phase SAR image in the form of windows in different directions, and the image of the directional window is processed in the difference map to obtain an image after multi-directional characterization enhancement processing; Step 103: Construct a change detection network based on self-attention and Senet; The change detection network includes an attention module and a multi-scale feature fusion layer; Step 104: using the dual-phase SAR image as first training data and the image after multi-directional characterization enhancement processing as second training data; Step 105: Input the first training data and the second training data into different input ports of the change detection network respectively to perform multi-scale feature extraction, scale fusion, self-attention module learning, and deep feature processing, and train the change detection network in combination with the cost function to obtain a trained change detection network; the self-attention module learning includes: sending the fused features into the self-attention module, so that the self-attention module sends the fused features to the matching blocks for self-attention learning; wherein the size of the block corresponds to the radius of the direction window; Step 106: Use the trained change detection network to identify the SAR image to achieve ground object change detection in different SAR images.

2. The SAR image change detection method based on multi-directional attention representation enhancement according to claim 1 is characterized in that: The step 102 includes: Step 201: input a dual-phase SAR image; Step 202: Slice the dual-phase SAR image pixel by pixel according to the corresponding position area to obtain each SAR slice image; Step 203: Calculate the difference between the features of the slice images corresponding to the dual-phase regions to obtain a difference map; Step 204: setting a plurality of direction windows with different directions and radii and the number of iterations, and selecting the target direction window closest to the center pixel in the difference map; Step 205: keep the difference value in the target direction window unchanged, set the difference values in other areas to 0, and use the difference map after setting to 0 as the image after multi-directional representation enhancement processing.

3. The SAR image change detection method based on multi-directional attention representation enhancement according to claim 2 is characterized in that: The difference graph in step 203 is: Among them, img1 and img2 represent the features of slice images in different phases, and Dif represents the difference map generated by the two phases.

4. The SAR image change detection method based on multi-directional attention representation enhancement according to claim 2 is characterized in that: Step 204 includes: Step 204a: Set eight directional windows of different sizes, where the directional window set is: S = {L, R, U, D, NW, NE, SW, SE}, and input the difference map; Step 204b: Calculate the pixel averages of the difference map in the eight direction windows; Step 204c: Calculate the similarity between the pixel average value of each direction window and the central pixel, and select the most similar direction window as the target direction window based on the difference value.

5. The SAR image change detection method based on multi-directional attention representation enhancement according to claim 3 is characterized in that: The pixel average value of the n-th direction window is: Among them, ω ij is the weight of pixel j in the neighborhood of center pixel i, q j is the pixel value of pixel j in the neighborhood, is the nth neighborhood of the center pixel i, and j is the jth pixel in the neighborhood; The similarity values are: Among them, I m Indicates the nearest direction window index, q i is the pixel value of the center pixel i, n is the direction window number, and S is the direction window set.

6. The SAR image change detection method based on multi-directional attention representation enhancement according to claim 1 is characterized in that: Step 103 includes: 301: Obtain sequentially connected convolutional layers of different scales and a Senet backbone network; 302: Add a feature fusion layer and a self-attention module between the convolutional layer and the Senet backbone network in sequence to obtain a change detection network based on self-attention and Senet.

7. The SAR image change detection method based on multi-directional attention representation enhancement according to claim 2 is characterized in that: Step 104 includes: Step 401: normalize the dual-phase SAR image and the image after multi-directional characterization enhancement processing; Step 402: using the normalized dual-phase SAR image as the first training data, and using the normalized image after multi-directional characterization enhancement processing as the second training data.

8. The SAR image change detection method based on multi-directional attention representation enhancement according to claim 1 is characterized in that: Step 105 includes: Step 501: input the first training data into the first convolutional layer of the change detection network, and convolve through multiple convolutional layers to obtain multi-scale information with different receptive fields; Step 502: Input the second training data into the feature fusion layer of the change detection network to perform alignment and addition fusion with the multi-scale information; Step 503: Send the fused features to the self-attention module, so that the self-attention module sends the fused features to the matching blocks for self-attention learning; Among them, the size of the block corresponds to the radius of the direction window; Step 504: Send the self-attention learning results to the backbone network to learn deep features, and obtain the classification probability from the linear layer, and select the one with the highest classification probability as the prediction result; Step 505: Calculate the cost function based on the original SAR image and the prediction result; Step 506: Using the gradient descent method in batches, solve the cost function minimization problem to obtain a trained change detection network.

9. The SAR image change detection method based on multi-directional attention representation enhancement according to claim 8, characterized in that: Step 503 includes: Step 601: Align each feature layer in the self-attention module with the feature size position of the four direction windows with smaller radius, so as to divide the feature layer of the self-attention module into 4 patches; Step 602: Flatten each patch and pass it through a linear layer, and feed the fused features into the patch to triple the number of original features fed into the patch, and divide it into three equal parts, namely q, k, and v. Step 603: Measure the similarity of q, k, v of each patch with the q, k, v of other patches, and then perform scaling, normalization, and weighting to effectively learn the importance of each patch and its relationship with other patches; Step 604: Restore the mapping output obtained by self-attention learning to the original patch size to complete the self-attention learning.

Images