An Unsupervised Domain Adaptation SAR Target Recognition Method for Fusing and Aligning Visual Features and Scattering Topological Features

Abstract

The present invention discloses an unsupervised domain adaptation SAR target recognition method for fusing and aligning visual features and scattering topological features, comprising the following steps: extracting scattering topological points of source domain images and target domain images respectively based on the ASC model and constructing corresponding scattering topological graphs; extracting scattering topological points of source domain images and target domain images respectively based on the SAR-SIFT algorithm and constructing corresponding scattering topological graphs; constructing a deep recognition network Ψ; training the deep recognition network Ψ using source domain data (including images and two different scattering topological graphs) and target domain data (including images and two different scattering topological graphs); inputting the target domain data to be recognized into the trained deep recognition network to obtain target recognition results. The present invention can make full use of the scattering characteristics of SAR images, reduce the visual differences and scattering topological differences between simulated images and measured images at the same time, and align simulated images and measured images class by class in the feature space, thereby significantly improving the recognition performance.

Description

Technical Field

[0001] The present invention belongs to the technical field of radar image processing, and particularly relates to an unsupervised domain adaptation SAR target recognition method for fusing and aligning visual features and scattering topological features. Background Art

[0002] Benefiting from its unique electromagnetic scattering imaging mechanism, synthetic aperture radar (SAR) can work all day and all weather, and can image with long distance and high resolution, so it is widely used in military and civilian fields. Compared with optical images, SAR images lack color information and are vulnerable to speckle noise, so it is more difficult to interpret SAR images. Automatic target recognition (ATR) is a key topic for intelligent interpretation of SAR images and has received extensive attention from researchers.

[0003] In recent years, with the rapid development of deep learning, researchers have proposed a variety of deep learning-based algorithms. Deep learning is a data-driven algorithm, and its high performance is often based on a large amount of training data. However, the acquisition and labeling of measured SAR images are both time-consuming and laborious. Compared with optical datasets, SAR image datasets often have fewer images and smaller image sizes. The insufficiency of samples seriously hinders the development of deep learning-based SAR ATR algorithms. Data augmentation is a mainstream theory that can alleviate the insufficiency of samples. Due to the unique imaging mechanism of SAR based on the target's backscattered echo, simulated SAR images can be generated through the computer-aided design (CAD) model of the target and electromagnetic simulation software. However, simulated SAR images cannot completely replace measured SAR images. In the simulation program, the CAD model of the target, the ground background where the target is located, the electromagnetic reflection parameters of the target surface material, and the propagation process of electromagnetic waves cannot accurately conform to the real physical environment, which leads to a domain distribution difference between simulated SAR images and measured SAR images, that is, they do not meet the independent and identically distributed condition. The distribution difference further causes the classifier trained only with simulated images to be difficult to generalize well to measured images, so the accuracy of the classifier is often low. How to bridge the distribution difference between simulated images and measured images and accurately identify measured images based on simulated images has become an engineering problem with practical application value. In response to this problem, some related research has been carried out currently.

[0004] Patent CN202111188558.3 ("Zero-shot SAR target recognition method fusing CNN and image similarity") proposed a two-step recognition algorithm. This method uses a convolutional neural network trained based on simulated data as a pre-classifier and a multi-similarity fusion classifier as a fine classifier, and improves the recognition performance to a certain extent through coarse classification and fine classification.

[0005] In 2019, Wang et al. achieved cross-task and cross-domain transfer learning by combining meta-learning and domain adversarial training in their published paper "SAR target recognition based on cross-domain and cross-task transfer learning" (IEEE Access). This algorithm first pre-trains the network using a large amount of simulated data and then fine-tunes the network using measured data.

[0006] In 2021, Sun et al. introduced the domain adaptation theory in their published paper "SAR target recognition using simulated data by an ensemble multi-scale deep domain adaption recognition framework" (Proc. CIE Int. Conf. Radar). Meanwhile, a multi-scale feature extraction module was designed to extract more generalizable features, and the recognition accuracy was further improved through ensemble learning.

[0007] In 2022, Du et al. proposed an adversarial encoding network by combining image reconstruction and domain adversarial training in their published paper "Physical-Related Feature Extraction from Simulated SAR Image Based on the Adversarial Encoding Network for Data Augmentation" (IEEE Geoscience and Remote Sensing Letters). This network can transfer the physical features contained in the simulated data to the measured data as a supplement.

[0008] The problems still existing in the above methods are as follows: Among the above four methods, the fine classifier of the first method is based on the recognition idea of template matching, which requires a large number of template images and has very low computational efficiency. The latter three methods all utilize the domain adaptation theory, but essentially treat SAR images as optical images for processing, that is, input the amplitude SAR images into a deep convolutional neural network to extract features, while ignoring the different imaging characteristics between SAR images and optical images. At the same time, these three methods only align the overall distributions between the simulated images and the measured images, while ignoring the distribution situations between sub-categories, which affects the cross-domain recognition performance. Summary of the Invention

[0009] To solve the above problems existing in the prior art, the present invention provides an unsupervised domain adaptation SAR target recognition method that fuses and aligns visual features and scattering topological features. The technical problems to be solved by the present invention are realized through the following technical solutions:

[0010] An unsupervised domain adaptation SAR target recognition method that fuses and aligns visual features and scattering topological features, comprising the following steps:

[0011] Step 100, respectively extract the scattering topological points of each SAR image X in the original source domain image and the original target domain image and construct a corresponding first scattering topological graph G ASC to obtain a first source domain scattering topological graph and a first target domain scattering topological graph

[0012] Step 200, respectively extract the scattering topological points of each SAR image X in the original source domain image and the original target domain image and construct a corresponding second scattering topological graph G SST to obtain a second source domain scattering topological graph and a second target domain scattering topological graph

[0013] Step 300, construct a deep recognition network Ψ including a visual feature extraction sub-network, an image reconstruction sub-network, an ASC-based scattering topological feature extraction sub-network, a SAR-SIFT-based scattering topological feature extraction sub-network, and a feature fusion sub-network;

[0014] Step 400, based on the original source domain image the first source domain scattering topological graph the second source domain scattering topological graph the original target domain image the first target domain scattering topological graph and the second target domain scattering topological graph train the deep recognition network Ψ to obtain a trained target deep recognition network Ψ';

[0015] Step 500, input the original target domain image to be recognized, the first target domain scattering topological graph to be recognized, and the second target domain scattering topological graph to be recognized into the target deep recognition network Ψ' to obtain a target recognition result.

[0016] In an embodiment of the present invention, the step 100 includes:

[0017] Step 110: For each SAR image X, extract multiple scattering topology points using the ASC model; where each scattering topology point corresponds to a feature vector, and the feature vector corresponding to the i-th scattering topology point is A i represents the complex amplitude, a i represents the frequency-dependent factor, x i and y i represent the position coordinates in the range direction and azimuth direction respectively, L i represents the length of the scattering topology point, and γ i represent the direction angle and azimuth-dependent factor of the scattering topology point respectively;

[0018] Step 120: Decompose the complex amplitude A in the feature vector corresponding to the scattering topology point of the SAR image X i into modulus |A i | and phase Remove the frequency-dependent factor a i and the azimuth-dependent factor γ i , and the feature vector corresponding to the i-th scattering topology point is expressed as

[0019]

[0020] Step 130: Remove the scattering topology points located in the background region in the SAR image X whose Euclidean distance from the centroid of the multiple scattering topology points is greater than the distance threshold, and obtain the first extracted scattering topology points of the SAR image X;

[0021] Step 140: Construct the first scattering topology graph G corresponding to the SAR image X according to the first extracted scattering topology points ASC ; The first scattering topology graphs of all SAR images in the original source domain image and the first scattering topology graphs of all SAR images in the original target domain image correspondingly constitute the first source domain scattering topology graph and the first target domain scattering topology graph

[0022] In an embodiment of the present invention, the step 130 includes:

[0023] Step 131: Calculate the centroid of the multiple scattering topology points of the SAR image X

[0024]

[0025]

[0026] where p represents the number of scattering topology points;

[0027] Step 132, calculate the Euclidean distance between the coordinates (x i , y i ) corresponding to each scattering topology point of the SAR image X and the centroid

[0028] Step 133, remove the scattering topology points located in the background region where the Euclidean distance d i is greater than the distance threshold τ dis to obtain the first extracted scattering topology points of the SAR image X.

[0029] In one embodiment of the present invention, the step 140 includes:

[0030] Step 141, use the first extracted scattering topology points of the SAR image X as the nodes of the graph structure of the first scattering topology graph G ASC ; where the node set q′ is the number of the first extracted scattering topology points;

[0031] Step 142, calculate the cosine similarity between the eigenvectors corresponding to any two nodes of the SAR image X. If the cosine similarity is greater than or equal to the similarity threshold τ sim then there is an edge between the two nodes;

[0032] The node set and the edge set constitute the first scattering topology graph G ASC =(V, M) corresponding to the SAR image X; the first scattering topology graphs of all SAR images in the original source domain image and the first scattering topology graphs of all SAR images in the original target domain image correspondingly constitute the first source domain scattering topology graph and the first target domain scattering topology graph

[0033] In one embodiment of the present invention, the step 200 includes:

[0034] Step 210, extract the corner points of each SAR image X based on the SAR-Harris algorithm;

[0035] Step 220, extract the strong scattering points of each SAR image X based on the clustering algorithm. The strong scattering points and corner points of a SAR image X together constitute the scattering topology points of the SAR image X;

[0036] ​Step 230: Remove the scattering topology points in the background area of SAR image X whose amplitude values are outside the range of the top 1% of the cumulative distribution function, to obtain multiple second-extracted scattering topology points of SAR image X;

[0037] Step 240: Remove the second-extracted scattering topology points in the background area of SAR image X whose Euclidean distance from the centroid of the multiple second-extracted scattering topology points is greater than the threshold, to obtain the third-extracted scattering topology points of SAR image X;

[0038] Step 250: Generate a SAR-SIFT descriptor as the feature vector of the third-extracted scattering topology points of SAR image X;

[0039] Step 260: Construct a second scattering topology graph G corresponding to SAR image X based on the third-extracted scattering topology points of SAR image X SST ; All the second scattering topology graphs of the SAR images in the original source domain image and all the second scattering topology graphs of the SAR images in the original target domain image correspondingly constitute the second source domain scattering topology graph and the second target domain scattering topology graph

[0040] In an embodiment of the present invention, the extraction of strong scattering points of each SAR image X based on a clustering algorithm includes:

[0041] Step 221: Deform the SAR image X from matrix form to vector form;

[0042] Step 222: Reorder the vector according to the descending order of the amplitude values of SAR image X;

[0043] Step 223: Initialize the clustering label of the first element to 0;

[0044] Step 224: Traverse each element from the second element in descending order;

[0045] Step 225: When traversing to the i-th element, determine whether the i-th element is within the 8-neighborhood of the previous i - 1 elements;

[0046] If the i-th element is not within the 8-neighborhood of any of the previous i - 1 elements, then the i-th element is a new category;

[0047] If the i-th element is only within the 8-neighborhood of a certain element among the previous i - 1 elements, then assign the category label of that element to the i-th element;

[0048] If the i-th element is within the 8-neighborhoods of multiple elements among the first i - 1 elements, the class label of the i-th element is marked as the class corresponding to the maximum amplitude value among the multiple elements.

[0049] Step 226, after the traversal is completed, restore the SAR image X from the vector form to the matrix form, and abstract each clustering cluster as a strong scattering point.

[0050] In an embodiment of the present invention, the step 230 includes:

[0051] Step 231, statistically calculate the cumulative distribution function of all pixel amplitudes in the SAR image X;

[0052] Step 232, remove the scattering topological points in the background area whose amplitude values are outside the maximum 1% range of the cumulative distribution function, and obtain multiple second extracted scattering topological points of the SAR image X.

[0053] In an embodiment of the present invention, the visual feature extraction sub-network includes a first extraction convolutional layer L arranged in sequence C1 , a second extraction activation layer L C2 , a third extraction convolutional layer L C3 , a fourth extraction activation layer L C4 , a fifth extraction convolutional layer L C5 , a sixth extraction activation layer L C6 , a seventh extraction convolutional layer L C7 , an eighth extraction activation layer L C8 , a ninth extraction convolutional layer L C9 and a tenth extraction activation layer L C10 ;

[0054] The input of the visual feature extraction sub-network is the SAR image X, and the output is a 512-dimensional visual feature vector;

[0055] The image reconstruction sub-network includes a first reconstruction convolutional layer L' arranged in sequence C1 , a second reconstruction activation layer L' C2 , a third reconstruction convolutional layer L' C3 , a fourth reconstruction activation layer L' C4 , a fifth reconstruction convolutional layer L' C5 , a sixth reconstruction activation layer L' C6 , a seventh reconstruction convolutional layer L' C7 , an eighth reconstruction activation layer L' C8 , a ninth reconstruction convolutional layer L' C9 and a tenth reconstruction activation layer L' C10 ;

[0056] The input of the image reconstruction sub-network is the 512-dimensional visual feature vector, and the output is the reconstructed SAR image

[0057] The ASC-based scattering topology feature extraction sub-network includes a first ASC graph convolutional layer L arranged in sequence G1 , a second ASC activation layer L G2 , a third ASC graph convolutional layer L G3 , a fourth ASC activation layer L G4 and a fifth ASC readout layer L G5 ;

[0058] The input of the ASC-based scattering topology feature extraction sub-network is the first scattering topology graph G ASC , and the output is a 256-dimensional scattering topology feature vector;

[0059] The SAR-SIFT-based scattering topology feature extraction sub-network includes a first SST graph convolutional layer L' arranged in sequence G1 , a second SST activation layer L' G2 , a third SST graph convolutional layer L' G3 , a fourth SST activation layer L' G4 and a fifth SST readout layer L' G5 ;

[0060] The input of the SAR-SIFT-based scattering topology feature extraction sub-network is the second scattering topology graph G SST , and the output is a 256-dimensional scattering topology feature vector;

[0061] The feature fusion sub-network includes a first fusion fully connected layer L F1 , a second fusion activation layer L F2 , a third fusion fully connected layer L F3 and a fourth fusion classifier layer L F4 ;

[0062] The input of the feature fusion sub-network is a 1024-dimensional feature vector jointly formed by splicing the 512-dimensional visual feature vector, the 256-dimensional scattering topology feature vector extracted by the ASC-based scattering topology feature extraction sub-network, and the 256-dimensional scattering topology feature vector extracted by the SAR-SIFT-based scattering topology feature extraction sub-network, and the output is a 10-dimensional class prediction label

[0063] The visual feature extraction sub-network is connected to the image reconstruction sub-network, and the visual feature extraction sub-network, the ASC-based scattering topology feature extraction sub-network, and the SAR-SIFT-based scattering topology feature extraction sub-network are all connected to the feature fusion sub-network.

[0064] In one embodiment of the present invention, step 400 includes:

[0065] Step 410, input the original source domain image The first source domain scattering topology map The second source domain scattering topology map The original target domain image The first target domain scattering topology map and the second target domain scattering topology map into the deep recognition network Ψ for forward propagation to obtain the reconstructed SAR image, the fused features output by the second fusion activation layer L F2 and the class prediction label;

[0066] Step 420, calculate the reconstruction loss, and the loss function of the reconstruction loss is:

[0067]

[0068] wherein, and respectively represent the i-th original source domain image and the j-th original target domain image, and respectively represent the i-th source domain image and the j-th target domain image obtained by reconstruction, n s and n t respectively represent the numbers of the source domain images and the target domain images, represents the square matrix F norm;

[0069] Step 430, calculate the classification loss, and the loss function of the classification loss is:

[0070]

[0071] wherein, represents the class true label in one-hot encoding form of the i-th source domain image, represents the corresponding class prediction label;

[0072] Step 440, calculate the domain adaptation loss, and the loss function of the domain adaptation loss is:

[0073]

[0074] wherein, Loss global represents the global domain adaptation loss, Loss local represents the local domain adaptation loss, τ ep represents the training iteration threshold;

[0075] Step 450, calculate the total loss, and the calculation formula is:

[0076] Loss total = Loss rec + Loss cls + Loss da ;

[0077] Step 460, update the parameters of the depth recognition network Ψ through backpropagation to obtain the target depth recognition network Ψ'.

[0078] Advantages of the present invention:

[0079] 1. The present invention fully exploits the scattering characteristics of SAR images and discovers that the same type of targets in simulated SAR images and measured SAR images not only have visual differences but also differences in scattering topological structures. Therefore, the present invention aligns the distributions of simulated images and measured images in both the visual space and the scattering topological space, further improving the cross-domain recognition performance of the network.

[0080] 2. The present invention uses a graph structure to model the scattering topological structure of the target and introduces a graph neural network (GNN) to extract the scattering topological features of the target. The graph structure can represent local scattering information through nodes and the spatial topological structure of the target through edges. The GNN can aggregate structure learning in the process of node feature learning, so the finally extracted scattering topological features contain both the local scattering information and the spatial topological structure of the target.

[0081] 3. The present invention uses two different methods, the attribute scattering center (ASC) model and the SAR-scale invariant feature transform (SIFT) algorithm, to extract and model the scattering topological structure of the target. The former reflects the physical attributes of the target, such as amplitude, position, shape, and orientation, and the latter mainly reflects the edge information of the target through gradient information. The present invention fuses these two methods at the feature level, providing more information for the network, thereby improving the cross-domain recognition performance of the network.

[0082] 4. The present invention not only aligns the overall distributions of simulated SAR images and measured SAR images but also aligns the distributions of corresponding subcategories through a two-stage domain adaptation loss function, enabling the network trained based on simulated SAR images to better generalize to measured SAR images, further improving the cross-domain recognition performance of the network.

[0083] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Brief description of the drawings

[0084] Figure 1 It is a schematic flow chart of an unsupervised domain adaptation SAR target recognition method that fuses and aligns visual features and scattering topological features provided by an embodiment of the present invention;

[0085] Figure 2a is a SAR image X;

[0086] Figure 2b is Figure 2a a schematic diagram of the scattered topological points extracted through step 110;

[0087] Figure 2c is Figure 2b a schematic diagram of the first extracted scattered topological points obtained through step 130;

[0088] Figure 2d is the first scattered topological graph G constructed according to the embodiment of the present invention ASC ;

[0089] Figure 3a is a SAR image X;

[0090] Figure 3b is Figure 3a a schematic diagram of the corner points extracted through step 210;

[0091] Figure 3c is Figure 3a a schematic diagram of the strong scattering points extracted through step 220;

[0092] Figure 3d is a schematic diagram of the scattered topological points jointly composed of the strong scattering points and the corner points provided by the embodiment of the present invention;

[0093] Figure 3e is Figure 3d a schematic diagram of the second extracted scattered topological points obtained through step 230;

[0094] Figure 3f is Figure 3e a schematic diagram of the third extracted scattered topological points obtained through step 240;

[0095] Figure 3g is the second scattered topological graph G constructed according to the embodiment of the present invention SST ;

[0096] Figure 4 is a schematic diagram of the structure and forward propagation of the depth recognition network Ψ provided by the embodiment of the present invention. Specific Embodiments

[0097] The present invention will be further described in detail below in conjunction with specific embodiments, but the embodiments of the present invention are not limited thereto.

[0098] At present, object recognition algorithms based on deep learning have achieved good results in SAR ATR tasks. However, deep learning is a data-driven algorithm, and its high performance often relies on a large amount of training data. Nevertheless, the acquisition and annotation of measured SAR images are both time-consuming and laborious. Compared with optical datasets, SAR image datasets usually have fewer images and smaller image sizes. When the samples are insufficient, the performance of recognition algorithms based on deep learning often drops significantly. Therefore, how to design recognition algorithms in the case of scarce measured samples is an engineering problem with practical application value. Benefiting from the characteristics of SAR actively emitting electromagnetic waves and imaging based on the backscattered echoes of targets, synthesizing simulated SAR images according to the CAD models of targets and electromagnetic simulation programs has become an effective way to augment SAR images. However, due to the fact that the CAD modeling and electromagnetic simulation processes do not fully conform to the real physical situation, there are certain differences between simulated SAR images and measured SAR images, and they do not meet the independent and identically distributed conditions. This results in the fact that directly using simulated images to train a classifier to recognize measured images often cannot achieve good performance.

[0099] Currently, some scholars have introduced the domain adaptation theory to align the distributions between simulated images and measured images in the feature space, so that the classifier trained based on simulated images can be generalized to measured images. However, most of the existing methods directly treat SAR images as optical images for processing, ignoring the imaging characteristics of SAR images themselves. At the same time, the existing methods often only align the distributions between the overall simulated images and measured images, while ignoring the distribution situations between sub-categories.

[0100] In order to improve the recognition performance, the present invention proposes an unsupervised domain adaptation SAR object recognition method that fuses and aligns visual features and scattering topological features, specifically as follows:

[0101] Embodiment 1

[0102] As Figure 1 shown, an unsupervised domain adaptation SAR object recognition method that fuses and aligns visual features and scattering topological features includes the following steps:

[0103] Step 100, respectively extract the scattering topological points of each SAR image X in the original source domain image and the original target domain image based on the ASC model and construct the corresponding first scattering topological graph G ASC , obtaining the first source domain scattering topological graph and the first target domain scattering topological graph

[0104] Step 200, respectively extract the original source domain image and the original target domain image based on the SAR-SIFT algorithm For each SAR image X, extract the scattering topological points and construct the corresponding second scattering topology graph G SST to obtain the second source domain scattering topology graph and the second target domain scattering topology graph

[0105] Step 300: Construct a deep recognition network Ψ including a visual feature extraction sub-network, an image reconstruction sub-network, an ASC-based scattering topology feature extraction sub-network, a SAR-SIFT-based scattering topology feature extraction sub-network, and a feature fusion sub-network;

[0106] Step 400: Based on the original source domain images the first source domain scattering topology graph the second source domain scattering topology graph the original target domain images the first target domain scattering topology graph and the second target domain scattering topology graph train the deep recognition network Ψ to obtain the trained target deep recognition network Ψ';

[0107] Step 500: Input the original target domain image to be recognized, the first target domain scattering topology graph to be recognized, and the second target domain scattering topology graph to be recognized into the target deep recognition network Ψ' to obtain the target recognition result.

[0108] In this embodiment, the visual features of the SAR image are extracted by a convolutional neural network (CNN), and the generalization ability of the visual features is enhanced based on image reconstruction. The scattering topological points of the SAR image are extracted by two different methods, the ASC model and the SAR-SIFT algorithm. Further, the graph structure is used to model the scattering topological structure of the target, and the scattering topological features of the target are extracted by the GNN. The visual features of the target are fused and re-encoded with the two different scattering topological features to remove redundant information. The distribution between the simulation image and the measured image is aligned category by category in the feature space through the improved two-stage domain adaptive loss, so that the classifier trained based on the simulation image can accurately recognize the measured image.

[0109] The method of the present invention can make full use of the scattering characteristics of SAR images, extract the scattering topological points of the target using different methods, model the scattering topological structure of the target using graph structure, and then extract the scattering topological features of the target through GNN. This feature can not only contain the local electromagnetic scattering information of the target, but also reflect the spatial topological structure information of the target to a certain extent. At the same time, the present invention also extracts the visual features of SAR images and further improves the generalization ability of visual features through image reconstruction. Then, the present invention also fuses and re-encodes the visual features and scattering topological features to generate more robust and effective features. Finally, the two-stage domain adaptation loss of the present invention can more accurately align the distributions of corresponding categories in simulated images and measured images in the feature space, further improving the recognition performance.

[0110] Embodiment 2

[0111] An unsupervised domain adaptation SAR target recognition method that fuses and aligns visual features and scattering topological features, comprising the following steps:

[0112] Step 100, respectively extract the scattering topological points of each SAR image X in the original source domain image (labeled simulated image) and the original target domain image (unlabeled measured image) and construct the corresponding scattering topological graph G ASC , obtaining the first source domain scattering topological graph and the first target domain scattering topological graph Specifically, as Figures 2a - 2d shown, step 100 includes: step 110 - step 140.

[0113] Step 110, use the ASC parameter estimation algorithm based on image domain sparse representation to extract multiple scattering topological points (i.e., scattering centers) for each SAR image X; in this embodiment, the number p of extracted scattering topological points is set to 20.

[0114] Among them, each scattering topological point corresponds to a feature vector, and the feature vector corresponding to the i-th scattering topological point is A i represents the complex amplitude, a i represents the frequency-dependent factor, x i and y i respectively represent the range and azimuth position coordinates, L i represents the length of the scattering topological point, and γ i respectively represent the direction angle and azimuth-dependent factor of the scattering topological point;

[0115] Step 120, the complex amplitude A in the feature vector corresponding to the scattering topological points of the SAR image Xi Decompose into modulus |A i and phase Ignore the frequency-dependent factor a i and the azimuth-dependent factor γ i , the eigenvector corresponding to the i-th scattering topological point is expressed as

[0116] Step 130, remove the scattering topological points mis-extracted in the background area according to the distance between the scattering topological points and the centroid in the SAR image X: remove the scattering topological points whose Euclidean distance from the centroid of the scattering topological points is greater than the distance threshold, and obtain the first extracted scattering topological points of the SAR image X; among them, step 130 includes steps 131-step 133:

[0117] Step 131, calculate the centroid of all scattering topological points in the SAR image X

[0118]

[0119]

[0120] where p represents the number of scattering topological points;

[0121] Step 132, calculate the coordinates (x i , y i ) corresponding to each scattering topological point in the SAR image X and the centroid between the Euclidean distances

[0122] Step 133, remove the scattering topological points in the background area where the Euclidean distance d i is greater than the distance threshold τ dis , and obtain the first extracted scattering topological points of the SAR image X. Among them, the distance threshold τ dis is set to 30.

[0123] Step 140, construct the first scattering topological graph G corresponding to the SAR image X according to the first extracted scattering topological points ASC ; specifically, step 140 includes steps 141-step 142:

[0124] Step 141, the graph structure consists of a node set and an edge set. Take the first extracted scattering topological points of the SAR image X as the nodes of the first scattering topological graph G ASC to obtain the node set where q′ is the number of the first extracted scattering topological points;

[0125] Step 142, the edge set is usually represented by an adjacency matrix to represent, m i,j= 1 indicates that there is an edge between node i and node j. Calculate the cosine similarity between the feature vectors corresponding to any two nodes of the SAR image X. If the cosine similarity is greater than or equal to the similarity threshold τ sim then there is an edge between the two nodes, that is:

[0126]

[0127] where <·,·> represents the vector inner product. After obtaining the node set V and the adjacency matrix M, the first scattering topology graph G corresponding to the SAR image X is obtained ASC = (V, M). Preferably, the similarity threshold τ sim is set to 0.0. Figure 2d shows the connection relationships between different nodes on the target. The original source domain The first scattering topology graphs of all SAR images in the image and the original target domain image correspondingly constitute the first source domain scattering topology graph and the first target domain scattering topology graph

[0128] It should be noted that since the original source domain image and the original target domain image both adopt the processing process of the above steps 110 - 140 to obtain the scattering topology graph. For the sake of concise description above, the original source domain image and the original target domain image Any one of the images is denoted as the SAR image X, where n s and n t are the numbers of the original source domain image and the original target domain image respectively. Specifically, any one SAR image of the original source domain image corresponding first scattering topology graph is Any one SAR image of the original target domain image corresponding first scattering topology graph is Then all the first scattering topology graphs of the SAR images of the original source domain image constitute the first source domain scattering topology graph Then all the first scattering topology graphs of the SAR images of the original target domain image constitute the first target domain scattering topology graph

[0129] Step 200, respectively extract the original source domain image based on the SAR - SIFT algorithm and the original target domain image Scattering topological points of each SAR image X and construct a corresponding scattering topology graph G SST , obtain the second source domain scattering topology graph and the second target domain scattering topology graph Specifically, as Figures 3a - 3g shown, step 200 includes steps 210 - 260:

[0130] Step 210, extract the corner points of each SAR image X based on the SAR - Harris algorithm;

[0131] Step 220, extract the strong scattering points of each SAR image X based on the clustering algorithm. The strong scattering points and corner points of a SAR image X together constitute the scattering topological points of the SAR image X. Specifically, extracting the strong scattering points of each SAR image X based on the clustering algorithm includes steps 221 - 226:

[0132] Step 221, deform the SAR image X from matrix form to vector form;

[0133] Step 222, re - sort the vector according to the amplitude values of the SAR image X from large to small;

[0134] Step 223, initialize the clustering label of the first element (i.e., the element with the largest amplitude value) to 0;

[0135] Step 224, start traversing each element from the second element (i.e., the element with the second - largest amplitude value) from large to small;

[0136] Step 225, when traversing to the i - th element, determine whether the i - th element is within the 8 - neighborhood of the previous i - 1 elements;

[0137] If the i - th element is not within the 8 - neighborhood of any of the previous i - 1 elements, then the i - th element is a new category;

[0138] If the i - th element is only within the 8 - neighborhood of a certain one of the previous i - 1 elements, then assign the category label of that element to the i - th element;

[0139] If the i - th element is within the 8 - neighborhood of multiple elements among the previous i - 1 elements, then the category label of the i - th element is marked as the category corresponding to the maximum amplitude value among these multiple elements;

[0140] Step 226, after all elements are traversed, restore the SAR image X from vector form to matrix form, abstract each clustering cluster as a strong scattering point, the coordinates of this point are the centroid of the clustering, and the amplitude value of this point is the amplitude value at the centroid.

[0141] Step 230: Remove the points mis-extracted in the background area according to the amplitude values of the scattering topology points of SAR image X: Remove the scattering topology points whose amplitude values are outside the range of the largest 1% of the cumulative distribution function, and obtain multiple second-extracted scattering topology points of SAR image X; Step 230 includes Step 231 - Step 232:

[0142] Step 231: Statistically calculate the cumulative distribution function of the amplitudes of all pixels in SAR image X;

[0143] Step 232: Retain the scattering topology points whose amplitude values are within the range of the largest 1% of the cumulative distribution function, and remove the scattering topology points not within the range, to obtain multiple second-extracted scattering topology points of SAR image X.

[0144] Step 240: Further remove the points mis-extracted in the background area according to the distance between the scattering topology points and the centroid of SAR image X: Remove the second-extracted scattering topology points in the background area whose Euclidean distance from the centroid of the second-extracted scattering topology points is greater than the threshold, to obtain the third-extracted scattering topology points of SAR image X; This step is the same as Step 130.

[0145] Step 250: Generate a SAR-SIFT descriptor as a feature vector for the third-extracted scattering topology points of SAR image X;

[0146] Step 260: Construct the second scattering topology graph G corresponding to SAR image X according to the third-extracted scattering topology points of SAR image X SST ; The specific steps of this step are similar to those of Step 140, the difference is that the node feature of this step is a 108-dimensional SAR-SIFT descriptor, and at the same time the similarity threshold τ sim is set to 0.6. Figure 3g shows the connection relationships between different nodes on the target. The original source domain image The second scattering topology graphs of all SAR images in and the second scattering topology graphs of all SAR images in the original target domain correspondingly constitute the second source domain scattering topology graph

[0147] It should be noted that since the original source domain image and the original target domain image both adopt the processing procedures of the above Step 210 - Step 260 to obtain the scattering topology graph, for the sake of concise description above, any one of the original source domain image and the original target domain image is denoted as SAR image X. Specifically, any one SAR image of the original source domain image of The corresponding second scattering topology diagram is The original target domain image Any SAR image of The corresponding second scattering topology diagram is Then the original source domain image The second scattering topology diagrams of all SAR images of form the second source domain scattering topology diagram Then the original target domain image The second scattering topology diagrams of all SAR images of form the second target domain scattering topology diagram

[0148] Step 300, construct a deep recognition network Ψ including a visual feature extraction sub-network, an image reconstruction sub-network, a scattering topology feature extraction sub-network based on ASC, a scattering topology feature extraction sub-network based on SAR-SIFT, and a feature fusion sub-network, as Figure 4 shown

[0149] The specific structures of each sub-network are as follows

[0150] The visual feature extraction sub-network is a convolutional encoding network Encoder, including five convolutional layers and five activation layers. The specific structure is the first extraction convolutional layer L C1 set in sequence, the second extraction activation layer L C2 the third extraction convolutional layer L C3 the fourth extraction activation layer L C4 the fifth extraction convolutional layer L C5 the sixth extraction activation layer L C6 the seventh extraction convolutional layer L C7 the eighth extraction activation layer L C8 the ninth extraction convolutional layer L C9 and the tenth extraction activation layer L C10 ;

[0151] The input of the visual feature extraction sub-network is the SAR image X, and the output is a 512-dimensional visual feature vector

[0152] Set the parameters of each layer as follows: the number of convolutional kernels of the five convolutional layers are set to 32, 64, 128, 256, 512 respectively, the convolutional kernel sizes are set to 5×5, 5×5, 3×3, 3×3, 4×4 respectively, and the convolutional kernel strides are set to 2, 2, 2, 2, 1 respectively; all five activation layers use the ReLU activation function

[0153] The image reconstruction sub-network is a convolutional decoding network Decoder, including five transposed convolutional layers and five activation layers. The specific structure is the first reconstruction transposed convolutional layer L′ C1 set in sequence, the second reconstruction activation layer L′ C2, the third reconstruction deconvolution layer L' C3 , the fourth reconstruction activation layer L' C4 , the fifth reconstruction deconvolution layer L' C5 , the sixth reconstruction activation layer L' C6 , the seventh reconstruction deconvolution layer L' C7 , the eighth reconstruction activation layer L' C8 , the ninth reconstruction deconvolution layer L' C9 and the tenth reconstruction activation layer L' C10 ;

[0154] The input of the image reconstruction subnetwork is a 512-dimensional visual feature vector, and the output is the reconstructed SAR image

[0155] Set the parameters of each layer as follows: the number of convolution kernels of the five deconvolution layers are set to 256, 128, 64, 32, and 1 respectively, the convolution kernel sizes are set to 4×4, 3×3, 3×3, 5×5, and 5×5 respectively, and the convolution kernel strides are set to 1, 2, 2, 2, and 2 respectively; the ReLU activation function is used in the first four of the five activation function layers, and the Tanh activation function is used in the last layer.

[0156] The scattering topology feature extraction subnetwork based on ASC is the graph neural network GNN ASC , including two graph convolution layers, two activation layers and one readout layer. The specific structure is the first ASC graph convolution layer L set in sequence G1 , the second ASC activation layer L G2 , the third ASC graph convolution layer L G3 , the fourth ASC activation layer L G4 and the fifth ASC readout layer L G5 ;

[0157] The input of the scattering topology feature extraction subnetwork based on ASC is the first scattering topology graph G ASC , and the output is a 256-dimensional scattering topology feature vector;

[0158] Set the parameters of each layer as follows: both graph convolution layers use graph attention convolution layers, the number of convolution heads is set to 4, and the features extracted by multiple heads are concatenated. The weight dimension of each convolution head in the first layer is 6×64, and the weight dimension of each convolution head in the second layer is 256×64; both activation layers use the ELU activation function; the readout layer uses the element-wise maximum pooling readout method.

[0159] The scattering topology feature extraction subnetwork based on SAR-SIFT is the graph neural network GNN SST , including two graph convolution layers, two activation layers and one readout layer. The specific structure is the first SST graph convolution layer L'set in sequence G1 , the second SST activation layer L'G2 , the third SST graph convolutional layer L′ G3 , the fourth SST activation layer L′ G4 and the fifth SST readout layer L′ G5 ;

[0160] The input of the scattering topology feature extraction sub-network based on SAR-SIFT is the second scattering topology graph G SST , and the output is a 256-dimensional scattering topology feature vector;

[0161] Set the parameters of each layer as follows: Both graph convolutional layers use graph attention convolutional layers, the number of convolutional heads is set to 4, and the features extracted by multiple heads are concatenated. The weight dimension of each convolutional head in the first layer is 108×64, and the weight dimension of each convolutional head in the second layer is 256×64; Both activation layers use the ELU activation function; The readout layer uses the element-wise max pooling readout method.

[0162] The feature fusion sub-network is a fully connected network FC, including two fully connected layers, one activation layer and one classifier layer. The specific structure is the first fusion fully connected layer L F1 , the second fusion activation layer L F2 , the third fusion fully connected layer L F3 and the fourth fusion classifier layer L F4 ;

[0163] The input of the feature fusion sub-network is a 1024-dimensional feature vector jointly concatenated by the 512-dimensional visual feature vector extracted by the convolutional encoding network Encoder, the 256-dimensional scattering topology feature vector extracted by the graph neural network GNN ASC , the 256-dimensional scattering topology feature vector extracted by the graph neural network GNN SST and the 256-dimensional scattering topology feature vector extracted by the extraction sub-network, and the output is a 10-dimensional class prediction label

[0164] Set the parameters of each layer as follows: The weight dimensions of the two fully connected layers are set to 1024×128 and 128×10 respectively; The activation layer uses the ReLU activation function; The classifier layer uses the softmax classifier.

[0165] In this embodiment, the visual feature extraction sub-network is connected to the image reconstruction sub-network, and the visual feature extraction sub-network, the scattering topology feature extraction sub-network based on ASC, and the scattering topology feature extraction sub-network based on SAR-SIFT are all connected to the feature fusion sub-network. That is, the five sub-networks are combined together in the order as Figure 4 shown to form the deep neural network Ψ.

[0166] Step 400, the original source domain image The first source domain scattering topology graph Second source domain scattering topology map Original target domain image First target domain scattering topology map and second target domain scattering topology map Input the depth recognition network Ψ for forward propagation, calculate the reconstruction loss, classification loss, and domain adaptation loss, and update the network parameters through backpropagation to obtain the trained target depth recognition network Ψ'. Specifically, as Figure 4 shown, step 400 includes steps 410 - 460:

[0167] Step 410, input the original source domain image First source domain scattering topology map Second source domain scattering topology map Original target domain image First target domain scattering topology map and second target domain scattering topology map into the depth recognition network Ψ for forward propagation to obtain the reconstructed SAR image, the fused features output by the second fusion activation layer L F2 and the class prediction label;

[0168] Step 420, calculate the reconstruction loss. The reconstruction loss uses the mean squared error loss function, and the loss function of the reconstruction loss is:

[0169]

[0170] where, and represent the i-th original source domain image and the j-th original target domain image respectively, and represent the reconstructed i-th source domain image and the j-th target domain image respectively, n s and n t represent the number of source domain images and target domain images respectively, represents the square matrix F norm;

[0171] Step 430, calculate the classification loss. The classification loss uses the cross-entropy loss function, and the loss function of the classification loss is:

[0172]

[0173] where, represents the true class label of the i-th source domain image in one-hot encoded form, represents the corresponding class prediction label;

[0174] Step 440, calculate the domain adaptation loss. The loss function of the domain adaptation loss is:

[0175]

[0176] Among them, Loss global represents the global domain adaptive loss, and Loss local represents the local domain adaptive loss, and τ ep represents the training iteration threshold; in this embodiment, the training iteration threshold is set to 60. In the first 60 rounds of training, the present invention uses the global domain adaptive loss to quickly align the overall distributions of the source domain data and the target domain data; in the later stage of training, after the overall distributions of the source domain data and the target domain data have been initially aligned, the present invention uses the local domain adaptive loss to align the distributions of the subclasses of the source domain data and the target domain data class by class, so that the classifier trained based on the source domain data can better generalize to the target domain data. Specifically, the global domain adaptive loss Loss global used in the present invention is the correlationalignment (CORAL) loss, as shown below,

[0177]

[0178] where C s and C t are the covariance matrices of the feature vectors F2 output by the source domain data and the target domain data passing through the second fusion activation layer L and respectively; d is the dimension of the feature vector, that is, 128. The local domain adaptive loss Loss local used in the present invention is the local maximum mean discrepancy (LMMD) loss, as shown below

[0179]

[0180] where L represents the total number of data categories, is the reproducing Hilbert space defined by the kernel function k, φ represents the feature mapping that maps the fused features into the reproducing Hilbert space, and there is Specifically, the kernel function k uses a Gaussian kernel function with a bandwidth set to the average pairwise squared distance between any two samples in the source domain and the target domain. and respectively represent the probabilities that the i-th source domain image and the i-th target domain image belong to the l-th class, and the calculation process is as follows,

[0181]

[0182]

[0183] Among them, represents the i-th source domain image of the true one-hot encoded label vector of the l-th dimension, represents the i-th target domain image of the predicted one-hot encoded label vector of the l-th dimension.

[0184] Step 450, calculate the total loss, and the calculation formula is:

[0185] Loss total = Loss rec + Loss cls + Loss da

[0186] Step 460, update the parameters of the deep recognition network Ψ through backpropagation to obtain the target deep recognition network Ψ'.

[0187] Step 500, input the original target domain image to be recognized, the first target domain scattering topology map to be recognized, and the second target domain scattering topology map to be recognized into the target deep recognition network Ψ' to obtain the target recognition result. Among them, the original target domain image to be recognized needs to first construct the first target domain scattering topology map to be recognized and the second target domain scattering topology map to be recognized according to Step 100 and Step 200, and then input them together into the target deep recognition network Ψ' for recognition to obtain the target recognition result.

[0188] The present invention fully exploits the scattering characteristics of SAR images and finds that there are differences not only in visual information but also in scattering topology between simulated images and measured images. Therefore, the present invention simultaneously reduces the distribution difference between simulated images and measured images in both the visual space and the scattering topology space, further improving the cross-domain recognition performance. The present invention not only aligns the overall distributions between simulated images and measured images but also aligns the distributions between corresponding sub-categories, significantly improving the cross-domain recognition performance. In addition, the present invention does not require known measured image labels during the training phase.

[0189] The effect of the target recognition method of the present invention will be further described through simulation experiments below.

[0190] Experimental conditions:

[0191] The hardware platform for the simulation experiment is: the processor is Intel(R) Xeon(R) Silver 4210R CPU@2.40GHz, the memory is 128GB, and the graphics card is GeForce RTX 3090. The software platform is: Ubuntu16.04 operating system, python 3.8, pytorch1.12, and matlab 2021a.

[0192] Experimental data: The experimental data uses the SAMPLE dataset, which includes paired measured SAR images and simulated SAR images. The measured SAR images are from the MSTAR dataset, and the simulated SAR images are generated by refined CAD models and progressive ray tracing techniques. The data includes 10 types of ground vehicle targets, namely: 2S1, BMP2, BTR70, M1, M2, M35, M548, M60, T72, and ZSU23. The image resolution is 0.3m × 0.3m, the image size is 128 × 128 pixels, the azimuth angle is 10° - 80°, the elevation angle is 14° - 17°, the band is the X-band, and the polarization mode is HH polarization. This dataset contains a total of 1345 measured SAR images and 1345 simulated SAR images. There are currently two experimental settings for this dataset: In the first setting, the source domain data consists of simulated images from 14° to 16°, and the target domain data consists of measured images at 17°. In the second setting, the source domain data consists of all simulated images, and the target domain data consists of all measured images. In the first experimental setting, the one-to-one correspondence between the simulated images and the measured images is artificially broken, so it is more in line with the actual situation and more difficult. In the second experimental setting, there is a one-to-one correspondence between the simulated images and the measured images, so it is relatively simple and can evaluate the upper limit of the potential of the simulated data. Tables 1 and 2 respectively show the data details of experimental setting 1 and experimental setting 2.

[0193] Table 1 Data details of experimental setting 1

[0194]

[0195] Table 2 Data details of experimental setting 2

[0196]

[0197] Comparative experiment method: The method proposed in the present invention and various existing algorithms are used to compare the results under the same conditions. The comparison methods include the literature "A SAR dataset for ATR development: The synthetic and measured paired labeled experiment (SAMPLE)" (Proc. SPIE) published by B. Lewis et al. in 2019, the literature "A deep learning approach to the synthetic and measured paired and labeled experiment (SAMPLE) challenge problem" (Proc. SPIE) published by T. Scarnati et al. in 2019, the literature "Bridging a gap in SAR-ATR: Training on fully synthetic and testing on measured data" (IEEE J. Sel. Topics Appl. Earth Observ. Remote Sens.) published by N. Inkawhich et al. in 2019, the literature "Convolutional and generative pairing for SAR cross-target transfer learning" (Proceedings of the Algorithms for Synthetic Aperture Radar Imagery XXVII, SPIE) published by A. Jennison et al. in 2021, the literature "A hierarchical receptive network oriented to target recognition in SAR images" (Pattern Recognit.) published by G. Dong et al. in 2022, and the literature "Exploring characteristics of neural network architecture computation for enabling SAR ATR" (Proceedings of the Algorithms for Synthetic Aperture Radar Imagery XXVII, SPIE) published by Melzer et al. in 2021.

[0198] Evaluation criterion: The accuracy of SAR image target recognition.

[0199] Experimental content:

[0200] The recognition experiment was carried out on the above experimental data using the method of the present invention and the existing method, and the comparative experimental results of the present invention and the existing method under two experimental settings were shown.

[0201] Experimental results and analysis:

[0202] Comparative test:

[0203] Tables 3 and 4 are the recognition accuracies of the method of the present invention and the existing method under experimental setting 1 and experimental setting 2. As shown in Tables 3 and 4, the method proposed by the present invention is significantly superior to the existing method, indicating the effectiveness of the method proposed by the present invention.

[0204] Table 3 Recognition accuracies of each method under experimental setting 1

[0205] Method Recognition accuracy Lewis(2019) 24.97％ Scarnati(2019) 55.62％ Inkawhich(2021) 95.06％ The method proposed by the present invention 99.15％

[0206] Table 4 Recognition accuracies of each method under experimental setting 2

[0207] Method Recognition accuracy Jennison(2019) 88.45％ Dong(2021) 93.66％ Melzer(2021) 96.88％ The method proposed by the present invention 98.18％

[0208] In summary, a SAR target recognition method that fuses and aligns visual features and scattering topological features proposed by the present invention improves the problems in the prior art that the scattering features of SAR images cannot be fully combined and only the overall distributions of simulated images and measured images are concerned. The implementation of the present invention includes: extracting the scattering topological points of the target based on the ASC model and constructing a scattering topological map, extracting the scattering topological points of the target based on the SAR-SIFT algorithm and constructing a scattering topological map, constructing a deep network, training the deep network, and recognizing unknown images. The present invention does not require labeled measured images during training, extracts not only the visual information of the target but also the scattering topological structure of the target, aligns the sub-domain distributions of simulated images and measured images in the feature space, and has more accurate classification, which has strong practical application significance and can be applied to the field of automatic recognition of vehicle targets in SAR images.

[0209] In addition, the terms "first" and "second" are only used for descriptive purposes and cannot be understood as indicating or implying relative importance or implicitly specifying the quantity of the indicated technical features. Thus, the features defined with "first" and "second" may explicitly or implicitly include one or more of such features. In the description of the present invention, "a plurality" means two or more unless otherwise specifically defined.

[0210] In the description of this specification, the description with reference to terms such as "one embodiment", "some embodiments", "examples", "specific examples", or "some examples" means that the specific features, structures, materials, or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms are not necessarily directed to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art can combine and combine the different embodiments or examples described in this specification.

[0211] The above content is a further detailed description of the present invention in combination with specific preferred embodiments, and it cannot be determined that the specific implementation of the present invention is only limited to these descriptions. For those of ordinary skill in the technical field to which the present invention pertains, without departing from the concept of the present invention, several simple deductions or substitutions can still be made, and all should be regarded as belonging to the protection scope of the present invention.

Claims

1. An unsupervised domain adaptation SAR target recognition method that fuses and aligns visual features and scattering topological features, characterized in that, It includes the following steps: Step 100, respectively extract the original source domain images based on the ASC model and the original target domain images for each SAR image X in the scattering topological points and construct the corresponding first scattering topology graph G ASC , obtaining the first source domain scattering topology graph and the first target domain scattering topology graph Step 200, respectively extract the scattering topological points of each SAR image X in the original source domain image and the original target domain image and construct the corresponding second scattering topology graph G SST to obtain the second source domain scattering topology graph and the second target domain scattering topology graph Step 300: Construct a deep recognition network Ψ including a visual feature extraction sub-network, an image reconstruction sub-network, a scattering topology feature extraction sub-network based on ASC, a scattering topology feature extraction sub-network based on SAR-SIFT, and a feature fusion sub-network; Step 400, based on the original source domain image The first source domain scattering topology map The second source domain scattering topology map The original target domain image The first target domain scattering topology map and the second target domain scattering topology map train the depth recognition network Ψ to obtain the trained target depth recognition network Ψ'; Step 500: Input the original target domain image to be recognized, the first target domain scattering topology map to be recognized, and the second target domain scattering topology map to be recognized into the target deep recognition network Ψ′ to obtain a target recognition result.

2. An unsupervised domain adaptation SAR target recognition method for fusing and aligning visual features and scattering topological features according to claim 1, characterized in that, The said step 100 includes: Step 110, for each SAR image X, extract multiple scattering topology points using the ASC model; where each scattering topology point corresponds to a feature vector, and the feature vector corresponding to the i-th scattering topology point is A i represents the complex amplitude, a i represents the frequency-dependent factor, x i and y i represent the range and azimuth position coordinates respectively, L i represents the length of the scattering topology point, and γ i represent the direction angle and azimuth-dependent factor of the scattering topology point respectively; Step 120, decompose the complex amplitude A in the eigenvector corresponding to the scattering topological points of the SAR image X i into the modulus |A i | and the phase Remove the frequency-dependent factor a i and the azimuth-dependent factor γ i , and the eigenvector corresponding to the i-th scattering topological point is expressed as Step 130: Remove the scattering topology points in the background area in the SAR image X whose Euclidean distance from the centroid of multiple scattering topology points is greater than the distance threshold to obtain the first extracted scattering topology points of the SAR image X; Step 140, construct a first scattering topology graph G corresponding to the SAR image X according to the first extracted scattering topology points ASC ; all the first scattering topology graphs of the SAR images in the original source domain image and all the first scattering topology graphs of the SAR images in the original target domain image correspondingly form a first source domain scattering topology graph and a first target domain scattering topology graph 3. An unsupervised domain adaptation SAR target recognition method for fusing and aligning visual features and scattering topological features according to claim 2, characterized in that, The said step 130 includes: Step 131, calculate the centroid of the multiple scattering topology points of the SAR image X Where p represents the number of scattering topology points; Step 132, calculate the Euclidean distance between the coordinates (x i , y i ) corresponding to each scattering topology point of the SAR image X and the centroid . Step 133, remove the Euclidean distance d i that is greater than the distance threshold τ dis of the scattering topological points located in the background area, and obtain the first extracted scattering topological points of the SAR image X.

4. An unsupervised domain adaptation SAR target recognition method for fusing and aligning visual features and scattering topological features according to claim 3, characterized in that, The said step 140 includes: Step 141, taking the first extracted scattering topological points of the SAR image X as the nodes of the graph structure of the first scattering topological graph G ASC wherein, the node set of the graph structure q' is the number of the first extracted scattering topological points; Step 142, calculate the cosine similarity between the eigenvectors corresponding to any two nodes of the SAR image X. If the cosine similarity is greater than or equal to the similarity threshold τ sim then there is an edge between the two nodes; The set of nodes and the set of edges constitute the first scattering topology graph G corresponding to the SAR image X ASC =(V,M); the original source domain image The first scattering topology graphs of all SAR images in and the first scattering topology graphs of all SAR images in the original target domain image correspondingly constitute the first source domain scattering topology graph 5. An unsupervised domain adaptation SAR target recognition method for fusing and aligning visual features and scattering topological features according to claim 1, characterized in that, The said step 200 includes: Step 210: Extract the corner points of each SAR image X based on the SAR-Harris algorithm; Step 220: Extract the strong scattering points of each SAR image X based on the clustering algorithm. The strong scattering points and corner points of a SAR image X together constitute the scattering topology points of the SAR image X; Step 230: Remove the scattering topology points in the background area in the SAR image X whose amplitude values are outside the 1% range with the largest cumulative distribution function to obtain multiple second extracted scattering topology points of the SAR image X; Step 240: Remove the second extracted scattering topology points in the background area in the SAR image X whose Euclidean distance from the centroid of multiple second extracted scattering topology points is greater than the threshold to obtain the third extracted scattering topology points of the SAR image X; Step 250: Generate a SAR-SIFT descriptor as the feature vector of the third extracted scattering topology points of the SAR image X; Step 260: Construct the second scattering topology graph G corresponding to the SAR image X according to the third extracted scattering topology points of the SAR image X SST ; All the second scattering topology graphs of the SAR images in the original source domain image and all the second scattering topology graphs of the SAR images in the original target domain image correspondingly constitute the second source domain scattering topology graph and the second target domain scattering topology graph 6. An unsupervised domain adaptation SAR target recognition method for fusing and aligning visual features and scattering topological features according to claim 5, characterized in that The extracting the strong scattering points of each SAR image X based on the clustering algorithm includes: Step 221: Deform the SAR image X from matrix form to vector form; Step 222: Reorder the vector according to the amplitude values of the SAR image X from large to small; Step 223: Initialize the clustering label of the first element as 0; Step 224: Traverse each element from the second element in descending order; Step 225: When traversing to the i-th element, determine whether the i-th element is within the 8-neighborhood of the previous i - 1 elements; If the i-th element is not within the 8-neighborhood of any of the previous i - 1 elements, then the i-th element is a new category; If the i-th element is only within the 8-neighborhood of a certain one of the previous i - 1 elements, then assign the category label of that element to the i-th element; If the i-th element is within the 8-neighborhood of multiple elements among the previous i - 1 elements, then the category label of the i-th element is marked as the category corresponding to the maximum amplitude value among these multiple elements; Step 226: After the traversal is completed, restore the SAR image X from vector form to matrix form, and abstract each clustering cluster as a strong scattering point.

7. An unsupervised domain adaptation SAR target recognition method for fusing and aligning visual features and scattering topological features according to claim 5, characterized in that, The said step 230 includes: Step 231: Statistically calculate the cumulative distribution function of all pixel amplitudes in the SAR image X; Step 232: Remove the scattering topology points in the background area whose amplitude values are outside the range of the top 1% of the cumulative distribution function, to obtain multiple second-extracted scattering topology points of the SAR image X.

8. An unsupervised domain adaptive SAR target recognition method for fusing and aligning visual features and scattering topological features according to claim 1, characterized in that, The visual feature extraction sub-network includes a first extraction convolutional layer L arranged in sequence C1 , a second extraction activation layer L C2 , a third extraction convolutional layer L C3 , a fourth extraction activation layer L C4 , a fifth extraction convolutional layer L C5 , a sixth extraction activation layer L C6 , a seventh extraction convolutional layer L C7 , an eighth extraction activation layer L C8 , a ninth extraction convolutional layer L C9 and a tenth extraction activation layer L C10 ; The input of the visual feature extraction sub-network is the SAR image X, and the output is a 512-dimensional visual feature vector; The image reconstruction sub-network includes a first reconstruction convolutional layer L' arranged in sequence C1 , a second reconstruction activation layer L' C2 , a third reconstruction convolutional layer L' C3 , a fourth reconstruction activation layer L' C4 , a fifth reconstruction convolutional layer L' C5 , a sixth reconstruction activation layer L' C6 , a seventh reconstruction convolutional layer L' C7 , an eighth reconstruction activation layer L' C8 , a ninth reconstruction convolutional layer L' C9 and a tenth reconstruction activation layer L' C10 ; The input of the image reconstruction sub-network is the 512-dimensional visual feature vector, and the output is the reconstructed SAR image The ASC-based scattering topological feature extraction sub-network includes a first ASC graph convolutional layer L arranged in sequence G1 , a second ASC activation layer L G2 , a third ASC graph convolutional layer L G3 , a fourth ASC activation layer L G4 and a fifth ASC readout layer L G5 ; The input of the scattering topology feature extraction sub-network based on ASC is the first scattering topology graph G ASC , and the output is a 256-dimensional scattering topology feature vector; The SAR-SIFT-based scattering topology feature extraction sub-network includes a first SST graph convolutional layer L' arranged in sequence G1 , a second SST activation layer L' G2 , a third SST graph convolutional layer L' G3 , a fourth SST activation layer L' G4 and a fifth SST readout layer L' G5 ; The input of the SAR-SIFT-based scattering topology feature extraction sub-network is the second scattering topology graph G SST , and the output is a 256-dimensional scattering topology feature vector; The feature fusion sub-network includes a first fusion fully-connected layer L F1 , a second fusion activation layer L F2 , a third fusion fully-connected layer L F3 , and a fourth fusion classifier layer L F4 ; The input of the feature fusion sub-network is a 1024-dimensional feature vector formed by concatenating the 512-dimensional visual feature vector, the 256-dimensional scattering topology feature vector extracted by the scattering topology feature extraction sub-network based on ASC, and the 256-dimensional scattering topology feature vector extracted by the scattering topology feature extraction sub-network based on SAR-SIFT. The output is a 10-dimensional class prediction label y; The visual feature extraction sub-network is connected to the image reconstruction sub-network. The visual feature extraction sub-network, the scattering topology feature extraction sub-network based on ASC, and the scattering topology feature extraction sub-network based on SAR-SIFT are all connected to the feature fusion sub-network.

9. An unsupervised domain adaptation SAR target recognition method for fusing and aligning visual features and scattering topological features, characterized in that, The step 400 includes: Step 410, input the original source domain image The first source domain scattering topology map The second source domain scattering topology map The original target domain image The first target domain scattering topology map and the second target domain scattering topology map into the depth recognition network Ψ for forward propagation to obtain the reconstructed SAR image, the second fusion activation layer L F2 the output fusion features and the class prediction label; Step 420: Calculate the reconstruction loss, and the loss function of the reconstruction loss is: Among them, and respectively represent the i-th original source domain image and the j-th original target domain image, and respectively represent the reconstructed i-th source domain image and the j-th target domain image, n s and n t respectively represent the numbers of source domain images and target domain images, represents the Frobenius norm of the square matrix F; Step 430: Calculate the classification loss, and the loss function of the classification loss is: Among them, represents the class ground truth label in one-hot encoding form of the i-th source domain image, represents the corresponding class prediction label; Step 440: Calculate the domain adaptation loss, and the loss function of the domain adaptation loss is: Among them, Loss global represents the global domain adaptive loss, and Loss local represents the local domain adaptive loss. τ ep represents the training iteration threshold; Step 450: Calculate the total loss, and the calculation formula is: Loss total = Loss rec + Loss cls + Loss da ; Step 460: Update the parameters of the depth recognition network Ψ through backpropagation to obtain the target depth recognition network Ψ'.

Images