SAR image few-sample target identification method driven by electromagnetic scattering characteristics

Abstract

The invention discloses an SAR image few-sample target identification method driven by electromagnetic scattering characteristics, and the method comprises the steps: selecting a scattering center parameter of an SAR aircraft image, and extracting an attribute scattering center parameter through employing an AML algorithm; utilizing attribute scattering center parameter clustering to generate an aircraft target subcomponent structure chart to make an aircraft classification data set; graph construction and graph aggregation are carried out, setting of edge weights between nodes is constrained by a structure connection relation of an aircraft target, and target local scattering topological structure features are extracted by using a graph neural network; and carrying out feature fusion on the local scattering topological structure features extracted by the graph neural network and the global depth features extracted by using the residual network to obtain a target classification result. According to the method, global visual features and local scattering structure features are combined, prior is provided for a data driving method by using known physical knowledge, the interpretability of the model is improved, and the overall classification precision is improved.

Description

Technical Field

[0001] The invention belongs to the field of radar image processing and target classification, and in particular relates to a target recognition and classification method suitable for SAR aircraft images. Background Art

[0002] Synthetic Aperture Radar (SAR) is an active microwave imaging system that can observe the Earth around the clock, in all weather conditions, regardless of weather or illumination. Its versatility in both military and civilian applications has led to its widespread use in a variety of fields, such as disaster monitoring, target identification, and detection. Among these applications, Automatic Target Recognition (ATR) plays a key role in both military and civilian SAR applications and has garnered extensive attention and research.

[0003] In recent years, deep learning technology has achieved breakthroughs in many fields, making it a valuable option for SAR target classification. However, this approach tends to overlook the physical information of SAR targets. Attribute scattering centers (ASCs) have become a research hotspot for SAR target recognition because they assign attribute parameters related to their physical properties and their distribution can represent physical structure. Liu et al. aggregated multiple attribute scattering centers (ASCs) into clustered data rather than individual points, effectively capturing the local electromagnetic properties and spatial structure of the target. Feng et al., based on the target component model extracted from the ASCs, employed a proportional dot product attention mechanism to associate the characteristics of the target components with the classification output, improving the interpretability of SAR ATR. While these methods link ASCs with neural networks, they lack understanding of the ASC distribution relationship and ignore the geometric structure of the target. Li et al. first combined convolutional neural networks and graph convolutional networks for SAR target recognition, modeling each scattering center point as a node to construct graph data to improve target classification accuracy. While this approach considers the local structure of the target, it is limited to a predefined graph structure and fails to further explore the complex distribution relationships of ASC features or the ability to extract features from small sample sizes. Therefore, a target recognition and classification method for SAR aircraft images that can address these technical issues is urgently needed. Summary of the Invention

[0004] Purpose of the invention: The purpose of the present invention is to provide a small-sample target recognition method for SAR images driven by electromagnetic scattering features, which is used for aircraft target recognition and classification in SAR images to improve classification accuracy.

[0005] Technical solution: The electromagnetic scattering feature-driven SAR image small-sample target recognition method described in the present invention specifically includes the following steps:

[0006] (1) Select the attribute scattering center parameters of the SAR aircraft image and use the AML algorithm to extract the attribute scattering center parameters;

[0007] (2) Using the attribute scattering center parameter clustering obtained in step (1) to generate the aircraft target sub-component structure diagram to produce the aircraft classification data set;

[0008] (3) Graph construction and graph aggregation are performed, and the structural connection relationship of the aircraft target is used to constrain the setting of edge weights between nodes, and the graph neural network is used to extract the local scattering topological structure features of the target;

[0009] (4) The local scattering topological structure features extracted by the graph neural network in step (3) are fused with the global depth features extracted by the residual network to obtain the target classification result.

[0010] Furthermore, the process of selecting the attribute scattering center parameters of the SAR aircraft image in step (1) is as follows:

[0011] Assume that the backscatter of a target is approximated as the sum of responses from individual scattering centers as follows:

[0012]

[0013] in, is the attribute set of the scattering center, x i ,y i Indicates the position of the scattering center in azimuth and range, A i Indicates the amplitude, α i represents the frequency dependence factor, L i is the scattering center length, φ i is the scattering center direction angle, γ i Represents the directional dependence of the scattering center on the direction angle; three parameters [A, x, y] that are directly related to the physical structure of the target are selected as the scattering center characteristics.

[0014] Furthermore, the process of extracting the attribute scattering center parameters using the AML algorithm in step (1) is as follows:

[0015] S1: Use the watershed algorithm to segment the image region containing the scattering center response from the image;

[0016] S2: Extract the attribute scattering center parameters corresponding to the scattering center area and use the Newton iteration method to optimize the parameters;

[0017] S3: Use the CLEAN method to eliminate the extracted scattering center area and segment the next scattering center area; until the remaining pixel value after segmentation is less than the set threshold, the estimated parameters are reconstructed to obtain the attribute scattering center model map.

[0018] Furthermore, the process of using the watershed algorithm to segment the image region containing the scattering center response from the image is as follows:

[0019] Calculate the amplitude of the SAR image, find the pixel with the highest amplitude, and determine the -3dB and -40dB regions below the peak. Mark the pixels between 0 and -3dB in descending order of amplitude, and the label value of the pixel to be segmented is determined by the label values of its eight neighboring pixels.

[0020] Determine the number of peaks in the area above the -3dB level of the highest peak, and mark multiple peaks corresponding to the same scattering center with the same value, which is the minimum value among these peak marks;

[0021] The area between the highest peak level of -40dB and the level of -3dB is segmented. The label value of the pixel to be segmented is determined by the label values of its eight neighboring pixels.

[0022] The area below the -40dB level of the highest peak is considered to be clutter and noise and is not segmented; the scattering center area corresponding to the highest peak, that is, the area with a mark value of 1, is extracted and recorded as ROI.

[0023] Furthermore, the implementation process of step (2) is as follows:

[0024] The estimated parameters are clustered into five categories using the K-Means clustering algorithm; then the image is reconstructed based on the parameters to obtain a component structure reconstruction diagram containing the five sub-components of the aircraft target: nose, fuselage, tail, left wing and right wing.

[0025] Furthermore, the graph construction implementation process in step (3) is as follows:

[0026] The five component images of the nose, fuselage, tail, left wing and right wing are regarded as topological nodes. The mask image of the corresponding SAR image is obtained according to the five component images of each SAR image. Then, the five target component areas on the mask image are mapped to the relevant areas on the shallow feature map output by CNN of the original SAR image for feature extraction. The extracted features are used as the node features of each image. The node features are expressed as H = {h1, h2....h N}, N = 5; node feature h i and h j The edge of ij ∈E, The initial weights of the edges between nodes are all set to 1, as the adjacency matrix A∈R N×N , the input value of N=5, that is, A ij =1;

[0027] The structural connection relationship of aircraft components is added as prior knowledge to the network to constrain the edge weight setting: SAR map and corresponding mask Figure 1 Starting from the input, each mask image contains 5 kinds of pixel values after removing the background. Each pixel value corresponds to a serial number, and each serial number corresponds to a component structure, that is, 1 corresponds to the fuselage, 2 corresponds to the nose, 3 corresponds to the right wing, 4 corresponds to the tail, and 5 corresponds to the left wing. Furthermore, the serial number of each node corresponds to an edge, and multiple two-node combinations (i, j)∈I, I={(4,5),(4,2),(4,3),(5,2),(5,3),(2,3)}, corresponding to the fixed edge e ij , and initialize the weight of the fixed edge to 0, that is, the parts are not connected, and the weight of the remaining edges is still 1, so the value of the adjacency matrix is updated as follows:

[0028]

[0029] The spatial relationship of the component structure is modeled using the graph G = {H, E}.

[0030] Furthermore, the graph aggregation implementation process in step (3) is as follows:

[0031] For each node feature h i Applying a shared linear transformation, we get Wh i , W is a learnable weight matrix; a self-attention mechanism is implemented for each node. The attention mechanism makes the edge weights in the graph adaptive, and the size of the weight depends on the feature similarity or correlation between the nodes; the attention coefficient is defined as:

[0032]

[0033] Among them, α is a single-layer feedforward network, which can represent the degree of attention of node i to node j;

[0034] Introduce softmax to normalize all neighbor nodes j of i:

[0035]

[0036] Among them, N i is the neighborhood of node i in the graph, and the above formula is expanded to the complete attention mechanism as follows:

[0037]

[0038] in, It is the new vector after splicing, which contains the node feature information of nodes i and j. Splicing combines the feature information of the two nodes together; It is a learnable weight vector that performs a linear transformation on the concatenated features. The resulting scalar represents the attention score of node i to node j.

[0039] After obtaining the above attention coefficients, predict the output features of each node in the lth layer:

[0040]

[0041] Among them, a ij is the dynamically changing adjacency matrix value;

[0042] A multi-head attention mechanism is introduced. Specifically, M independent attention mechanisms perform the transformation above, and then connect their features to obtain the following output feature representation:

[0043]

[0044] All node features are connected, and then the pooling layer is used to read the features, and the result is represented as f1.

[0045] Furthermore, the implementation process of step (4) is as follows:

[0046] Feature fusion is used to simultaneously learn the deep features f2 extracted by the residual network ResNet50 and the scattering topology structure features f1 extracted by GAT; the fused feature f is:

[0047] f=concat(f1f2)

[0048] Among them, concat represents the connection operation; the fused features are then input into the fully connected layer, and then softmax is used for classification.

[0049] Beneficial effects: Compared with the prior art, the present invention has the following beneficial effects: 1. The present invention uses multi-scale feature enhancement to fully exploit the global depth features and local scattering topological structure features of SAR aircraft targets; 2. The present invention introduces physical prior information; first, the electromagnetic scattering characteristics of SAR images are utilized to define the SAR aircraft component graph as a topological node, and the edge weights represent the component connection relationship. The spatial relationship and semantic information interaction of different groups of components at different positions are modeled. Finally, by fixing the connection relationship of aircraft components as a physical prior constraint for the learning and updating of the graph attention network, the topological structure features of the aircraft target skeleton are better utilized, which increases the interpretability of the model; 3. The present invention constructs a SAR aircraft category dataset, and after experimental comparison, it can be seen that the proposed method obtains superior classification performance under small sample conditions, ensuring the richness and stability of feature expression under this condition. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 is a flow chart of the present invention;

[0051] Figure 2 This is a flowchart of extracting SAR image attribute scattering centers in the present invention;

[0052] Figure 3 This is a flowchart for extracting aircraft target subcomponents in the present invention. DETAILED DESCRIPTION

[0053] The present invention will be further described in detail below with reference to the accompanying drawings.

[0054] The present invention proposes a novel scattering feature-driven aircraft target recognition and classification method. Different from previous methods, the proposed method integrates the electromagnetic scattering characteristics of the components, introduces physical prior information, and utilizes the spatial topological structure relationship of the aircraft components. First, the component image of the aircraft target is constructed based on the attribute scattering center parameter clustering, which utilizes the unique physical information of SAR. Secondly, the method takes into account the spatial topological structure, defines each component image as a topological node, and models the connection relationship between different nodes as an edge, and constrains the setting of edge weights according to the component structure relationship of the aircraft target; uses the graph attention network to convert the scattering information and spatial relationship into a discriminant object representation, extracts the local scattering topological structure features, and improves the robustness and adaptive expression ability of the features in the case of few samples; finally, combines the global visual features extracted by the residual network and the local scattering structure features extracted by the graph attention network, uses known physical knowledge to provide a priori for the data-driven method, improves the interpretability of the model, and improves the overall classification accuracy. Figure 1 The specific steps shown include:

[0055] Step 1: Based on the Approximate Maximum Likelihood (AML) algorithm, the attribute scattering center parameters of the SAR aircraft image are extracted and the electromagnetic scattering characteristics are integrated.

[0056] Based on the SAR aircraft amplitude image, the AML algorithm is used to extract the attribute scattering center parameters of the SAR image. The attribute scattering center model is based on the geometric diffraction theory model and is now widely used in the field of target recognition. This model assumes that the backscattering of a target can be well approximated as the sum of the responses from a single scattering center, as shown below:

[0057]

[0058]

[0059] in, is the attribute set of the scattering center. i ,y i Indicates the position of the scattering center in azimuth and range, A i Indicates the amplitude, α i represents the frequency dependence factor, L i is the scattering center length, φ i is the scattering center direction angle, γ i Indicates the directional dependence of the scattering center on the direction angle. In order to improve the recognition performance of the target, three parameters [A, x, y] directly related to the physical structure of the target are selected as scattering center features for parameter estimation and extraction.

[0060] The AML algorithm is used to estimate the ASC parameters. The flowchart of the AML algorithm is shown in the attached figure. Figure 2 As shown, the iterative process of the AML algorithm can be expressed as:

[0061] S1: Use the watershed algorithm to segment the image region containing the scattering center response from the image:

[0062] Calculate the amplitude of the SAR image, find the pixel with the highest amplitude, and determine the -3dB and -40dB regions below this peak. Then, label the pixels between 0 and -3dB in descending order of amplitude, with the label value of the pixel to be segmented determined by the label values of its eight neighboring pixels.

[0063] Determine the number of peaks contained in the area above the -3dB level of the highest peak, and mark multiple peaks corresponding to the same scattering center with the same value, which is the minimum value among these peak marks.

[0064] The area between the highest peak level of -40dB and the level of -3dB is segmented, and the label value of the pixel to be segmented is determined by the label values of its eight neighboring pixels.

[0065] The area below the highest peak level of -40dB is considered to be clutter and noise and is not segmented.

[0066] The scattering center area corresponding to the highest peak value is extracted, that is, the area with a mark value of 1, and recorded as ROI.

[0067] S2: Extract the attribute scattering center parameters corresponding to the scattering center area and use the Newton iteration method to optimize the parameters.

[0068] S3: Use the CLEAN method to eliminate the extracted scattering center area and segment the next scattering center area. When the remaining pixel value after segmentation is less than the set threshold, the estimated parameters are reconstructed to obtain the attribute scattering center model map.

[0069] Step 2: Construct an aircraft classification dataset containing aircraft target SAR images and corresponding target sub-component mask images.

[0070] The parameter amplitude A and relative position x, y estimated during the iterative process of the ASC parameter calculation in the AML algorithm are extracted. The estimated parameters are clustered into 5 categories using the K-Means clustering algorithm. Then, the image is reconstructed based on the parameters to obtain a component structure reconstruction diagram containing the 5 sub-components of the aircraft target: nose, fuselage, tail, left wing and right wing. The specific flow chart of the method is shown in the attached figure. Figure 3 shown.

[0071] Step 3: Model the graph structure, use the structural connection relationship of the aircraft target to constrain the setting of the edge weights between nodes, and use the graph neural network to extract the local scattering topological structure features of the target.

[0072] (3.1) Use graph neural network to map the correspondence between scattering centers and target structures, and first construct the graph.

[0073] The location and semantic information of different scattering points are crucial for exploring the spatial structure of targets in SAR images. After extracting the attribute scattering centers and clustering and reconstructing them according to their intensity and relative position, five component images of the aircraft target, namely the nose, fuselage, tail, left wing and right wing, are generated, and these five component images are regarded as topological nodes. The mask image of the corresponding SAR image is obtained based on the five sub-component images of each SAR image. The five target component areas on the mask image are then mapped to the relevant areas on the shallow feature map output by CNN of the original SAR image for feature extraction. The extracted features are used as the node features of each image. The node features can be expressed as H = {h1, h2....h N}. Node feature h iand h j The edge of ij ∈E, The initial weights of the edges between nodes are all set to 1, as the adjacency matrix A∈R N×N , the input value of N=5, that is, A ij = 1. In addition, inspired by the topological structure and skeleton characteristics of the aircraft target, the structural connection relationship of the aircraft components is added as prior knowledge. The specific steps are as follows: SAR image and corresponding mask Figure 1 Each mask image, minus the background, contains five pixel values. Each pixel value corresponds to a serial number, and each serial number corresponds to a component structure, i.e., 1 corresponds to the fuselage, 2 corresponds to the nose, 3 corresponds to the right wing, 4 corresponds to the tail, and 5 corresponds to the left wing. Furthermore, each node serial number corresponds to an edge, and multiple two-node combinations (i, j)∈I, I={(4,5),(4,2),(4,3),(5,2),(5,3),(2,3)}, corresponding to a fixed edge e ij , and initialize the weights of these fixed edges to 0, that is, the parts are not connected, and the weights of the remaining edges are still 1, so the value of the adjacency matrix can be updated as follows:

[0074]

[0075] The spatial relationship of the component structure is modeled using a graph G = {H, E}. In the experiment, the adjacency matrix and the calculated initial node features are used as input to the graph neural network.

[0076] (3.2) Use graph neural network to map the correspondence between scattering centers and target structures, and then perform graph aggregation.

[0077] For each node feature h i Applying a shared linear transformation, we get Wh i , where W is a learnable weight matrix. A self-attention mechanism is implemented for each node. The attention mechanism makes the edge weights in the graph adaptive, and the size of the weight depends on the feature similarity or correlation between the nodes. The attention coefficient can be defined as:

[0078]

[0079] Among them, α is a single-layer feedforward network, which can represent the degree of attention of node i to node j.

[0080] In order to make the attention coefficient easier to calculate and compare, softmax is introduced to normalize all neighbor nodes j of i, as shown in the following figure:

[0081]

[0082] N i is the neighborhood of node i in the graph. The above formula can be expanded to a complete attention mechanism as follows:

[0083]

[0084] in, It is a new vector after splicing, which contains the node feature information of nodes i and j. Splicing combines the feature information of the two nodes, allowing the model to consider their features at the same time, so as to better capture the relationship between the nodes and calculate the attention coefficient more accurately. is a learnable weight vector that linearly transforms the concatenated features. The resulting scalar represents the attention score of node i to node j. The LeakyReLU activation function is used to retain some negative information to avoid the vanishing gradient problem during training.

[0085] After obtaining the above attention coefficients, the output features of each node in the lth layer can be predicted:

[0086]

[0087] From the above formula, we can see that the output feature of the node is related to all its adjacent nodes. ij It can be viewed as a dynamically changing adjacency matrix value. In order to stabilize the self-attention learning process and learn different feature representations, a multi-head attention mechanism is introduced. Specifically, the above transformation is performed on M independent attention mechanisms, and their features are then concatenated to obtain the following output feature representation:

[0088]

[0089] This shows that dynamically updating edge weights allows for more accurate node feature learning. All node features are concatenated, then read out using a pooling layer. The result is represented as f1, extracting local scattering topology features.

[0090] Step 4: Feature fusion, the aircraft classification dataset is used as the input data of the proposed target recognition classification model for training, and the target classification result is output.

[0091] Since both the global depth features extracted by the residual network and the local scattering structure features extracted by GAT are necessary, we use the feature fusion module to learn these two features at the same time. The fused features can be expressed as follows f:

[0092] f = concat(f1, f2)

[0093] Here, f2 represents the output of the deep feature extraction module, and concat represents the concatenation operation. The fused features are then fed into the fully connected layer, where softmax is used for classification. The SAR aircraft classification dataset obtained in step S2 is used as input for training the target recognition classification model, resulting in aircraft classification results.

[0094] Table 1 compares the target classification results of the constructed SAR aircraft classification dataset using the classic deep learning method and the new few-sample SAR target recognition method proposed in this invention.

[0095] Table 1 compares the target classification results of the present invention with other deep learning methods.

[0096] method Alexnet VGG16 Vit Vitamin + GAT ResNet50 Proposed method Training 2.5% (10 images per class) 52.83％ 43.87％ 51.83％ 53.07％ 57.91％ 68.68％ Training 5% (20 images per class) 57.54％ 60.55％ 52.59％ 58.72％ 59.90％ 77.69％ Training 10% (40 images per class) 67.21％ 74.42％ 61.79％ 65.80％ 74.76％ 84.01％ Training 15% (60 images per class) 81.54％ 80.62％ 73.58％ 75.48％ 84.90％ 89.38％ Training 20% (80 images per class) 82.84％ 85.79％ 74.29％ 77.89％ 86.98％ 94.51％

[0097] Compared with traditional deep learning methods, the target recognition and classification algorithm proposed in this invention has better algorithm performance, increases the interpretability of the model, ensures the stability of feature expression, and improves classification accuracy.

[0098] The present invention has been described in detail above with reference to specific embodiments. However, these descriptions should not be construed as limiting the present invention. Those skilled in the art will appreciate that various equivalent substitutions, modifications, or improvements may be made to the technical solutions and implementations of the present invention without departing from the spirit and scope of the present invention, all of which fall within the scope of the present invention. The scope of protection of the present invention shall be determined by the appended claims.

Claims

1. A method for target recognition in SAR images with a small number of samples driven by electromagnetic scattering features, characterized in that: The following steps are involved: (1) Select the attribute scattering center parameters of the SAR aircraft image and use the AML algorithm to extract the attribute scattering center parameters; (2) Using the attribute scattering center parameter clustering obtained in step (1) to generate the aircraft target sub-component structure diagram to produce the aircraft classification data set; (3) Graph construction and graph aggregation are performed, and the structural connection relationship of the aircraft target is used to constrain the setting of edge weights between nodes, and the graph neural network is used to extract the local scattering topological structure features of the target; (4) The local scattering topological structure features extracted by the graph neural network in step (3) are fused with the global depth features extracted by the residual network to obtain the target classification result.

2. The electromagnetic scattering feature-driven SAR image small-sample target recognition method according to claim 1, characterized in that: The process of selecting the attribute scattering center parameters of the SAR aircraft image in step (1) is as follows: Assume that the backscatter of a target is approximated as the sum of responses from individual scattering centers as follows: in, is the attribute set of the scattering center, x i ,y i Indicates the position of the scattering center in azimuth and range, A i Indicates the amplitude, α i represents the frequency dependence factor, L i is the scattering center length, φ i is the scattering center direction angle, γ i Represents the directional dependence of the scattering center on the direction angle; three parameters [A, x, y] that are directly related to the physical structure of the target are selected as the scattering center characteristics.

3. The electromagnetic scattering feature-driven SAR image small-sample target recognition method according to claim 1, characterized in that: The process of extracting the attribute scattering center parameters using the AML algorithm in step (1) is as follows: S1: Use the watershed algorithm to segment the image region containing the scattering center response from the image; S2: Extract the attribute scattering center parameters corresponding to the scattering center area and use the Newton iteration method to optimize the parameters; S3: Use the CLEAN method to eliminate the extracted scattering center area and segment the next scattering center area; until the remaining pixel value after segmentation is less than the set threshold, the estimated parameters are reconstructed to obtain the attribute scattering center model map.

4. The electromagnetic scattering feature-driven SAR image small-sample target recognition method according to claim 3, characterized in that: The process of using the watershed algorithm to segment the image region containing the scattering center response from the image is as follows: Calculate the amplitude of the SAR image, find the pixel with the highest amplitude, and determine the -3dB and -40dB regions below the peak. Mark the pixels between 0 and -3dB in descending order of amplitude, and the label value of the pixel to be segmented is determined by the label values of its eight neighboring pixels. Determine the number of peaks in the area above the -3dB level of the highest peak, and mark multiple peaks corresponding to the same scattering center with the same value, which is the minimum value among these peak marks; The area between the highest peak level of -40dB and the level of -3dB is segmented. The label value of the pixel to be segmented is determined by the label values of its eight neighboring pixels. For the area below the highest peak level of -40dB, it is considered as clutter and noise and is not segmented; The scattering center area corresponding to the highest peak value is extracted, that is, the area with a mark value of 1, and recorded as ROI.

5. The electromagnetic scattering feature-driven SAR image small-sample target recognition method according to claim 1, characterized in that: The implementation process of step (2) is as follows: The estimated parameters are clustered into five categories using the K-Means clustering algorithm; then the image is reconstructed based on the parameters to obtain a component structure reconstruction diagram containing the five sub-components of the aircraft target: nose, fuselage, tail, left wing and right wing.

6. The electromagnetic scattering feature-driven SAR image small-sample target recognition method according to claim 1, characterized in that: The graph construction implementation process in step (3) is as follows: The five component images of the nose, fuselage, tail, left wing and right wing are regarded as topological nodes. The mask image of the corresponding SAR image is obtained according to the five component images of each SAR image. Then, the five target component areas on the mask image are mapped to the relevant areas on the shallow feature map output by CNN of the original SAR image for feature extraction. The extracted features are used as the node features of each image. The node features are expressed as H = {h1, h2....h N }, N = 5; node feature h i and h j The edge of ij ∈E, The initial weights of the edges between nodes are all set to 1, as the adjacency matrix A∈R N×N , the input value of N=5, that is, A ij =1; The structural connection relationship of aircraft components is added as prior knowledge to the network to constrain the edge weight setting: the SAR image and the corresponding mask image are input together. Each mask image contains 5 pixel values ​​excluding the background. Each pixel value corresponds to a serial number, and each serial number corresponds to a component structure, that is, 1 corresponds to the fuselage, 2 corresponds to the nose, 3 corresponds to the right wing, 4 corresponds to the tail, and 5 corresponds to the left wing; then, the serial number of each node corresponds to an edge, and multiple two-node combinations (i, j)∈I, I={(4, 5), (4, 2), (4, 3), (5, 2), (5, 3), (2, 3)}, corresponding to the fixed edge e ij , and initialize the weight of the fixed edge to 0, that is, the parts are not connected, and the weight of the remaining edges is still 1, so the value of the adjacency matrix is updated as follows: The spatial relationship of the component structure is modeled using the graph G = {H, E}.

7. The electromagnetic scattering feature-driven SAR image small-sample target recognition method according to claim 1, characterized in that: The graph aggregation implementation process in step (3) is as follows: For each node feature h i Applying a shared linear transformation, we get Wh i , W is a learnable weight matrix; a self-attention mechanism is implemented for each node. The attention mechanism makes the edge weights in the graph adaptive, and the size of the weight depends on the feature similarity or correlation between the nodes; The attention coefficient is defined as: Among them, α is a single-layer feedforward network, which can represent the degree of attention of node i to node j; Introduce softmax to normalize all neighbor nodes j of i: Among them, N i is the neighborhood of node i in the graph, and the above formula is expanded to the complete attention mechanism as follows: in, It is the new vector after splicing, which contains the node feature information of nodes i and j. Splicing combines the feature information of the two nodes together; It is a learnable weight vector that performs a linear transformation on the concatenated features. The resulting scalar represents the attention score of node i to node j. After obtaining the above attention coefficients, predict the output features of each node in the lth layer: Among them, a ij is the dynamically changing adjacency matrix value; A multi-head attention mechanism is introduced. Specifically, M independent attention mechanisms perform the transformation above, and then connect their features to obtain the following output feature representation: All node features are connected, and then the pooling layer is used to read the features, and the result is represented as f1.

8. The electromagnetic scattering feature-driven SAR image small-sample target recognition method according to claim 1, characterized in that: The implementation process of step (4) is as follows: Feature fusion is used to simultaneously learn the deep features f2 extracted by the residual network ResNet50 and the scattering topology structure features f1 extracted by GAT; the fused feature f is: f = concat(f1, f2) Among them, concat represents the connection operation; the fused features are then input into the fully connected layer, and then softmax is used for classification.

Images