Maneuvering Space Target Recognition Method Based on Complex Region Graph Transformer

Abstract

A method for recognizing maneuvering space targets based on a complex region graph transformer relates to the field of inverse synthetic aperture radar image processing. The present invention aims to solve the problem that rapid maneuvers of space targets can cause large posture changes in ISAR images, and since ISAR images usually exist in the form of complex values, general real domain networks can lead to the loss of key phase information in ISAR images. First, the present invention uses a contrastive learning module to train ISAR image block pairs to adapt to changes in the target ISAR image. Secondly, the PGT feature extraction module extracts local and global context features of the image blocks through a graph reasoning method and a Transformer framework to obtain a more effective representation of the target. Finally, the graph recognition module updates the features of the nodes and edges between nodes formed by feature embedding, and outputs the classification probability and classification results of the target.

Description

Technical Field

[0001] The present invention belongs to the field of Inverse Synthetic Aperture Radar (ISAR) image processing. Background Art

[0002] ISAR offers advantages such as high resolution, all-day, long-range imaging, insensitivity to target materials, and the ability to operate despite atmospheric interference. These advantages give it broad application prospects in areas such as space target detection and space situational awareness. However, traditional classification methods rely on artificially designed features, which can lead to problems such as an inability to effectively represent target features and poor robustness to new categories. In space target classification tasks, rapid maneuvers of space targets can cause significant attitude changes. This further exacerbates the high sensitivity to attitude changes inherent in the ISAR imaging mechanism, resulting in greater variability in the resulting images. Furthermore, ISAR images are typically complex-valued, and the real-valued representation of current deep learning networks results in the loss of phase information. This phase information indicates important features such as the target's outline, attitude, and maneuverability. Consequently, this loss of phase information can reduce the accuracy of space target classification. Summary of the Invention

[0003] The present invention aims to solve the problem that rapid maneuvers of space targets will cause large attitude changes in ISAR images, and since ISAR images usually exist in the form of complex values, this will lead to the loss of key phase information in ISAR images. A maneuvering space target recognition method based on complex area graph Transformer is now provided.

[0004] A method for target recognition in maneuvering space based on a complex area graph transformer is proposed. Specifically, the complex area graph transformer module is used to extract a recognition feature matrix of an ISAR image to be recognized in the maneuvering space. Each recognition feature in the recognition feature matrix is ​​concatenated with a one-hot encoding and used as a node to be recognized. The graph recognition module is used to recognize the node to be recognized, and the type of the target in the ISAR image is obtained.

[0005] The complex region graph Transformer module includes a context feature extraction unit, an embedding feature extraction unit and a feature fusion unit;

[0006] The context feature extraction unit is used to segment the ISAR image into N 2 image blocks, N 2 The image blocks are input into the trained complex domain Resnet18 network framework to obtain the node feature matrix. 2The image blocks containing the maneuvering target pixels are screened from the image blocks to create a proximity matrix, and the target local feature graph data created by the node feature matrix and the proximity matrix are updated using the graph reasoning method. The updated graph data is used as the local context feature, and the global context feature in the local context feature is extracted using the Transformer framework, where N is a positive integer;

[0007] The embedded feature extraction unit is used to extract the overall embedded features of the ISAR image using a CNN network;

[0008] The feature fusion unit is used to fuse the global context feature and the overall embedded feature to obtain an identification feature matrix of the ISAR image to be identified in the maneuvering space.

[0009] Furthermore, the ISAR image is segmented into N 2 Image blocks, including:

[0010] The pixels containing the target in the ISAR image are located to obtain the outer rectangular outline of the target, and N overlapping parts are segmented from the outer rectangular outline. 2 image blocks.

[0011] Furthermore, the training process of the complex domain Resnet18 network framework is as follows:

[0012] Each training sample in the training set is simulated and expanded using different imaging parameters to obtain 2M image blocks, and the two image blocks obtained by the simulation expansion of the same training sample are used as the positive image block pair of the training sample;

[0013] The complex domain Resnet18 network framework is trained using 2M image patches;

[0014] The positive image block pair (p i ,p j )’s loss function L i,j The expression is:

[0015]

[0016] in,

[0017] exp[] represents the exponential function with the natural constant e as the base, T is the temperature coefficient, u i 、u j 、u m are image blocks p i 、p j 、p m Features, and denote the real and imaginary parts of the complex domain data respectively, M is the number of training samples in the training set, m=1,2,...,M, i=1,2,...,M, j=1,2,...,M,

[0018] The imaging parameters include: an accumulation angle that affects the resolution of the ISAR image, a rotation acceleration that affects the defocus degree of the ISAR image, and a random image rotation angle.

[0019] Furthermore, based on the set threshold in N 2 The image blocks containing the maneuvering target pixels are filtered out from the image blocks, and the node features corresponding to each filtered image block are taken as a node. α With node N β Create an edge A between αβ , thereby creating the target local feature map data, the node N α and node N β The corresponding image block p α and image patch p β Satisfies: image block p α Located in image block p β All edges created within the 8-neighborhood form the proximity matrix

[0020] Furthermore, the node N in the target local feature map data is transformed into β Update to Node

[0021]

[0022] Among them, CvReLU[] represents the ReLU function operation on the real and imaginary parts of the input data respectively, and CvBN1d() represents the one-dimensional complex BatchNorm function.

[0023]

[0024] w β is the weight, and Represents the real and imaginary parts of complex domain data, α=1,2,...,N 2 , β=1,2,...,N 2 .

[0025] Furthermore, the Transformer framework is used to extract global context features from local context features, including:

[0026] A multiple-head self-attention mapping mechanism is used to query the global context features of each image block among all image blocks.

[0027] Furthermore, the multi-head self-attention mapping mechanism function MSA() in the Transformer framework is expressed as:

[0028]

[0029] SA(x)=A w v,

[0030]

[0031] Among them, x represents the input of the Transformer framework, q, k, and v are the query words, keywords, and numerical words of the plural multi-head self-attention mapping mechanism respectively. To generate the weights of the complex domain fully connected layer of q, k, v, SA() represents the self-attention function, A w is the self-attention weight, is the weight of the complex domain fully connected layer that generates the multi-head self-attention result, D k is the number of self-attention operations in the multi-head self-attention mapping mechanism.

[0032] Furthermore, the mapping relationship between the target local feature map data and the Transformer framework is:

[0033]

[0034] in, MLP[] represents the multi-layer perception function, CvLN() represents the complex LayerNorm function, L is the number of layers in the Transformer framework, l = 1, 2, ..., L, o l+1 and o l Respectively represent the output of the l+1th and lth layers in the Transformer framework. When l=1, Indicates the Nth 2 Local features of image patches.

[0035] Furthermore, the global context features and the overall embedding features are fused by the following formula:

[0036]

[0037] Among them, E f is the fusion result of global context features and embedded features, E w is the overall embedding feature, E p is the global context feature, is the weight of the complex domain fully connected layer that generates fusion features.

[0038] Furthermore, the step of using the graph recognition module to identify the node to be identified includes:

[0039]

[0040] Among them, η is the loss, y k is the label represented by the one-hot encoding, Υ is the target true category, P() represents the probability of target classification, Y * is the node of the graph data, and k is the number of categories.

[0041] Beneficial effects of the present invention:

[0042] The present invention proposes a CvPGT (Complex-valued Patch Graph Transformer) recognition network framework for recognizing ISAR images of small-sample maneuvering space targets. First, the comparative learning module trains ISAR images to adapt to the changes in the target ISAR images. Secondly, the PGT feature extraction module extracts local and global context features of image patches through graph reasoning methods and the Transformer framework to obtain a more effective representation of the target. Finally, the graph recognition module updates the features of nodes and edges between nodes formed by feature embedding, and outputs the classification probability and classification results of the target. Compared with other small-sample recognition models, the proposed CvPGT model shows stronger robustness and higher classification accuracy for space target images with large posture changes. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 Schematic diagram of the training process of the contrastive learning module;

[0044] Figure 2 This is a schematic diagram of the structure of the CvPGT feature extraction module;

[0045] Figure 3 Schematic diagram of the overall feature embedding network;

[0046] Figure 4 (a) Figure 4 (b) Figure 4 (c) The original ISAR images generated by Sentinel-5, Skynet-5, and CRS-1 simulations, respectively. Figure 4 (d) Figure 4 (e) Figure 4 (f) are the attention maps of the network with PGT extraction module, Figure 4 (g) Figure 4 (h) Figure 4(i) Attention maps of the network with PGT extraction module respectively. DETAILED DESCRIPTION

[0047] The following will be combined with the accompanying drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. It should be noted that the embodiments of the present invention and the features in the embodiments can be combined with each other in the absence of conflict.

[0048] This embodiment is a method for identifying maneuverable space targets based on a complex area graph Transformer. The specific implementation is as follows:

[0049] Step 1: A contrastive learning module based on a self-supervised learning strategy is used to extract the embedding features of ISAR image patches.

[0050] use Figure 1 Imaging parameters in The ISAR image of the space target is simulated, and then the ISAR image blocks are segmented to achieve sample expansion. Figure 1 As shown, represents the accumulation angle, which affects the resolution of the ISAR image; ω r represents the rotational acceleration that affects the defocus degree of the ISAR image, θ r Represents the image rotation angle generated randomly in a small range. Its change allows the network to adapt to the posture transformation of the target within a certain range.

[0051] Different imaging parameters are used to simulate and expand each training sample in the training set. Each simulated image block is expanded into two image blocks (p i and p j ), and the two image blocks obtained by the simulated expansion of the same training sample are regarded as the positive image block pair of the training sample. For a training set with N samples, there are a total of 2N expanded image blocks, of which the i-th sample has two positive image blocks, and the remaining 2N-2 expanded image blocks are regarded as negative image blocks of the i-th sample.

[0052] The complex domain Resnet18 network framework is constructed in the contrastive learning module, and the maximum distance between the extended block pairs of the same ISAR image block is calculated through contrastive learning loss, so as to effectively embed the image block into the latent space (u i and u j )middle.

[0053] The positive image block pair (p i ,p j )’s loss function L i,j The expression is defined as:

[0054]

[0055] Where,

[0056]

[0057]

[0058] exp[] represents the exponential function with the natural constant e as the base, T is the temperature coefficient, u i and u j are image blocks p i and p j Features, and Represent the real and imaginary parts of complex domain data respectively.

[0059] After convergence, the parameters of the complex domain ResNet18 network framework are retained and used to extract embedded features in the PGT module. Through the contrastive learning process, the network can obtain more robust feature embeddings despite the target's pose changes and maneuvers.

[0060] Step 2: Extract local and global context features of the image block through the graph reasoning method and the Transformer framework to obtain a more effective representation of the target. The feature extraction process is processed in parallel through two routes, namely, region-based feature extraction for image blocks and global feature extraction for the entire image, as shown in Figure 2 shown.

[0061] The first route is to locate the pixels containing the target in the ISAR image during training and testing to obtain the outer rectangular outline of the target, and segment N pixels with overlapping parts in the rectangular outline. 2 The network uses the feature vectors obtained from these image patches as node features to construct graph data.

[0062] In the contrastive learning process, the pre-trained model extracts node features to form a feature matrix where f i Represents a D-dimensional embedding vector. In this approach, the embedding vector corresponding to each image block is regarded as a node in the graph data, and then the edges between image blocks are calculated based on the spatial relationship and the information contained in the image blocks.

[0063] If the image block P i Located in image block P jIf the sum of the pixel amplitudes of the two image blocks is greater than the preset threshold, it proves that there are some parts of the spatial target in the two image blocks, rather than pure background. Then, the PGT module is at node N. i and node N j Create an edge A between ij =1,A ji =1, otherwise A ij =0,A ji = 0. The above node feature matrix and the proximity matrix composed of edges are used to create graph data that can fully represent the local features of the spatial target.

[0064] The module implements a graph reasoning approach (GRA) to update graph data. The information propagation and aggregation process can be expressed as:

[0065]

[0066] in,

[0067]

[0068]

[0069] In the formula, CvReLU[] represents the ReLU function operation on the real and imaginary parts of the input data respectively. is the proximity matrix formed by all edges in the target local feature map data, N j is the input of the jth layer of GRA, and the first input is initialized by the node feature matrix, w j Represents a learnable weight, CvBN1d() represents a one-dimensional complex BatchNorm function, described as:

[0070]

[0071] Represents the i-th output result of the GRA l+1 layer network. The covariance matrix V l Expressed as:

[0072]

[0073] in, It is the i-th input of the l-th layer of GRA, E() means to calculate the mean of the data in the batch size dimension, and Cov() means to calculate the covariance of the data.

[0074] In addition, this embodiment adds a complex domain Transformer layer in the PGT module to extract global context features to solve the problem that the above-mentioned node update process easily causes the classification results to fall into local optimality.

[0075] The Transformer uses a self-attention (SA) mechanism to determine which patches in the ISAR image should be given more global attention by allowing them to interact. The Transformer framework internally uses a complex-valued qkv mapping mechanism, which receives as input the one-dimensional node features generated by the graph inference process described above. This mechanism constructs a query term q, a keyword k, and a numerical term v through a complex-domain linear mapping function to query the importance of a particular patch of interest to the overall recognition task among all patches. Assuming x is the input set of one-dimensional node features, the self-attention computation process in the Transformer framework can be described as follows:

[0076]

[0077]

[0078] SA(x)=A w v (3)

[0079]

[0080] Among them, A w Is the attention weight, which can represent the similarity between two image blocks in the image block set, which is calculated by their corresponding q and k. MSA() represents the multi-head self-attention mapping mechanism function in the Transformer framework. The network performs D k By concatenating their outputs and multiplying them by weights, we can obtain richer global importance information. is the weight of the complex domain fully connected layer that generates the multi-head self-attention results, is the weight of the complex domain fully connected layer that generates q, k, and v.

[0081] Finally, the mapping output from graph space to Transformer space can be expressed as:

[0082]

[0083]

[0084]

[0085] Where G represents the graph space, CvLN[] is the complex LayerNorm function, and MLP[] represents the multi-layer perceptron function. L is the number of layers in the Transformer framework, where each layer consists of an MSA layer and an MLP layer, and l = 1, 2, ..., L.

[0086] In the second route of the PGT module, the network uses an embedding layer to extract the features of the entire ISAR image. The detailed structure is as follows: Figure 3 As shown. Then embed the feature E w With the output feature E of the first route p Connect together to form the final representation feature E f , which can be expressed as:

[0087]

[0088]

[0089] Among them, the function fla() means flattening the output of the Transformer space to one dimension. is the weight of the fully connected layer whose flattened output becomes the complex domain feature, It is the weight of the complex domain fully connected layer where the connected features become the final output features.

[0090] By combining the features of local image patches with the embedded features of the entire image, the network can avoid falling into local optimality while continuing to focus on the local region features of the target.

[0091] Step 3:

[0092] The final output features are first concatenated with a one-hot encoding, which represents the relative category of the target in the recognition task data. The concatenated result is considered a node in the recognition process. Then, the proximity matrix between the nodes is constructed as follows:

[0093]

[0094] in, It is the i-th node of the k-th layer in the graph recognition module, abs() means calculating the absolute value, is a multi-layer perceptron, Represents the corresponding relationship of nodes. The processed adjacency matrix B is obtained by performing a Softmax function operation on the weight values ​​of the previous layer. The node update process is as follows:

[0095]

[0096] Among them, G c() is the graph recognition layer, θ is the learnable parameter, and ρ() is the nonlinear activation layer, in which the Leaky-ReLU function is used.

[0097] The cross entropy loss function is used to identify nodes, which is expressed as:

[0098]

[0099] η is the loss, y k is the label represented by the one-hot encoding, Υ is the target true category, P() represents the probability of target classification, Y * is the node of the graph data, and k is the number of categories.

[0100] Finally, the model predicts the label Y corresponding to the target and obtains the recognition result x∈Υ.

[0101] The network framework proposed in this embodiment uses eight types of space target models to generate ISAR images, including MRO, Sentinel-5, Skynet-5, trace, Yantar-4KS1, CRS-1, SZ-5, and CZ-1. During the simulation, the carrier frequency of the transmitted linear frequency modulation (LFM) signal is set to 10 GHz, the bandwidth is set to 5 GHz, the pulse repetition frequency (PRF) is set to 5000, and the observation radar is positioned 2000 meters from the target. To verify the robustness of the proposed CvPGT framework to large attitude changes, the relative angle between the target and the observation radar is continuously varied from 0 to 360 degrees, and a 224×224 ISAR image is generated every 12 degrees. 30 images of each target type are used for the experiment.

[0102] To verify the robustness of the proposed method, four different datasets were simulated using four combinations of two imaging parameters, including rotation speed (RS) and rotation acceleration (RA). These two parameters affect the cumulative angle of ISAR imaging, resulting in different resolutions of ISAR images. In addition, rotation acceleration can also cause defocusing of ISAR images, which affects the target recognition results. This patent divides the dataset into a training set and a test set, which contain different categories of space targets. The training set only contains a dataset with one imaging parameter combination, and the test set contains a dataset with all the above imaging parameter combinations; this allows testing data with different image resolutions and defocus levels to demonstrate the effectiveness and robustness of the proposed CvPGT framework.

[0103] In the small sample recognition task of this embodiment, the training and testing processes share the same C-way, K-shot task mode. trainWhen , we first randomly select C target categories from the training set and randomly select K images for each category to form a support set. Then we randomly select a class from the above classes, and in this class, we select one more image as the labeled query set Q train The labels are calculated together with the network output to generate the supervised learning loss. The above process is repeated B times to generate training tasks, where B represents the batch size.

[0104] In the construction of the test task, four different test tasks are generated using datasets with different imaging parameters. test and query set Q test The construction of is similar to the above process. The difference is that Q test The images in are unlabeled.

[0105] The above parameters include: C is set to 3, and K is set to 1 or 5. The deep learning training process is based on the Pytorch framework, using the Xavier method to initialize weights and a hierarchical learning rate setting strategy. For the pre-trained feature embedding network, the learning rate is set to 0.00005 and the weight decay is set to 0; for other layers, the learning rate is set to 0.001 and the weight decay is set to 1e-6. A learning rate decay strategy is adopted, with the learning rate decaying to 0.8 times the original value every 100 iterations. The number of training iterations is set to 3000. The batch size is set to 32, and the AdamW optimizer is used during optimization. During testing, 100 independent repeated experiments were conducted, and the reliability of the model was verified by observing the distribution of test results.

[0106] For four different datasets, when the task changes from 1-shot mode to 5-shot mode, the minimum recognition accuracy increases from 42.22% to 81.11%. In 1-shot mode, the average classification accuracy varies between 70% and 80%. In 5-shot mode, the average classification accuracy exceeds 95%, which proves the effectiveness of the proposed CvPGT model.

[0107] The variance of the recognition accuracy distribution in the 5-shot mode is smaller than that in the 1-shot mode. The accuracy in the 5-shot mode is more concentrated above 95%, indicating that increasing the amount of data in the support set can significantly improve recognition accuracy. This also shows that the proposed CvPGT model is highly robust to the resolution and defocus of ISAR data, demonstrating the advantages of the proposed contrastive learning and its effectiveness for data with large pose variations.

[0108] In addition, combined Figure 4In order to further intuitively see the role of the complex Transformer framework in the PGT extraction module in the recognition process, the Grad-CAM method is used to calculate the network attention graph before and after the complex domain Transformer framework is added. Figure 4 As shown in the figure, it can be seen that when the Transformer framework is not used, the proposed network's attention to ISAR images is relatively narrow and focused on the local area of ​​the target. After adding multiple Transformer layers, the network's attention is more evenly distributed on the target, resulting in a more global view. In addition, the Transformer layer can also avoid recognition errors caused by the network's excessive focus on the local area of ​​the target.

[0109] In summary, this embodiment addresses the problem of large changes in the target's posture on ISAR images caused by maneuvering space targets and the problem of small sample recognition caused by imaging conditions. A complex domain graph data-based recognition network framework, the complex region graph transformer, is proposed to recognize ISAR images of maneuvering space targets. The network can extract effective features within the image, solving the problem of poor space target recognition, and the complex domain network can avoid the loss of key phase information on ISAR images. This embodiment constructs a feature extraction module that can extract rich spatial information and effective feature expressions from the internal regions of the image; uses graph information reasoning methods and complex domain transformer structures to extract contextual features between image regions to overcome the recognition difficulties caused by large changes in the posture of space targets; in addition, the proposed PGT feature extraction module also incorporates a contrastive learning method to reduce the impact of image defocus caused by maneuvering and image resolution changes caused by changes in imaging parameters on the recognition results. This embodiment achieves high-precision classification and recognition of ISAR images of maneuvering space targets with large posture changes under small sample conditions.

[0110] Although the present invention is described herein with reference to specific embodiments, it should be understood that these embodiments are merely illustrative of the principles and applications of the invention. It should be understood that many modifications may be made to the illustrative embodiments, and that other arrangements may be devised, without departing from the spirit and scope of the invention as defined by the appended claims. It should be understood that the various dependent claims and features described herein may be combined in ways other than those described in the original claims. It should also be understood that features described in conjunction with individual embodiments may be used in conjunction with other described embodiments.

Claims

1. A maneuverable space target recognition method based on a complex region graph transformer, characterized by: Specifically, the complex area graph Transformer module is used to extract the recognition feature matrix of the ISAR image to be identified in the maneuvering space, each recognition feature in the recognition feature matrix is ​​concatenated with a one-hot encoding and used as a node to be identified, and the graph recognition module is used to identify the node to be identified to obtain the type of the target in the ISAR image; The complex region graph Transformer module includes a context feature extraction unit, an embedding feature extraction unit and a feature fusion unit; The context feature extraction unit is used to segment the ISAR image into N 2 image blocks, N 2 The image blocks are input into the trained complex domain Resnet18 network framework to obtain the node feature matrix. 2 The image blocks containing the maneuvering target pixels are screened from the image blocks to create a proximity matrix, and the target local feature graph data created by the node feature matrix and the proximity matrix are updated using the graph reasoning method. The updated graph data is used as the local context feature, and the global context feature in the local context feature is extracted using the Transformer framework, where N is a positive integer; The embedded feature extraction unit is used to extract the overall embedded features of the ISAR image using a CNN network; The feature fusion unit is used to fuse the global context feature and the overall embedded feature to obtain an identification feature matrix of the ISAR image to be identified in the maneuvering space.

2. The method for maneuvering space target recognition based on complex region graph transformer according to claim 1, characterized in that: The ISAR image is segmented into N 2 Image blocks, including: The pixels containing the target in the ISAR image are located to obtain the outer rectangular outline of the target, and N overlapping parts are segmented from the outer rectangular outline. 2 image blocks.

3. The method for maneuvering space target recognition based on complex region graph Transformer according to claim 1, characterized in that: The training process of the complex domain Resnet18 network framework is as follows: Each training sample in the training set is simulated and expanded using different imaging parameters to obtain 2M image blocks, and the two image blocks obtained by the simulation expansion of the same training sample are used as the positive image block pair of the training sample; The complex domain Resnet18 network framework is trained using 2M image patches; The positive image block pair (p i ,p j )’s loss function L i,j The expression is: in, exp[] represents the exponential function with the natural constant e as the base, T is the temperature coefficient, u i 、u j 、u m are image blocks p i 、p j 、p m Features, and denote the real and imaginary parts of the complex domain data respectively, M is the number of training samples in the training set, m=1,2,...,M, i=1,2,...,M, j=1,2,...,M, The imaging parameters include: an accumulation angle that affects the resolution of the ISAR image, a rotation acceleration that affects the defocus degree of the ISAR image, and a random image rotation angle.

4. The method for maneuvering space target recognition based on complex region graph Transformer according to claim 1, characterized in that: Based on the set threshold in N 2 The image blocks containing the maneuvering target pixels are filtered out from the image blocks, and the node features corresponding to each filtered image block are taken as a node. α With node N β Create an edge A between αβ , thereby creating the target local feature map data, the node N α and node N β The corresponding image block p α and image patch p β Satisfies: image block p α Located in image block p β All edges created within the 8-neighborhood form the proximity matrix 5. The method for maneuvering space target recognition based on complex region graph transformer according to claim 4, characterized in that: The node N in the target local feature map data is converted into β Update to Node Among them, CvReLU[] represents the ReLU function operation on the real and imaginary parts of the input data respectively, and CvBN1d() represents the one-dimensional complex BatchNorm function. w β is the weight, and Represents the real and imaginary parts of complex domain data, α=1,2,...,N 2 , β=1,2,...,N 2 .

6. The method for maneuvering space target recognition based on complex region graph transformer according to claim 4, characterized in that: The Transformer framework is used to extract global context features from local context features, including: A multiple-head self-attention mapping mechanism is used to query the global context features of each image block among all image blocks.

7. The method for maneuvering space target recognition based on complex region graph transformer according to claim 6, characterized in that: The multi-head self-attention mapping mechanism function MSA() in the Transformer framework is expressed as: SA(x)=A w in, Among them, x represents the input of the Transformer framework, q, k, and v are the query words, keywords, and numerical words of the plural multi-head self-attention mapping mechanism respectively. To generate the weights of the complex domain fully connected layer of q, k, v, SA() represents the self-attention function, A w is the self-attention weight, is the weight of the complex domain fully connected layer that generates the multi-head self-attention result, D k is the number of self-attention operations in the multi-head self-attention mapping mechanism.

8. The method for maneuvering space target recognition based on complex region graph Transformer according to claim 6, characterized in that: The mapping relationship between the target local feature map data and the Transformer framework is: in, MLP[] represents the multi-layer perception function, CvLN() represents the complex LayerNorm function, L is the number of layers in the Transformer framework, l = 1, 2, ..., L, o l+1 and o l Respectively represent the output of the l+1th and lth layers in the Transformer framework. When l=1, Indicates the Nth 2 Local features of image patches.

9. The method for maneuvering space target recognition based on complex region graph Transformer according to claim 1, characterized in that: The global context features and the overall embedding features are fused through the following formula: Among them, E f is the fusion result of global context features and embedded features, E w is the overall embedding feature, E p is the global context feature, is the weight of the complex domain fully connected layer that generates fusion features.

10. The method for maneuvering space target recognition based on complex region graph Transformer according to claim 1, characterized in that: The method of using the graph recognition module to identify the node to be identified includes: Among them, η is the loss, y k is the label represented by the one-hot encoding, Υ is the target true category, P() represents the probability of target classification, Y * is the node of the graph data, and k is the number of categories.

Images