Small-Sample HRRP Radar Target Recognition Method, Device, Equipment and Medium

Abstract

The present application relates to a small-sample HRRP radar target recognition method, device, equipment and medium. After normalizing the one-dimensional high-resolution range image sequence of unknown targets with a small number of samples, a static adjacency matrix is first constructed by using radar scattering physical prior knowledge, and then a dynamic adjacency matrix is generated through a feature enhancement network. After fusing the static and dynamic adjacency matrices, a graph neural network with hierarchical adaptive meta-learning is used for the training of target recognition, so as to achieve high-precision recognition of position targets with less sample data.

Description

Technical Field

[0001] The present application relates to the technical field of radar signal processing, and particularly to a small-sample HRRP radar target recognition method, device, equipment, and medium. Background Technique

[0002] High-resolution range profile (HRRP) forms a one-dimensional feature signal by capturing the electromagnetic scattering center distribution of a target along the radar line of sight. It has advantages such as all-weather operation, strong anti-interference ability, and compact data structure, and is the core means of radar target recognition. Traditional methods such as template matching and Principal Component Analysis-Support Vector Machine (PCA-SVM) rely on artificial feature extraction and are difficult to handle the sensitivity of HRRP to target attitude angle, range migration, and noise.

[0003] However, although deep learning methods (such as convolutional neural networks and recurrent neural networks) have improved the feature expression ability, they face two major challenges in practical applications: one is that the dynamic migration of scattering centers due to target attitude changes makes it difficult for static models to capture non-Euclidean space relationships; the other is the scarcity of labeled data for new target categories, and overfitting is prone to occur in conventional training. Summary of the Invention

[0004] Based on this, it is necessary to provide a small-sample HRRP radar target recognition method, device, equipment, and medium that can accurately recognize unknown targets under the premise of scarce samples for the above technical problems.

[0005] A small-sample HRRP radar target recognition method, the method includes:

[0006] Obtain a small-sample HRRP data set about the same unknown target, where the small-sample HRRP data set includes a finite number of one-dimensional high-resolution range profile sequences, and perform normalization processing on each of the one-dimensional high-resolution range profile sequences to obtain a normalized sequence;

[0007] For each of the normalized sequences, use the prior of radar scattering physics to construct a corresponding static adjacency matrix;

[0008] Adopt a feature enhancement network to enhance the features of each of the normalized sequences to obtain an enhanced feature matrix, and then generate a dynamic adjacency matrix according to each of the enhanced feature matrices;

[0009] After fusing the relevant dynamic adjacency matrix and static adjacency matrix, perform normalization processing to obtain a mixed adjacency matrix;

[0010] Input each of the mixed adjacency matrices into a graph neural network in sequence and perform forward propagation. Hierarchical adaptive meta-learning is carried out in each graph convolutional layer of the graph neural network to identify the unknown target through a small-sample HRRP data set.

[0011] In one embodiment, when normalizing the one-dimensional high-resolution range image sequence, the following operations are performed on the samples at each range cell in the one-dimensional high-resolution range image sequence:

[0012] ;

[0013] In the above formula, represents the one-dimensional high-resolution range image sequence, where represents the number of range cells, represents the set of real numbers, , .

[0014] In one embodiment, the static adjacency matrix includes multiple elements, where the element is expressed as:

[0015] ;

[0016] In the above formula, the element represents the static association strength between each range cell and , represents the static mixing coefficient, represents the local neighborhood window size, represents the indicator function.

[0017] In one embodiment, the feature enhancement network includes three sequentially connected one-dimensional convolutional layers, and a batch normalization layer and a LeakyReLU activation function are connected after each one-dimensional convolutional layer;

[0018] The first one-dimensional convolutional layer is provided with 16 filters, and the kernel size of each filter is 5, and the padding value is 2;

[0019] The second one-dimensional convolutional layer is provided with 32 filters, and the kernel size of each filter is 3, and the padding value is 1;

[0020] The third one-dimensional convolutional layer is provided with 64 filters, and the kernel size of each filter is 3, and the padding value is 1.

[0021] In one embodiment, when generating a dynamic adjacency matrix according to each of the enhanced feature matrices:

[0022] Use the multi-head attention mechanism to calculate the query matrix, key matrix, and value matrix according to the enhanced feature matrix;

[0023] Calculate the in - head attention score matrix according to the query matrix and the key matrix, and then generate the dynamic adjacency matrix according to the in - head attention score matrix.

[0024] In one embodiment, generating the dynamic adjacency matrix according to the in - head attention score matrix is expressed as:

[0025] ;

[0026] In the above formula, represents the in - head attention score matrix, represents the number of attention heads in the multi - head attention mechanism.

[0027] In one embodiment, fusing the relevant dynamic adjacency matrix and static adjacency matrix is expressed as:

[0028] ;

[0029] In the above formula, represents the dynamic mixing coefficient, represents the dynamic adjacency matrix, represents the static adjacency matrix.

[0030] This application also provides a small - sample HRRP radar target recognition device, and the device includes:

[0031] A small - sample data set acquisition module, configured to acquire a small - sample HRRP data set about the same unknown target. The small - sample HRRP data set includes a finite number of one - dimensional high - resolution range image sequences, and perform normalization processing on each of the one - dimensional high - resolution range image sequences to obtain a normalized sequence;

[0032] A static adjacency matrix construction module, configured to construct a corresponding static adjacency matrix for each of the normalized sequences by using the prior knowledge of radar scattering physics;

[0033] A dynamic adjacency matrix generation module, configured to perform feature enhancement on each of the normalized sequences by using a feature enhancement network to obtain an enhanced feature matrix, and then generate a dynamic adjacency matrix according to each of the enhanced feature matrices;

[0034] A mixed adjacency matrix obtaining module, configured to fuse the relevant dynamic adjacency matrix and static adjacency matrix, and then perform normalization processing to obtain a mixed adjacency matrix;

[0035] The small-sample unknown target recognition module is used to sequentially input each of the mixed adjacency matrices into a graph neural network and perform forward propagation, and perform hierarchical adaptive meta-learning in each graph convolutional layer in the graph neural network to recognize the unknown target through a small-sample HRRP data set.

[0036] A computer device includes a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, the steps of the above-mentioned small-sample HRRP radar target recognition method are implemented.

[0037] A computer-readable storage medium stores a computer program thereon, and when the computer program is executed by a processor, the steps of the above-mentioned small-sample HRRP radar target recognition method are implemented.

[0038] For the above-mentioned small-sample HRRP radar target recognition method, device, equipment and medium, after normalizing the one-dimensional high-resolution range image sequence of the unknown target with a small number of samples, first construct a static adjacency matrix using the prior of radar scattering physics, then generate a dynamic adjacency matrix through a feature enhancement network, fuse the static and dynamic adjacency matrices, and then adopt a graph neural network with hierarchical adaptive meta-learning for target recognition training, so as to achieve high-precision recognition of the position target with less sample data. Description of the Drawings

[0039] Figure 1 It is a schematic flowchart of the small-sample HRRP radar target recognition method in one embodiment;

[0040] Figure 2 It is a schematic framework diagram of the small-sample HRRP radar target recognition method in one embodiment;

[0041] Figure 3 It is a schematic diagram of the comparison of recognition performance under different signal-to-noise ratios in a simulation experiment;

[0042] Figure 4 It is a structural block diagram of the small-sample HRRP radar target recognition device in one embodiment;

[0043] Figure 5 It is an internal structure diagram of a computer device in one embodiment. Detailed Embodiments

[0044] In order to make the purpose, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0045] In view of the problems existing in the existing HRRP recognition methods, such as insufficient modeling of dynamic scattering characteristics, low optimization efficiency of meta-learning, and poor noise robustness in the small-sample scenario, in this application, as Figure 1 shown, a small-sample HRRP radar target recognition method is provided, including the following steps:

[0046] Step S100: Obtain a small-sample HRRP data set of the same unknown target. The small-sample HRRP data set includes a finite number of one-dimensional high-resolution range profile sequences, and perform normalization processing on each one-dimensional high-resolution range profile sequence to obtain a normalized sequence.

[0047] Step S110: For each normalized sequence, construct a corresponding static adjacency matrix by using the prior knowledge of radar scattering physics.

[0048] Step S120: Use a feature enhancement network to enhance the features of each normalized sequence to obtain an enhanced feature matrix, and then generate a dynamic adjacency matrix according to each enhanced feature matrix.

[0049] Step S130: After fusing the relevant dynamic adjacency matrix and static adjacency matrix, perform normalization processing to obtain a mixed adjacency matrix.

[0050] Step S140: Input each mixed adjacency matrix into the graph neural network in turn and perform forward propagation, and perform hierarchical adaptive meta-learning in each graph convolutional layer in the graph neural network to identify the unknown target through the small-sample HRRP data set.

[0051] In step S100, a small-sample HRRP data set of the same unknown target is obtained. The small-sample HRRP data set includes a finite number of one-dimensional high-resolution range profile sequences. Here, 'finite number' or'small sample' specifically refers to that compared with the large-scale data set required for training traditional deep learning models, the number of available labeled samples for each target category is extremely limited. For example, in a single task of meta-learning, there may be only a single-digit or double-digit number of sample sequences, which is exactly the technical difficulty to be solved by the present invention.

[0052] In this embodiment, when performing normalization processing on the one-dimensional high-resolution range profile sequence, the following operations are performed on the samples in each range cell of the one-dimensional high-resolution range profile sequence:

[0053] (1)

[0054] [[ID=,30]]In formula (1), represents the one-dimensional high-resolution range profile sequence, where represents the number of range cells, represents the set of real numbers, , The normalized signal is used as the input for subsequent steps.

[0055] In step S110, based on the physical prior of radar scattering, a static adjacency matrix is defined , which is constructed to incorporate the physical prior knowledge of radar scattering. The physical structure of the target determines that there are often inherent physical correlations among spatially adjacent range cells. The static adjacency matrix aims to encode this relatively stable structural information based on physical distance and scattering mechanism, providing a basic graph structure with physical meaning for the model, which is particularly important in small-sample scenarios with sparse training data and can play a certain regularization and guiding role. The static adjacency matrix includes multiple elements for representing the static correlation strength between range cells , expressed as:

[0056] (2)

[0057] In formula (2), the element represents the static correlation strength between each range cell and , represents the static mixing coefficient, represents the local neighborhood window size, that is, the maximum range cell span of adjacent scattering centers. represents the indicator function, which outputs 1 when , otherwise 0.

[0058] In one embodiment, the static mixing coefficient is 0.5. The local neighborhood window size is 5.

[0059] Furthermore, in step S120, a corresponding dynamic adjacency matrix is generated based on the normalized sequence. First, a feature enhancement network is used to enhance the features of the normalized sequence, and then a multi-head attention mechanism is used to generate the dynamic adjacency matrix. Feature enhancement is to extract deeper, more robust, and more discriminative feature representations. The original one-dimensional HRRP sequence may contain noise and has relatively primary features. Directly calculating the dynamic relationship based on the original sequence may not work well or be unstable enough. The enhanced features can better capture the internal patterns and connections between range cells, laying a foundation for accurately generating the dynamic adjacency matrix subsequently. For the dynamic adjacency matrix, its purpose should be emphasized to capture the dynamic scattering characteristics that change with observation conditions (such as target pose angle, pitch angle) in HRRP data and the complex relationships inherent in the data itself. The static matrix is based on fixed physical priors and is difficult to fully adapt to the dynamic changes such as the migration of target scattering centers with the viewing angle. By using the enhanced sequence and combining with the multi-head attention mechanism, the dynamic adjacency matrix can adaptively learn the association strength between range cells according to the features of each HRRP sample, dynamically adjust the connection relationship of the graph, and thus more flexibly and accurately model the dynamic characteristics of the target, making up for the deficiencies of the static graph in this regard.

[0060] In this embodiment, the feature enhancement network includes three sequentially connected one-dimensional convolutional layers (Conv1D), and a batch normalization layer (Batch Normalization, BN) and a LeakyReLU activation function (the negative slope can be set to 0.2) are connected after each one-dimensional convolutional layer. Among them, for the first one-dimensional convolutional layer, 16 filters are set, and the kernel size of each filter is 5, and the padding value is 2; for the second one-dimensional convolutional layer, 32 filters are set, and the kernel size of each filter is 3, and the padding value is 1; for the third one-dimensional convolutional layer, 64 filters are set, and the kernel size of each filter is 3, and the padding value is 1. After being processed by the feature enhancement network, the output enhanced feature matrix , where, is the hidden layer dimension (default value 64).

[0061] Furthermore, when generating the dynamic adjacency matrix according to each enhanced feature matrix: the multi-head attention mechanism is used to calculate the query matrix, key matrix, and value matrix according to the enhanced feature matrix, calculate the intra-head attention score matrix according to the query matrix and the key matrix, and then generate the dynamic adjacency matrix according to the intra-head attention score matrix.

[0062] Specifically, using the multi-head attention mechanism, set the number of attention heads , and the dimension of each head ; for each head , calculate the query matrix , key matrix , value matrix :

[0063] (3)

[0064] In formula (3), are trainable parameters.

[0065] Furthermore, calculate the intra-head attention score matrix , expressed as:

[0066] (4)

[0067] Next, generate the dynamic adjacency matrix according to the intra-head attention score matrix, expressed as:

[0068] (5)

[0069] In formula (5), represents the intra-head attention score matrix, represents the number of attention heads in the multi-head attention mechanism.

[0070] In step S130, fuse the dynamic adjacency matrix and the static adjacency matrix corresponding to the same one-dimensional high-resolution range image sequence, expressed as:

[0071] (6)

[0072] In the above formula, represents the dynamic mixing coefficient, represents the dynamic adjacency matrix, represents the static adjacency matrix. Among them, the dynamic mixing coefficient is set to 0.3.

[0073] Furthermore, perform normalization processing on the fused mixed graph to obtain the mixed adjacency matrix, expressed as:

[0074] (7)

[0075] In formula (7), represents the degree matrix, and the diagonal element .

[0076] In step S140, input the mixed adjacency matrix into the graph neural network and perform hierarchical adaptive meta-learning. Under the training of the mixed adjacency matrix composed of a small number of one-dimensional high-resolution range image sequences, obtain the final unknown target recognition result.

[0077] In this embodiment, layers of graph convolution are stacked in the graph neural network, and the node features , and its process is expressed as:

[0078] (8)

[0079] In formula (8), represents the LeakyReLU activation function, and its negative slope is set to 0.2. are trainable weights. Dropout is a regularization technique that randomly sets the outputs of some neurons to zero with probability p during training to prevent the model from overfitting on small-sample data and enhance the generalization ability of the model. This probability p is usually set empirically. For example, it can be taken as 0.1, but it can also be adjusted according to specific applications and network structures.

[0080] Furthermore, calculate the attention weights and aggregate the global features , and its process is expressed as:

[0081] (9)

[0082] (10)

[0083] (11)

[0084] In formulas (9) to (11), both e and h represent the aggregated global features, where h represents each dimensional component, and the subscript i represents the i-th , , , are trainable parameters.

[0085] In this embodiment, in the graph neural network, hierarchical adaptive meta-learning is performed. During the meta-training stage, first, the internal loop task adaptation is performed, and the input support set i.e., the training set, where is the number of classes, is the number of samples per class, represents a one-dimensional high-resolution range image sequence or its initial feature representation, is its corresponding target class label. The mixed adjacency matrix is constructed based on the input sequence through static and dynamic mechanisms. When identifying unknown targets (i.e., meta-testing), the model first loads the parameters obtained from meta-training, then uses a small number of support samples provided by the current task to quickly fine-tune and adapt the model parameters, and finally uses the adapted model to classify the query samples. Among them, steps of gradient descent are performed, and at the step, the parameter is updated:

[0086] (12)

[0087] In formula (12), is the layer-specific learning rate, which is set according to the layer type. In this embodiment, the layer-specific learning rate of the graph convolutional layer is set to 0.004 - 0.006, and the layer-specific learning rate of the feature extraction layer is set to 0.002 - 0.004. represents the Hadamard product.

[0088] Furthermore, perform external loop meta-update to calculate the weighted multi-step loss on the query set , which is expressed as:

[0089] (13)

[0090] In formula (13), the weight . Use the Adam optimizer to update the meta-parameters, with the initial learning rate , and adopt the cosine annealing schedule. Finally, perform classification prediction testing to calculate the class probability for the query sample :

[0091] (14)

[0092] In formula (14), , are the classification layer parameters.

[0093] After going through the meta-training stage, the trained graph neural network can be used to perform online hierarchical adaptive meta-learning based on the mixed adjacency matrix corresponding to the one-dimensional high-resolution range image sequence of small samples to identify the category of unknown targets.

[0094] In this embodiment, all the parameters of the feature enhancement network and the graph neural network part are jointly optimized and trained end-to-end through the above meta-learning framework.

[0095] In another embodiment, a schematic diagram of the process framework of the small sample unknown target recognition method is shown in Figure 3 , where, further, this method is divided into two stages. One stage is the generation stage of the mixed adjacency graph, and the second stage is meta-training and meta-testing, that is, the unknown target recognition stage.

[0096] To verify the performance of the method of this application, it is illustrated through a simulation experiment. In this simulation experiment, the graph structure hyperparameter settings involved include , , . In terms of network structure settings, set Layer graph convolutional layer, hidden layer dimension , Rate 0.1. In terms of the meta-learning strategy, the number of steps in the inner loop , the number of task batches , the initial learning rate , and cosine annealing scheduling is adopted. The specific training and testing processes are as follows: 1) Meta-training stage: Sample 3-class 5-sample tasks from the training set. Each task contains a support set and a query set. After quickly adapting through the inner loop, the meta-parameters are updated; 2) Meta-testing stage: Evaluate the model on unseen test categories. Only a small number of support samples are used for fine-tuning and then classification is performed. The dataset is based on an electromagnetic simulation software, and the X-band is selected for simulation. The radar signal bandwidth is set to , the step size of the progress is ; the azimuth angle range is , the step size is ; the elevation angle range is , the step size is . Six different aviation target models are selected for simulation to obtain an HRRP dataset with a size of , and it is randomly divided at a ratio of into categories to obtain the training set and the test set.

[0097] As Figure 3 shown, the results of the simulation experiments in the test under the signal-to-noise ratio (SNR) from 20 dB to -5 dB show that the present invention still maintains an accuracy of 59.4% at an extremely low signal-to-noise ratio (-5 dB), which is 15.2% higher than the optimal comparison model. The comparison methods include MAML, MAML++, template matching, PCA-SVM, GCN, GAT, HRRPGraphNet, and CNN. The accuracy is the average accuracy on 600 query samples of 3-way 1-shot test tasks.

[0098] In the above small-sample HRRP radar target recognition method, first, the obtained HRRP data is preprocessed and normalized. Second, a static adjacency matrix is constructed. Third, a dynamic adjacency matrix is generated. Fourth, a hybrid graph is generated and normalized. Then, forward propagation is completed in the graph neural network. Finally, hierarchical adaptive meta-learning is performed. Among them, the generation and normalization of the static, dynamic, and hybrid graphs are the HRRP hybrid graph generation stage. The forward propagation of the graph neural network generates recognition results, and meta-learning helps the network obtain optimal initial parameters and optimize them on the support samples. This method synchronously models the physical association and data-driven relationship of range cells through the hybrid graph structure, improving the adaptability to pose changes and noise interference. At the same time, a hierarchical-specific meta-learning optimization strategy is designed to accelerate the rapid adaptation of the model to new target categories, so that this method can still maintain high recognition accuracy in a harsh environment where the Signal-to-Noise Ratio (SNR) is lower than 0 dB.

[0099] It should be understood that although Figure 1 the steps in the flowchart of Figure 1 are shown sequentially according to the arrows, these steps are not necessarily executed sequentially in the order indicated by the arrows. Unless otherwise clearly stated in this article, the execution of these steps has no strict order limit, and these steps can be executed in other orders. Moreover,

[0100] In one embodiment, as Figure 4 shown, a small-sample HRRP radar target recognition device is provided, including: a small-sample data set acquisition module 200, a static adjacency matrix construction module 210, a dynamic adjacency matrix generation module 220, a hybrid adjacency matrix obtaining module 230, and a small-sample unknown target recognition module 240, where:

[0101] The small-sample data set acquisition module 200 is used to acquire a small-sample HRRP data set about the same unknown target. The small-sample HRRP data set includes a finite number of one-dimensional high-resolution range image sequences, and each of the one-dimensional high-resolution range image sequences is normalized to obtain a normalized sequence;

[0102] The static adjacency matrix construction module 210 is used to construct a corresponding static adjacency matrix for each of the normalized sequences by using the prior knowledge of radar scattering physics;

[0103] The dynamic adjacency matrix generation module 220 is configured to enhance the features of each of the normalized sequences by using a feature enhancement network to obtain an enhanced feature matrix, and then generate a dynamic adjacency matrix according to each of the enhanced feature matrices;

[0104] The hybrid adjacency matrix obtaining module 230 is configured to fuse the relevant dynamic adjacency matrix and static adjacency matrix, and then perform normalization processing to obtain a hybrid adjacency matrix;

[0105] The few-shot unknown target recognition module 240 is configured to sequentially input each of the hybrid adjacency matrices into a graph neural network and perform forward propagation, and perform hierarchical adaptive meta-learning in each graph convolutional layer in the graph neural network to recognize the unknown target by using a few-shot HRRP data set.

[0106] For the specific limitations of the few-shot HRRP radar target recognition device, reference may be made to the limitations of the few-shot HRRP radar target recognition method in the foregoing text, which will not be elaborated here. Each module in the foregoing few-shot HRRP radar target recognition device may be implemented in whole or in part by software, hardware, and their combination. The foregoing modules may be embedded in or independent of a processor in a computer device in the form of hardware, or may be stored in a memory in the computer device in the form of software, so as to facilitate the processor to call and execute the operations corresponding to the foregoing modules.

[0107] In one embodiment, a computer device is provided. The computer device may be a terminal, and its internal structure diagram may be as Figure 5 shown. The computer device includes a processor, a memory, a network interface, a display screen, and an input device connected through a system bus. Wherein, the processor of the computer device is configured to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the computer device is configured to communicate with an external terminal through a network connection. The computer program, when executed by the processor, implements a few-shot HRRP radar target recognition method. The display screen of the computer device may be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device may be a touch layer covered on the display screen, or may be a button, a trackball, or a touchpad provided on the housing of the computer device, or may also be an external keyboard, touchpad, or mouse, etc.

[0108] Those skilled in the art can understand, Figure 5The structure shown is only a block diagram of some structures related to the solution of this application, and does not constitute a limitation on the computer device to which the solution of this application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine some components, or have different component arrangements.

[0109] In one embodiment, a computer device is provided, including a memory and a processor. A computer program is stored in the memory, and when the processor executes the computer program, the following steps are implemented:

[0110] Obtain a small-sample HRRP data set about the same unknown target. The small-sample HRRP data set includes a finite number of one-dimensional high-resolution range image sequences, and perform normalization processing on each of the one-dimensional high-resolution range image sequences to obtain a normalized sequence;

[0111] For each of the normalized sequences, use the prior knowledge of radar scattering physics to construct a corresponding static adjacency matrix;

[0112] Use a feature enhancement network to enhance the features of each of the normalized sequences to obtain an enhanced feature matrix, and then generate a dynamic adjacency matrix according to each of the enhanced feature matrices;

[0113] After fusing the relevant dynamic adjacency matrix and static adjacency matrix, perform normalization processing to obtain a hybrid adjacency matrix;

[0114] Input each of the hybrid adjacency matrices into a graph neural network in sequence and perform forward propagation, and perform hierarchical adaptive meta-learning in each graph convolutional layer in the graph neural network to identify the unknown target through the small-sample HRRP data set.

[0115] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the following steps are implemented:

[0116] Obtain a small-sample HRRP data set about the same unknown target. The small-sample HRRP data set includes a finite number of one-dimensional high-resolution range image sequences, and perform normalization processing on each of the one-dimensional high-resolution range image sequences to obtain a normalized sequence;

[0117] For each of the normalized sequences, use the prior knowledge of radar scattering physics to construct a corresponding static adjacency matrix;

[0118] Use a feature enhancement network to enhance the features of each of the normalized sequences to obtain an enhanced feature matrix, and then generate a dynamic adjacency matrix according to each of the enhanced feature matrices;

[0119] After fusing the relevant dynamic adjacency matrix and static adjacency matrix, perform normalization processing to obtain a hybrid adjacency matrix;

[0120] Input each of the hybrid adjacency matrices into a graph neural network in sequence and perform forward propagation. Perform hierarchical adaptive meta-learning in each graph convolutional layer in the graph neural network to identify the unknown target through a small-sample HRRP data set.

[0121] Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the above method embodiments. Among them, any reference to a memory, storage, database, or other medium used in the various embodiments provided in the present application can include non-volatile and / or volatile memories. Non-volatile memories can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memories can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0122] The technical features of the above embodiments can be combined arbitrarily. For the sake of concise description, not all possible combinations of the technical features in the above embodiments are described. However, as long as these combinations of technical features do not conflict, they should be considered as within the scope described in this specification.

[0123] The above-described embodiments merely represent several implementation manners of the present application. Their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the invention. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present application, several modifications and improvements can be made, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the appended claims.

Claims

1. A small-sample HRRP radar target recognition method, characterized in that, The method includes: Obtaining a small-sample HRRP data set regarding the same unknown target, where the small-sample HRRP data set includes a finite number of one-dimensional high-resolution range profile sequences, and performing normalization processing on each of the one-dimensional high-resolution range profile sequences to obtain a normalized sequence; For each of the normalized sequences, constructing a corresponding static adjacency matrix by using the prior knowledge of radar scattering physics; Using a feature enhancement network to perform feature enhancement on each of the normalized sequences to obtain an enhanced feature matrix, and then generating a dynamic adjacency matrix according to each of the enhanced feature matrices. Specifically, using a multi-head attention mechanism to calculate a query matrix, a key matrix, and a value matrix according to the enhanced feature matrix, calculating an intra-head attention score matrix according to the query matrix and the key matrix, and then generating the dynamic adjacency matrix according to the intra-head attention score matrix. Among them, generating the dynamic adjacency matrix according to the intra-head attention score matrix is expressed as: In the above formula, represents the intra-head attention score matrix, represents the number of attention heads in the multi-head attention mechanism; After fusing the relevant dynamic adjacency matrix and the static adjacency matrix, performing normalization processing to obtain a mixed adjacency matrix; Sequentially inputting each of the mixed adjacency matrices into a graph neural network and performing forward propagation, and performing hierarchical adaptive meta-learning in each graph convolutional layer in the graph neural network to identify the unknown target through the small-sample HRRP data set.

2. The small-sample HRRP radar target recognition method according to claim 1, wherein When performing normalization processing on the one-dimensional high-resolution range profile sequence, the following operations are performed on the samples on each range cell in the one-dimensional high-resolution range profile sequence: In the above formula, represents a one-dimensional high-resolution range profile sequence, where represents the number of range cells, represents the set of real numbers, , .

3. The small-sample HRRP radar target recognition method according to claim 2, wherein, The static adjacency matrix includes multiple elements, where the element is expressed as: In the above formula, the element represents the static association strength between each distance unit and . The static mixing coefficient is represented by , the local neighborhood window size is represented by , and the indicator function is represented by .

4. The small-sample HRRP radar target recognition method according to claim 3, characterized in that The feature enhancement network includes three sequentially connected one-dimensional convolutional layers, and a batch normalization layer and a LeakyReLU activation function are connected after each of the one-dimensional convolutional layers; The first one-dimensional convolutional layer is provided with 16 filters, and the kernel size of each filter is 5, and the padding value is 2; The second one-dimensional convolutional layer is provided with 32 filters, and the kernel size of each filter is 3, and the padding value is 1; The third one-dimensional convolutional layer is provided with 64 filters, and the kernel size of each filter is 3, and the padding value is 1.

5. The small-sample HRRP radar target recognition method according to claim 4, wherein Fusing the relevant dynamic adjacency matrix and the static adjacency matrix is expressed as: In the above formula, represents the dynamic mixing coefficient, represents the dynamic adjacency matrix, represents the static adjacency matrix.

6. A small-sample HRRP radar target recognition device, characterized in that, The device includes: A small-sample data set acquisition module, configured to obtain a small-sample HRRP data set regarding the same unknown target, where the small-sample HRRP data set includes a finite number of one-dimensional high-resolution range profile sequences, and perform normalization processing on each of the one-dimensional high-resolution range profile sequences to obtain a normalized sequence; A static adjacency matrix construction module, configured to construct a corresponding static adjacency matrix for each of the normalized sequences by using the prior knowledge of radar scattering physics; A dynamic adjacency matrix generation module, which is used to enhance the features of each of the normalized sequences by means of a feature enhancement network to obtain an enhanced feature matrix, and then generate a dynamic adjacency matrix according to each of the enhanced feature matrices. Specifically, a query matrix, a key matrix, and a value matrix are calculated according to the enhanced feature matrix by using a multi-head attention mechanism, an intra-head attention score matrix is calculated according to the query matrix and the key matrix, and then the dynamic adjacency matrix is generated according to the intra-head attention score matrix. Among them, generating the dynamic adjacency matrix according to the intra-head attention score matrix is expressed as: In the above formula, represents the head-in attention score matrix, represents the number of attention heads in the multi-head attention mechanism; A hybrid adjacency matrix obtaining module, which is used to fuse the relevant dynamic adjacency matrix and static adjacency matrix, and then perform normalization processing to obtain a hybrid adjacency matrix; A few-shot unknown target recognition module, which is used to sequentially input each of the hybrid adjacency matrices into a graph neural network and perform forward propagation, and perform hierarchical adaptive meta-learning in each graph convolutional layer in the graph neural network to identify the unknown target through a few-shot HRRP data set.

7. A computer device, comprising a memory and a processor, the memory storing a computer program, characterized in that, When the processor executes the computer program, the steps of the method according to any one of claims 1 to 5 are implemented.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, the steps of the method according to any one of claims 1 to 5 are implemented.

Images