Small sample HRRP identification method, device and equipment integrating Gram angle field

Abstract

The present application relates to a small-sample HRRP recognition method, device and equipment that integrates Gram angular field. The method converts one-dimensional high-resolution range image samples into Gram angular field two-dimensional images after normalization. Multiple support sets and query sets are constructed according to the labeled and unlabeled Gram angular field two-dimensional images according to the target categories, and then multiple training batches are constructed. The graph neural network is trained according to the multiple training batches and meta-learning methods to enable it to have the ability to recognize small-sample unknown targets. In the graph neural network, all Gram angular field two-dimensional images in a training batch are mapped to nodes of the graph structure. The weights of the edges between the nodes are learned to predict the target categories of the corresponding unlabeled nodes. The trained graph neural network is used to recognize small-sample one-dimensional high-resolution range images of unknown targets to overcome the serious overfitting problem caused by insufficient training samples in small-sample HRRP recognition.

Description

Technical Field

[0001] The present application relates to the field of radar automatic target recognition technology, and in particular to a small sample HRRP recognition method, device and equipment integrating Gram angle field. Background Art

[0002] Radar plays a vital role in target recognition due to its non-contact detection and stability in harsh environments. High-resolution range profile (HRRP), a key radar target feature, is the amplitude representation of the coherent superposition of complex echo signals from the target's scattering center along the radar's line of sight (LOS). It reflects the target's spatial scattering structure along the radar's line of sight. HRRP contains rich geometric features, such as the target's size, shape, and scattering distribution, making it crucial for target recognition. Compared to synthetic aperture radar (SAR) or inverse synthetic aperture radar (ISAR) images, HRRP data is simpler to acquire and requires less computational overhead for storage and processing. These advantages have made radar automatic target recognition (RATR) based on HRRP an important research direction over the past few decades.

[0003] Traditional HRRP RATR methods, while having few parameters and simple implementation, often suffer from the inherent sensitivity of HRRP data to time shift, object pose, and amplitude scale. In recent years, deep learning techniques, such as one-dimensional convolutional neural networks (CNNs), recurrent neural networks (RNNs), long short-term memory networks (LSTMs), and one-dimensional residual networks (ResNets), have provided powerful end-to-end solutions for complex recognition tasks and have been widely used in the field of HRRP RATR.

[0004] However, these deep learning models suffer from severe overfitting when training samples are scarce, i.e., in so-called small-sample HRRP recognition scenarios. Limited training samples make it difficult for the models to learn feature representations with good generalization capabilities, leading to a sharp decline in performance. This significantly limits their deployment in practical applications such as non-cooperative object recognition, where obtaining large numbers of labeled HRRP samples is often difficult. Summary of the Invention

[0005] Based on this, it is necessary to provide a small sample HRRP recognition method, device and equipment that can fully utilize the information in limited samples to overcome the problem of insufficient training samples in order to solve the above technical problems.

[0006] A small sample HRRP identification method integrating Gram angle field, the method comprising:

[0007] Acquire a training sample set, wherein the training sample set includes labeled and unlabeled one-dimensional high-resolution range image samples;

[0008] Normalizing each one-dimensional high-resolution range image sample in the training sample set and converting the result into a two-dimensional Gram angle field image;

[0009] Based on the labeled and unlabeled Gram angle field two-dimensional maps in the training sample set, the two-dimensional maps are divided by target category to construct multiple support sets and query sets, and sampling is performed in the multiple support sets and query sets to construct multiple training batches, each of the training batches including a support task set and a query task set;

[0010] A graph neural network is trained based on multiple training batches and a meta-learning approach to obtain the ability to recognize one-dimensional high-resolution range images of unknown targets in a small sample. In the graph neural network, all two-dimensional Gram angle field graphs in a training batch are mapped into nodes of a graph structure, and the weights of the edges between the nodes are learned to predict the target categories of the corresponding unlabeled nodes.

[0011] A small sample of one-dimensional high-resolution range images of unknown targets are obtained, each of the one-dimensional high-resolution range images is normalized and converted into a two-dimensional Gram angular field map, which is then input into the trained graph neural network to realize the recognition of unknown targets.

[0012] In one embodiment, the amplitude value of each range unit in the one-dimensional high-resolution range image sample is normalized using the following formula:

[0013]

[0014] In the above formula, Represents the normalized The amplitude value of the distance unit, and are the minimum and maximum amplitude values in the one-dimensional high-resolution range image samples, respectively.

[0015] In one embodiment, in each of the training batches:

[0016] The support task set includes a support task of performing target recognition on multiple support sets;

[0017] The query task set includes query tasks for performing target recognition on multiple query sets.

[0018] In one embodiment, the graph neural network includes a node feature mapping layer, an MLP layer, a graph convolution layer, and a Softmax layer;

[0019] The node feature mapping layer maps the labeled and unlabeled Gram angle field two-dimensional graphs to the initialization features of the corresponding nodes;

[0020] The MLP layer and graph convolution layer are used to update the weights of the edges between nodes, and the Softmax layer is used to output the probability distribution of different target categories corresponding to each node, thereby predicting the target category of the unlabeled corresponding node.

[0021] In one embodiment, when the node feature mapping layer maps the labeled and unlabeled Gram angle field two-dimensional graphs to the initialization features of the corresponding nodes:

[0022] For the labeled Gram angle field two-dimensional graph, extract the embedded features and generate a one-hot encoding based on the corresponding category label. The embedded features and the one-hot encoding are concatenated as the initialization features of the node.

[0023] For the unlabeled Gram's angle field two-dimensional graph, embedding features are extracted and the uniform distribution over multiple target categories is replaced with one-hot encoding to obtain the initialization features of the corresponding nodes.

[0024] In one embodiment, when training the graph neural network:

[0025] According to each supporting task in the supporting task set, preliminarily updating the adjustable parameters in the graph neural network by predicting the cross entropy loss;

[0026] The query tasks in the query task set are used to evaluate the performance of the graph neural network after the adjustable parameters are initially updated, and the adjustable parameters in the graph neural network are updated again according to the total loss function.

[0027] In one embodiment, the adjustable parameters in the graph neural network include adjustable parameters in the node feature mapping layer, the MLP layer, and the graph convolution layer.

[0028] The present application also provides a small sample HRRP recognition device integrating Gram angle field, the device comprising:

[0029] A training sample set acquisition module is used to acquire a training sample set, wherein the training sample set includes labeled and unlabeled one-dimensional high-resolution range image samples;

[0030] a Gram angular field two-dimensional image conversion module, configured to normalize each one-dimensional high-resolution range image sample in the training sample set and convert the result into a Gram angular field two-dimensional image;

[0031] A training batch construction module is used to construct multiple support sets and query sets based on the labeled and unlabeled Gram angle field two-dimensional maps in the training sample set, and to divide them by target category, and to sample the multiple support sets and query sets to construct multiple training batches, each of which includes a support task set and a query task set;

[0032] A graph neural network training module is used to train the graph neural network based on multiple training batches and a meta-learning method to obtain the ability to recognize one-dimensional high-resolution range images of small-sample unknown targets, wherein, in the graph neural network, all Gram angle field two-dimensional graphs in a training batch are mapped to nodes of the graph structure, and the weights of the edges between each node are learned to predict the target category of the corresponding unlabeled node;

[0033] The small sample unknown target recognition module is used to obtain a small sample one-dimensional high-resolution range image of the unknown target, normalize each one-dimensional high-resolution range image, convert it into a two-dimensional Gram angular field image, and input it into the trained graph neural network to realize the recognition of the unknown target.

[0034] A computer device includes a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the following steps are implemented:

[0035] Acquire a training sample set, wherein the training sample set includes labeled and unlabeled one-dimensional high-resolution range image samples;

[0036] Normalizing each one-dimensional high-resolution range image sample in the training sample set and converting the result into a two-dimensional Gram angle field image;

[0037] Based on the labeled and unlabeled Gram angle field two-dimensional maps in the training sample set, the two-dimensional maps are divided by target category to construct multiple support sets and query sets, and sampling is performed in the multiple support sets and query sets to construct multiple training batches, each of the training batches including a support task set and a query task set;

[0038] A graph neural network is trained based on multiple training batches and a meta-learning approach to obtain the ability to recognize one-dimensional high-resolution range images of unknown targets in a small sample. In the graph neural network, all two-dimensional Gram angle field graphs in a training batch are mapped into nodes of a graph structure, and the weights of the edges between the nodes are learned to predict the target categories of the corresponding unlabeled nodes.

[0039] A small sample of one-dimensional high-resolution range images of unknown targets are obtained, each of the one-dimensional high-resolution range images is normalized and converted into a two-dimensional Gram angular field map, which is then input into the trained graph neural network to realize the recognition of unknown targets.

[0040] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the following steps:

[0041] Acquire a training sample set, wherein the training sample set includes labeled and unlabeled one-dimensional high-resolution range image samples;

[0042] Normalizing each one-dimensional high-resolution range image sample in the training sample set and converting the result into a two-dimensional Gram angle field image;

[0043] Based on the labeled and unlabeled Gram angle field two-dimensional maps in the training sample set, the two-dimensional maps are divided by target category to construct multiple support sets and query sets, and sampling is performed in the multiple support sets and query sets to construct multiple training batches, each of the training batches including a support task set and a query task set;

[0044] A graph neural network is trained based on multiple training batches and a meta-learning approach to obtain the ability to recognize one-dimensional high-resolution range images of unknown targets in a small sample. In the graph neural network, all two-dimensional Gram angle field graphs in a training batch are mapped into nodes of a graph structure, and the weights of the edges between the nodes are learned to predict the target categories of the corresponding unlabeled nodes.

[0045] A small sample of one-dimensional high-resolution range images of unknown targets are obtained, each of the one-dimensional high-resolution range images is normalized and converted into a two-dimensional Gram angular field map, which is then input into the trained graph neural network to realize the recognition of unknown targets.

[0046] The above-mentioned small-sample HRRP recognition method, device and equipment integrating Gram angular field normalizes each one-dimensional high-resolution range image sample in the training sample set and converts it into a Gram angular field two-dimensional map. According to the labeled and unlabeled Gram angular field two-dimensional maps in the training sample set, sampling is performed in multiple support sets and query sets constructed by dividing according to target categories, and multiple training batches including a support task set and a query task set are constructed. The graph neural network is trained according to the multiple training batches and the meta-learning method to obtain the recognition ability of the one-dimensional high-resolution range image of a small-sample unknown target. In the graph neural network, all the Gram angular field two-dimensional maps in a training batch are mapped to nodes of the graph structure, and the weights of the edges between the nodes are learned to predict the target category of the corresponding unlabeled node. Finally, the trained graph neural network is used to recognize the small-sample one-dimensional high-resolution range image of the unknown target to overcome the serious overfitting problem caused by insufficient training samples in small-sample HRRP recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 1. It is a flowchart of a small sample HRRP identification method integrating Gram angle field in one embodiment;

[0048] Figure 2 A schematic diagram of the structure of a neural network in one embodiment;

[0049] Figure 3 is a structural block diagram of a small sample HRRP identification device integrating Gram angle field in one embodiment;

[0050] Figure 4 FIG. 1 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0051] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0052] In view of the serious overfitting problem of existing deep learning methods in small sample HRRP recognition due to insufficient training samples, in this application, Figure 1 As shown, a small sample HRRP identification method integrating Gram angle field is provided, comprising the following steps:

[0053] Step S100: Acquire a training sample set, where the training sample set includes labeled and unlabeled one-dimensional high-resolution range image samples.

[0054] In step S110 , each one-dimensional high-resolution range image sample in the training sample set is normalized and converted into a two-dimensional Gram angle field image.

[0055] In step S120, the labeled and unlabeled Gram angle field two-dimensional images in the training sample set are divided by target category to construct multiple support sets and query sets, and sampling is performed on the multiple support sets and query sets to construct multiple training batches, each of which includes a support task set and a query task set.

[0056] Step S130: Train the graph neural network based on multiple training batches and meta-learning to obtain the ability to recognize one-dimensional high-resolution range images of small sample unknown targets. In the graph neural network, all Gram angular field two-dimensional graphs in a training batch are mapped to nodes of the graph structure, and the weights of the edges between the nodes are learned to predict the target categories of the unlabeled corresponding nodes.

[0057] In step S140, a small sample of one-dimensional high-resolution range images of the unknown target is obtained, each one-dimensional high-resolution range image is normalized, and converted into a two-dimensional Gram angular field image, which is then input into the trained graph neural network to realize the recognition of the unknown target.

[0058] In this example, to fully utilize the information in limited samples, a novel small-sample high-resolution range profile recognition framework, called GAF-MLGNN, is proposed based on the Gram Angular Field and a meta-learning graph neural network. The core idea of this framework is to simultaneously utilize the internal spatial correlation between HRRP range units (intra-sample information) and the interactive relationship structure between different HRRP samples (inter-sample information) to improve the recognition efficiency and generalization ability of the model under small sample conditions.

[0059] In step S100, the training sample set includes one-dimensional high-resolution range image samples acquired from multiple different classes of targets. This sample data includes a portion with class labels, i.e., labeled samples, and a portion without class labels, i.e., unlabeled samples. The training sample set is briefly described here and will be further described in subsequent method steps.

[0060] In step S110, the one-dimensional high-resolution range image samples are represented as a sequence ,in, Represents the range unit (or dimension) of the HRRP sequence (one-dimensional high-resolution range image sample). The amplitude value of each range unit in the one-dimensional high-resolution range image is normalized using the following formula:

[0061] (1)

[0062] In formula (1), Represents the normalized The amplitude value of the distance unit, and are the minimum and maximum amplitude values in the one-dimensional high-resolution range image sample. After being processed by formula (1), the amplitude value on each range unit is scaled to In the interval.

[0063] In this embodiment, the normalized 1D high-resolution range image is subjected to a Gram Angular Field (GAF) transformation to retain the spatial correlation between its internal distance units, provide a more suitable input for the graph neural network, and convert it into a two-dimensional image containing rich structural information, helping the model to better capture the sample interaction relationship, improve performance in small sample recognition, and alleviate overfitting.

[0064] Specifically, the normalized one-dimensional HRRP sequence Convert from Cartesian coordinate system to polar coordinate system, the calculation formula is as follows:

[0065] (2)

[0066] In formula (2), is the angle in polar coordinates, through the arc cosine function Get, its value range is , which is a monotone mapping, is the radius in polar coordinates, indexed by time step (or distance unit) Divide by the total dimensions get.

[0067] Furthermore, the spatial correlation between HRRP points is encoded by calculating the cosine value of the sum of the angles corresponding to different distance units, and the Gram Angular Field (GAF) matrix (image) is generated. The calculation formula is as follows:

[0068] (3)

[0069] The two-dimensional Gram angular field obtained by the Gram angular field (GAF) transformation is a The two-dimensional matrix, that is, the GAF image , which contains the correlation information between the internal distance units of the original HRRP sequence and serves as one of the input node feature sources of the subsequent graph neural network.

[0070] In step S120, multiple training batches are constructed from the training dataset after the Gram Angular Field (GAF) transformation. First, multiple support sets are constructed based on the labeled Gram Angular Field two-dimensional graph according to different target categories, that is, each support set corresponds to a single target category. Similarly, multiple query sets are constructed based on the unlabeled Gram Angular Field two-dimensional graph according to different target categories, that is, each query set corresponds to a single target category. Next, multiple training batches are constructed from these support sets and query sets according to a preset ratio. Each training batch includes a support task set and a query task set, wherein the support task set includes support tasks for performing target recognition on multiple support sets, and the query task set includes query tasks for performing target recognition on multiple query sets.

[0071] Specifically, each training batch is actually N -way K -shot small sample recognition task. This task generally includes a supporting task set and a query task set . Supported task sets Depend on N There are samples of different categories, each category containsK labeled samples, total samples, represented by ,in It is The image of samples after GAF transformation, is its corresponding category label, Query task set Include There are unlabeled samples to be classified, represented as .

[0072] In this embodiment, the query task set , usually focus on That is, each task only classifies one query sample.

[0073] Furthermore, each training batch It can be formally expressed as:

[0074] (4)

[0075] The subsequent output of the graph neural network is the label of the sample in the predicted query set ,in :

[0076] (5)

[0077] In step S130, a graph neural network (GNN) is used to associate the entire small-sample recognition task with a fully connected graph. The nodes of the graph correspond to GAF image samples (i.e., the two-dimensional Gram Angular Field graph), and the edges correspond to the relationships between samples. Through a multi-layer message passing mechanism, the GNN continuously updates the feature representations of the nodes and mines multi-level interactive information between samples.

[0078] Specifically, in the graph neural network, the entire training batch task is associated with a fully connected graph , where the node set of the graph is Corresponding to the task All GAF image samples in the task set (including GAF image samples in the support task set and the query task set), that is, each sample or Each is a node in the graph or The set of edges between nodes Initially, it is fully connected, and the edge weights will be learned in subsequent steps.

[0079] In this embodiment, the graph neural network includes a node feature mapping layer, an MLP layer, a graph convolution layer, and a Softmax layer. The node feature mapping layer maps the labeled and unlabeled Gram angle field two-dimensional graphs to the initial features of the corresponding nodes. The MLP layer and graph convolution layer are used to update the weights of the edges between each node. The Softmax layer then outputs the probability distribution of different target categories for each node, thereby predicting the target category of the corresponding unlabeled node.

[0080] In this embodiment, when the node feature mapping layer maps the labeled and unlabeled Gram's angular field two-dimensional graphs to the initialization features of the corresponding nodes respectively: for the labeled Gram's angular field two-dimensional graph, the embedded features are extracted, and a one-hot encoding is generated according to the corresponding category label, and the embedded features and the one-hot encoding are spliced as the initialization features of the node; for the unlabeled Gram's angular field two-dimensional graph, the embedded features are extracted, and the one-hot encoding is replaced by a uniform distribution on multiple target categories to obtain the initialization features of the corresponding node.

[0081] Specifically, for the labeled samples in the support set , its initial node characteristics It is composed of two parts: one part is through a two-dimensional convolutional neural network (2D CNN) That is, the image embedding features extracted by the node feature mapping layer , and the other part is its category label One-hot encoding of Then the initial node features of the two-dimensional graph with labeled Gram angle field are Expressed as:

[0082] (6)

[0083] Specifically, for the unlabeled samples in the query set , since its label is unknown, use a Uniform distribution over categories Instead of one-hot encoding ,in It is a full 1 dimensional vector, then the initial node features of the unlabeled Gram angle field two-dimensional graph Expressed as:

[0084] (7)

[0085] In this embodiment, the point feature map layer An embedding network for extracting GAF image features can be used. In one embodiment, the point feature map layer It is a network structure containing several convolutional layers, batch normalization layers, activation layers and pooling layers.

[0086] Furthermore, the graph convolution layer is usually composed of a series of graph convolution layers. In the Before the graph convolution layer, an adjacency matrix needs to be calculated , which represents the relationship strength between nodes (i.e., the weight of the edge). Unlike traditional GNN, the adjacency matrix here is based on the current layer Hidden node representation of Dynamically learned.

[0087] Specifically, node and nodes The edge weights between Through a multi-layer perceptron (MLP) Acting on the absolute difference of their feature vectors, the MLP has trainable parameters :

[0088] (8)

[0089] From formula (8), we can see that MLP is a symmetric similarity metric function that satisfies and The adjacency matrix calculated Normalization is usually required, for example by applying a Softmax function to each row, making it a random kernel. Then, Graph convolutional layers Receive input signal ,in is the number of nodes, It is The feature dimension of the layer, and according to the adjacency matrix and a trainable weight matrix Aggregate neighbor information and perform nonlinear transformation to generate node features of the next layer The calculation formula is as follows:

[0090] (9)

[0091] In formula (9), It is the element-wise nonlinear activation function LeakyReLU.

[0092] In this embodiment, two groups of the above-mentioned MLP layers and graph convolution layers are stacked to capture the multi-order interaction relationships between nodes, that is, the interaction relationships between the one-dimensional high-resolution range images.

[0093] like Figure 2 As shown, in one embodiment, a specific structure of a graph neural network is proposed. A 2D CNN (called Conv64F) with four convolutional blocks is used. Each block contains a convolutional layer (64 filters, 3x3 kernel), a batch normalization layer, a LeakyReLU activation layer, and a 2x2 max pooling layer. The graph neural network contains two graph convolution modules, each of which consists of an MLP for calculating the adjacency matrix (see the structure). Figure 2 ) and a graph convolution layer (Gc Layer, see the structure Figure 2 ). The node feature dimension is kept at 64 between GNN layers. Contains two fully connected layers with LeakyReLU activation in the middle.

[0094] Since the training process of the graph neural network is based on the entire task Constructing a graph and updating parameters is different from the traditional small sample learning method (such as MAML) which is based on the support set update and evaluation on the query set. In this embodiment, a meta-learning strategy based on the task set for GNN is proposed. The goal of this strategy is to learn a set of initial model parameters. , including the node feature map layer Parameters of graph convolutional layer Parameters and MLP layers The parameters of , enable the model to achieve good performance on new small sample tasks starting from this set of parameters.

[0095] In this embodiment, when training the graph neural network, in each training batch, the adjustable parameters in the graph neural network are initially updated by predicting the cross-entropy loss based on each supporting task in the supporting task set. The performance of the graph neural network after the initial update of the adjustable parameters is then evaluated using the query tasks in the query task set, and the adjustable parameters in the graph neural network are updated again based on the total loss function. After training and learning for multiple training batches, the training is completed when the graph neural network meets the convergence requirements.

[0096] Specifically, define the distribution of meta-learning tasks In each iteration, instead of sampling a single task, a “meta-task batch” is sampled. That is, the training dataset is divided into multiple training batches in step S120, and each training batch contains a supporting task set. and a query task set ,in and is the number of tasks contained in each set. Both task sets are from obtained by sampling in .

[0097] Furthermore, for the supported task set Each task in , computational model Losses on this mission This loss is usually the cross entropy loss of the prediction for the query node in the task:

[0098] (10)

[0099] In formula (10), is the true label of the query node, is its one-hot encoded elements, The graph neural network predicts that the query node belongs to the category The probability is obtained by outputting the last layer of the graph neural network through Softmax.

[0100] Then, based on the support task set The total loss (or average loss) on the , calculate a simulated gradient update step, and get a set of adapted parameters :

[0101] (11)

[0102] In formula (11), represents the internal learning rate.

[0103] Next, use this set of adapted parameters In the query task set Evaluate the performance on the query task set and calculate the total loss (or average loss). The goal of meta-learning is to minimize this loss on the query task set:

[0104] (12)

[0105] In formula (12), The meta-task batches that represent the samples are the training batches.

[0106] Finally, by calculating the meta-objective function with respect to the original parameters The gradient of , and use an external learning rate To update :

[0107] (13)

[0108] The above training process is repeated using multiple training batches until the graph neural network converges. Through this training method, the adjustable parameters in the node feature mapping layer, MLP layer, and graph convolution layer in the graph neural network converge, thereby achieving good performance in recognizing unknown targets with small samples.

[0109] After the above meta-training phase is completed, a set of optimized graph neural network parameters will be obtained , in the meta-test phase, when encountering a new small sample recognition task When using this fixed set of parameters To make predictions. The specific process is: All samples (support set and query set) in the , initialize the features of all nodes (support set samples use real labels, query set samples use uniform distribution), and then use the trained graph neural network Perform forward propagation and calculate the final feature representation of all nodes. , its predicted label It is determined by the probability distribution of the corresponding graph node output in the last layer (after Softmax), and the category with the highest probability is usually selected as the prediction result.

[0110] Furthermore, the performance of the graph neural network is evaluated by calculating the average recognition accuracy on a large number of test tasks. The accuracy calculation formula is:

[0111] (14)

[0112] In formula (14), is the total number of query samples in all test tasks, It is The predicted labels of query samples, Is its real label, is an indicator function that takes the value 1 if the condition is true and 0 otherwise.

[0113] In step S140, when using the trained graph neural network to identify unknown targets under small sample conditions, a small number of one-dimensional high-resolution range images obtained by actual measurement are subjected to GAF transformation and input into the trained graph neural network to obtain the category prediction result of the unknown target.

[0114] In order to verify the effectiveness of this method, simulation results are also given in this paper. In this simulation experiment, a HRRP data set generated based on electromagnetic simulation of twelve types of aircraft three-dimensional models was selected. The radar used in the simulation operates in the X-band, with a frequency range of 9.5GHz to 10.5GHz and a step frequency interval of 10MHz. The radar observation angle covers the pitch angle of 75° to 105° (step size 3°) and the azimuth angle of 0° to 60° (step size 0.05°). In the simulation experiment, the HRRP data corresponding to the first pitch angle (for example, 75°) in the HH polarization mode was used. For each aircraft category, a total of 1201 HRRP samples were obtained within the covered azimuth range, and each HRRP sample contained 1000 distance units. The size of the data set thus constructed is .

[0115] In the data preprocessing stage, the HRRP samples are first normalized and then converted into GAF images using the above formulas (1) to (3). The GAF image is resampled to In order to reduce the computational burden, we randomly select from all 12 categories when conducting N-way K-shot experiments. N categories as the categories in the test task, and the remaining 12- N The data of categories is used in the meta-training phase (from which the training tasks are sampled).

[0116] The network structure adopts Figure 2 The graph neural network shown in Figure 1. When training the graph neural network, the Adam optimizer is used for meta-training. The internal learning rate and the external learning rate are all set to 0.001. The model is trained for 100 epochs. In MLGNN, the size of the supported task set included in each meta-task batch is and query task set size It can be set as needed, for example, it can be set to 4 or more. Performance evaluation is performed on multiple randomly sampled test tasks, and the average recognition accuracy is calculated as the final result. The accuracy is calculated using the above formula (14).

[0117] Experimental results, shown in Table 1, show that our method achieves significantly better recognition accuracy than existing state-of-the-art methods across various N-way K-shot configurations (e.g., 4-way 1-shot, 5-way 1-shot, 6-way 1-shot, and their corresponding 2-shot and 5-shot counterparts). For example, under the 5-way 1-shot condition, GAF-MLGNN achieves an accuracy of 84.88%, significantly exceeding the baseline GNN method (65.42%) and other comparative methods such as GAF-MAML (58.60%) and Improved ProtoNet (57.23%). This demonstrates that by combining GAF's use of intra-sample information, GNN's use of inter-sample information, and the MLGNN meta-learning strategy, our method effectively overcomes the overfitting problem in small-sample HRRP recognition, significantly improving the model's generalization and recognition accuracy. Furthermore, ablation experiments on the number of GNN layers, the number of meta-training samples, and the effect of GAF transformations further validate the effectiveness and robustness of each component of our framework. For example, a comparison of using GAF image input with direct HRRP sequence input (which requires a 1D CNN embedding network) shows that using GAF significantly improves performance (e.g., 5-way 5-shot accuracy increases from 84.36% to 91.63%), demonstrating the importance of leveraging spatial correlation within samples. Furthermore, MLGNN consistently outperforms the baseline GNN across various settings, validating the effectiveness of the proposed meta-learning strategy.

[0118] Table 1 Experimental results of small sample target recognition using this method and various existing methods

[0119]

[0120] In the aforementioned Gram Angular Field-integrated small-shot HRRP recognition method, the Gram Angular Field (GAF) transform technique is first used to encode one-dimensional HRRP sequences into two-dimensional images. This effectively preserves and leverages the inherent spatial correlation between distance units within HRRP, thereby exploiting the spatial correlation within samples. Secondly, a graph neural network (GNN) is introduced to explicitly model and learn the complex relationships between HRRP samples. The small-shot learning problem is formulated as a node classification task on a graph, and the interaction information between samples is captured through node feature aggregation. Furthermore, given the unique training mechanism of GNNs (which construct a graph based on the entire task and perform parameter updates), traditional model-agnostic meta-learning (MAML) strategies are not directly applicable. Therefore, this method proposes a task-set-based meta-learning strategy. This strategy draws on the ideas of MAML but is adapted at the task sampling and gradient update levels. This strategy aims to learn a more optimal initial parameter state for the GNN model, enabling it to quickly adapt to new small-shot tasks, thereby significantly improving the performance of GNNs on small-shot HRRP recognition. The effectiveness of the proposed method was verified through a large number of experiments on simulation datasets. The results showed that the proposed method has significant performance advantages over the current optimal meta-learning and small-sample HRRP RATR methods, effectively alleviates the overfitting phenomenon, and improves the recognition accuracy.

[0121] It should be understood that although Figure 1 The steps in the flowchart are shown in sequence as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. In addition, Figure 1 At least part of the steps may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least part of the sub-steps or stages of other steps.

[0122] In one embodiment, Figure 3 As shown, a small sample HRRP recognition device integrating Gram angular field is provided, comprising: a training sample set acquisition module 200, a Gram angular field two-dimensional graph conversion module 210, a training batch construction module 220, a graph neural network training module 230 and a small sample unknown target recognition module 240, wherein:

[0123] The training sample set acquisition module 200 is used to acquire a training sample set, wherein the training sample set includes one-dimensional high-resolution range image samples with labels and without labels.

[0124] The Gram angular field two-dimensional image conversion module 210 is used to normalize each one-dimensional high-resolution range image sample in the training sample set and convert the normalized image into a Gram angular field two-dimensional image.

[0125] The training batch construction module 220 is used to construct multiple support sets and query sets based on the labeled and unlabeled Gram angle field two-dimensional maps in the training sample set, and divide them by target category, and sample the multiple support sets and query sets to construct multiple training batches, each of the training batches including a support task set and a query task set.

[0126] The graph neural network training module 230 is used to train the graph neural network based on multiple training batches and a meta-learning method to obtain the ability to recognize one-dimensional high-resolution range images of small sample unknown targets. In the graph neural network, all Gram angular field two-dimensional graphs in a training batch are mapped to nodes of the graph structure, and the weights of the edges between the nodes are learned to predict the target categories of the unlabeled corresponding nodes.

[0127] The small sample unknown target recognition module 240 is used to obtain a small sample one-dimensional high-resolution range image of the unknown target, normalize each of the one-dimensional high-resolution range images, convert them into a two-dimensional Gram angular field image, and input them into the trained graph neural network to realize the recognition of the unknown target.

[0128] Regarding the specific definition of the small sample HRRP identification device that integrates the Gram angular field, please refer to the definition of the small sample HRRP identification method that integrates the Gram angular field above, which will not be repeated here. The various modules in the above-mentioned small sample HRRP identification device that integrates the Gram angular field can be implemented in whole or in part by software, hardware, and a combination thereof. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.

[0129] In one embodiment, a computer device is provided. The computer device may be a terminal, and its internal structure diagram may be as follows: Figure 4As shown. The computer device includes a processor, a memory, a network interface, a display screen and an input device connected via a system bus. The processor of the computer device is used 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 used to communicate with an external terminal via a network connection. When the computer program is executed by the processor, a small sample HRRP identification method that integrates the Gram angle field is implemented. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device can be a touch layer covering the display screen, or a key, trackball or touchpad provided on the computer device housing, or an external keyboard, touchpad or mouse.

[0130] Those skilled in the art will understand that Figure 4 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0131] In one embodiment, a computer device is provided, including a memory and a processor, wherein a computer program is stored in the memory, and when the processor executes the computer program, the following steps are implemented:

[0132] Acquire a training sample set, wherein the training sample set includes labeled and unlabeled one-dimensional high-resolution range image samples;

[0133] Normalizing each one-dimensional high-resolution range image sample in the training sample set and converting the result into a two-dimensional Gram angle field image;

[0134] Based on the labeled and unlabeled Gram angle field two-dimensional maps in the training sample set, the two-dimensional maps are divided by target category to construct multiple support sets and query sets, and sampling is performed in the multiple support sets and query sets to construct multiple training batches, each of the training batches including a support task set and a query task set;

[0135] A graph neural network is trained based on multiple training batches and a meta-learning approach to obtain the ability to recognize one-dimensional high-resolution range images of unknown targets in a small sample. In the graph neural network, all two-dimensional Gram angle field graphs in a training batch are mapped into nodes of a graph structure, and the weights of the edges between the nodes are learned to predict the target categories of the corresponding unlabeled nodes.

[0136] A small sample of one-dimensional high-resolution range images of unknown targets are obtained, each of the one-dimensional high-resolution range images is normalized and converted into a two-dimensional Gram angular field map, which is then input into the trained graph neural network to realize the recognition of unknown targets.

[0137] 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:

[0138] Acquire a training sample set, wherein the training sample set includes labeled and unlabeled one-dimensional high-resolution range image samples;

[0139] Normalizing each one-dimensional high-resolution range image sample in the training sample set and converting the result into a two-dimensional Gram angle field image;

[0140] Based on the labeled and unlabeled Gram angle field two-dimensional maps in the training sample set, the two-dimensional maps are divided by target category to construct multiple support sets and query sets, and sampling is performed in the multiple support sets and query sets to construct multiple training batches, each of the training batches including a support task set and a query task set;

[0141] A graph neural network is trained based on multiple training batches and a meta-learning approach to obtain the ability to recognize one-dimensional high-resolution range images of unknown targets in a small sample. In the graph neural network, all two-dimensional Gram angle field graphs in a training batch are mapped into nodes of a graph structure, and the weights of the edges between the nodes are learned to predict the target categories of the corresponding unlabeled nodes.

[0142] A small sample of one-dimensional high-resolution range images of unknown targets are obtained, each of the one-dimensional high-resolution range images is normalized and converted into a two-dimensional Gram angular field map, which is then input into the trained graph neural network to realize the recognition of unknown targets.

[0143] Those skilled in the art will understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the 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 embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in many forms such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), Synchronous Link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0144] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0145] The above-described embodiments merely represent several implementation methods of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that a person of ordinary skill in the art may make various modifications and improvements without departing from the spirit of the present application, and such modifications and improvements fall within the scope of protection of the present application. Therefore, the scope of protection of the present application shall be determined by the appended claims.

Claims

1. A small sample HRRP identification method integrating Gram angle field, characterized in that: The method comprises: Acquire a training sample set, wherein the training sample set includes labeled and unlabeled one-dimensional high-resolution range image samples; Normalizing each one-dimensional high-resolution range image sample in the training sample set and converting the result into a two-dimensional Gram angle field image; According to the labeled and unlabeled Gram angle field two-dimensional maps in the training sample set, the two-dimensional maps are divided by target category to construct multiple support sets and query sets, and sampling is performed in the multiple support sets and query sets to construct multiple training batches, each of the training batches includes a support task set and a query task set, wherein the support task set includes a support task of performing target recognition on multiple support sets, and the query task set includes a query task of performing target recognition on multiple query sets; The graph neural network is trained according to multiple training batches and a meta-learning method to obtain the ability to recognize one-dimensional high-resolution range images of unknown targets in small samples, wherein, in the graph neural network, all Gram angular field two-dimensional graphs in a training batch are mapped to nodes of a graph structure, and the weights of the edges between the nodes are learned to predict the target categories of the unlabeled corresponding nodes. When training the graph neural network: according to each supporting task in the supporting task set, the adjustable parameters in the graph neural network are preliminarily updated by predicting the cross entropy loss, and the performance of the graph neural network after the preliminarily updated adjustable parameters is evaluated using the query tasks in the query task set, and the adjustable parameters in the graph neural network are updated again according to the total loss function. The graph neural network includes a graph convolution layer, which is composed of a series of graph convolution layers. Composition, in Before the graph convolution layer, calculate an adjacency matrix , which represents the relationship strength between nodes, that is, the weight of the edge, where the adjacency matrix Based on the current layer Hidden node representation of Dynamically learned, specific, nodes and nodes The edge weights between Through a multi-layer perceptron (MLP) Acting on the absolute difference of their feature vectors, the MLP has trainable parameters : ; A small sample of one-dimensional high-resolution range images of unknown targets are obtained, each of the one-dimensional high-resolution range images is normalized and converted into a two-dimensional Gram angular field map, which is then input into the trained graph neural network to realize the recognition of unknown targets.

2. The small sample HRRP identification method based on the fusion of Gram angle field according to claim 1 is characterized in that: The amplitude value of each range unit in the one-dimensional high-resolution range image sample is normalized using the following formula: ； In the above formula, Represents the normalized The amplitude value of the distance unit, and are the minimum and maximum amplitude values in the one-dimensional high-resolution range image samples, respectively.

3. The small sample HRRP identification method based on the fusion of Gram angle field according to claim 2 is characterized in that: The graph neural network includes a node feature mapping layer, an MLP layer, a graph convolution layer and a Softmax layer; The node feature mapping layer maps the labeled and unlabeled Gram angle field two-dimensional graphs to the initialization features of the corresponding nodes; The MLP layer and graph convolution layer are used to update the weights of the edges between nodes, and the Softmax layer is used to output the probability distribution of different target categories corresponding to each node, thereby predicting the target category of the unlabeled corresponding node.

4. The small sample HRRP identification method based on the fusion of Gram angle field according to claim 3 is characterized in that: When the node feature mapping layer maps the labeled and unlabeled Gram angle field two-dimensional graphs to the initialization features of the corresponding nodes: For the labeled Gram angle field two-dimensional graph, extract the embedded features and generate a one-hot encoding based on the corresponding category label. The embedded features and the one-hot encoding are concatenated as the initialization features of the node. For the unlabeled Gram's angle field two-dimensional graph, embedding features are extracted and the uniform distribution over multiple target categories is replaced with one-hot encoding to obtain the initialization features of the corresponding nodes.

5. The small sample HRRP identification method based on the fusion of Gram angle field according to claim 4 is characterized in that: The adjustable parameters in the graph neural network include the adjustable parameters in the node feature mapping layer, the MLP layer and the graph convolution layer.

6. A small sample HRRP recognition device integrating Gram angle field, characterized in that: The device comprises: A training sample set acquisition module is used to acquire a training sample set, wherein the training sample set includes one-dimensional high-resolution range image samples with labels and without labels; a Gram angular field two-dimensional image conversion module, configured to normalize each one-dimensional high-resolution range image sample in the training sample set and convert the result into a Gram angular field two-dimensional image; A training batch construction module is used to construct multiple support sets and query sets based on the labeled and unlabeled Gram angle field two-dimensional maps in the training sample set, and to divide them by target category, and to sample the multiple support sets and query sets to construct multiple training batches, each of the training batches including a support task set and a query task set, wherein the support task set includes a support task of performing target recognition on multiple support sets, and the query task set includes a query task of performing target recognition on multiple query sets; A graph neural network training module is used to train the graph neural network based on multiple training batches and a meta-learning method to obtain the ability to recognize one-dimensional high-resolution range images of unknown targets in small samples. In the graph neural network, all Gram angular field two-dimensional graphs in a training batch are mapped to nodes of a graph structure, and the weights of the edges between the nodes are learned to predict the target categories of the unlabeled corresponding nodes. When training the graph neural network: according to each supporting task in the supporting task set, the adjustable parameters in the graph neural network are preliminarily updated by predicting the cross entropy loss, and the performance of the graph neural network after the preliminarily updated adjustable parameters is evaluated using the query tasks in the query task set, and the adjustable parameters in the graph neural network are updated again according to the total loss function. The graph neural network includes a graph convolution layer, which is composed of a series of graph convolution layers. Composition, in Before the graph convolution layer, calculate an adjacency matrix , which represents the relationship strength between nodes, that is, the weight of the edge, where the adjacency matrix Based on the current layer Hidden node representation of Dynamically learned, specific, nodes and nodes The edge weights between Through a multi-layer perceptron (MLP) Acting on the absolute difference of their feature vectors, the MLP has trainable parameters : ; The small sample unknown target recognition module is used to obtain a small sample one-dimensional high-resolution range image of the unknown target, normalize each one-dimensional high-resolution range image, convert it into a two-dimensional Gram angular field image, and input it into the trained graph neural network to realize the recognition of the unknown target.

7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 5 are implemented.

Images