A cross-domain HRRP small sample target recognition method based on adversarial meta-transfer learning
By using adversarial meta-transfer learning, virtual meta-tasks are generated, which solves the problem of unstable recognition performance in cross-domain small sample scenarios in radar HRRP target recognition, and achieves better recognition performance and stability, which is applicable to the field of radar target recognition.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- XIDIAN UNIV
- Filing Date
- 2023-03-09
- Publication Date
- 2026-06-23
AI Technical Summary
Existing radar HRRP target recognition methods suffer from unstable recognition performance in small sample scenarios due to significant differences in data distribution between the source and target domains, making them difficult to apply effectively in cross-domain small sample scenarios.
We employ an adversarial meta-transfer learning approach, which constructs a virtual meta-task generation network to adaptively generate more challenging virtual tasks for the model, thereby reducing the data distribution differences between the source and target domains. We also utilize meta-learning concepts for double-layer sampling to enhance the diversity of the mapping between the category space and the data space.
It improves recognition performance in cross-domain small sample scenarios, has better generalizability, stability and feasibility, reduces the requirement for similarity between source and target domain data, and reduces the risk of model overfitting.
Smart Images
Figure CN116466312B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of radar technology, specifically relating to a cross-domain HRRP small sample target recognition method based on adversarial meta-transfer learning. Background Technology
[0002] Radar target recognition refers to determining the type of a target using its radar echo signals. Compared to traditional target detection tasks, target recognition requires measuring more in-depth target details, consuming more radar resources such as time and spectrum. High-resolution radar range profiles (HRRPs) contain rich structural information such as target geometry, making them valuable for target recognition. Currently, research on recognition methods based on radar HRRPs is one of the important approaches to achieving radar target recognition. Most existing radar HRRP target recognition methods are based on the assumption of a complete dataset, requiring a sufficient number of training samples during the classifier training phase to extract essential target features and achieve strong generalization ability.
[0003] In real-world scenarios, the targets and environments detected by radar are extremely complex, and the limitations of radar system performance often make it difficult to acquire a sufficient number of target HRRP samples, thus hindering the construction of complex and complete datasets. In small-sample scenarios, classifiers tend to focus excessively on the limited training samples in the dataset, easily leading to overfitting and causing performance degradation or even failure of the recognition method, severely mismatched with the small-sample target recognition problem. To address this issue, researchers typically employ small-sample learning algorithms based on knowledge transfer, such as transfer learning and meta-learning. These algorithms use existing large-sample datasets as source domain datasets for pre-training, leveraging domain knowledge learned from the source domain to reduce the difficulty of performing recognition tasks on small-sample datasets in the target domain.
[0004] The aforementioned knowledge transfer methods are applicable when the source and target domains have similar data distribution spaces. However, in the field of HRRP recognition, due to the complexity of the detection environment and the variety of non-cooperative targets, the category spaces and data distribution spaces of the source and target domains generally differ significantly. Most existing HRRP small-sample recognition methods ignore this inter-domain distribution difference, resulting in unstable recognition performance in cross-domain small-sample scenarios. Summary of the Invention
[0005] To address the aforementioned problems in existing technologies, this invention provides a cross-domain HRRP few-shot target recognition method based on adversarial meta-transfer learning. The technical problem to be solved by this invention is achieved through the following technical solution:
[0006] A cross-domain HRRP few-shot target recognition method based on adversarial meta-transfer learning includes the following steps:
[0007] Step one, obtaining a source domain data set and a target domain fine-tuning data set; wherein the source domain data set includes radar HRRP echo signals of N categories of targets, N≥5; the target domain fine-tuning data set includes radar HRRP echo signals of M categories of targets, 3≤M<N, the N categories and the M categories being different;
[0008] Step two, respectively constructing a first convolutional neural network, a virtual meta-task generation network and a second convolutional neural network;
[0009] The virtual meta-task generation network includes a random scale convolutional layer, a full connection layer and a Softmax layer, the number of convolution kernels of the random scale convolutional layer is 1, the size of the random scale convolution kernel is 1*H, the convolution kernel moving step is 1, the number of neurons of the full connection layer is Q, Q=M;
[0010] Wherein, H∈Set={3,5,7,9,11,13}, in each call of the virtual meta-task generation network, the value of H is uniformly extracted from the scale set Set and the size of the random scale convolution kernel is set;
[0011] Step three, training the first convolutional neural network based on the source domain data set and the virtual meta-task generation network, to obtain a trained first convolutional neural network;
[0012] Step four, migrating the parameters of the convolutional layer of the trained first convolutional neural network to the second convolutional neural network, to obtain a recognition convolutional neural network;
[0013] Step five, training the recognition convolutional neural network based on the target domain fine-tuning data set, to obtain a trained recognition convolutional neural network;
[0014] Step six, inputting each target domain echo signal to be recognized into the trained recognition convolutional neural network, to output a recognition result.
[0015] In an embodiment of the present application, the uniformly extracting the value of H from the scale set Set and setting the size of the random scale convolution kernel comprises:
[0016] According to the probability distribution P, one sampling is performed to obtain the selected element sequence number id, and then the size of the random scale convolution kernel H=Set{id} is set;
[0017] Wherein, Z represents the total number of elements in the scale set Set, id represents the sequence number of the element in the scale set Set, 1≤id≤Z, and Set{id} represents the idth element in the scale set Set.
[0018] In one embodiment of the present invention, the first convolutional neural network and the second convolutional neural network have the same structure, both including an input layer, a first convolutional layer, a first pooling layer, a first activation layer, a second convolutional layer, a second pooling layer, a second activation layer, a third convolutional layer, a third pooling layer, a third activation layer, a first fully connected layer, a second fully connected layer, and a Softmax layer arranged sequentially.
[0019] In one embodiment of the present invention, the number of convolutional kernels in the first convolutional layer, the second convolutional layer and the third convolutional layer of the first convolutional neural network are 8, 16 and 32 respectively, the kernel size is 1×3 and the kernel stride is set to 1.
[0020] The pooling strategies of the first pooling layer, the second pooling layer and the third pooling layer of the first convolutional neural network all adopt the max pooling method, the pooling kernel size is 1×2 and the pooling kernel moving step size is 2.
[0021] The first activation layer, the second activation layer, and the third activation layer of the first convolutional neural network all use the ReLU activation function;
[0022] The number of neurons in the first fully connected layer and the second fully connected layer of the first convolutional neural network are 256 and P, respectively, where P = M;
[0023] The Softmax layer of the first convolutional neural network uses the Softmax function to calculate the probability that an input sample is identified as each class.
[0024] In one embodiment of the present invention, the number of convolutional kernels in the first, second, and third convolutional layers of the second convolutional neural network are 8, 16, and 32, respectively, the kernel size is 1×3, and the kernel stride is set to 1.
[0025] The pooling strategies of the first, second, and third pooling layers of the second convolutional neural network all adopt the max pooling method, with a pooling kernel size of 1×2 and a pooling kernel step size of 2.
[0026] The first, second, and third activation layers of the second convolutional neural network all use the ReLU activation function;
[0027] The number of neurons in the first fully connected layer and the second fully connected layer of the second convolutional neural network are 128 and R, respectively, where R = M;
[0028] The Softmax layer of the second convolutional neural network uses the Softmax function to calculate the probability that an input sample is identified as each class.
[0029] In one embodiment of the present invention, step three includes the following steps:
[0030] S classes are randomly selected from the source domain dataset, and 2*L echo signals are taken from each class to form a meta-task. The meta-task is divided into two groups according to the number of samples. Each group contains S classes and L echo signals. One group is used as the support set of the meta-task, and the other group is used as the query set of the meta-task.
[0031] T echo signals are extracted from the source domain dataset to obtain T sets of support and T sets of query. The T sets of support and T sets of query constitute a T-group task, where 1 ≤ T < 5.
[0032] The T sets of support are sequentially input into the virtual meta-task generation network. After passing through a random-scale convolutional layer, T sets of data representation vectors X are generated, and the T sets of data representation vectors X are saved as T sets of virtual support.
[0033] The T sets of data representation vectors X are passed through the fully connected layer and the Softmax layer of the virtual meta-task generation network to generate corresponding predicted labels, and the loss value of the predicted label and the true label of the data representation vector X is calculated using the cross-entropy loss function.
[0034] The parameters of the first convolutional neural network are updated based on the loss values of the predicted labels and the true labels of the data representation vector X to maximize the loss values of the predicted labels and the true labels of the data representation vector X, and the updated first convolutional neural network is obtained. The data dimension of the data representation vector X is the same as the dimension of the original echo signal data.
[0035] The T groups of virtual support sets and the T groups of query sets are combined into T groups of virtual meta-tasks;
[0036] The T-group meta-tasks and the T-group virtual meta-tasks are input into the updated first convolutional neural network. The first convolutional neural network is trained using the support set of the meta-tasks and the virtual support set. The loss values of its predicted labels and true labels are obtained according to the query set. The network parameters are iteratively updated using the backpropagation algorithm.
[0037] Determine whether the loss value of the first convolutional neural network on the query set has converged;
[0038] If so, then the first convolutional neural network after training is obtained.
[0039] In one embodiment of the present invention, the specific steps of step four are as follows:
[0040] The weight parameters of the first, second, and third convolutional layers of the trained first convolutional neural network are assigned to the weight parameters of the first, second, and third convolutional layers of the second convolutional neural network, respectively.
[0041] By keeping the weight parameters of the first, second, and third convolutional layers in the second convolutional neural network unchanged, a recognition convolutional neural network is obtained.
[0042] In one embodiment of the present invention, step five includes the following steps:
[0043] W echo signals are randomly selected from the target domain fine-tuning dataset and input into the recognition convolutional neural network. The Softmax layer of the recognition convolutional neural network outputs the predicted label of each echo signal. The cross-entropy loss function is used to calculate the loss value between the predicted label and the true class label of each echo signal currently input into the recognition convolutional neural network, where W≥8.
[0044] The weight parameters of the first and second fully connected layers of the recognition convolutional neural network are iteratively updated using the backpropagation algorithm.
[0045] Determine whether the cross-entropy loss function of the current iteration has converged. If so, obtain the trained recognition convolutional neural network.
[0046] In one embodiment of the present invention, the specific steps of step six are as follows:
[0047] The echo signal of each target domain to be identified is input into the trained recognition convolutional neural network. The probability of the target being identified being classified into each category is calculated through the Softmax layer of the recognition convolutional neural network, and the category corresponding to the highest probability is selected as the recognition result output.
[0048] The beneficial effects of this invention are:
[0049] This invention adaptively generates more challenging virtual meta-tasks for the model through a virtual meta-task generation network, enabling the model to learn broader data distribution representations, reducing the data distribution differences between the source and target domains, and lowering the requirement for high similarity between source and target domain data in practical engineering. Compared with existing technologies, this invention's method has better generalizability, stability, and feasibility in cross-domain small sample scenarios.
[0050] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description
[0051] Figure 1 The flowchart illustrates a cross-domain HRRP few-shot target recognition method based on adversarial meta-transfer learning, as provided in this embodiment of the invention. Detailed Implementation
[0052] The present invention will be further described in detail below with reference to specific embodiments, but the implementation of the present invention is not limited thereto.
[0053] Example 1
[0054] like Figure 1 As shown, a cross-domain HRRP few-shot target recognition method based on adversarial meta-transfer learning includes the following steps:
[0055] Step 1: Obtain the source domain dataset and the target domain fine-tuning dataset. The source domain dataset includes radar HRRP echo signals of N target categories, where N ≥ 5. The target domain fine-tuning dataset includes radar HRRP echo signals of M target categories, where 3 ≤ M < N. The N categories and M categories are different.
[0056] Step 2: Construct the first convolutional neural network, the virtual meta-task generation network, and the second convolutional neural network respectively;
[0057] The virtual meta-task generation network consists of: random-scale convolutional layers, fully connected layers, and softmax layers. The random-scale convolutional layers have 1 convolutional kernel, a kernel size of 1*H, and a kernel stride of 1. The fully connected layers have Q neurons, where Q = M. H ∈ Set = {3, 5, 7, 9, 11, 13}. Each time the virtual meta-task generation network is invoked, the value of H is uniformly extracted from the scale set Set, and the random-scale convolutional kernel size is set.
[0058] Step 3: Train the first convolutional neural network based on the source domain dataset and the virtual meta-task generation network to obtain the trained first convolutional neural network;
[0059] Step 4: Transfer the parameters of the convolutional layers of the first convolutional neural network to the second convolutional neural network to obtain the recognition convolutional neural network.
[0060] Step 5: Train the recognition convolutional neural network based on the target domain fine-tuning dataset to obtain the trained recognition convolutional neural network;
[0061] Step 6: Input the echo signal of each target domain to be identified into the trained recognition convolutional neural network and output the recognition result.
[0062] Compared to traditional full-class, batch sample learning, this embodiment performs two-layer sampling of radar echo signals in the source domain from both data category and sample subset perspectives. This enhances the diversity of the mapping between the category space and the data space and reduces the risk of model overfitting. By constructing a virtual meta-task, this embodiment expands the data distribution range of the source domain, reduces the data distribution difference between the source and target domains to some extent, and lowers the requirement for high similarity between source and target domain data in practical engineering.
[0063] Example 2
[0064] A cross-domain HRRP few-shot target recognition method based on adversarial meta-transfer learning includes the following steps:
[0065] Step 10, Generate source domain dataset:
[0066] Extract radar HRRP echo signals containing N target categories as the source domain dataset, with each category containing at least 800 radar echo signals, where N≥5.
[0067] Step 20: Generate the target domain fine-tuning dataset:
[0068] Extract radar HRRP echo signals containing M target categories as the target domain fine-tuning dataset. Each category contains at least 30 radar echo signals, where 3 ≤ M < N. The target domain dataset and the source domain dataset do not have the same target category.
[0069] Step 30, Construct the first convolutional neural network:
[0070] The first convolutional neural network is constructed, and its structure includes, in sequence, an input layer, a first convolutional layer, a first pooling layer, a first activation layer, a second convolutional layer, a second pooling layer, a second activation layer, a third convolutional layer, a third pooling layer, a third activation layer, a first fully connected layer, a second fully connected layer, and a softmax layer.
[0071] The parameter settings for each layer are as follows:
[0072] The first convolutional neural network has 8, 16, and 32 convolutional kernels in its first, second, and third convolutional layers, respectively. The kernel size is 1×3, and the kernel stride is set to 1.
[0073] The first, second, and third pooling layers of the first convolutional neural network all use the max pooling method, with a pooling kernel size of 1×2 and a pooling kernel step size of 2.
[0074] The first, second, and third activation layers of the first convolutional neural network all use the ReLU activation function.
[0075] The first fully connected layer and the second fully connected layer of the first convolutional neural network have 256 and P neurons respectively. The number of neurons P in the second fully connected layer is equal to the total number of target domain categories M, that is, P = M.
[0076] The Softmax layer of the first convolutional neural network uses the Softmax function to calculate the probability that an input sample is identified as each class.
[0077] Step 40, Construct the virtual meta-task generation network:
[0078] A virtual metatask generation network is constructed, whose structure includes: random-scale convolutional layers, fully connected layers, and softmax layers.
[0079] The parameters for each layer are set as follows: the number of convolution kernels in the random-scale convolutional layer is set to 1, the size of the random-scale convolution kernel is set to 1*H, and the kernel stride is set to 1; the number of neurons in the fully connected layer is set to Q, and the number of neurons Q in the fully connected layer is equal to the total number of target domain categories M; the Softmax layer uses the Softmax activation function to calculate the probability that the input echo signal is classified into each category.
[0080] Where H∈Set={3,5,7,9,11,13}, when the virtual metatask generation network is called each time, the value of H is uniformly extracted from the scale set Set and the size of the random scale convolution kernel is set.
[0081] The specific steps for uniformly extracting values of H from the scale set Set and setting the convolution kernel size are as follows:
[0082] If a sample is taken according to the probability distribution P, the selected element index id is obtained. Then, the random scale convolution kernel size H = Set{id}.
[0083]
[0084] Where Z represents the total number of elements in the scale set Set, id represents the index of the element in the scale set Set, 1≤id≤Z, and Set{id} represents the id-th element in the scale set Set.
[0085] Step 50, Construct the second convolutional neural network: The structure of the second convolutional neural network is the same as that of the first convolutional neural network, but the parameters are different.
[0086] A second convolutional neural network is constructed, the structure of which includes, in sequence, an input layer, a first convolutional layer, a first pooling layer, a first activation layer, a second convolutional layer, a second pooling layer, a second activation layer, a third convolutional layer, a third pooling layer, a third activation layer, a first fully connected layer, a second fully connected layer, and a softmax layer.
[0087] The second convolutional neural network has 8, 16, and 32 convolutional kernels in the first, second, and third convolutional layers, respectively. The kernel size is 1×3, and the kernel stride is set to 1.
[0088] The pooling strategy of the first, second, and third pooling layers of the second convolutional neural network all adopts the max pooling method, with a pooling kernel size of 1×2 and a pooling kernel step size of 2.
[0089] The first, second, and third activation layers of the second convolutional neural network all use the ReLU activation function.
[0090] The number of neurons in the first fully connected layer and the second fully connected layer of the second convolutional neural network are 128 and R, respectively. The number of neurons R in the second fully connected layer is equal to the total number of target domain categories M, that is, R = M.
[0091] The Softmax layer of the second convolutional neural network uses the Softmax function to calculate the probability that an input sample is identified as each class.
[0092] Step 60: Train the first convolutional neural network based on the source domain dataset and the virtual meta-task generation network to obtain the trained first convolutional neural network. The specific steps of step 60 include:
[0093] Step 61: Randomly select S classes from the source domain dataset, and take 2*L echo signals from each class to form a meta-task. Divide the meta-task into two groups according to the number of samples. Each group contains S classes and each class contains L echo signals. Use one group as the support set of the meta-task and the other group as the query set of the meta-task.
[0094] Step 62: Repeat step 61 to extract T times, obtaining T sets of support sets and T sets of query sets. The T sets of support sets and T sets of query sets constitute T sets of meta-tasks, where 1 ≤ T < 5.
[0095] Step 63: Input the T sets of support sets into the virtual meta-task generation network in sequence, generate new data representation vectors X after passing through random scale convolutional layers, and save the T sets of data representation vectors X as the T sets of virtual support sets;
[0096] Step 64: Generate corresponding predicted labels by passing the T sets of data representation vectors X through a fully connected layer and a Softmax layer. Calculate the loss values of the predicted labels and true labels of the data representation vectors X using the cross-entropy loss function, and update the parameters of the first convolutional neural network to maximize the loss values of the predicted labels and true labels of the data representation vectors X. The data dimension of X is the same as the data dimension of the original echo signal.
[0097] Step 65: Combine the virtual support set of group T and the query set of group T into virtual meta-task of group T;
[0098] Step 66: Input the T sets of meta-tasks obtained in Step 62 and the T sets of virtual meta-tasks obtained in Step 65 into the updated first convolutional neural network. Train the network using the support set of the meta-tasks and the virtual support set of the virtual meta-tasks. Obtain the loss values of the predicted labels and the true labels from the query set. Iteratively update the network parameters using the backpropagation algorithm.
[0099] Step 67: Determine whether the loss value of the first convolutional neural network on the query set has converged. If yes, proceed to step 70 after obtaining the trained first convolutional neural network; otherwise, continue to step 62.
[0100] Step 70: Transfer the parameters of the convolutional layers of the first convolutional neural network after training to the second convolutional neural network to obtain the recognition convolutional neural network.
[0101] Specifically, the weight parameters of the first, second, and third convolutional layers of the trained first convolutional neural network are assigned to the weight parameters of the first, second, and third convolutional layers of the second convolutional neural network.
[0102] By keeping the weight parameters of the first, second, and third convolutional layers in the second convolutional neural network unchanged, a recognition convolutional neural network is obtained.
[0103] Step 80: Train the recognition convolutional neural network based on the target domain fine-tuning dataset to obtain the trained recognition convolutional neural network. The specific steps of step 80 include:
[0104] Step 81: Randomly select W echo signals from the target domain fine-tuning dataset and input them into the recognition convolutional neural network. The Softmax layer of the recognition convolutional neural network outputs the predicted label of each echo signal. Using the cross-entropy loss function, calculate the loss value between the predicted label and the true class label of each echo signal currently input into the recognition convolutional neural network, where W≥8.
[0105] Step 82: Iteratively update the weight parameters from the first fully connected layer to the second fully connected layer in step 81 using the backpropagation algorithm.
[0106] Step 83: Determine whether the cross-entropy loss function of the current iteration has converged. If it has, obtain the trained recognition convolutional neural network and then execute step 90. Otherwise, continue to execute step 81.
[0107] Step 90: Input the echo signal of each target domain to be identified into the trained recognition convolutional neural network, calculate the probability of the target being identified being classified into each category through the Softmax layer, and select the category corresponding to the highest probability as the recognition result.
[0108] This invention employs the concept of meta-learning to conduct model pre-training. Compared with traditional full-class and batch sample learning, this invention performs dual-layer sampling task extraction of radar echo signals in the source domain from both the data category and sample perspectives. This enhances the diversity of the mapping between the category space and the data space, reduces the risk of model overfitting, and can generate more promising pre-trained models with better generalization performance. Therefore, the method of this invention has better stability.
[0109] This invention introduces virtual meta-tasks, which adversarially generate more challenging virtual meta-tasks for the model through a virtual meta-task generation network. This expands the data distribution range of the tasks during the training phase, thereby reducing the data distribution differences between the source and target domains and lowering the requirement for high similarity between source and target domain data in practical engineering. Therefore, the method of this invention has better generalizability.
[0110] In the implementation of this invention, the virtual meta-task is adaptively generated by the network according to the corresponding source task distribution. The implementation effect of the method does not depend on a certain key parameter weight, thus eliminating the pre-training parameter optimization process. Therefore, the method of this invention has better feasibility.
[0111] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.
[0112] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. In addition, those skilled in the art can combine and integrate the different embodiments or examples described in this specification.
[0113] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.
Claims
1. A cross-domain HRRP few-shot target recognition method based on adversarial meta-transfer learning, characterized in that, Includes the following steps: Step 1: Obtain the source domain dataset and the target domain fine-tuning dataset; wherein, the source domain dataset includes radar HRRP echo signals of N categories of targets, N≥5; the target domain fine-tuning dataset includes radar HRRP echo signals of M categories of targets, 3≤M<N, and the N categories and M categories are different; Step 2: Construct the first convolutional neural network, the virtual meta-task generation network, and the second convolutional neural network respectively; The virtual metatask generation network includes: random-scale convolutional layers, fully connected layers, and softmax layers. The random-scale convolutional layers have 1 convolutional kernel, a kernel size of 1*H, and a kernel stride of 1. The fully connected layers have Q neurons, where Q=M. in, Each time the virtual meta-task generation network is invoked, from the scale set Uniform extraction H The values are set and the random scale convolution kernel size is set; Step 3: Train the first convolutional neural network based on the source domain dataset and the virtual meta-task generation network to obtain the trained first convolutional neural network; Step four: Transfer the parameters of the convolutional layers of the trained first convolutional neural network to the second convolutional neural network to obtain the recognition convolutional neural network; Step 5: Train the recognition convolutional neural network based on the target domain fine-tuning dataset to obtain the trained recognition convolutional neural network; Step 6: Input the echo signal of each target domain to be identified into the trained recognition convolutional neural network and output the recognition result; Step three includes the following steps: S classes are randomly selected from the source domain dataset, and 2*L echo signals are selected from each class to form a meta-task. The meta-task is divided into two groups according to the number of samples. Each group contains S classes and each class contains L echo signals. One group is used as the support set of the meta-task, and the other group is used as the query set of the meta-task. T echo signals are extracted from the source domain dataset to obtain T sets of support sets and T sets of query sets. The T sets of support sets and the T sets of query sets constitute T sets of meta-tasks, where 1 ≤ T < 5. The T sets of support are sequentially input into the virtual meta-task generation network. After passing through a random-scale convolutional layer, T sets of data representation vectors X are generated, and the T sets of data representation vectors X are saved as T sets of virtual support. The T sets of data representation vectors X are passed through the fully connected layer and the Softmax layer of the virtual meta-task generation network to generate corresponding predicted labels, and the loss value of the predicted label and the true label of the data representation vector X is calculated using the cross-entropy loss function. The parameters of the first convolutional neural network are updated based on the loss values of the predicted labels and the true labels of the data representation vector X to maximize the loss values of the predicted labels and the true labels of the data representation vector X, and the updated first convolutional neural network is obtained. The data dimension of the data representation vector X is the same as the dimension of the original echo signal data. The T groups of virtual support sets and the T groups of query sets are combined into T groups of virtual meta-tasks; The T-group meta-tasks and the T-group virtual meta-tasks are input into the updated first convolutional neural network. The first convolutional neural network is trained using the support set of the meta-tasks and the virtual support set. The loss values of its predicted labels and true labels are obtained according to the query set. The network parameters are iteratively updated using the backpropagation algorithm. Determine whether the loss value of the first convolutional neural network on the query set has converged; If so, then the first convolutional neural network after training is obtained.
2. The method for cross-domain HRRP small sample target recognition based on adversarial meta-transfer learning according to claim 1, characterized in that, The scale set Uniform extraction H The values of and the size of the random-scale convolution kernel are set, including: A single sampling is performed based on the probability distribution P to obtain the selected element index. Then the random-scale convolution kernel size ; in, , Representation scale set The total number of elements in Representation scale set The index of the element in the middle. , Representation scale set The first in Each element.
3. The method for cross-domain HRRP few-shot target recognition based on adversarial meta-transfer learning according to claim 1, characterized in that, The first convolutional neural network and the second convolutional neural network have the same structure, both including an input layer, a first convolutional layer, a first pooling layer, a first activation layer, a second convolutional layer, a second pooling layer, a second activation layer, a third convolutional layer, a third pooling layer, a third activation layer, a first fully connected layer, a second fully connected layer, and a softmax layer arranged sequentially.
4. The method for cross-domain HRRP few-shot target recognition based on adversarial meta-transfer learning according to claim 1, characterized in that, The first convolutional neural network has 8, 16 and 32 convolutional kernels in its first, second and third convolutional layers, respectively. The kernel size is 1×3 and the kernel stride is set to 1. The pooling strategies of the first pooling layer, the second pooling layer and the third pooling layer of the first convolutional neural network all adopt the max pooling method, the pooling kernel size is 1×2 and the pooling kernel moving step size is 2. The first activation layer, the second activation layer, and the third activation layer of the first convolutional neural network all use the ReLU activation function; The first fully connected layer and the second fully connected layer of the first convolutional neural network have 256 and P neurons, respectively, where P=M; The Softmax layer of the first convolutional neural network uses the Softmax function to calculate the probability that an input sample is identified as each class.
5. The method for cross-domain HRRP few-shot target recognition based on adversarial meta-transfer learning according to claim 1, characterized in that, The number of convolutional kernels in the first, second, and third convolutional layers of the second convolutional neural network are 8, 16, and 32, respectively. The kernel size is 1×3, and the kernel stride is set to 1. The pooling strategies of the first, second, and third pooling layers of the second convolutional neural network all adopt the max pooling method, with a pooling kernel size of 1×2 and a pooling kernel step size of 2. The first, second, and third activation layers of the second convolutional neural network all use the ReLU activation function; The number of neurons in the first fully connected layer and the second fully connected layer of the second convolutional neural network are 128 and R, respectively, where R=M; The Softmax layer of the second convolutional neural network uses the Softmax function to calculate the probability that an input sample is identified as each class.
6. The method for cross-domain HRRP few-shot target recognition based on adversarial meta-transfer learning according to claim 1, characterized in that, The specific steps of step four are as follows: The weight parameters of the first, second, and third convolutional layers of the trained first convolutional neural network are assigned to the weight parameters of the first, second, and third convolutional layers of the second convolutional neural network, respectively. By keeping the weight parameters of the first, second, and third convolutional layers in the second convolutional neural network unchanged, a recognition convolutional neural network is obtained.
7. The method for cross-domain HRRP few-shot target recognition based on adversarial meta-transfer learning according to claim 1, characterized in that, Step five includes the following steps: W echo signals are randomly selected from the target domain fine-tuning dataset and input into the recognition convolutional neural network. The Softmax layer of the recognition convolutional neural network outputs the predicted label of each echo signal. The cross-entropy loss function is used to calculate the loss value between the predicted label and the true class label of each echo signal currently input into the recognition convolutional neural network, where W≥8. The weight parameters of the first and second fully connected layers of the recognition convolutional neural network are iteratively updated using the backpropagation algorithm. Determine whether the cross-entropy loss function of the current iteration has converged. If so, obtain the trained recognition convolutional neural network.
8. The method for cross-domain HRRP few-shot target recognition based on adversarial meta-transfer learning according to claim 1, characterized in that, The specific steps of step six are as follows: The echo signal of each target domain to be identified is input into the trained recognition convolutional neural network. The probability of the target being identified being classified into each category is calculated through the Softmax layer of the recognition convolutional neural network, and the category corresponding to the highest probability is selected as the recognition result output.