Small sample radar HRRP recognition method based on feature generation
By constructing a target recognition network model and using a self-attention mechanism and a maximum mean difference loss function to decompose HRRP data into identity features and angle features, the problem of small sample size in automatic radar target recognition is solved, and the robustness and accuracy of recognition are improved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- XIAN UNIV OF TECH
- Filing Date
- 2026-01-20
- Publication Date
- 2026-06-09
AI Technical Summary
Existing radar automatic target recognition technologies struggle to extract robust features under small sample conditions, resulting in poor recognition performance. Furthermore, existing methods suffer from overfitting and recognition instability.
A feature-based small-sample radar HRRP identification method is adopted. By constructing a target recognition network model, a feature extraction module, a feature decoupling module, an identity recognition module, an angle decoding module, and an angle encoding module are used. Combined with a self-attention mechanism and a maximum mean difference loss function, the HRRP data is decomposed into identity features and angle features, and mixed features under arbitrary angles are generated.
It enhances the model's ability to perceive details of local distance units, breaks through angle limitations, effectively supplements the feature space of small sample data, and improves the robustness and accuracy of recognition.
Smart Images
Figure CN122173879A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of radar automatic target recognition technology, and more specifically to a small-sample radar HRRP recognition method based on feature generation. Background Technology
[0002] Automatic target identification (ADI) technology for radar builds upon traditional radar detection and tracking by mining and modeling features from the reflected echoes of targets to automatically identify their attributes, categories, and types. Existing ADI technologies primarily include one-dimensional high-resolution range profile (HRRP) target identification and two-dimensional high-resolution image target identification. Compared to high-resolution images, HRRP, as a one-dimensional vector, not only contains rich information such as target structure and scattering characteristics but also offers advantages such as ease of acquisition, storage, and processing, thus finding widespread application in target type identification tasks.
[0003] However, for targets such as aircraft, ships, and automobiles, due to their non-cooperative nature, it is difficult to obtain sufficient samples with accurate category labels and complete azimuth domains. Therefore, high-resolution range image target recognition suffers from the small sample problem. When radar target recognition suffers from the small sample problem, the model cannot fit the distribution of real data and struggles to extract sufficiently robust features, resulting in poor recognition performance on unknown test samples and overfitting. Therefore, the small sample problem becomes crucial for improving the robustness, adaptability, and scalability of high-resolution range image recognition technology.
[0004] To address this issue, existing research largely focuses on either the feature level or the model level. Feature-level approaches extract high-level features that differ from traditional statistical features, ultimately enabling target recognition. However, this method reduces the accuracy of describing HRRP data characteristics, leading to the loss of target information and impacting target recognition performance. Model-level approaches aim to reduce the model's sample size requirements by using improved recognition networks or training strategies, enhancing the model's robustness under small sample conditions. However, this approach suffers from high computational complexity and unstable recognition performance. Summary of the Invention
[0005] The purpose of this invention is to solve the problem of inaccurate identification of radar targets when there are few samples, which leads to a decrease in radar target identification accuracy.
[0006] The technical solution adopted in this invention is a small-sample radar HRRP recognition method based on feature generation, and the specific operation steps are as follows: Step 1: Acquire the radar echo signal of the target, divide the radar echo signal into multiple sub-echo signals, and convert the sub-echo signals into HRRP data sample sets; and preprocess the HRRP data; the target to be identified includes, but is not limited to, any one of automobiles, airplanes, and ships, and the collected radar echo signal can be in a stationary state or in a moving state.
[0007] Step 2: Construct a target recognition network model, which includes a feature extraction module, a feature decoupling module, an identity recognition module, a feature generation module, an angle decoding module, and an angle encoding module; Step 3: Input the HRRP data preprocessed in Step 1 into the target recognition network model for training. The HRRP data is processed by feature extraction to obtain hybrid features. The hybrid features are then sent to the feature decoupling module to decouple them into identity features and angle features. Step 4: The identity features are identified by the identity recognition module and stored in the identity category feature library; the angle features are calculated by the angle decoding module, the desired angle is input, and the desired angle features are obtained by the angle encoding module. Step 5: Then, the target's identity features and the desired angle features are concatenated to fuse the information from the two features; the hybrid feature generation module can then generate new hybrid features at the desired angle. Step 6: Repeat steps 3-5 until the model parameters are gradually optimized and tend to converge, which means the angle-supervised feature generation and recognition training is complete. Then test the model to obtain the recognition results.
[0008] The invention is further characterized in that, In the feature extraction module, firstly, a one-dimensional convolution with a large kernel is used to capture local waveform features and short-term change patterns in the original signal; then, the time length is gradually reduced through multiple layers of small kernels and a one-dimensional convolution with a stride of 2, which enhances feature robustness while extracting higher-level temporal structure information; after pooling, a 1×1 one-dimensional convolution is finally used to complete the integration of channel-dimensional information. In the identity recognition module, the extracted identity features are subjected to a Flatten operation to obtain a fixed-length feature vector; then, a linear mapping layer is used to extract high-level feature information, and a ReLU activation function is connected to enhance the non-linear expressive ability of the network; on this basis, a linear layer and a Softmax layer are passed to achieve the final identity category probability output. The angle decoding module consists of an angle decoder, used to further decode angle features into regressible angle values. The specific process is as follows: First, the extracted angle features are subjected to a Flatten operation to obtain a fixed-length representation vector; then, this vector is sequentially input into two linear mapping layers, and a Tanh nonlinear activation function is introduced after each linear mapping layer to enhance the expressive power and nonlinear fitting ability of the angle features; finally, the output result is normalized by the Sigmoid activation function to obtain the corresponding normalized angle prediction value.
[0009] In step 1, the HRRP data sample set is divided into a training set and a test set in a 3:7 ratio.
[0010] Furthermore, methods for preprocessing HRRP data include: Norm normalization and bary alignment; Perform HRRP data Norm normalization yields:
[0011] in: It is the processed data; This is the original HRRP data; Represents the 2-norm operation; Align the HRRP data with the center of gravity:
[0012] in: W As the center of gravity; N HRRP sample length; for n The amplitude of the HRRP signal at the point; Then, the center of each HRRP sample is shifted to the centroid, thus completing the centroid alignment operation.
[0013] Furthermore, in step 3, the preprocessed HRRP data is input and processed by the feature extraction module to obtain mixed features:
[0014] in, This is the feature extraction module, which obtains the mixed features of HRRP. Represents mixed characteristics; This represents HRRP data.
[0015] Furthermore, in step 3, the mixed features are fed into the feature decoupling module and decoupled into identity features and angle features; Decompose the mixed features into:
[0016] in: This indicates multiplication at corresponding positions; This represents the self-attention mechanism; Indicates identity characteristics; Indicates angular characteristics.
[0017] Identity features and perspective features are consistent with the hybrid features in terms of dimension, which facilitates feature decoupling and subsequent processing.
[0018] Furthermore, in step 3, to decouple the separated angular features and identity features as much as possible, the identity features and angular features of the HRRP data are treated as two different domains. The identity features and angular features are decoupled by maximizing the distribution difference between these two domains. Therefore, the Maximum Mean Difference (MMD) loss function is used to update the network parameters.
[0019] in, C Indicates the total number of categories. Indicates the first j The number of class samples, It is a nonlinear feature mapping function that maps input features to a reproducing kernel Hilbert space. It is a positive number, with a value between 0 and 1.
[0020] Furthermore, the mathematical expression of the self-attention mechanism is:
[0021] in: Q , K , V These represent the query matrix, key matrix, and value matrix, respectively, with dimensions of [dimension number missing]. :
[0022] in: They are respectively for Q , K , V The learnable weight matrix.
[0023] The SoftMax function is used to normalize the calculated similarity and finally output the weights.
[0024] Furthermore, the specific process of step 4 is as follows: the identity features are processed by the identity recognition module to obtain the recognition result, and then stored in the identity category feature library; The angle feature is processed by the angle decoding module to calculate its angle value. The desired angle is input, and the angle encoding module then outputs the desired angle feature.
[0025] in, Indicates the angle encoding module, Indicates the desired perspective; identifies the target's characteristics. Expected characteristics of the target The two features are concatenated to fuse their information; the mixed feature generation module generates mixed features at the desired angle; once the model is trained, the subsequent generation process will not depend on the angle range of the samples, and the module can generate mixed features from any input angle.
[0026] in, This indicates the hybrid feature generation module.
[0027] The beneficial effects of this invention are as follows: 1. The small-sample radar HRRP recognition method based on feature generation proposed in this invention can decompose the HRRP hybrid features into angle features and identity features. Through the self-attention mechanism, attention weights can be dynamically allocated according to the features of the distance unit, thereby enhancing the model's ability to perceive the details of local distance units while maintaining the overall feature structure, and providing a more accurate feature representation for subsequent feature decoupling.
[0028] 2. The feature generation enhancement method of the present invention differs from the traditional data sample generation method. It breaks through the angular limitations of the original sample and can generate HRRP data features at any angle, effectively supplementing the feature space of the original small sample data. Attached Figure Description
[0029] Figure 1 This is an overall flowchart of the feature-based small-sample radar HRRP recognition method of the present invention; Figure 2 This is the network structure model of the feature-based small-sample radar HRRP identification method of the present invention; Figure 3(a) shows the waveform of the HRRP sample before normalization preprocessing; Figure 3(b) shows the waveform of the HRRP sample after normalization preprocessing; Figure 3(c) shows the amplitude diagram before preprocessing using the centroid alignment method; Figure 3(d) shows the amplitude diagram after preprocessing using the centroid alignment method; Figure 4 This is a comparison chart of the recognition accuracy of the method of the present invention with other traditional recognition methods. Detailed Implementation
[0030] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments, but the embodiments of the present invention are not limited thereto.
[0031] Example 1 See Figure 1 and Figure 2 , Figure 1 This is an overall flowchart of the feature-based small-sample radar HRRP recognition method of the present invention. Figure 2 This invention presents the network structure model of a feature-based small-sample radar HRRP recognition method. The angle-supervised feature generation recognition method comprises two stages: a training stage and a recognition stage. The training stage includes five steps: converting radar echo signals into an HRRP training sample set, preprocessing the training samples, extracting features from the training samples, decoupling features, and training the classifier. The recognition stage includes five steps: converting radar echo signals into HRRP recognition samples, preprocessing the recognition samples, extracting features from the recognition samples, decoupling features from the recognition samples, and outputting the recognition result from the classifier.
[0032] Training phase: Step 1.1: The radar echo signal is converted into an HRRP training sample set. The data used comes from a publicly available dataset published by AFRL Labs, which contains simulation data of different types of civilian vehicles. During training, different numbers of training samples are used to construct the training set, resulting in the HRRP data training set.
[0033] Step 1.2: Preprocess the HRRP data, mainly including... Norm normalization and barycentric alignment. This is performed on HRRP data. Norm normalization yields:
[0034] in, This is the processed HRRP data. It is the original HRRP data. Represents the 2-norm operation; Align the HRRP data with the center of gravity:
[0035] in, W As the center of gravity, NFor HRRP sample length, for n The amplitude of the HRRP signal at the point; Then, the center of each HRRP sample is shifted to the centroid to complete the centroid alignment operation. Figure 3 shows a comparison before and after preprocessing.
[0036] Step 1.3: Input the preprocessed HRRP data, and obtain the mixed features through the feature extraction module:
[0037] in, As a feature extraction module, in this embodiment, the module first uses a one-dimensional convolution with a large kernel to capture local waveform features and short-term change patterns in the original signal; then, it gradually reduces the time length through multiple layers of small kernels and a one-dimensional convolution with a stride of 2, thereby enhancing feature robustness and extracting higher-level temporal structure information; after pooling, it finally uses a 1×1 one-dimensional convolution to complete the integration of channel-dimensional information.
[0038] Step 1.4: Send the mixed features into the feature decoupling module to decouple them into identity features and angle features.
[0039] Decompose the mixed features into:
[0040] in, This indicates multiplication at corresponding positions. This indicates the self-attention mechanism. Indicates identity characteristics, The angular feature is represented. In this embodiment, the identity feature and the angular feature are consistent with the hybrid feature in terms of dimension, which facilitates feature decoupling and subsequent processing.
[0041] To decouple the angular and identity features as much as possible, this embodiment treats the identity and angular features of the HRRP data as two distinct domains. The identity and angular features are decoupled by maximizing the distribution difference between these two domains; therefore, the Maximum Mean Difference (MMD) loss function is used to update the network parameters.
[0042] in, C Indicates the total number of categories. Indicates the first j The number of class samples, It is a nonlinear feature mapping function that maps input features to a reproducing kernel Hilbert space. It is a small positive number.
[0043] Step 1.5: The identity features are processed by the identity recognition module to obtain the recognition results and are stored in the identity category feature library; the identity features indicate the different models of the same type of target (such as which ship is a cruise ship and which is a fishing boat), and the angle features show the different attitude angles of the target. The angle features further assist in accurately identifying the target model category.
[0044] Example 2 Based on Example 1, the identity recognition module structure in this example is as follows: First, the identity features are flattened to obtain a fixed-length feature vector; then, a linear mapping layer is used to extract high-level feature information, and a ReLU activation function is applied to enhance the non-linear expressive power of the network; based on this, a linear layer and a softmax layer are then applied to achieve the final identity category probability output. Cross-entropy loss is used here.
[0045] in, N For sample size, identity features The identity recognition module obtains logits (unnormalized classification scores), which are then processed by softmax to obtain the classification probabilities. , N The cross-entropy loss is finally calculated for each sample. .
[0046] The angle feature is processed by the angle decoding module to calculate its angle value. The desired angle is input, and the angle encoding module then outputs the desired angle feature.
[0047] in, Indicates the angle encoding module, This represents the desired angle. In this embodiment, the angle decoding module consists of an angle decoder, used to further decode the angle features into regressible angle values. The specific process is as follows: First, the extracted angle features are subjected to a Flatten operation to obtain a fixed-length representation vector; then, this vector is sequentially input into two linear mapping layers, and a Tanh nonlinear activation function is introduced after each linear layer to enhance the expressive power and nonlinear fitting ability of the angle features; finally, the output result is normalized using a Sigmoid activation function to obtain the corresponding normalized angle prediction value.
[0048] The target's identity characteristics Expected characteristics of the target The two features are concatenated to fuse their information, and the resulting blended feature generation module generates a blended feature at the desired angle. Once the model is trained, subsequent generation processes will not depend on the angle range of the samples; any input angle can generate its blended features from the blended feature generation module.
[0049] in, This refers to the hybrid feature generation module; in this embodiment, the hybrid feature generation module first generates the target's identity features. Angular features obtained from the expected angle encoder The features are concatenated to fuse the conditional information of the two types of features. Subsequently, the features are upsampled step by step through a three-layer ConvTranspose + BatchNorm + ReLU structure. Deconvolution (transposed convolution) is used to gradually restore the resolution of the feature map, while BatchNorm and ReLU further stabilize the training process and enhance the non-linear expressive power of the features, so that the fused information is fully refined and enhanced during the generation process.
[0050] Ideally, the generated mixed features should have the same distribution in the feature space as the mixed features of the original data. Therefore, this embodiment uses center loss:
[0051] in: Represented as the original sample number j Mixed feature center of class, Represents the generation of the first j Mixed feature centers of the class.
[0052] The formula for calculating the feature center is:
[0053] in, For the first j The mean of the mixed features of the class samples, For the first j Number of class samples It is a mixed feature center.
[0054] This completes the training phase of the feature-based small-sample radar HRRP recognition method. Example 3 The testing phase includes five stages: conversion of radar echo signals into HRRP recognition samples, recognition sample preprocessing, recognition sample feature extraction, recognition sample feature decoupling, and classifier output of recognition results. Specifically, it includes the following steps: Step 2.1: Following the method in Step 1.1, the data used comes from the publicly available dataset published by AFRL Labs, which contains simulation data of different types of civilian vehicles. During testing, a test set is constructed using samples different from the training set, resulting in the HRRP data test set.
[0055] Step 2.2: Following the method in Step 1.2, preprocess the HRRP test set data to obtain the preprocessed test HRRP data.
[0056] Step 2.3: Following the method in Step 1.3, input the preprocessed test HRRP data, and obtain the mixed features of the test data through the trained feature extraction module.
[0057] Step 2.4: Following the method in Step 1.4, send the mixed features of the test data into the feature decoupling module to decouple them into identity features and angle features.
[0058] Step 2.5: Following the method in Step 1.5, input the identity features of the test sample into the classification module to obtain the recognition accuracy.
[0059] Example 4 The method of this invention is compared with conditional data generation methods, feature-only decoupling methods, and traditional CNN methods under different numbers of training samples. The structure is as follows: Figure 4 As shown in the figure, the conditional data generation method refers to generating data using CGAN as the baseline model, and then mixing the generated data with the original data to train the recognition network. The feature-only decoupling method simply trains the network using a self-attention-based feature decoupling structure without performing feature-level data augmentation. The traditional CNN method is the ordinary CNN network recognition method; the proposed method is the method of this invention. As can be seen from the figure, the method of this invention has the highest recognition accuracy with the same number of samples, especially under small sample conditions, its recognition accuracy is even more significant.
[0060] Figures 3(a) and 3(b) show the HRRP's range cell number and the corresponding target's different scattering points along the range dimension on the horizontal axis; the vertical axis represents the HRRP's amplitude value. Before normalization, the HRRP amplitude range is large and fluctuates wildly, easily affected by factors such as target range and radar parameters. After normalization, the amplitude is uniformly compressed into a fixed range (0-1), resulting in a smoother overall amplitude. Although the absolute amplitude changes, the relative strength relationships at different range cells and the positions of the main scattering peaks are preserved. Therefore, normalization not only highlights the structural characteristics of HRRP and reduces interference caused by amplitude differences, but also facilitates subsequent network training, improving the algorithm's stability and comparability.
[0061] Figures 3(c) and 3(d) show a two-dimensional visualization comparison of HRRP before and after preprocessing using the "centroid alignment method" (the horizontal axis represents distance units, the vertical axis represents sample indices, and the colors indicate amplitude strength). The HRRP centroid alignment method eliminates the uncertainty caused by distance translation, ensuring that the main scattering energy of different samples remains consistent along the distance dimension. This highlights the stable structural features of the target, enhances the comparability between samples, and provides a more robust and stable data foundation for subsequent feature extraction and target recognition.
[0062] The results show that the proposed method outperforms other methods under different sample sizes, demonstrating the rationality of the proposed method design. When the sample size is small, the proposed method shows a more significant improvement over other methods, indicating that the proposed method is well-suited for small sample conditions.
[0063] Example 5 The present invention provides a small-sample radar HRRP recognition method based on feature generation, and the specific operation steps are as follows: Step 1: Acquire the radar echo signal of the target, divide the radar echo signal into multiple sub-echo signals, and convert the sub-echo signals into HRRP data sample sets; and preprocess the HRRP data. Step 2: Construct a target recognition network model, which includes a feature extraction module, a feature decoupling module, an identity recognition module, a feature generation module, an angle decoding module, and an angle encoding module; Step 3: Input the HRRP data preprocessed in Step 1 into the target recognition network model for training. The HRRP data is processed by feature extraction to obtain hybrid features. The hybrid features are then sent to the feature decoupling module to decouple them into identity features and angle features. Step 4: The identity features are identified by the identity recognition module and stored in the identity category feature library; the angle features are calculated by the angle decoding module, the desired angle is input, and the desired angle features are obtained by the angle encoding module. Step 5: Then, the target's identity features and the desired angle features are concatenated to fuse the information from the two features; the hybrid feature generation module can then generate new hybrid features at the desired angle. Step 6: Repeat steps 3-5 until the model parameters are gradually optimized and tend to converge, which means the angle-supervised feature generation and recognition training is complete. Then test the model to obtain the recognition results.
[0064] Example 6 Based on Example 5, In the feature extraction module, firstly, one-dimensional convolution with large kernels is used to capture local waveform features and short-term change patterns in the original signal; then, the time length is gradually reduced through multiple layers of small kernels and one-dimensional convolution with a stride of 2, which enhances feature robustness while extracting higher-level temporal structure information; after pooling layers, 1×1 one-dimensional convolution is finally used to complete the integration of channel-dimensional information. In the identity recognition module, the extracted identity features are subjected to a Flatten operation to obtain a fixed-length feature vector; then, a linear mapping layer is used to extract high-level feature information, and a ReLU activation function is connected to enhance the non-linear expressive ability of the network; on this basis, a linear layer and a Softmax layer are passed to achieve the final identity category probability output. The angle decoding module consists of an angle decoder, used to further decode angle features into regressible angle values. The specific process is as follows: First, the extracted angle features are subjected to a Flatten operation to obtain a fixed-length representation vector; then, this vector is sequentially input into two linear mapping layers, and a Tanh nonlinear activation function is introduced after each linear mapping layer to enhance the expressive power and nonlinear fitting ability of the angle features; finally, the output result is normalized by the Sigmoid activation function to obtain the corresponding normalized angle prediction value.
Claims
1. A feature-based small-sample radar HRRP recognition method, specifically including the following steps: Step 1: Acquire the radar echo signal of the target, divide the radar echo signal into multiple sub-echo signals, and convert the sub-echo signals into HRRP data sample sets; and preprocess the HRRP data. Step 2: Construct a target recognition network model, which includes a feature extraction module, a feature decoupling module, an identity recognition module, a feature generation module, an angle decoding module, and an angle encoding module; Step 3: Input the HRRP data preprocessed in Step 1 into the target recognition network model for training. The HRRP data is processed by feature extraction to obtain hybrid features. The hybrid features are then sent to the feature decoupling module to decouple them into identity features and angle features. Step 4: The identity features are identified by the identity recognition module and stored in the identity category feature library; the angle features are calculated by the angle decoding module, the desired angle is input, and the desired angle features are obtained by the angle encoding module. Step 5: Then, the target's identity features and the desired angle features are concatenated to fuse the information from the two features; the hybrid feature generation module can then generate new hybrid features at the desired angle. Step 6: Repeat steps 3-5 until the model parameters are gradually optimized and tend to converge, which means the angle-supervised feature generation and recognition training is complete. Then test the model to obtain the recognition results.
2. The small-sample radar HRRP recognition method based on feature generation according to claim 1, characterized in that: In the feature extraction module, firstly, one-dimensional convolution with large kernels is used to capture local waveform features and short-term change patterns in the original signal; then, the time length is gradually reduced through multiple layers of small kernels and one-dimensional convolution with a stride of 2, which enhances feature robustness while extracting higher-level temporal structure information; after pooling layers, 1×1 one-dimensional convolution is finally used to complete the integration of channel-dimensional information. In the identity recognition module, the extracted identity features are subjected to a Flatten operation to obtain a fixed-length feature vector; then, a linear mapping layer is used to extract high-level feature information, and a ReLU activation function is connected to enhance the non-linear expressive ability of the network; on this basis, a linear layer and a Softmax layer are passed to achieve the final identity category probability output. The angle decoding module consists of an angle decoder, which is used to further decode the angle features into regressible angle values. The specific process is as follows: First, the extracted angle features are subjected to a Flatten operation to obtain a fixed-length representation vector; then, the vector is sequentially input into two linear mapping layers, and a Tanh nonlinear activation function is introduced after each linear mapping layer to enhance the expressive power and nonlinear fitting ability of the angle features. Finally, the output is normalized using the Sigmoid activation function to obtain the corresponding normalized angle prediction value.
3. The small-sample radar HRRP recognition method based on feature generation according to claim 1, characterized in that, In step 1, the HRRP data sample set is divided into a training set and a test set in a 3:7 ratio.
4. The small-sample radar HRRP recognition method based on feature generation according to claim 1, characterized in that: Methods for preprocessing HRRP data include: Norm normalization and bary alignment; Perform HRRP data Norm normalization yields: in: It is the processed data; This is the original HRRP data; Represents the 2-norm operation; Align the HRRP data with the center of gravity: in: W As the center of gravity; N HRRP sample length; for n The amplitude of the HRRP signal at the point; Then, the center of each HRRP sample is shifted to the centroid, thus completing the centroid alignment operation.
5. The small-sample radar HRRP recognition method based on feature generation according to claim 1, characterized in that: In step 3, the preprocessed HRRP data is input, and the feature extraction module obtains the mixed features: in, This is the feature extraction module, which obtains the mixed features of HRRP. Represents mixed characteristics; This represents HRRP data.
6. The small-sample radar HRRP recognition method based on feature generation according to claim 5, characterized in that: In step 3, the mixed features are sent to the feature decoupling module to decouple them into identity features and angle features; Decompose the mixed features into: in: This indicates multiplication at corresponding positions; This represents the self-attention mechanism; Indicates identity characteristics; Indicates angular characteristics.
7. The small-sample radar HRRP recognition method based on feature generation according to claim 6, characterized in that: In step 3, to decouple the separated angular features and identity features as much as possible, the identity features and angular features of the HRRP data are treated as two different domains. The decoupling of identity features and angular features is achieved by maximizing the distribution difference between these two domains. Therefore, the Maximum Mean Difference (MMD) loss function is used to update the network parameters. in, C Indicates the total number of categories. Indicates the first j The number of class samples, It is a nonlinear feature mapping function that maps input features to a reproducing kernel Hilbert space. It is a positive number, with a value between 0 and 1.
8. The small-sample radar HRRP recognition method based on feature generation according to claim 7, characterized in that: The mathematical expression for the self-attention mechanism is: in: Q , K , V These represent the query matrix, key matrix, and value matrix, respectively, with dimensions of [dimension number missing]. : in: They are respectively for Q , K , V The learnable weight matrix; The SoftMax function is used to normalize the calculated similarity and finally output the weights.
9. The small-sample radar HRRP recognition method based on feature generation according to claim 7, characterized in that: The specific process of step 4 is as follows: the identity features are processed by the identity recognition module to obtain the recognition result, and then stored in the identity category feature library; The angle feature is processed by the angle decoding module to calculate its angle value. The desired angle is input, and the angle encoding module then outputs the desired angle feature. in, Indicates the angle encoding module, Indicates the desired perspective; identifies the target's characteristics. Expected characteristics of the target The two features are spliced together to fuse their information; the hybrid feature generation module can then generate hybrid features at the desired angle. in, This indicates the hybrid feature generation module.