A method and device for radar target recognition with azimuth enhancement under physical scattering constraints

Abstract

The application discloses a radar target recognition method and device with azimuth enhancement under physical scattering constraints. The core is that a feature level azimuth expansion module (FAEM) is constructed, sparse view angle features are augmented in a feature space, and effective expansion of azimuth information is realized; meanwhile, a scattering center reconstruction constraint module (SCRCM) is innovatively introduced, physical authenticity and semantic fidelity of generated features are improved; in addition, multi-view perception enhancement modules (MVFM) are used to fuse generated multi-azimuth features, and intra-class compactness constraints are applied to optimize feature space structures, and the recognition ability of the model is further improved. The method realizes optimization from feature generation, physical constraints to semantic enhancement, and effectively improves the target recognition performance under sparse view angles.

Description

Technical Field

[0001] This invention belongs to the field of radar signal processing and artificial intelligence technology, specifically relating to a target recognition method for high-resolution range profiles (HRRP) radar under conditions of sparse azimuth coverage in training data. In particular, to achieve azimuth augmentation in the feature space, this invention proposes scattering center physical model constraints and multi-view semantic consistency enhancement to improve the generalization ability of the recognition model. Background Technology

[0002] The target echo of an ultra-narrow pulse radar contains multiple ultra-narrow pulses, each corresponding to a different scattering point on the target. This represents the distribution of the target's scattering points along the radar's line-of-sight, and is often referred to as the high-resolution range profile (HRRP) of the target. HRRP reflects the physical structural characteristics of the target and is of significant value in automatic radar target identification. However, HRRP is extremely sensitive to changes in the target's azimuth angle, exhibiting an "attitude sensitivity" problem. Theoretically, omnidirectional training data is required to build a robust identification model, but in practical applications, especially for non-cooperative targets, only sparse azimuth training data is often available, leading to a significant decrease in the model's identification performance when facing untrained viewpoints.

[0003] Existing solutions mainly fall into two categories: few-shot learning-based methods and generative methods. Few-shot learning methods improve generalization ability through model-level improvements such as transfer learning or meta-learning. Generative methods expand the training data by synthesizing new samples using generative models.

[0004] However, existing methods have the following key technical limitations: First, the performance of traditional few-shot learning-based methods depends on the similarity between the source and target domains, and their performance is limited when the azimuth interval is large. Second, generation-based methods suffer from unstable generation quality and a lack of targeted optimization for the recognition task. Furthermore, existing methods often neglect semantic consistency across multiple viewpoints, resulting in insufficient feature discriminative power.

[0005] To address the aforementioned problems, this invention proposes a novel radar target recognition method with feature-level azimuth enhancement. The core of this method lies in elevating data augmentation from the raw signal level to the feature space. Through the proposed physical scattering constraints and multi-view semantic consistency enhancement method, the view coverage of the training set is effectively expanded at the feature level, providing the model with more comprehensive attitude change information. First, this invention designs a Scattering Center Reconstruction Constraint Module (SCRCM), deeply integrating the physical mechanism of electromagnetic scattering into the feature generation process, effectively solving the semantic distortion problem existing in traditional generation methods and significantly improving the physical fidelity of the generated features. Second, this invention proposes a Multi-View Perception Feature Enhancement Module (MFAM), which achieves adaptive fusion of multi-view features through spatial and channel interactive attention, and, combined with intra-class multi-view compactness constraints, explicitly optimizes the feature space structure, effectively enhancing the semantic consistency of similar samples under different azimuth angles, significantly improving feature discriminativeness and model robustness. Summary of the Invention

[0006] The purpose of this invention is to construct an end-to-end recognition network that integrates feature generation, physical constraints, and semantic enhancement. This network mainly consists of a Feature-Level Azimuth Extension Module (FAEM), a Scattering Center Reconstruction Constraint Module (SCRCM), and a Multi-View Perception Feature Enhancement Module (MFAM). By proposing physical structure consistency constraints and intra-class semantic compactness constraints, it significantly improves the accuracy and robustness of radar HRRP target recognition under sparse viewpoints.

[0007] The technical solution of the method of the present invention is as follows: This invention proposes a radar target HRRP identification method based on physical structure constraints and multi-view semantic enhancement, comprising the following steps: Step 1: Construct an extremely narrow pulse echo dataset. The inversion process of its HRRP data is as follows: Figure 5 As shown; then the data will be processed according to... Figure 6 The training and testing scenarios shown are divided into source domain training set, target domain training set and test set, and normalization preprocessing is performed. Step 2: Construct the Feature-Level Azimuth Extension Module (FAEM), the overall structure of which is as follows: Figure 1 As shown, this module includes a feature extraction network E and a feature generation network G (see details in [link]). Figure 2 ) and feature discriminator D (see details) Figure 2 ), and construct the relevant adversarial loss and azimuth consistency loss; Step 3: Construct the Scattering Center Reconstruction Constraint Module (SCRCM), its location as shown in the figure. Figure 1 As shown, the scattering center reconstruction network structure included in this module is shown below. Figure 3And construct a physical consistency loss that integrates Dice loss, mean squared error loss and sparse regularization loss; Step 4: Construct the Multi-View Perception Feature Enhancement Module (MFAM), whose location is as follows: Figure 1 As shown, the specific structure of the fusion network based on Spatial and Channel Interaction Attention (SCIA) included in this module is shown in [link to diagram]. Figure 4 And classifiers, and construct related fusion loss and intra-class multi-view compactness constraint loss; Step 5: Input the full-view HRRP data and its corresponding labels from the source domain into the network, and pass them through FAEM, SCRCM and MFAM in sequence. The entire network is pre-trained through the corresponding joint loss function to complete parameter optimization. Step 6, Figure 6 The sparse viewpoint HRRP data and their corresponding labels in the target domain training set are input into the pre-trained network, and the entire network is trained and optimized again to complete the training of the model. Step 7, Figure 6 The HRRP data in the target domain test set shown is input into the trained network. After feature expansion, physical constraints and semantic enhancement, the recognition rate of the model on the target domain test set is finally obtained.

[0008] Beneficial effects: (1) Existing methods for solving sparse view recognition through generative models generally lack physical guidance, resulting in low semantic fidelity of generated features. This invention innovatively designs a physical constraint mechanism for scattering centers, integrating the electromagnetic scattering mechanism of HRRP into the feature generation process, ensuring the physical authenticity and semantic reliability of generated features, and fundamentally improving the generation quality.

[0009] (2) Existing feature enhancement methods mostly focus on feature fusion and fail to explicitly constrain the semantic structure of the feature space. This invention proposes intra-class multi-view compactness constraints, which enhance the robustness of the model to viewpoint changes by explicitly reducing the distribution range of similar samples in the feature space, and further improves the recognition accuracy.

[0010] (3) The present invention places the feature generation, physical constraint and semantic enhancement modules in an end-to-end framework for joint optimization, so that physical prior knowledge and recognition task objectives work together to guide feature generation, ensuring that the generated features maximize the utility of downstream recognition tasks. Attached Figure Description

[0011] Figure 1 The diagram below shows the overall architecture of the Feature-Level Azimuth Enhancement Network (FLAAN). Figure 2 1. Detailed structural diagram of the feature generation network G and the discriminator D; Figure 3A schematic diagram of the scattering center reconstruction network (SCRN); Figure 4 The diagram shows the structure of the Spatial and Channel Interaction Attention (SCIA) module within the Multi-View Perception Feature Enhancement (MFAM) module. Figure 5 A detailed diagram illustrating the export of HRRP data from MSTAR SAR images; Figure 6 Hypothetical diagram of the experimental scenario. Detailed Implementation

[0012] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0013] This invention proposes a feature-level physical enhancement recognition method for sparse viewpoint target recognition. The overall model structure is as follows: Figure 1 As shown, this method mainly consists of three core modules: Feature-level Azimuth Extension Module (FAEM), Scattering Center Reconstruction Constraint Module (SCRCM), and Multi-view Sensing Feature Enhancement Module (MFAM).

[0014] FAEM is responsible for extracting HRRP features and mapping sparse view HRRP data to multi-view features. A one-dimensional VGG network serves as the backbone for feature extraction, mapping the input azimuth angle θ HRRP data to deep features, and then... Figure 2 The feature generation network G shown learns the variation patterns between the center feature and neighboring features, outputting predictions of multi-view features. The detailed network structure and data flow of this module are as follows: Figure 2 As shown.

[0015] SCRCM is a key module of the physical constraints proposed in this invention, through which... Figure 3 The scattering center reconstruction network shown decodes the generated features into the scattering center parameter space and uses physical consistency loss constraints to generate features that conform to the physical scattering structure of the target. This module is based on the physical formation mechanism of HRRP and provides physical guidance for the feature generation process by reconstructing the position and amplitude parameters of the scattering centers. Figure 3 The specific structure of the residual deconvolution blocks and range-dimensional multi-head attention layers of the network is clearly shown.

[0016] MFAM utilizes generated multi-view features, through Figure 4 The Spatial and Channel Interactive Attention (SCIA) mechanism is used for adaptive fusion and intra-class multi-view compactness constraints are applied to improve the discriminativeness and robustness of features. Finally, the recognition result is output through a classifier. Figure 4 The parallel processing flow of spatial attention path and channel attention path and their interactive weighting method are described in detail.

[0017] The technical approach of this invention is as follows: It constructs an end-to-end feature-level physical enhancement recognition framework, systematically solving the sparse-view HRRP recognition problem through the synergistic effect of feature space generation, physical mechanism constraints, and semantic consistency enhancement. Specifically, firstly, the traditional data-level generation is upgraded to a more efficient feature-level generation through a feature-level azimuth extension module (FAEM). A generator with a one-dimensional U-Net structure learns the feature evolution rules between viewpoints, achieving feature extension from sparse to dense viewpoints. Secondly, an innovative scattering center reconstruction constraint module (SCRCM) is proposed, mapping the generated features back to the scattering center parameter space. Physical consistency loss forces the generated features to conform to the target electromagnetic scattering mechanism, ensuring the physical authenticity of its feature semantics. Finally, the spatial and channel interactive attention in the multi-view perceptual feature enhancement module (MFAM) achieves adaptive fusion of multi-view features, and the intra-class multi-view compactness constraint explicitly shapes the feature space structure, improving the model's robustness to viewpoint changes. The entire system adopts a domain-adaptive training strategy, first pre-training on omnidirectional source domain data, then transferring to sparse viewpoint target domain for fine-tuning, and finally achieving a deep fusion of physical priors and data-driven approaches, significantly improving recognition performance under sparse viewpoints.

[0018] Specifically, the steps include the following: Step 1: Construct an extremely narrow pulse echo dataset. The inversion process of its HRRP data is as follows: Figure 5 As shown; then the data will be processed according to... Figure 6 The training and testing scenarios shown are divided into source domain training set, target domain training set and test set, and normalization preprocessing is performed. 1.1 Obtain HRRP data and corresponding class labels for all-view targets of class M, and use them as the training dataset for the source domain; 1.2 Obtain HRRP data and corresponding category labels for N types of targets with different pitch angles and full field of view, and perform data intervals at certain pitch angles. The azimuth angle sampling yields sparse viewpoint HRRP data and corresponding category labels as the training set for the target domain, while the omnidirectional HRRP data and corresponding category labels at other pitch angles are used as the test set for the target domain. 1.3. Perform maximum amplitude normalization on all datasets in the source and target domains.

[0019] Maximum amplitude normalization uses the maximum value of the extremely narrow pulse echo in each frame to normalize the range image. Let... To extract a single frame of extremely narrow pulse echo from the dataset, its amplitude-normalized extremely narrow pulse echo is denoted as... , can be represented as: Step 2: Construct the Feature-Level Azimuth Extension Module (FAEM), the overall structure of which is as follows: Figure 1 As shown, this module includes a feature extraction network E and a feature generation network G (see details in [link]). Figure 2 ) and feature discriminator D (see details) Figure 2 ), and construct the relevant adversarial loss and azimuth consistency loss; 2.1 Construct a one-dimensional convolutional neural network as a feature extractor. The network adopts the 1D-VGG architecture consisting of 12 convolutional blocks as described in the paper. Each convolutional block contains a one-dimensional convolutional layer with a kernel size of 3, a batch normalization layer, and a ReLU activation function.

[0020] 2.2 Construct a feature generation network based on a one-dimensional U-Net architecture, specifically comprising three downsampling layers (DO), one intermediate layer (IPU), and three upsampling layers (UO). The generated feature set... For multiple side angles, For input In azimuth Feature embedding at the location, Features generated at a certain azimuth angle. and These represent the number of generated multi-views and the dimension of the feature vectors, respectively. The generated feature set can be represented as: 2.3 Constructing a discriminator network and loss function: Discriminator Network Composed of three convolutional layers, it distinguishes between ground truth and generated features. Regarding the loss function, this invention proposes adversarial loss and aspect consistency loss. These guide the parameter learning of G from two aspects: the distribution level and consistency between the generated multi-view features and the ground truth. For the generator network G, the optimization objective is to minimize the adversarial loss. For the discriminator network D, the optimization objective is to maximize the adversarial loss. This represents the corresponding ground truth features sampled from the true feature distribution. Let represent the expectation operator, and represent the average loss over the sampled data. The adversarial loss can be expressed as: : The orientation consistency loss can be expressed as: in This represents the calculation of the L1 norm.

[0021] Therefore, the loss function It can be represented as: Step 3, construct the Scattering Center Reconstruction Constraint Module (SCRCM), its location is as follows: Figure 1 As shown, the scattering center reconstruction network structure included in this module is shown below. Figure 3 And construct a physical consistency loss that integrates Dice loss, mean squared error loss and sparse regularization loss; 3.1 Construct a scattering center reconstruction network, such as Figure 3 The network shown first upsamples layer by layer through six residual deconvolutional blocks (RDBs) to ensure dimensionality alignment between the output feature map and the target parameter space. These blocks merge residual connections within the deconvolutional layers to ensure smooth propagation of feature information throughout the network. Let represent the input feature map of the l-th RDB, ReLU(·) be the ReLU activation function, Upsample represent the upsampling operation, and Conv and DeConv represent the standard convolution and transposed convolution operations, respectively. The output of the l-th RDB layer can be described as: A range-dimensional multi-head attention (RW-MHA) layer is applied after the RDB layer to refine and enhance the reconstructed feature map. Here, Q, K, and V are all set as feature maps of the output of the last RDB layer. , Let represent the projection matrix of the i-th head. Finally, each attention head output from the linear layer after merging can be represented as: Attention operations are defined as follows: Finally, through the output projection matrix The reconstructed scattering center can be represented as: 3.2 Construct the physical consistency loss function, which is as follows: The weighted sum of loss, mean squared error loss, and sparse regularization loss. Indicates the true scattering center of the ground. This represents the reconstructed scattering center. The one mentioned here... The loss can be expressed as: The coefficient ranges from 0 to 1, where 1 indicates a perfect match between the prediction and the ground truth, and 0 indicates no overlap. Therefore, loss It can be defined as: L is the sequence length. The i-th element of the true value. Let be the i-th element of the predicted value. The proposed mean squared error loss... It can be represented as: Furthermore, considering the sparse distribution of scattering centers along the distance dimension, we propose a sparsity regularization loss function. To calculate the absolute difference between the predicted value and the true value, it can be expressed as: Therefore, the physical consistency loss of the scattering center Defined as the sum of the three losses mentioned above. This combination can efficiently reconstruct the location parameters of the scattering center, where... It is a hyperparameter that controls the strength of sparsity regularization by adjusting its relative importance in the overall loss, and guides the generation of features with clearer rational meaning and improved semantic structure consistency. It can be represented as: Step 4, construct the Multi-View Perception Feature Enhancement Module (MFAM), its location is as follows: Figure 1 As shown, the specific structure of the fusion network based on Spatial and Channel Interaction Attention (SCIA) included in this module is shown in [link to diagram]. Figure 4 And classifiers, and construct related fusion loss and intra-class multi-view compactness constraint loss; 4.1 Construct a fusion network based on spatial and channel interaction attention. This network dynamically performs adaptive weighted fusion of multi-view features and outputs fused features: For spatial attention ( In depthwise convolution, information is processed independently for each viewpoint. For channel attention (…),… This approach employs a self-attention mechanism to calculate the importance of each viewpoint, thereby achieving adaptive fusion based on the relevance of each view channel. Finally, the spatial attention map and channel attention map interact to dynamically assign weights to the input feature map, resulting in a refined fused feature map. , where σ(·) represents the sigmoid function, which can be expressed as:

[0022] 4.2 Construct a classifier consisting of fully connected layers, which performs label pre-processing based on fused features. The test uses standard cross-entropy loss for training. It also utilizes the generated features. and the characteristics of integration ,in Indicates the number of views generated. These represent the outputs of classifier C corresponding to the i-th generated view. , Indicates fusion features The output of C. y represents the true label, and CE represents the cross-entropy loss used for classification supervision. Its loss function can be expressed as: 4.3 Construct an intra-class multi-view compactness constraint loss function that minimizes the average distance from features of samples of the same class to their class centers: For a batch of features generated by generator G, these features are first grouped into subsets according to their class labels. For each k classes, the corresponding feature subset is defined as:

[0023] in It contains all feature vectors belonging to class k in the batch processing.

[0024] Then, for each k class, This represents the number of feature vectors belonging to class k. The corresponding class center vectors are calculated as follows:

[0025] Finally, intra-class tightness constraint loss Designed to minimize the distance between each feature and its corresponding class center within each batch, it can be expressed as follows:

[0026] Step 5: Input the full-view HRRP data and its corresponding labels from the source domain into the network, and pass them through FAEM, SCRCM and MFAM in sequence. The entire network is pre-trained through the corresponding joint loss function to complete parameter optimization. 5.1 Input the full-view HRRP data and labels from the source domain into FAEM to obtain multi-view features; 5.2 Input the multi-view features into SCRCM to calculate the physical consistency loss; 5.3 Input the multi-view features into MFAM to obtain the predicted labels; 5.4 Update the network parameters of FAEM, SCRCM, and MFAM using the joint loss function; 5.5 Repeat the above steps to perform pre-training until the network converges.

[0027] Step 6, Figure 6 The sparse viewpoint HRRP data and their corresponding labels in the target domain training set are input into the pre-trained network, and the entire network is trained and optimized again to complete the training of the model. 6.1 Input the sparse viewpoint HRRP data and labels of the target domain training set into the pre-trained network; 6.2 Optimize and update network parameters using classification loss and intra-class compactness constraint loss; 6.3 Repeat the above steps until the network converges.

[0028] Step 7, Figure 6 The HRRP data in the target domain test set shown is input into the trained network. After feature expansion, physical constraints and semantic enhancement, the recognition rate of the model on the target domain test set is finally obtained.

[0029] Example This embodiment illustrates the effectiveness of the present invention through experiments using measured data. To evaluate the performance of the proposed feature-level physical enhancement recognition method, HRRP data obtained by SAR image inversion from the publicly available radar dataset MSTAR was experimentally verified.

[0030] Dataset and parameter settings: The MSTAR dataset contains synthetic aperture radar (SAR) images of ten categories of military targets. This dataset was acquired using a high-resolution spotlight SAR radar with a resolution of 0.3 × 0.3 m, operating in the X-band and employing HH polarization. By inverting the SAR images from the MSTAR dataset, the corresponding HRRP data can be obtained. The specific inversion process is as follows: Figure 5 As shown.

[0031] This invention employs a specific data partitioning strategy. Seven target classes (BTR-60, 2S1, BRDM-2, D7, T-62, ZIL-131, and ZSU-23 / 4) from the MSTAR dataset are used as the source domain for pre-training network parameters. The remaining three target classes (BMP2, BTR-70, and T-72) are used as non-cooperative targets in the target domain. For the target domain data, data sampled at different azimuth intervals (10°, 20°, and 30°) at a 17° pitch angle are selected to form a sparse viewpoint training set, while omnidirectional data at a 15° pitch angle are used as the test set. The specific data configuration is shown in Table 1.

[0032] Table 1 Dataset Details

[0033] Regarding the experimental parameter settings, this invention uses the Adam optimizer for training, with an initial learning rate set to 2×10⁻⁶.-4 The learning rate is decayed to 70% of its original value every 20 training epochs. The batch size for source domain training is set to 128, and the batch size for target domain fine-tuning is set to 16. The sparse regularization coefficient in scattering center reconstruction... Set to 100, intra-class compactness constraint coefficient Set it to 0.5.

[0034] Experimental results: Table 2 shows the recognition results of the method of this invention on the MSTAR dataset. Through comparative analysis with baseline methods, few-shot learning methods, and deep generative model methods, it can be clearly seen that the method of this invention achieves the best recognition performance under three different sparse sampling conditions.

[0035] Table 2 Comparison of experimental results from different methods

[0036] Especially in the case of sparse sampling with a 20° interval, the method of this invention improves the performance by approximately 12 percentage points compared to the best comparison method (MAML), which fully demonstrates the effectiveness of feature-level enhancement combined with physical constraints. Even in the most extreme case of a 30° interval, the method of this invention still maintains a recognition accuracy of 67.65%, exhibiting strong generalization ability.

[0037] To further explore the contributions of each module, a systematic ablation experiment was conducted. As shown in Table 3, when only the Feature-Level Azimuth Extension Module (FAEM) was used, the recognition accuracy at 20° intervals was 71.90%; after adding the Scattering Center Reconstruction Constraint Module (SCRCM), the performance improved to 73.74%; and when all modules were used, the recognition accuracy reached 76.34%. This result verifies the effectiveness and synergistic effect of each module.

[0038] Table 3 Ablation Experiment Results

[0039] Feature visualization analysis reveals that the enhanced features generated by this invention highly overlap with the distribution of real features in the feature space, and the physically constrained features exhibit better intra-class clustering and inter-class separation. This further confirms the advantages of this invention in feature generation quality and feature space optimization from a visualization perspective.

[0040] In summary, this embodiment verifies the effectiveness and advancement of the present invention in the sparse viewpoint HRRP recognition task through detailed experimental data and systematic analysis. Compared with existing methods, the present invention demonstrates significant advantages in recognition accuracy, generalization ability, and feature quality.

[0041] In summary, the above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A radar target identification method and apparatus with physical scattering constraint-based location enhancement, characterized in that, include: Step 1: Construct an extremely narrow pulse echo dataset. Divide the data into source domain training set, target domain training set, and test set according to training and testing scenarios, and perform normalization preprocessing. Step 2: Construct the Feature-Level Azimuth Extension Module (FAEM), which includes a feature extraction network E, a feature generation network G, and a feature discriminator D, and constructs the relevant adversarial loss and azimuth consistency loss. Step 3: Construct the Scattering Center Reconstruction Constraint Module (SCRCM) and construct a physical consistency loss that integrates Dice loss, mean square error loss, and sparse regularization loss; Step 4: Construct the Multi-View Perceptual Feature Enhancement Module (MFAM) and the classifier, and construct the relevant fusion loss and intra-class multi-view compactness constraint loss; Step 5: Input the full-view HRRP data and its corresponding labels from the source domain into the network, and pass them through FAEM, SCRCM and MFAM in sequence. The entire network is pre-trained through the corresponding joint loss function to complete parameter optimization. Step 6: Input the sparse viewpoint HRRP data and their corresponding labels from the target domain training set into the pre-trained network, and perform training and optimization of the entire network again to complete the training of the model. Step 7: Input the HRRP data from the target domain test set into the trained network. After feature expansion, physical constraints and semantic enhancement, the recognition rate of the model on the target domain test set is finally obtained.

2. The radar target identification method and apparatus with physical scattering constraint-based location enhancement as described in claim 1, characterized in that, Step 1 includes: 1.1 Obtain HRRP data and corresponding category labels for all targets of class M, as the source domain. Training dataset; 1.2 Obtain HRRP data and corresponding category labels for N types of targets at different pitch angles across the entire field of view. Data interval at a certain pitch angle The azimuth angle sampling yields sparse viewpoint HRRP data and corresponding category labels as the training set for the target domain, while the omnidirectional HRRP data and corresponding category labels at other pitch angles are used as the test set for the target domain. 1.

3. Perform maximum amplitude normalization on all datasets in the source and target domains.

3. The radar target identification method and apparatus with physical scattering constraint-based location enhancement as described in claim 1, characterized in that, Step 2 includes: 2.1 Construct a one-dimensional convolutional neural network as a feature extractor. The network adopts the 1D-VGG architecture consisting of 12 convolutional blocks as described in the paper. Each convolutional block contains a one-dimensional convolutional layer with a kernel size of 3, a batch normalization layer, and a ReLU activation function. 2.2 Construct a feature generation network, which is based on a one-dimensional U-Net architecture, specifically consisting of three downsampling layers (DO), one intermediate layer (IPU), and three upsampling layers (UO); 2.3 Constructing a discriminator network and loss function: Discriminator Network It consists of three convolutional layers to distinguish between real features and generated features; in terms of loss function, this invention proposes adversarial loss and aspect consistency loss; the parameters of G are guided from the aspects of distribution level and consistency between generated multi-view features and ground truth.

4. The radar target identification method and apparatus with physical scattering constraint-based location enhancement as described in claim 1, characterized in that, Step 3 includes: 3.1 Constructing a scattering center reconstruction network, as shown in Figure 3. This network first upsamples layer by layer through six residual deconvolutional blocks (RDB) to ensure dimensional alignment between the output feature map and the target parameter space; 3.2 Construct the physical consistency loss function, which is as follows: Loss, mean square error loss The weighted sum of sparse regularization loss.

5. The radar target identification method and apparatus with physical scattering constraint-based location enhancement as described in claim 1, characterized in that, Step 4 includes: 4.1 Construct a fusion network based on spatial and channel interaction attention, which dynamically handles multiple perspectives. The features are adaptively weighted and fused to output the fused features; 4.2 Construct a classifier consisting of fully connected layers, which performs label pre-processing based on fused features. For testing, standard cross-entropy loss is used for training; 4.3 Construct an intra-class multi-view compactness constraint loss function, which minimizes the characteristics of samples of the same class. The average distance from the center of its class.

6. The radar target identification method and apparatus with physical scattering constraint-based location enhancement as described in claim 1, characterized in that, Step 5 includes: 5.1 Input the full-view HRRP data and labels from the source domain into FAEM to obtain multi-view features; 5.2 Input the multi-view features into SCRCM to calculate the physical consistency loss; 5.3 Input the multi-view features into MFAM to obtain the predicted labels; 5.4 Update the network parameters of FAEM, SCRCM, and MFAM using the joint loss function; 5.5 Repeat the above steps to perform pre-training until the network converges.

7. The radar target identification method and apparatus with physical scattering constraint-based location enhancement as described in claim 1, characterized in that, Step 6 includes: 6.1 Input the sparse viewpoint HRRP data and labels of the target domain training set into the pre-trained network; 6.2 Optimize and update network parameters using classification loss and intra-class compactness constraint loss; 6.3 Repeat the above steps until the network converges.

8. The radar target identification method and apparatus with physical scattering constraint-based location enhancement as described in claim 3, characterized in that, In step 2.3, the orientation consistency loss is expressed as: ； loss function Represented as: ； in This represents finding the L1 norm; This represents the corresponding ground truth features sampled from the true feature distribution; Let represent the expectation operator, and let represent the average loss over the sampled data.

9. The radar target identification method and apparatus with physical scattering constraint-based location enhancement as described in claim 4, characterized in that, In step 3.2, Indicates the true scattering center of the ground. Indicates the reconstructed scattering center; in which the mentioned The loss is represented as: ； The coefficient ranges from 0 to 1, where 1 indicates a perfect match between the prediction and the ground truth, and 0 indicates no overlap. loss Defined as: ； L is the sequence length. The i-th element of the true value. The i-th element of the predicted value; mean squared error loss Represented as: ； Furthermore, considering the sparse distribution of scattering centers along the distance dimension, a sparsity regularization loss function is proposed. The absolute difference between the predicted value and the true value is calculated and expressed as: ； Therefore, the physical consistency loss of the scattering center Defined as the sum of the three losses mentioned above; this combination can efficiently reconstruct the scattering center location parameters, where It is a hyperparameter that controls the strength of sparsity regularization by adjusting its relative importance in the overall loss, and guides the generation of features with clearer rational meaning and improved semantic structure consistency. Represented as: 。 10. The radar target identification method and apparatus with physical scattering constraint-based location enhancement as described in claim 5, characterized in that, In step 4.3, the intra-class tightness constraint loss Represented as: 。

Images