HRRP target identification method and system based on angle guidance
By introducing an angle-conditional kernel hybrid convolution module and a cross-attention mechanism into HRRP target recognition, dynamic convolution kernel parameters are generated, which solves the performance degradation problem of HRRP target recognition when the azimuth angle changes, and improves recognition accuracy and robustness.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- XIAN UNIV OF TECH
- Filing Date
- 2026-01-22
- Publication Date
- 2026-05-12
AI Technical Summary
Existing high-resolution range image HRRP target recognition methods exhibit significantly reduced recognition performance and insufficient robustness when the azimuth angle of the vehicle target observation changes, making it difficult to effectively utilize azimuth information.
A target recognition network model is constructed. The azimuth information is introduced into the convolution kernel generation process through the angle conditional kernel hybrid convolution module to generate dynamic convolution kernel parameters. Combined with the dual-branch feature extraction module and the cross attention mechanism, global and local features are integrated, and the model is optimized using the center loss function.
It improves the accuracy and generalization ability of HRRP target recognition, alleviates the impact of orientation sensitivity on recognition performance, and reduces the dependence on traditional orientation-invariant feature extraction.
Smart Images
Figure CN122017779A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of radar target detection technology, specifically to an angle-guided HRRP target recognition method and an angle-guided HRRP target recognition system. Background Technology
[0002] Due to the long range and all-weather operation of radar, research on vehicle target recognition driven by radar measurement data is of great significance in traffic monitoring. High-resolution range profiles (HRRPs) are one-dimensional echo characteristics of radar targets obtained under high range resolution conditions. Essentially, they reflect the range and energy distribution characteristics of each scattering center of the target along the radar's line-of-sight. Because of its advantages such as low data dimensionality, fast acquisition speed, and strong adaptability to complex environments, HRRPs have been widely used in the field of automatic radar target recognition.
[0003] Currently, radar-based vehicle target recognition methods have made some progress, but some key shortcomings remain. First, the echo characteristics of vehicle targets vary significantly under different observation azimuths, speeds, and attitudes, leading to decreased recognition performance. Second, radar echo data is noisy and has high annotation costs; existing methods rely heavily on large-scale labeled data, limiting their generalization ability. Meanwhile, as a typical vehicle radar echo signal, high-resolution range profiles (HRRPs) are widely used in vehicle target recognition research due to their clear structure and high information density. Currently, HRRP target recognition methods can be broadly classified into two categories: traditional HRRP target recognition methods and deep learning-based HRRP target recognition methods. Traditional HRRP target recognition methods mainly include feature engineering-based pattern classification methods and probabilistic graphical model-based statistical inference methods. Feature engineering-based methods typically extract statistical features or transform domain features from HRRP data and train a classifier based on these features to achieve target recognition. Probability density estimation-based methods, on the other hand, typically assume that range cells are independent, construct a joint probability model for each HRRP range cell, and determine the target category by combining the probability distributions of each range cell. With the development of deep learning technology, HRRP target recognition methods based on deep neural networks have gradually become a research hotspot. These methods construct multi-layer nonlinear mapping networks to achieve end-to-end joint optimization of HRRP feature representation learning and classification decisions. Compared to traditional methods, they can automatically learn discriminative feature representations through backpropagation, thus significantly reducing the reliance on manual prior feature design. Although these methods improve HRRP target recognition performance to some extent, due to the inherent azimuth sensitivity of HRRP data, the HRRP values obtained for the same target under different observation azimuths or poses often show significant differences, posing a serious challenge to stable recognition under different azimuth conditions. Furthermore, azimuth angle, as an important auxiliary prior information, can theoretically effectively improve the accuracy of target recognition, but in practical applications, it is often difficult to obtain training samples covering a complete azimuth angle range, limiting the model's generalization ability under unseen azimuth conditions. Therefore, how to fully utilize azimuth angle information under incomplete sample azimuth conditions and improve the model's adaptability to azimuth changes has become a key problem that urgently needs to be solved in the field of HRRP target recognition. Summary of the Invention
[0004] The purpose of this invention is to provide an angle-guided HRRP target recognition method, which solves the problem that existing high-resolution range image HRRP target recognition methods suffer from significant performance degradation and insufficient robustness when the azimuth angle of vehicle target observation changes in multi-type target radar recognition.
[0005] Another objective of this invention is to provide an angle-guided HRRP target recognition system.
[0006] The first technical solution adopted in this invention is an angle-guided HRRP target recognition method. This method addresses the issue that the scattering characteristics of vehicle targets change significantly under different observation azimuths, leading to substantial differences between the range cell echo amplitude distribution and the projection structure of the scattering center in HRRP. Existing recognition methods typically rely on azimuth-sensitive feature representations, resulting in significant differences in the HRRP features of the same vehicle under different observation azimuths. This makes it difficult for the model to learn azimuth-invariant discriminative features, easily leading to misclassification. To address these problems, this invention employs the following technical means:
[0007] Step 1: Construct a target recognition network model, which includes a preprocessing module, an angle-conditional kernel hybrid convolution module, a dual-branch feature extraction module, a feature fusion module, and a classification and recognition module; Step 2: Acquire radar echoes of the target vehicle in a stationary state and generate HRRP data to construct a high-resolution range profile HRRP dataset. =( ， ), i =1,2,… P, P Indicates the number of samples. Indicates azimuth angle ， Indicates HRRP distance; Step 3: Normalize and align the HRRP distance, and perform two-dimensional encoding on the azimuth. Step 4: Input the azimuth angle after the encoding process in Step 3 into the angle conditional kernel hybrid convolution module, map the azimuth angle into a Gaussian distribution with variable mean and variance, and then use the reparameterization method to generate dynamic convolution kernel parameters; Step 5: Use the dynamic convolution kernel parameters generated in Step 4 to extract local features in the dual-branch feature extraction module; at the same time, the global branch in the dual-branch feature extraction module extracts global features. Step 6: Utilize the cross-attention mechanism of the feature fusion module to fuse global and local features; use global features as queries and local features as keys and values to guide the network to selectively focus on the features most critical to the discrimination, thereby obtaining fused features; Step 7: Input the fused features into the classification module to obtain the classification result; Step 8: Set the classification loss based on the classification results, and use the fused features to set the center loss to train the model; use the trained target recognition model to identify the HRRP of the vehicle to be identified.
[0008] The invention is further characterized by: Step 3 is as follows: Step 3.1: Normalize and centroid-align each HRRP sample, resulting in the following normalized sample: ;in, This represents the normalized sample. This indicates the calculation of the 2-norm; Each HRRP distance Includes n One distance unit, , ; Indicates the first i In the nth sample p One distance unit; Center of gravity alignment processing
[0009] in, C This indicates the centroid of the HRRP sample. n This indicates the number of distance cells. Then, the center of each HRRP sample is shifted to the centroid, thus completing the centroid alignment operation. Step 3.2, for the first i Azimuth angle of each sample Perform radian conversion, calculate its sine and cosine values, and generate azimuth encoding vectors for all samples. ; in, It is the azimuth angle. This is the processed azimuth angle.
[0010] Step 4, generating dynamic convolution kernel parameters, is as follows: Step 4.1: Convert the preprocessed azimuth angle As a conditional input, the mean vector of the output dynamic convolution kernel parameters is generated through a mean generation branch composed of a three-level fully connected network. ;in, and These represent the weights and biases of the fully connected layer, respectively. The mean generation branch consists of a three-layer fully connected network and two ReLU activations; the mean vector... The output dimension is preset to ; in, This indicates the number of parallel dynamic convolution kernels. =8, This represents the number of channels in the input feature map. This indicates the number of channels in the output feature map. This represents the size of the one-dimensional convolution kernel; And according to the number of parallel dynamic convolution kernels Will Reconstructed In which, the mean parameter of each group Corresponding to a set of candidate dynamic convolution kernel parameter values; Among them, each set of mean parameters Corresponding to a set of candidate dynamic convolution kernel parameter values Step 4.2, the preprocessed azimuth angle As conditional parallel inputs to the variance generation branch, the variance generation branch network packet...
[0011] in, This refers to the Softplus activation function. This represents the Sigmoid function. To predetermine the variance constraint parameters, the variance generation branch consists of a three-layer fully connected network, two Softplus activations, and a final Sigmoid activation. The variance vector Also reconstructed as ; Variance parameters for each group Used to characterize the discreteness of the corresponding dynamic convolution kernel parameters; Step 4.3: Using the mean parameter and variance parameter as the distribution parameters of the dynamic convolution kernel random variable, for the first... Group dynamic convolution kernels are used for resampling and gradient updates:
[0012] in, Indicates the first The dynamic convolution kernel parameters of the group, Represents a standard normal distribution; Step 4.4, preprocess the input samples Perform global average pooling and input it into the attention network to obtain... Each weight,
[0013] The attention network includes a global average pooling layer, a flattening layer, a fully connected layer, a ReLU activation layer, a fully connected layer, and a sigmoid activation layer. Indicates the first The group of dynamic convolutional kernels' response weights to the current input. , , , These are the weights and parameters of the fully connected layer; Step 4.5, will The group of dynamic convolutional kernels are fused together according to attention weights to obtain the final convolutional kernel. , Step 4.6: Perform dynamic convolution calculation.
[0014] in, Let y represent the input feature, y represent the output feature, and C represent the dynamic convolution operation.
[0015] In step 5, the dual-branch feature extraction module includes a CNN local feature extraction branch composed of dynamic convolutions and a global feature extraction branch composed of a Transformer encoder; The local feature extraction branch consists of three convolutional blocks connected sequentially and by residuals. The first and third are static convolutional blocks, while the second is a dynamic convolutional kernel generated by a conditional kernel hybrid convolution module. The dynamic convolutional block introduces an azimuth parameter into the convolution operation, giving the convolution weights azimuth-aware characteristics; the output of this module is local features. ; The global feature extraction module consists of a Transformer encoder, comprising a multi-head self-attention layer, two fully connected feedforward layers (FFN), and two LayerNorm layers. It sequentially processes the preprocessed samples through position encoding, the multi-head self-attention mechanism, layer normalization, and feedforward network connection layers to obtain a global feature representation specific to the current sample. .
[0016] Step 6 uses cross-attention fusion as the feature fusion method, which consists of the following steps: Step 6.1, the feature fusion module is constructed in two parts: a multi-head cross-attention mechanism and a feedforward network module (FFN). The cross-attention can be represented as:
[0017] Where Q, K, and V represent the query matrix, key matrix, and value matrix, respectively, which are obtained by performing three different linear transformations on the input, with dimensions of . The cross-attention mechanism incorporates the concept of cross-interaction in the expression of Q, K, and V, allowing Q, which calculates attention, to be derived from the extracted global features, while K and V are derived from local features.
[0018] in, Let Q, K, and V be the learnable weight matrix. , Features extracted from different modules; cross-attention calculation results are:
[0019] in, Its function is to perform a linear transformation on the input, transforming the input vector into a high-dimensional space. It contains 4 fully connected layers, 1 layer normalization layer, 1 Dropout layer, and 1 Softmax operation. Step 6.2: By stacking multiple cross-attention mechanisms to form MHCA, firstly, Q, K, and V are partitioned into multiple heads along the second dimension, resulting in... h Size , Then, perform a cross-attention calculation for each head:
[0020] The output matrices of each head are concatenated along the last dimension:
[0021] in, The concatenation operation consists of three linear transformation layers and a Softmax attention layer; finally, the output of the feature fusion module is... ;; Step 6.3: Introduce a center loss into the fused feature space of the cross-attention output. ; The center loss is calculated by measuring the squared Euclidean distance between the feature vector and its corresponding class center, as shown in the following expression:
[0022] in, For the first i The fusion features of each sample For its corresponding category label, For the tag Category center, This represents the L2 norm.
[0023] The classification module described in step 7 includes a fully connected layer and a softmax activation layer; The cross-entropy classification loss used is:
[0024] Where N is the number of categories, Real label one-hot encoding, Predict class probabilities for the model; The overall loss function is:
[0025] in, This represents the trade-off parameter, which is set to 0.2.
[0026] The second technical solution adopted in this invention is an HRRP target recognition system based on angle-supervised feature generation, including a target recognition network model: the target recognition network model includes: The preprocessing module is used to preprocess the samples in the dataset and perform two-dimensional processing on the azimuth angles to obtain preprocessed samples and two-dimensional angles. Angle-conditional kernel hybrid convolution module is used to map the azimuth angle to a Gaussian distribution with variable mean and variance through the encoding network, and then use a reparameterization method to generate a dynamic convolution kernel; this convolution kernel is used for feature extraction in convolution; The dual-branch feature extraction module is used to extract local features from CNN local features composed of dynamic convolution kernels from the angle conditional kernel hybrid convolution module and to extract global features from global features composed of Transformer encoders. The classification module is used to obtain classification results; The loss function module is used to set the classification loss based on the classification results and to set the center loss using the fused features. The target recognition module is used to train the target recognition model by setting a total loss function based on the classification loss and center loss to obtain a trained target recognition model; and to use the trained target recognition model to identify the HRRP of the vehicle to be identified.
[0027] The third technical solution adopted in this invention is a computer device, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of the above-mentioned identification method.
[0028] The beneficial effects of this invention are as follows: 1. This invention introduces azimuth information into the convolution kernel generation process through an angle-conditional kernel hybrid convolution module, enabling the convolution kernel to adaptively adjust according to the azimuth angle, effectively characterizing the variation law of HRRP with azimuth angle, without relying on traditional azimuth-invariant feature extraction or angular domain division, thus alleviating the impact of azimuth sensitivity on recognition performance.
[0029] 2. This invention adopts an end-to-end training method, which eliminates the need for manually designing complex feature extraction criteria, reducing human intervention. At the same time, through joint loss function optimization, it further improves recognition accuracy and generalization ability. Attached Figure Description
[0030] Figure 1 This invention relates to an angle-guided HRRP target recognition method and system network result model.
[0031] Figure 2 This is a comparison chart of the recognition accuracy of the method of the present invention with other traditional recognition methods. Detailed Implementation
[0032] 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.
[0033] Example 1 Figure 1 This is a flowchart of the angle-guided HRRP recognition method of the present invention; it includes constructing a target recognition network model, acquiring HRRP data of the target vehicle in a stationary state, dynamically convolving the HRRP data by introducing azimuth information to extract angle-adaptive local scattering features; further utilizing a global attention module to obtain global structural features; finally, cross-fusing the local dynamic convolution features and global attention features, and outputting the target category through a classification module; specifically including the following steps: Step 1: Construct a target recognition network model, which includes a preprocessing module, an angle-conditional kernel hybrid convolution module, a dual-branch feature extraction module, a feature fusion module, and a classification and recognition module; Step 2: Obtain HRRP data of the target vehicle in a stationary state and construct a high-resolution distance-image HRRP dataset. =( ， ), i =1,2,… P, P Indicates the number of samples. Indicates azimuth angle ， Indicates HRRP distance; Step 3: Normalize and align the HRRP distance, and perform two-dimensional encoding on the azimuth. Step 4: Input the azimuth angle after the encoding process in Step 3 into the angle conditional kernel hybrid convolution module, map the azimuth angle into a Gaussian distribution with variable mean and variance, and then use the reparameterization method to generate dynamic convolution kernel parameters; Step 5: Use the dynamic convolution kernel parameters generated in Step 4 to extract local features in the dual-branch feature extraction module; at the same time, the global branch in the dual-branch feature extraction module extracts global features. Step 6: Utilize the cross-attention mechanism of the feature fusion module to fuse global and local features; use global features as queries and local features as keys and values to guide the network to selectively focus on the features most critical to the discrimination, thereby obtaining fused features; Step 7: Input the fused features into the classification module to obtain the classification result; Step 8: Set the classification loss based on the classification results, and use the fused features to set the center loss to train the model; use the trained target recognition model to identify the HRRP of the vehicle to be identified.
[0034] Example 2 The angle-guided HRRP recognition method comprises two stages: a training stage and a recognition stage. The training stage includes six steps: constructing the training dataset, preprocessing training samples and angles, generating dynamic convolutional kernels for training samples, extracting global and local features from training samples, fusing features from training samples, and training the classifier. The recognition stage includes six steps: constructing the test dataset, preprocessing recognition samples and angles, generating dynamic convolutional kernels for recognition samples, extracting global and local features from recognition samples, fusing features from recognition samples, and outputting the recognition results from the classifier.
[0035] Example 3 Based on Example 2, the training phase of this invention, specifically includes the following steps: The angle-guided HRRP target recognition method of this invention: Step 1: Obtain HRRP data of the target vehicle in a stationary state and construct a high-resolution distance profile HRRP dataset. =( ， ), i =1,2,… P, P Indicates the number of samples. Indicates azimuth angle ， This represents the HRRP distance, and each HRRP distance... Includes n One distance unit, , ; Indicates the first i In the nth sample p 10000 distance units; the data used comes from a publicly available dataset published by AFRL Labs.
[0036] Step 2: Normalize and align the HRRP distance, and perform two-dimensional encoding on the azimuth. Step 2.1: Normalize and centroid align each HRRP sample, resulting in the following normalized sample: ;in, This represents the normalized sample. This indicates the calculation of the 2-norm; Each HRRP distance Includes n One distance unit, , ; Indicates the first i In the nth sample p One distance unit; Center of gravity alignment processing
[0037] in, C This indicates the centroid of the HRRP sample. n This indicates the number of distance cells. Then, the center of each HRRP sample is shifted to the centroid, thus completing the centroid alignment operation. Step 3: Input the encoded azimuth angle into the angle conditional kernel hybrid convolution module, map the azimuth angle into a Gaussian distribution with variable mean and variance, and then use the reparameterization method to generate dynamic convolution kernel parameters; Step 3.1, the preprocessed azimuth angle As a conditional input, the mean generation branch is composed of a three-level fully connected network. , in and These represent the weights and biases of the fully connected layer, respectively. This mean generation branch consists of a three-layer fully connected network and two ReLU activations, and finally outputs the mean vector of the dynamic convolution kernel parameters.
[0038] The mean vector The output dimension is preset to .
[0039] in, This indicates the number of parallel dynamic convolution kernels. =8, This represents the number of channels in the input feature map. This indicates the number of channels in the output feature map. This represents the size of the one-dimensional convolution kernel.
[0040] And according to the number of parallel dynamic convolution kernels Reconstructed into
[0041] Among them, each set of mean parameters This corresponds to the values of a set of candidate dynamic convolution kernel parameters.
[0042] Step 3.2, the preprocessed azimuth angle As conditional parallel inputs to the variance generation branch, the network sequentially includes
[0043] in, This refers to the Softplus activation function. This represents the Sigmoid function. The preset variance constraint parameters are used. This branch consists of a three-layer fully connected network, two Softplus activations, and a final Sigmoid activation.
[0044] The variance vector Also reconstructed as .
[0045] Among them, each set of variance parameters Used to characterize the discreteness of the corresponding dynamic convolution kernel parameters.
[0046] Step 3.3: Using the mean parameter and variance parameter as the distribution parameters of the dynamic convolution kernel random variable, for the first... Group dynamic convolution kernels are used for resampling and gradient updates:
[0047] in, Indicates the first The dynamic convolution kernel parameters of the group, This represents a standard normal distribution.
[0048] Step 3.4, preprocess the input samples Perform global average pooling and input it into the attention network to obtain... Each weight,
[0049] The attention network includes global average pooling, flattening, fully connected layers, ReLU activation, fully connected layers, and Sigmoid activation. Indicates the first The group of dynamic convolutional kernels' response weights to the current input. , , , These are the weights and parameters of the fully connected layer.
[0050] Step 3.5, will The group of dynamic convolutional kernels are fused together according to attention weights to obtain the final convolutional kernel. , Step 3.6: Perform dynamic convolution calculation.
[0051] in, Let y represent the input feature, y represent the output feature, and C represent the dynamic convolution operation.
[0052] Step 4: Use the dynamic convolution kernel parameters generated in Step 3 to extract local features in the dual-branch feature extraction module; at the same time, the global branch in the dual-branch feature extraction module extracts global features. Step 4.1, the local feature extraction branch consists of three convolutional blocks in sequence and residual connections, where the first and third are static convolutional blocks, and the second is a dynamic convolutional block. The preprocessed samples... First, it goes through the first static convolutional block, which includes a standard convolutional layer, a normalization layer, and a ReLU activation function. The output features of the first convolutional block are then passed through a second dynamic convolutional layer, followed by dynamic convolution calculation, a normalization layer, and the ReLU activation function. The output features of the second dynamic convolutional layer are input into the third static convolutional block, and dynamic convolution calculations, normalization layers, and ReLU activation functions are performed sequentially. The residuals are then concatenated with the input to obtain local features. .
[0053] Step 4.2, the global feature extraction module consists of a Transformer encoder, which comprises a multi-head self-attention layer, two fully connected feedforward layers (FFN), and two LayerNorm layers. The preprocessed samples are sequentially processed through position encoding, the multi-head self-attention mechanism, layer normalization, and the feedforward network connection layers to obtain a global feature representation specific to the current sample. .
[0054] Step 5: Use the cross-attention mechanism of the feature fusion module to fuse global and local features; use global features as queries and local features as keys and values to guide the network to selectively focus on the features most critical to the discrimination, thus obtaining fused features; Step 5.1, the feature fusion module is constructed in two parts: a multi-head cross-attention mechanism and a feedforward network module (FFN). The cross-attention can be represented as:
[0055] Where Q, K, and V represent the query matrix, key matrix, and value matrix, respectively, which are obtained by performing three different linear transformations on the input, with dimensions of . The cross-attention mechanism incorporates the concept of cross-interaction in the expression of Q, K, and V, allowing Q, which calculates attention, to be derived from the extracted global features, while K and V are derived from local features.
[0056] in, Let Q, K, and V be the learnable weight matrix. , Features extracted from different modules; cross-attention calculation results are:
[0057] in, Its function is to perform a linear transformation on the input, transforming the input vector into a high-dimensional space. It contains 4 fully connected layers, 1 layer normalization layer, 1 Dropout layer, and 1 Softmax operation.
[0058] Step 5.2: Form MHCA by stacking multiple cross-attention mechanisms. First, perform multi-head partitioning of Q, K, and V along the second dimension to obtain... h Size , Then, perform a cross-attention calculation for each head:
[0059] The output matrices of each head are concatenated along the last dimension:
[0060] in, This represents the splicing operation, which consists of three linear transformation layers and a Softmax attention layer.
[0061] Finally, the output of the feature fusion module is .
[0062] Step 5.3: Introduce a center loss into the fused feature space of the cross-attention output. ; The center loss is calculated by measuring the squared Euclidean distance between the feature vector and its corresponding class center, as shown in the following expression:
[0063] in, For the first i The fusion features of each sample, y i For its corresponding category label, c yi For the label y i Category center, Describing the L2 norm, Step 6: The recognition results obtained by fusing the data from the classification module are then used to update the model parameters using cross-entropy classification loss.
[0064] The classifier module in this embodiment mainly includes a fully connected layer and a softmax activation layer.
[0065] At this point, the training process is complete, and the optimal model parameters have been obtained.
[0066] Example 4 Based on Examples 2 and 3, the testing phase: Step 7.1: Following the method in Step 1, the data used comes from the publicly available dataset published by AFRL Labs, "Cvdomes: A data set of simulated x-band signatures of civilian vehicles," which contains simulation data of different types of civilian vehicles. During testing, samples different from the training set are used to construct the test set, resulting in the HRRP data test set.
[0067] Step 7.2: Following the method in Step 2, preprocess the test set samples and angles to obtain the preprocessed test HRRP data.
[0068] Step 7.3: Following the method in Step 3, input the preprocessed test angle data, and obtain the test data convolution kernel parameters through the trained conditional angle kernel hybrid convolution module.
[0069] Step 7.4: Following the method in Step 4, input the test sample into the global feature extraction module and the local feature extraction module to obtain their features.
[0070] Step 7.5: Following the method in Step 5, input the global features and local features into the feature fusion module to obtain the test fused features.
[0071] Step 7.6: Following the method in Step 6, send the fused features of the test samples into the classification module to obtain the predicted category.
[0072] Example 5 The HRRP target recognition system based on angle-supervised feature generation of the present invention includes a target recognition network model: the target recognition network model includes: The preprocessing module is used to preprocess the samples in the dataset and perform two-dimensional processing on the azimuth angles to obtain preprocessed samples and two-dimensional angles. Angle-conditional kernel hybrid convolution module is used to map the azimuth angle to a Gaussian distribution with variable mean and variance through the encoding network, and then use a reparameterization method to generate a dynamic convolution kernel; this convolution kernel is used for feature extraction in convolution; The dual-branch feature extraction module is used to extract local features from CNN local features composed of dynamic convolution kernels from the angle conditional kernel hybrid convolution module and to extract global features from global features composed of Transformer encoders. The classification module is used to obtain classification results; The loss function module is used to set the classification loss based on the classification results and to set the center loss using the fused features. The target recognition module is used to train the target recognition model by setting a total loss function based on the classification loss and center loss to obtain a trained target recognition model; and to use the trained target recognition model to identify the HRRP of the vehicle to be identified.
[0073] Example 6 The method of this invention is compared with the hybrid convolutional module without angle conditional kernels, the Transformer method, the Lstm method, and the CNN method under different angular intervals. The results are as follows: Figure 2 As shown in the figure, the horizontal axis represents the relative angle density, which is defined as the maximum angle interval divided by the current angle interval. It reflects the sparseness or density of the observation viewpoints: the smaller the angle interval, the greater the relative angle density, indicating that the observation viewpoints are denser and the information that can be obtained is richer; the vertical axis represents the recognition accuracy.
[0074] The data used came from electromagnetic simulation data of 10 categories of civilian vehicles published by AFRL. The center frequency of the radar signal was 9.6 GHz, and the frequency spacing was 10.48 MHz. Ten types of targets were selected as experimental subjects, corresponding to categories 0 to 9.
[0075] The method of this invention specifically identifies 10 types of civilian vehicles, namely Toyota Camry, Honda Civic 4dr, 1993 Jeep, 1999 Jeep, Nissan Maxima, Mazda MPV, Mitsubishi, Nissan Sentra, Toyota Avalon, and Toyota Tacoma. Experimental verification shows that this invention has excellent practicality and reliability in complex scenarios with multi-directional changes.
[0076] As shown in the figure, the recognition accuracy of each method steadily increases with the increase of relative angle density, indicating that denser observation azimuths provide more sufficient information, thereby improving recognition performance. Under all test conditions, the proposed method consistently achieves the highest accuracy, with a particularly significant advantage under low angle density conditions, demonstrating its ability to effectively model and discriminate features even with limited observation information. In contrast, the comparative methods without the angle-conditional kernel hybrid convolution module show a performance decrease under all conditions, especially in low angle density regions, verifying the crucial role of the angle-conditional kernel hybrid convolution module. The Transformer method performs well under high angle density, but its performance degrades rapidly under sparse observation conditions; CNN and LSTM methods have limited ability to model cross-view feature differences, resulting in lower overall accuracy than the proposed method. In summary, the proposed method demonstrates superior stability and robustness, effectively adapting to recognition tasks with varying view densities.
Claims
1. An angle-guided HRRP target recognition method, characterized in that, This includes constructing a target recognition network model, using HRRP echo data of stationary target vehicles in the dataset, and extracting angle-adaptive local scattering features by dynamically convolving the HRRP data with azimuth information; further utilizing a global attention module to obtain global structural features; and finally, cross-fusing the local dynamic convolution features with the global attention features and outputting the target category through a classification module.
2. The HRRP target recognition method based on angle guidance according to claim 1, characterized in that, Specifically, the steps include the following: Step 1: Construct a target recognition network model, which includes a preprocessing module, an angle-conditional kernel hybrid convolution module, a dual-branch feature extraction module, a feature fusion module, and a classification and recognition module; Step 2: Obtain HRRP data of the target vehicle in a stationary state and construct a high-resolution distance-image HRRP dataset. =( ， ), i =1,2,… P, P Indicates the number of samples. Indicates azimuth angle ， Indicates HRRP distance; Step 3: Normalize and align the HRRP distance, and perform two-dimensional encoding on the azimuth. Step 4: Input the azimuth angle after the encoding process in Step 3 into the angle conditional kernel hybrid convolution module, map the azimuth angle into a Gaussian distribution with variable mean and variance, and then use the reparameterization method to generate dynamic convolution kernel parameters; Step 5: Use the dynamic convolution kernel parameters generated in Step 4 to extract local features in the dual-branch feature extraction module; at the same time, the global branch in the dual-branch feature extraction module extracts global features. Step 6: Utilize the cross-attention mechanism of the feature fusion module to fuse global and local features; use global features as queries and local features as keys and values to guide the network to selectively focus on the features most critical to the discrimination, thereby obtaining fused features; Step 7: Input the fused features into the classification module to obtain the classification result; Step 8: Set the classification loss based on the classification results, and use the fused features to set the center loss to train the model; use the trained target recognition model to identify the HRRP of the vehicle to be identified.
3. The angle-guided HRRP target recognition method and system according to claim 2, characterized in that, Step 3 is as follows: Step 3.1: Normalize and centroid-align each HRRP sample, resulting in the following normalized sample: ;in, This represents the normalized sample. This indicates the calculation of the 2-norm; Each HRRP distance Includes n One distance unit, , ; Indicates the first i In the nth sample p One distance unit; Center of gravity alignment processing in, C Indicates the centroid of the HRRP sample. n This indicates the number of distance cells. Then, the center of each HRRP sample is shifted to the centroid, thus completing the centroid alignment operation. Step 3.2, for the first i Azimuth angle of each sample Perform radian conversion, calculate its sine and cosine values, and generate azimuth encoding vectors for all samples. ; in, It is the azimuth angle. This is the processed azimuth angle.
4. The HRRP target recognition method and system based on angle guidance according to claim 3, characterized in that: Step 4, generating dynamic convolution kernel parameters, is as follows: Step 4.1: Convert the preprocessed azimuth angle As a conditional input, the mean vector of the output dynamic convolution kernel parameters is generated through a mean generation branch composed of a three-level fully connected network. ;in, and These represent the weights and biases of the fully connected layer, respectively. The mean generation branch consists of a three-layer fully connected network and two ReLU activations; the mean vector... The output dimension is preset to ; in, This indicates the number of parallel dynamic convolution kernels. =8, This represents the number of channels in the input feature map. This indicates the number of channels in the output feature map. This represents the size of the one-dimensional convolution kernel; And according to the number of parallel dynamic convolution kernels Will Reconstructed In which, the mean parameter of each group Corresponding to a set of candidate dynamic convolution kernel parameter values; Among them, each set of mean parameters Corresponding to a set of candidate dynamic convolution kernel parameter values Step 4.2, the preprocessed azimuth angle As conditional parallel inputs to the variance generation branch, the variance generation branch network packet... in, This refers to the Softplus activation function. This represents the Sigmoid function. To predetermine the variance constraint parameters, the variance generation branch consists of a three-layer fully connected network, two Softplus activations, and a final Sigmoid activation. The variance vector Also reconstructed as ; Variance parameters for each group Used to characterize the discreteness of the corresponding dynamic convolution kernel parameters; Step 4.3: Using the mean parameter and variance parameter as the distribution parameters of the dynamic convolution kernel random variable, for the first... Group dynamic convolution kernels are used for resampling and gradient updates: in, Indicates the first The dynamic convolution kernel parameters of the group, Represents a standard normal distribution; Step 4.4, preprocess the input samples Perform global average pooling and input it into the attention network to obtain... Each weight, The attention network includes a global average pooling layer, a flattening layer, a fully connected layer, a ReLU activation layer, a fully connected layer, and a sigmoid activation layer. Indicates the first The group of dynamic convolutional kernels' response weights to the current input. , , , These are the weights and parameters of the fully connected layer; Step 4.5, The group of dynamic convolutional kernels are fused together according to attention weights to obtain the final convolutional kernel. , Step 4.6: Perform dynamic convolution calculation. in, Let y represent the input feature, y represent the output feature, and C represent the dynamic convolution operation.
5. The HRRP target recognition method based on angle-supervised feature generation according to claim 4, characterized in that: In step 5, the dual-branch feature extraction module includes a CNN local feature extraction branch composed of dynamic convolutions and a global feature extraction branch composed of a Transformer encoder; The local feature extraction branch consists of three convolutional blocks connected sequentially and by residuals. The first and third are static convolutional blocks, while the second is a dynamic convolutional kernel generated by a conditional kernel hybrid convolution module. The dynamic convolutional block introduces an azimuth parameter into the convolution operation, giving the convolution weights azimuth-aware characteristics; the output of this module is local features. ; The global feature extraction module consists of a Transformer encoder, comprising a multi-head self-attention layer, two fully connected feedforward layers (FFN), and two LayerNorm layers. It sequentially processes the preprocessed samples through position encoding, the multi-head self-attention mechanism, layer normalization, and feedforward network connection layers to obtain a global feature representation specific to the current sample. .
6. The HRRP target recognition method based on angle-supervised feature generation according to claim 1, characterized in that: Step 6 uses cross-attention fusion as the feature fusion method, which consists of the following steps: Step 6.1, the feature fusion module is constructed in two parts: a multi-head cross-attention mechanism and a feedforward network module (FFN). The cross-attention can be represented as: Where Q, K, and V represent the query matrix, key matrix, and value matrix, respectively, which are obtained by performing three different linear transformations on the input, with dimensions of . The cross-attention mechanism incorporates the concept of cross-interaction in the expression of Q, K, and V, allowing Q, which calculates attention, to be derived from the extracted global features, while K and V are derived from local features. in, Let Q, K, and V be the learnable weight matrix. , Features extracted from different modules; cross-attention calculation results are: in, Its function is to perform a linear transformation on the input, transforming the input vector into a high-dimensional space. It contains 4 fully connected layers, 1 layer normalization layer, 1 Dropout layer, and 1 Softmax operation. Step 6.2: By stacking multiple cross-attention mechanisms to form MHCA, firstly, Q, K, and V are partitioned into multiple heads along the second dimension, resulting in... h Size , Then, perform a cross-attention calculation for each head: The output matrices of each head are concatenated along the last dimension: in, The concatenation operation consists of three linear transformation layers and a Softmax attention layer; finally, the output of the feature fusion module is... ;; Step 6.3: Introduce a center loss into the fused feature space of the cross-attention output. ; The center loss is calculated by measuring the squared Euclidean distance between the feature vector and its corresponding class center, as shown in the following expression: in, For the first i The fusion features of each sample For its corresponding category label, For the tag Category center, This represents the L2 norm.
7. The HRRP target recognition method based on angle-supervised feature generation according to claim 1, characterized in that: The classification module described in step 7 includes a fully connected layer and a softmax activation layer; The cross-entropy classification loss used is: Where N is the number of categories, Real label one-hot encoding, Predict class probabilities for the model; The overall loss function is: in, This represents the trade-off parameter, which is set to 0.
2.
8. The HRRP target recognition system based on angle-supervised feature generation has the following features: The system includes a target recognition network model: the target recognition network model includes: The preprocessing module is used to preprocess the samples in the dataset and perform two-dimensional processing on the azimuth angles to obtain preprocessed samples and two-dimensional angles. Angle-conditional kernel hybrid convolution module is used to map the azimuth angle to a Gaussian distribution with variable mean and variance through the encoding network, and then use a reparameterization method to generate a dynamic convolution kernel; this convolution kernel is used for feature extraction in convolution; The dual-branch feature extraction module is used to extract local features from CNN local features composed of dynamic convolution kernels from the angle conditional kernel hybrid convolution module and to extract global features from global features composed of Transformer encoders. The classification module is used to obtain classification results; The loss function module is used to set the classification loss based on the classification results and to set the center loss using the fused features. The target recognition module is used to train the target recognition model by setting a total loss function based on the classification loss and center loss to obtain a trained target recognition model; and to use the trained target recognition model to identify the HRRP of the vehicle to be identified.
9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.