A dual-band radar cross-frequency feature learning target recognition method and device

Abstract

The present application belongs to the technical field of radar signal processing and target identification, and particularly relates to a method and device for dual-band radar cross-frequency feature learning target identification, step 1, dual-band HRRP data is acquired and preprocessed and dataset is divided to obtain a training set and a validation set; step 2, a cross-frequency feature learning CFFL network is constructed, including a hand-crafted feature encoding HFE module, a deep difference representation DDR module, a feature fusion FF module and a classifier; step 3, a difference search attention mechanism and a difference search loss function are introduced into the DDR module to enhance the discriminative difference between the dual bands, and a loss function is constructed in combination with a cross-entropy classification loss to perform joint optimization; step 4, the dual-band HRRP training set and validation set samples are input into the constructed cross-frequency feature learning network, and end-to-end training is performed by using the loss function constructed in step 3; step 5, the trained CFFL network is used for target classification and identification.

Description

Technical Field

[0001] This invention belongs to the field of radar signal processing and target recognition technology, specifically relating to a method and apparatus for target recognition using cross-frequency feature learning in dual-band radar. Background Technology

[0002] Automatic radar target identification is a technology that uses radar echoes to extract stable features of targets and then automatically determines the target type and attributes. Dual-band high-resolution range profiles (HRRPs) are the projected distribution of the target's scattering center along the radar's line-of-sight, obtained by broadband radar. They reflect the relative position and intensity information of the target's scattering points. Due to their simplicity in acquisition and ease of processing, they are of great value in radar target identification.

[0003] Traditional HRRP target recognition methods primarily rely on single-band data, manually designing geometric, statistical, or transform domain features and combining them with classifiers such as support vector machines and random forests to achieve recognition. In recent years, deep learning models such as convolutional neural networks, recurrent neural networks, and Transformers have been introduced into HRRP recognition tasks, enabling end-to-end feature learning and classification, thus improving recognition performance. However, in complex scenarios with varying target poses, noise interference, or high inter-class waveform similarity, the information contained in a single-band HRRP is limited, and recognition performance remains challenging.

[0004] Expanding radar frequency bands and acquiring target HRRP at different carrier frequencies can provide richer scattering information. Since electromagnetic scattering is frequency-dependent, radar waves in different frequency bands interact with targets, triggering different scattering mechanisms: low-frequency components mainly reflect the overall outline of the target, while high-frequency components are more sensitive to surface details. Therefore, complementary and discriminative differences exist between dual-band HRRPs, helping to enhance inter-class separability and improve recognition robustness. However, current research on dual-band HRRP recognition is insufficient, particularly lacking systematic methods for effectively extracting and fusing cross-band discriminative features.

[0005] Therefore, there is an urgent need for a target recognition method that can fully utilize cross-frequency difference information of dual-band HRRP and integrate physical features and deep semantic features to improve recognition performance under complex conditions such as high inter-class similarity and large pose changes. Summary of the Invention

[0006] This invention provides a method and apparatus for target recognition using cross-frequency feature learning for dual-band radar. Cross-frequency feature learning (CFFL) is a machine learning paradigm for dual-band radar data. It focuses on mining and enhancing the discriminative "subtraction" information between the two frequency band HRRPs, thereby effectively improving the recognition capability in scenarios with high inter-class similarity.

[0007] This invention is achieved through the following technical solution:

[0008] In a first aspect, the present invention provides a method for target recognition using cross-frequency feature learning in dual-band radar, comprising the following steps: Step 1: Obtain dual-band HRRP data and perform preprocessing and dataset partitioning to obtain training and validation sets; Step 2 involves constructing a cross-frequency feature learning (CFFL) network, comprising a handcrafted feature encoding (HFE) module, a deep differential representation (DDR) module, a feature fusion (FF) module, and a classifier. The HFE module extracts cross-frequency handcrafted features from dual-band HRRP data and performs feature re-representation using the coding rate reduction principle. The DDR module extracts multi-scale features from dual-band HRRP data through deep wavelet packet decomposition and performs feature interaction to obtain deep interactive features. The FF module performs nonlinear fusion of the two types of features through deep discriminative correlation analysis. The classifier classifies the nonlinearly fused features to obtain class prediction probabilities. Step 3: Introduce a difference search attention mechanism and a difference search loss function into the DDR module to enhance the discriminative difference between the two frequency bands, and combine it with the cross-entropy classification loss to construct a loss function for joint optimization; Step 4: Input the dual-band HRRP training set and validation set samples into the constructed cross-frequency feature learning network, and perform end-to-end training using the loss function constructed in Step 3; Step 5: Use the trained CFFL network to perform target classification and recognition.

[0009] Optionally, the data partitioning described in this invention adopts different strategies according to the specific needs of the recognition task: under standard operating conditions, the omnidirectional data is randomly divided into training set and test set according to a set ratio; under extended operating conditions, the training set and test set are constructed by partitioning by azimuth angle, separating by batch by acquisition time, or grouping by different observation conditions.

[0010] Optionally, the HFE module of the present invention includes a cross-frequency feature extraction submodule and a feature re-representation submodule; the cross-frequency feature extraction submodule extracts two types of features from HRRP data. The first type is multi-dimensional cross-frequency band comparative statistical features, including geometric features and statistical features. The same feature value extracted from both frequency bands is summed and subtracted respectively to form a multi-dimensional cross-frequency band comparative statistical feature set; the second type is cross-frequency distance features; the first and second types of features are concatenated to obtain... S A handcrafted feature vector; the feature re-representation submodule, based on the principles of sparse coding and coding rate reduction, maximizes the difference between the overall coding rate and the intra-class coding rate, and performs the re-representation on the... S Optimize the handcrafted feature vectors to achieve feature re-representation, and output the optimized handcrafted features. This improves the intra-class compactness and inter-class separability of features.

[0011] Optionally, the feature re-representation module of the present invention implements the K-SVD algorithm to... S Optimize the 3D handcrafted feature vectors.

[0012] Optionally, the DDR module of this invention consists of multi-level deep wavelet packet decomposition layers, with each DWPD layer containing two parallel wavelet blocks. Each decomposition layer performs three-level wavelet packet decomposition on the input dual-band HRRP to obtain multi-scale sub-waveform features. A difference search attention mechanism is used to dynamically enhance the difference regions of the dual-band features at multiple scales, and a difference search loss function is used to supervise the process, thereby strengthening the extraction of cross-frequency discriminative information. The module also includes a deep inter-frequency interaction module to realize information interaction and complementary fusion of dual-band deep features, obtaining deep interactive features. .

[0013] Optionally, this invention uses a difference search loss function. To supervise, for the first k The loss of the WaveBlock layer is:

[0014] in, and Frequency band 1 and frequency band 2 respectively after the first k The first layer obtained after layer decomposition i Sub-band feature sequences, The corresponding number calculated by the DSA mechanism i The attention weight sequence of each sub-band This indicates the calculation of cosine similarity; the total difference search loss is the sum of the losses of each layer:

[0015] in, LThis represents the total number of WaveBlock layers.

[0016] Optionally, the FF module of this invention includes an MLP-CNN unit that performs nonlinear mapping on the two types of features, mapping the features to a high-dimensional space to make the features sparsified, and then obtains the left and right singular vector matrices and the diagonal singular value matrix through singular value decomposition; further, eigenvalue decomposition is performed on the singular value matrix to decouple the singular value matrix into eigenvalue and eigenvector representations; based on the result of eigenvalue decomposition, transformation matrices are calculated for the handmade features and deep features respectively, realizing the mapping of the original features to a new discriminative subspace, and finally obtaining the fused features, which are then concatenated along the spatial dimension to obtain the final fused features.

[0017] Optionally, the present invention also extracts scattering center information based on a preset threshold and encodes it into an auxiliary feature vector, and inputs it into a classifier in conjunction with the fused features.

[0018] Optionally, the total loss function of this invention is composed of a weighted sum of difference search loss and cross-entropy loss.

[0019] In a second aspect, the present invention provides a target recognition device for cross-frequency feature learning in dual-band radar, comprising a manual feature encoding (HFE) module, a depth difference representation (DDR) module, a feature fusion (FF) module, and a classifier. The HFE module is used to extract cross-frequency handcrafted features from dual-band HRRP data and perform feature re-representation using the coding rate reduction principle to obtain handcrafted features. ; The DDR module extracts multi-scale features from dual-band HRRP data through deep wavelet packet decomposition and performs feature interaction to obtain deep interactive features. ; The FF module is used to perform deep discriminative correlation analysis on two types of features. and Perform nonlinear fusion; A classifier is used to classify nonlinear fused features and obtain the predicted probability of the category.

[0020] Beneficial effects: First, this invention constructs an end-to-end dual-band HRRP cross-frequency feature learning and recognition model. This model can automatically extract and fuse cross-frequency handcrafted features and deep semantic features from dual-band data, directly outputting recognition results, realizing integrated processing from dual-band data input to target category determination.

[0021] Second, this invention proposes a feature learning mechanism for explicitly modeling cross-frequency discriminative differences. Unlike traditional dual-band methods that focus on information "addition" or complementary fusion, this invention focuses on mining and enhancing the discriminative "subtraction" information between dual-band HRRPs through manual feature level coding rate optimization and deep feature level difference search attention and loss constraints, thereby effectively improving the recognition ability in scenarios with high inter-class similarity.

[0022] Third, this invention achieves a deep fusion of physically interpretable features and data-driven features. By introducing scattering structure descriptors with clear physical meaning through a manual feature encoding module, the interpretability and robustness of the model are enhanced; by extracting high-level semantic features driven by data through a deep difference representation module, the representational power of the features is improved; and finally, nonlinear fusion is performed through deep discriminative correlation analysis, fully utilizing the complementary advantages of the two types of features.

[0023] Fourth, the effectiveness of the proposed method has been verified on multiple simulation and experimental datasets. Experiments show that, under both standard operating conditions and extended operating conditions (such as pose changes), the proposed method significantly outperforms traditional handcrafted feature methods, deep learning models, and existing dual-band fusion methods, especially in distinguishing structurally similar targets.

[0024] This invention is particularly suitable for the automatic classification and identification of targets using dual-band radar echo data. Attached Figure Description

[0025] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0026] Figure 1 This is the overall flowchart of the CFFL framework.

[0027] Figure 2 This is a structural diagram of the HFE module.

[0028] Figure 3 This is a schematic diagram of the DDR module.

[0029] Figure 4 This is a structural diagram of the FF module and its classification.

[0030] Figure 5 This is a schematic diagram of the structure of a deep wavelet packet decomposition layer. Detailed Implementation

[0031] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0032] It should be noted that, in the absence of conflict, the following embodiments and features can be combined with each other; and, based on the embodiments of this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.

[0033] It should be noted that various aspects of embodiments within the scope of the appended claims are described below. It will be apparent that the aspects described herein can be embodied in a wide variety of forms, and any particular structure and / or function described herein is merely illustrative. Based on this disclosure, those skilled in the art will understand that one aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number of aspects set forth herein can be used to implement the device and / or practice the method. Additionally, this device and / or method can be implemented using structures and / or functionalities other than one or more of the aspects set forth herein.

[0034] For dual-band radar target HRRP, the scattering mechanisms of electromagnetic waves of different frequencies interacting with the target exhibit significant differences: lower-frequency electromagnetic waves (such as the Ku band) primarily interact with the macroscopic structure of the target, such as its overall outline, while higher-frequency electromagnetic waves (such as the W band) are more sensitive to microscopic details such as the fine roughness of the target surface. This frequency-dependent scattering behavior leads to discriminative differences in the waveforms of dual-band HRRP. Based on the aforementioned characteristics of dual-band HRRP, this invention proposes a new cross-frequency difference-driven paradigm. Unlike existing technologies that focus on the complementarity and superposition of multi-band information ('addition'), this paradigm uses the discriminative differences between radar echoes of different frequency bands as key features actively mined, enhanced, and utilized by machine learning models ('subtraction'), thereby fundamentally improving the ability to distinguish targets with similar scattering characteristics.

[0035] This invention provides a method for target recognition using cross-frequency feature learning in dual-band radar. The overall process of this method corresponds to... Figure 1 The CFFL model block diagram shown below illustrates the specific process as follows: First, dual-band HRRP data is acquired, and necessary preprocessing and dataset partitioning are performed. Preprocessing includes range cell registration of the dual-band HRRP using an alignment method, amplitude normalization to eliminate scale differences, and extraction of scattering center information based on a preset threshold, which is then encoded into auxiliary feature vectors.

[0036] Subsequently, a CFFL network was constructed and combined with a classifier to form a cross-frequency feature learning target recognition model. The CFFL network structure mainly consists of three core modules: the HFE module, the DDR module, and the FF module. The HFE module extracts cross-frequency handcrafted features from the dual-band HRRP data and uses the coding rate reduction principle for feature re-representation to reduce feature heterogeneity and enhance discriminability. In the DDR module, multi-scale features of the dual-band HRRP data are extracted through Deep Wavelet Packet Decomposition (DWPD), and the Difference Search Attention (DSA) mechanism and difference search loss function are used to enhance the discriminative difference between the two bands. In the FF module, the two types of features are nonlinearly fused through Deep Discriminant Correlation Analysis (DDCA). The classifier consists of a fully connected layer plus a Softmax layer; the nonlinearly fused features are input into the classifier to obtain the class prediction probability.

[0037] Finally, the cross-frequency feature learning target recognition model was trained using the training dataset, and the trained model was used for classification and recognition of test set samples.

[0038] Specifically, the steps include the following: Step 1: Data Preprocessing and Dataset Partitioning 101. Acquire dual-band HRRP data for radar targets. Assume there are C types of targets, and each sample contains HRRP data for two frequency bands, denoted as follows: and , where L is the number of distance units (i.e., HRRP length).

[0039] 102. Amplitude normalization is performed on the dual-band HRRP to eliminate the influence of factors such as radar transmit power and target distance on the absolute value of the amplitude. A common method is energy normalization.

[0040] in, This represents the L2 norm.

[0041] 103. Split the dataset. Randomly divide all the data into training and test sets according to a certain ratio (e.g., 70% and 30%) to ensure that the samples of each category are evenly distributed in the split.

[0042] Data partitioning employs different strategies depending on the specific needs of the recognition task: under standard operating conditions, omnidirectional data can be randomly divided into training and test sets according to a certain proportion; under extended operating conditions, training and test sets can be constructed by partitioning by azimuth angle, separating by batches of acquisition time, or grouping by different observation conditions, so as to fully verify the generalization and robustness performance of the model under multi-sample distribution and complex scenarios.

[0043] Step 2: Construct the Cross-Frequency Feature Learning (CFFL) network. The overall process of the CFFL network is as follows: Figure 1 It consists of a Handcrafted Feature Encoding (HFE) module, a Deep Difference Representation (DDR) module, and a Feature Fusion (FF) module, in sequence.

[0044] 201. Construct the HFE module. This module is responsible for extracting physically meaningful cross-frequency handcrafted features from dual-band HRRP and optimizing its feature space through feature re-representation. First, two classes of shared features are designed... S Multidimensional handcrafted features. The first type is cross-band comparative statistical features. The method is to extract geometric features (such as symmetry and dispersion) and statistical features (such as peak count and energy accumulation ratio) from the HRRP of each frequency band. Then, the same feature values ​​extracted from the two frequency bands are summed and subtracted respectively to finally form a multidimensional cross-band comparative statistical feature set.

[0045] Taking symmetry as an example: for frequency bands k ( k HRRP (1 or 2) Its symmetry characteristics The calculation is as follows:

[0046] in, This represents the cross-band comparison statistical characteristic operator. yes The i Each distance unit amplitude value.

[0047] The second category is cross-frequency distance characteristics, which directly calculate the differences between the two-band HRRP waveforms, such as Euclidean distance:

[0048] in, This represents the distance feature extraction operator. and The first two are respectively the first two bands of frequency band 1 and frequency band 2. i Each distance unit amplitude value.

[0049] By concatenating all features S 3D handcrafted feature vectors ,in N This represents the number of samples.

[0050] Feature re-representation module S 3D handcrafted feature vectors Feature re-representation, the specific process is as follows: First, sparse coding is performed using the K-SVD algorithm. The objective function is:

[0051] in, Denotes the Frobenius norm. It is a sparse coefficient matrix Z The i List, T The K-SVD algorithm is used to iteratively update the dictionary, with sparsity control parameters. D sparse matrix Z .

[0052] Subsequently, through L Layer coding rate reduction layer pair Z Perform iterative transformations, each layer of transformation Defined as:

[0053] in, This represents the normalization operator projected onto the unit sphere. η Step size, Used to increase the overall coding rate Used to reduce the j Class coding rate σ It is a nonlinear operator.

[0054] go through L Layer iterations yield the final mapping

[0055] in, This indicates the composition of functions.

[0056] Finally, through the dictionary D Transformed features Map back to the vicinity of the original space and output optimized handcrafted features. :

[0057] 202. Construct the DDR module. This module is responsible for extracting deep, multi-scale discriminative differential features from the dual-band HRRP, with the DWPD layer at its core. (The text abruptly ends here, likely due to an incomplete sentence or missing information.) , Parallel input DWPD layers. Each DWPD layer contains two parallel waveblocks, structured as follows: Figure 5 (As shown). Wavelet packet decomposition follows the Mallat algorithm, for the _ j Layer nodes The coefficient of the next layer is:

[0058] in, h and g These are the coefficients of the low-pass and high-pass filters, respectively. For approximate components, For detailed components. After... M After stacking several DWPD layers, the output is a deep multi-scale feature:

[0059]

[0060] in, Indicates the first k The first frequency band m WaveBlock of the layer.

[0061] In each DWPD layer, the DSA mechanism is used. Simultaneously, the difference search loss function is employed. Supervision shall be conducted. Regarding the first... k The loss of the WaveBlock layer is:

[0062] in, and Frequency band 1 and frequency band 2 respectively after the first k The first layer obtained after layer decomposition i Sub-band feature sequences, The corresponding number calculated by the DSA mechanism i The attention weight sequence of each sub-band This indicates the calculation of cosine similarity. The total difference search loss is the sum of the losses from each layer:

[0063] in, L This represents the total number of WaveBlock layers. and The input is a Deep Frequency Interaction (DFI) block. This block achieves feature interaction through operations such as full-dimensional dynamic convolution, and finally outputs deep interactive features. :

[0064] in, This represents the nonlinear transformation function of the DFI block.

[0065] 203. Build the FF module.

[0066] This module is responsible for handling manual features. with depth features For effective fusion, the DDCA algorithm is employed. This algorithm first performs nonlinear mapping on the two types of features using an MLP-CNN unit, sparsifying the features. Next, to ensure that the two types of features achieve maximum intra-class correlation and minimum inter-class correlation after transformation, this invention first performs singular value decomposition (SVD) on the covariance matrix to reduce inter-class correlation, obtaining its left and right singular vector matrices and diagonal singular value matrix. Based on this, eigenvalue decomposition is further performed on the singular value matrix, decoupling it into eigenvalue and eigenvector representations to construct an optimal linear transformation space. Based on the eigenvalue decomposition results, transformation matrices are calculated for both handcrafted and deep features, mapping the original features to a new discriminative subspace, ultimately yielding the fused features. and By splicing them along the spatial dimension, the final fused feature is obtained. .Will Input a classifier consisting of a fully connected layer and a softmax layer to obtain the predicted class probability. .

[0067] Step 3: Constructing and Optimizing the Loss Function The total loss function of the CFFL model consists of two parts:

[0068] in, For difference search loss, The standard cross-entropy classification loss is used. α and β These are the weights of the differential search loss and the cross-entropy loss, respectively.

[0069] Step 4: Model Training 401. First, independently train the handcrafted feature encoding (HFE) module. Extract handcrafted features using the training set data. The parameters of the coding rate reduction layer in the HFR block are updated through forward propagation. θ, thus obtaining optimized manual features .

[0070] 402. Training the DDR and FF modules. This involves training the dual-band HRRP. and corresponding Input network. Employ stochastic gradient descent or Adam optimizer, with a total loss of [missing information]. Perform backpropagation to the target and update all weight parameters of the network.

[0071] Step 5: Model Testing 501. For the test set samples, first extract their handcrafted features. And obtained through a trained hand-crafted feature encoding module (sparse encoding, encoding simplification layer iteration, reconstruction) .

[0072] 502. The dual-band HRRP of the test samples Input the trained DDR module to obtain deep features. .

[0073] 503, will and Input the data into the FF module for fusion and classification, and output the final category prediction result. .

[0074] Example This embodiment uses a simulation dataset (SD), measured dataset 1 (MD1), and measured dataset 2 (MD2) to verify the method of the present invention. SD contains simulated HRRP data for five types of vehicle targets (sedans, SUVs, pickup trucks, trucks, and minibuses) in the Ku band (16GHz) and W band (93GHz). The target attitude covers 360° azimuth angles (1° intervals) and two pitch angles of 29° and 30°. For each azimuth-pitch combination, 10 samples are generated by adding 20dB Gaussian white noise, for a total of 36,000 dual-band samples. MD1 contains measured Ku / W band data for three types of vehicle targets (trucks, SUVs, and vans), collected at eight evenly distributed azimuth angles, containing 20,000 samples. MD2 contains measured Ku / Ka band data for three types of vehicle targets (SUVs, pickup trucks, and sedans), covering all azimuth angles and two pitch angles of 0° and 30°. All datasets were split according to Standard Operating Conditions (SOC), that is, 70% of the samples were randomly selected as the training set and 30% as the test set.

[0075] To comprehensively evaluate performance, this embodiment sets up multiple control methods. These include traditional classifiers based on handcrafted features (Support Vector Machines (SVM), K-Nearest Neighbors (KNN), deep learning models (Convolutional Neural Networks (CNN), VGGNet, Target Attention Convolutional Neural Networks (TACNN), etc.), and methods that fuse handcrafted features with deep features (Multi-Input Convolutional Gated Recurrent Units (MIConvGRU), Multi-Frequency Feature Fusion Networks (MFCFNet), etc.). All methods were adapted to handle dual-band HRRP inputs and experimented on the same training-test split. Each method was run independently for five trials, and the final recognition rate was reported as "mean ± standard deviation" to assess the stability of the method.

[0076] The dual-band HRRP is input into the CFFL network of this invention. Physically meaningful cross-frequency features are extracted and optimized through a manual feature encoding module. Multi-scale discriminative differences are extracted through wavelet packet decomposition and attention mechanisms in the deep discriminative characterization module. Finally, the two types of features are fused via deep discriminative correlation analysis and input into the classifier to complete target recognition. The recognition performance of different methods on the three SOC datasets is compared in the table below.

[0077] Table 1. Comparison of recognition performance of different methods on three SOC datasets. (mean ± standard deviation, %)

[0078] As shown in Table 1, the CFFL method proposed in this invention achieved the highest average recognition accuracy on all three datasets, and its standard deviation was generally lower than or comparable to the best baseline method, indicating stable performance. Specifically, on SD, MD1, and MD2, CFFL improved the recognition rate by 0.79%, 2.30%, and 3.11% respectively compared to the second-best baseline method MFCFNet. Furthermore, in more challenging test scenarios such as azimuth extension conditions, the method of this invention also demonstrated optimal performance, with an average recognition rate improvement of 5% to 8% compared to other baseline methods. This verifies that the method effectively improves the generalization ability and robustness of the model under conditions of limited training data perspective by explicitly mining and enhancing cross-frequency discrimination differences.

[0079] In summary, the CFFL framework, by deeply fusing the physical and semantic features of dual-band HRRP and focusing on extracting cross-frequency discrimination information, can significantly and stably improve the performance of radar automatic target recognition systems under various operating conditions.

[0080] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for target recognition using cross-frequency feature learning in dual-band radar, characterized in that, Includes the following steps: Step 1: Obtain dual-band HRRP data and perform preprocessing and dataset partitioning to obtain training and validation sets; Step 2 involves constructing a cross-frequency feature learning (CFFL) network, comprising a handcrafted feature encoding (HFE) module, a deep differential representation (DDR) module, a feature fusion (FF) module, and a classifier. The HFE module extracts cross-frequency handcrafted features from dual-band HRRP data and performs feature re-representation using the coding rate reduction principle. The DDR module extracts multi-scale features from dual-band HRRP data through deep wavelet packet decomposition and performs feature interaction to obtain deep interactive features. The FF module performs nonlinear fusion of the two types of features through deep discriminative correlation analysis. The classifier classifies the nonlinearly fused features to obtain class prediction probabilities. Step 3: Introduce a difference search attention mechanism and a difference search loss function into the DDR module to enhance the discriminative difference between the two frequency bands, and combine it with the cross-entropy classification loss to construct a loss function for joint optimization; Step 4: Input the dual-band HRRP training set and validation set samples into the constructed cross-frequency feature learning network, and perform end-to-end training using the loss function constructed in Step 3; Step 5: Use the trained CFFL network to perform target classification and recognition.

2. The target recognition method for cross-frequency feature learning in dual-band radar according to claim 1, characterized in that, The data partitioning adopts different strategies according to the specific needs of the recognition task: under standard operating conditions, the omnidirectional data is randomly divided into training set and test set according to a set ratio; under extended operating conditions, the training set and test set are constructed by partitioning by azimuth angle, separating by batch by acquisition time, or grouping by different observation conditions.

3. The target recognition method for cross-frequency feature learning in dual-band radar according to claim 1 or 2, characterized in that, The HFE module includes a cross-frequency feature extraction submodule and a feature re-representation submodule. The cross-frequency feature extraction submodule extracts two types of features from the HRRP data. The first type is multi-dimensional cross-frequency band comparative statistical features, which include geometric features and statistical features. The same feature value extracted from the two frequency bands is summed and subtracted respectively to form a multi-dimensional cross-frequency band comparative statistical feature set. The second type is cross-frequency distance features. By concatenating the first and second types of features, we obtain... S A handcrafted feature vector; the feature re-representation submodule, based on the principles of sparse coding and coding rate reduction, maximizes the difference between the overall coding rate and the intra-class coding rate, and performs the re-representation on the... S Optimize the handcrafted feature vectors to achieve feature re-representation, and output the optimized handcrafted features. .

4. The target recognition method for cross-frequency feature learning in dual-band radar according to claim 3, characterized in that, The feature re-representation module implements the K-SVD algorithm to... S Optimize the 3D handcrafted feature vectors.

5. The target recognition method for cross-frequency feature learning in dual-band radar according to claim 1 or 2, characterized in that, The DDR module consists of multi-level deep wavelet packet decomposition layers, with each DWPD layer containing two parallel wavelet blocks. Each decomposition layer performs three-level wavelet packet decomposition on the input dual-band HRRP to obtain multi-scale sub-waveform features. A difference search attention mechanism is used to dynamically enhance the difference regions of the dual-band features at multiple scales, and a difference search loss function is used to supervise the process to strengthen the extraction of cross-frequency discrimination information. The module also includes a deep inter-frequency interaction module to realize the information interaction and complementary fusion of dual-band deep features to obtain deep interactive features. .

6. The target recognition method for cross-frequency feature learning in dual-band radar according to claim 5, characterized in that, Using the difference search loss function To supervise, for the first k The loss of the WaveBlock layer is: in, and Frequency band 1 and frequency band 2 respectively after the first k The first layer obtained after layer decomposition i Sub-band feature sequences, The corresponding number calculated by the DSA mechanism i The attention weight sequence of each sub-band This indicates the calculation of cosine similarity; the total difference search loss is the sum of the losses of each layer: in, L This represents the total number of WaveBlock layers.

7. The method for target recognition using cross-frequency feature learning in dual-band radar according to claim 1 or 2, characterized in that, The FF module includes an MLP-CNN unit that performs nonlinear mapping on the two types of features, mapping the features to a high-dimensional space to make the features sparsified. Then, singular value decomposition is used to obtain the left and right singular vector matrices and the diagonal singular value matrix. Furthermore, eigenvalue decomposition is performed on the singular value matrix to decouple the singular value matrix into eigenvalue and eigenvector representations. Based on the results of eigenvalue decomposition, transformation matrices are calculated for handcrafted features and deep features respectively, realizing the mapping of the original features to a new discriminative subspace. Finally, the fused features are obtained and concatenated along the spatial dimension to obtain the final fused features.

8. The target recognition method for cross-frequency feature learning in dual-band radar according to claim 7, characterized in that, Furthermore, scattering center information is extracted based on a preset threshold and encoded into an auxiliary feature vector, which is then combined with the fused features and input into the classifier.

9. The target recognition method for cross-frequency feature learning in dual-band radar according to claim 6, characterized in that, The total loss function is a weighted sum of the difference search loss and the cross-entropy loss.

10. A target recognition device for cross-frequency feature learning in dual-band radar, characterized in that, It includes a hand-crafted feature encoding (HFE) module, a deep difference representation (DDR) module, a feature fusion (FF) module, and a classifier; The HFE module is used to extract cross-frequency handcrafted features from dual-band HRRP data and perform feature re-representation using the coding rate reduction principle to obtain handcrafted features. ; The DDR module extracts multi-scale features from dual-band HRRP data through deep wavelet packet decomposition and performs feature interaction to obtain deep interactive features. ; The FF module is used to perform deep discriminative correlation analysis on two types of features. and Perform nonlinear fusion; A classifier is used to classify nonlinear fused features and obtain the predicted probability of the category.

Images