Radar target HRRP identification method based on sparse view angle feature generation fusion

Abstract

The invention discloses a radar target HRRP recognition method based on sparse view angle feature generation fusion. In order to solve the problem that the performance of an HRRP target recognition method is limited due to the fact that HRRP data of a non-cooperative target is usually difficult to cover a full view, the invention provides a generative multi-view HRRP recognition (GMVR) model. Firstly, a visual angle feature extension module (VFEM) is provided by utilizing the mode learning and generation capability of a depth generation model and is used for generating multi-visual angle features near a sparse visual angle; then, based on feature level fusion, a multi-view feature fusion module (MVFM) is provided. By means of an attention mechanism, the generated features can be adaptively integrated, thereby reducing decision deviation caused by a single-view HRRP sample. In addition, transfer learning is adopted to relieve the problem of small sample size caused by a sparse view angle.

Description

Technical Field

[0001] The present invention belongs to the technical field of ultra-narrow pulse radar target recognition, and specifically relates to an adjacent view feature expansion method of an ultra-narrow pulse radar target echo based on a generative adversarial network (GAN) and a feature adaptive fusion recognition method based on an attention mechanism. Background Art

[0002] Extremely narrow pulse radar target echo, also known as High Resolution Range Profile (HRRP), is an important radar signal processing technology. Since the width of a single pulse echo is much smaller than the target size, it represents the projection of the target scattering point along the radar line of sight. By analyzing the frequency, phase and amplitude of the target reflection signal, the high-resolution characteristics of the target are obtained, providing rich information about the target structure. Extremely narrow pulses reflect the shape and size of the target, provide a high-resolution feature description of the target, and help to achieve more accurate target recognition and positioning in radar target recognition and tracking. Due to its easy real-time processing characteristics, extremely narrow pulses play a key role in the research of radar automatic target recognition (RATR), and are one of the hot spots in the research of radar automatic target recognition based on extremely narrow pulses.

[0003] The ultra-narrow pulse RATR method based on deep learning has been proven to achieve state-of-the-art performance. However, in actual application scenarios, data collection requires a lot of manpower, material resources and financial resources, resulting in high costs. In addition, in real application environments, non-cooperative targets often need to be identified, and it is even more difficult to obtain actual measured data for non-cooperative targets, resulting in sparse collection of target ultra-narrow pulse data in azimuth. Given the strong attitude sensitivity of ultra-narrow pulses, this sparse perspective data will limit the performance of the ultra-narrow pulse RATR method based on deep learning.

[0004] Data enhancement is a common method to solve the problem of reduced recognition performance due to the sensitivity of extremely narrow pulse gestures. With the rapid development of deep learning and neural networks, the richness of the representation of extremely narrow pulse data can be enhanced through deep generative models.

[0005] Multi-view recognition uses consistent and complementary information between multiple viewpoints to make decisions. Although such methods cannot expand the viewpoint information of the data, multi-view decision-making reduces the decision bias caused by a single viewpoint and effectively improves the recognition rate. However, in scenes with sparse viewpoints, it can be challenging to obtain multi-view data of the target, making multi-view recognition methods impractical.

[0006] Therefore, it is extremely important to study how to solve the problem of reduced radar automatic target recognition performance due to the sparse viewing angle and extremely narrow pulses of non-cooperative targets. One solution is to combine the generative model with the recognition model and complete the multi-view expansion of features through the generative model.

[0007] In view of the above problems, the present invention proposes a method for generating a multi-view extremely narrow pulse recognition model for sparse view target recognition. The method first extracts features from the original extremely narrow pulse data through an encoder, then completes the multi-view expansion of the features through a generative model, then integrates the multi-view features through an attention mechanism, and finally recognizes the integrated multi-view features. The method mainly consists of two modules: a view feature expansion module (VFEM) and a multi-view fusion module (MVFM), which enables the network to reduce complexity through end-to-end cascade generation and fusion structures. In addition, considering that training samples are often limited in the case of sparse view, making them prone to overfitting in the generation network and recognition network, the present invention proposes a multi-domain transfer training strategy: by learning the view change pattern of cooperative targets and applying it to non-cooperative targets, the model effectively expands the view information of the data and reduces the distribution decision boundary deviation. Experiments show that this method can significantly improve the recognition accuracy of extremely narrow pulses under sparse view. Summary of the invention

[0008] The purpose of the present invention is to construct an extremely narrow pulse end-to-end cascade generation, fusion and recognition network, which is mainly composed of two main modules: a view feature expansion module (VFEM) and a multi-view fusion module (MVFM), so that the network can reduce complexity through an end-to-end cascade generation and fusion structure, thereby improving the performance of radar automatic target recognition under sparse view.

[0009] The technical solution of the method of the present invention is:

[0010] A radar target HRRP recognition method based on sparse viewing angle feature generation and fusion includes the following steps:

[0011] Step 1: Construct an extremely narrow pulse echo dataset, divide it into a source domain training set, a target domain training set, and a test set, and perform normalization preprocessing;

[0012] Step 2: Construct a view feature extension module (VFEM), including a feature extraction network E, a feature generation network G, and a feature discriminator D, and construct related adversarial losses and reconstruction losses;

[0013] Step 3: Construct a multi-view fusion module (MVFM), including the fusion network F based on the attention mechanism and the fully connected network classifier C, and construct the relevant classification loss;

[0014] Step 4: Input the extremely narrow pulse echoes of all viewing angles in the source domain and their corresponding labels into the viewing feature expansion module. The expanded features enter the multi-view fusion module. The entire network is pre-trained through the corresponding loss function to complete parameter optimization.

[0015] Step 5: Input the extremely narrow pulse echoes and their corresponding labels in the training set of the target domain into the view feature expansion module of the pre-trained network, and perform training optimization of the entire network again to complete the training of the model;

[0016] Step 6: Input the extremely narrow pulse echoes and their corresponding labels in the target domain test set into the trained network. After view feature expansion and feature fusion, the recognition rate of the model for the target domain test set is finally obtained.

[0017] Furthermore, the step 1 includes the following sub-steps:

[0018] Step 1.1, obtain extremely narrow pulse echo data and corresponding category labels of M types of full-view targets as training data sets in the source domain;

[0019] Step 1.2, obtain the extremely narrow pulse echo data and corresponding category labels of N types of full-view targets at different pitch angles, sample the azimuth angle of the data interval Δθ at a certain pitch angle to obtain the extremely narrow pulse echo data under sparse view angle and the corresponding category labels as the training set of the target domain, and use the all-round extremely narrow pulse data and corresponding category labels at another pitch angle as the test set of the target domain;

[0020] Step 1.3: Perform maximum amplitude normalization on all data sets in the above source domain and target domain.

[0021] Furthermore, the step 2 includes the following sub-steps:

[0022] Step 2.1: Construct a one-dimensional (1D) VGG network as a feature extractor E, which maps the input extremely narrow pulse data with an azimuth angle of θ into deep features f θ ;

[0023] Step 2.2: Build Figure 2 The feature generation network G shown is based on a one-dimensional U-Net architecture, consisting of three upsampling layers, three downsampling layers, and an intermediate layer;

[0024] Step 2.3: Construct the discriminator network D and loss function.

[0025] Furthermore, the step 3 includes the following sub-steps:

[0026] Step 3.1: Build Figure 3 The attention-based fusion network F shown in the figure consists of a channel attention block A. Cand spatial attention block A S Composition, input perspective expanded multi-view feature F, output fusion feature f us ;

[0027] Step 3.2: Construct a classifier C, which consists of two fully connected layers. us Perform label prediction;

[0028] Step 3.3: Construct a cross entropy loss function for parameter optimization of the multi-view fusion module.

[0029] Furthermore, the step 4 includes the following sub-steps:

[0030] Step 4.1, input the full-view extremely narrow pulse data and labels of the source domain into the view feature expansion module to obtain the multi-view feature F;

[0031] Step 4.2: Input the multi-view feature F into the multi-view fusion module to obtain the predicted label

[0032] Step 4.3: Update the parameters of the feature extractor E, discriminator D and generator network G through adversarial loss optimization, and update the parameters of the generator network G through reconstruction loss optimization;

[0033] Step 4.4, update the feature extractor E through cross entropy loss, generate network G, feature fusion network F and classifier network C;

[0034] Step 4.5: Repeat steps 4.1, 4.2, 4.3, and 4.4 for pre-training until the network converges to obtain a generative multi-view recognition model that is well trained on the source domain dataset.

[0035] Furthermore, the step 5 includes the following sub-steps:

[0036] Step 5.1, input the sparse view extremely narrow pulse data and corresponding category labels of the target domain training set into the view feature expansion module to obtain the multi-view feature F;

[0037] Step 5.2: Input the multi-view feature F into the multi-view fusion module to obtain the predicted label

[0038] Step 5.3, update the feature extractor E through cross entropy loss optimization, generate network G and feature fusion network F;

[0039] Step 5.4: Repeat steps 5.1, 5.2, and 5.3 until the network converges to obtain a generative multi-view recognition model that is optimized and fine-tuned with sparse view data.

[0040] Furthermore, in step 2, the present invention introduces adversarial loss and reconstruction loss to guide the parameter update of the generation network G from the distribution level and pixel level respectively.

[0041] Furthermore, in step 2, the adversarial loss can be expressed as: L adv

[0042]

[0043] in For the generated network G, the optimization goal is to minimize the adversarial loss. For the discriminator network D, the optimization goal is to maximize the adversarial loss, and F represents the characteristics of the real multi-view extremely narrow pulse. represents F obeys p data The probability distribution of (F) and the mathematical expectation of the formula are calculated. Represents obedience fθ (f θ ) and find the mathematical expectation of the formula.

[0044] Furthermore, in step 2, the reconstruction loss can be expressed as: L recon

[0045] L recon =min‖FF‖1

[0046] Among them, min represents the minimum value, and ‖‖1 represents the L1 norm.

[0047] Furthermore, in step 3, the output fusion feature f us , the calculation process is as follows:

[0048]

[0049] in represents the channel-by-channel summation operation, c represents the number of channels, and N represents the number of multi-views generated.

[0050] Furthermore, in step 3, the cross entropy loss function is expressed as:

[0051]

[0052] Where M represents the number of categories, Indicates calculating the formula from 1 to M and accumulating, y i and represent the true label and the predicted label respectively.

[0053] Beneficial effects:

[0054] (1) Existing methods for solving sparse viewpoint recognition through generative models often separate the generative model from the recognition model, resulting in a cumbersome process and difficulty in bringing into play the powerful fitting ability of the neural network. The present invention cascades the generative model and the recognition model to achieve an end-to-end generation and recognition fusion structure, reducing the model complexity. At the same time, during the parameter optimization process, the generative model can be optimized according to the results of the recognition network, the optimization efficiency of the generative model is improved, the interpretability is enhanced, and finally the target recognition rate under sparse viewpoint is improved.

[0055] (2) The existing data enhancement method to solve the problem of low recognition rate of sparse view angles is mainly to generate extremely narrow pulse data through a generative network. This method is difficult and the generated data is often of low quality. The present invention uses a generative model to expand the features of extremely narrow pulses, which is easy to implement. By expanding the potential features of the data, multi-view features are obtained, which gives full play to the characteristics of the neural network. Multi-view joint decision-making reduces the distribution decision boundary deviation and significantly improves the recognition accuracy under sparse view angles.

[0056] (3) The present invention proposes a multi-domain transfer training strategy, assuming that the target exhibits similar perspective change patterns within a certain angle range. By considering the cooperative target as the source domain and the sparse perspective non-cooperative target as the target domain, the network can learn these patterns in the source domain, and then these learned rules are transferred to the target domain through parameters. This method can solve the problem of reduced generalization ability and overfitting of the model due to the small number of sparse perspective training sets. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 To generate-fusion cascade network model structure diagram;

[0058] Figure 2 It is the structure diagram of generator G;

[0059] Figure 3 It is the structural diagram of the fusion network F;

[0060] Figure 4 The flowchart of the radar target HRRP recognition method based on sparse view feature generation fusion;

[0061] Figure 5 To divide the target into source domain and target domain;

[0062] Figure 6 for training and testing scenarios. DETAILED DESCRIPTION

[0063] The present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0064] The present invention proposes a method for generating a multi-view extremely narrow pulse recognition model for sparse view target recognition. The overall structure of the model is as follows: Figure 1 As shown in Figure 2, the method mainly consists of two main modules: view feature expansion module (VFEM) and multi-view fusion module (MVFM). In order to solve the problem of limited training samples, the present invention proposes a multi-domain migration training strategy.

[0065] VFEM is responsible for extracting the features of extremely narrow pulses and mapping the sparse view extremely narrow pulse data to multi-view features. The one-dimensional VGG network is used as the backbone of feature extraction to map the input extremely narrow pulse data of azimuth angle θ to deep features. Figure 2 The feature generation network G shown learns the correlation information between the central features and the adjacent features, and outputs the prediction of multi-view features.

[0066] MVFM is composed of Figure 3 The attention-based fusion network F and the fully connected network classifier C are shown. F consists of a channel attention block A. C and spatial attention block A S Composition, using a dual attention mechanism operating on both the channel dimension and the spatial dimension, F dynamically assigns weights to the input feature maps and outputs fused features.

[0067] The multi-domain transfer training strategy assumes that the targets exhibit similar perspective change patterns within a certain angle range. By considering cooperative targets as the source domain and sparse perspective non-cooperative targets as the target domain, the model can learn these patterns in the source domain and transfer the learned rules to the target domain through parameters.

[0068] The technical idea for implementing the present invention is: first, obtain the extremely narrow pulse echo data of M types of full-view targets, and use these full-view data as the source domain training data set. Then obtain the extremely narrow pulse echo data of N types of targets at different pitch angles, and use the data of different pitch angles for the target domain training set and test set respectively, wherein the training set data is collected at a certain interval between all angles, and the test set contains samples of all azimuth angles. The source domain and target domain data sets are uniformly normalized to the maximum amplitude to obtain the final source domain and target domain data sets. Construct a view feature expansion module and a multi-view fusion module to form the generative multi-view extremely narrow pulse recognition model of the present invention. Input the samples and corresponding category labels in the source domain training data set into the network for pre-training. After the pre-training is completed, parameter migration is performed, and then the target domain training set is input into the network for training to obtain the final trained model. Finally, the target domain test set is used for testing to obtain the recognition rate.

[0069] The specific steps include:

[0070] Step 1: Construct an extremely narrow pulse echo dataset, divide it into source domain training set, target domain training set and test set, and perform normalization preprocessing:

[0071] 1.1. Obtain extremely narrow pulse echo data and corresponding category labels of M types of full-view targets as the training data set in the source domain.

[0072] 1.2. Obtain the extremely narrow pulse echo data and corresponding category labels of N types of full-view targets at different pitch angles, sample the azimuth angle of the data interval Δθ at a certain pitch angle to obtain the extremely narrow pulse echo data under sparse view and the corresponding category labels as the training set of the target domain, and use the all-round extremely narrow pulse data and corresponding category labels at another pitch angle as the test set of the target domain.

[0073] 1.3. Perform amplitude maximum normalization on all the data sets in the above source domain and target domain.

[0074] The maximum amplitude normalization uses the maximum value of each frame of the extremely narrow pulse echo to normalize the range image. Let x be a frame of extremely narrow pulse echo extracted from the data set, then the extremely narrow pulse echo after amplitude normalization is recorded as It can be expressed as:

[0075]

[0076] Among them, max means taking the maximum value, and |x| means taking the absolute value of x.

[0077] Step 2: Construct a view feature extension module (VFEM), including a feature extraction network E, a feature generation network G, and a feature discriminator D, and construct related adversarial losses and reconstruction losses:

[0078] 2.1. Construct a one-dimensional (1D) VGG network as a feature extractor E, which maps the input extremely narrow pulse data with an azimuth angle of θ into deep features f θ .

[0079] 2.2. Construction Figure 2 The feature generation network G shown is based on a one-dimensional U-Net architecture and consists of three layers of upsampling and three layers of downsampling as well as an intermediate layer:

[0080] In the upsampling layer, the basic operation is convolution, followed by ReLU activation and batch normalization, denoted as CRB. In the intermediate layer, a combination of convolution, ReLU activation, convolution, and batch normalization is used to further increase the network depth, denoted as CRCB. In the downsampling layer, the basic operation is deconvolution, followed by ReLU activation and batch normalization, denoted as DCRB. In addition, residual connections are used to connect the upsampling and downsampling layers to alleviate gradient vanishing. In this way, multi-scale representations of the input features can be gradually extracted and combined, ensuring that the network has sufficient generation capacity.

[0081] Input single view feature f θ After generating the feature network G, the expanded feature F is obtained.

[0082] 2.3. Construct the discriminator network D and loss function:

[0083] The discriminator network D consists of three convolutional layers, which is used to discriminate the authenticity of the input expanded extremely narrow pulse multi-view features. In terms of loss function, the present invention introduces adversarial loss and reconstruction loss to guide the parameter update of the generation network G from the distribution level and pixel level respectively.

[0084] The adversarial loss can be expressed as: L adv

[0085]

[0086] in For the generated network G, the optimization goal is to minimize the adversarial loss. For the discriminator network D, the optimization goal is to maximize the adversarial loss, and F represents the characteristics of the real multi-view extremely narrow pulse. represents F obeys p data The probability distribution of (F) and the mathematical expectation of the formula are calculated. Represents obedience fθ (f θ ) and find the mathematical expectation of the formula.

[0087] The discriminant network D and the generative network G compete with each other, update parameters, and finally reach a Nash equilibrium.

[0088] The reconstruction loss can be expressed as: L recon

[0089] L recon =min‖FF‖1

[0090] Among them, min represents the minimum value, and ‖‖1 represents the L1 norm.

[0091] The reconstruction loss updates the parameters of the generative network G by minimizing the L1 norm between the generated features and the true features.

[0092] Step 3: Construct a multi-view fusion module (MVFM), including the fusion network F based on the attention mechanism and the fully connected network classifier C, and construct the relevant classification loss:

[0093] 3.1. Construction Figure 3 The attention-based fusion network F shown in the figure consists of a channel attention block A. C and spatial attention block A S Composition, input perspective expanded multi-view feature F, output fusion feature f us , the calculation process is as follows:

[0094]

[0095] in represents the channel-by-channel summation operation, c represents the number of channels, and N represents the number of multi-views generated.

[0096] 3.2. Construct the classifier C, which consists of two fully connected layers. According to the fusion feature f output by the fusion network us Perform label prediction.

[0097] 3.3. Construct a cross entropy loss function for parameter optimization of the multi-view fusion module, expressed as:

[0098]

[0099] Where M represents the number of categories, Indicates calculating the formula from 1 to M and accumulating, y i and represent the true label and the predicted label respectively.

[0100] Step 4: Input the extremely narrow pulse echoes of all viewing angles in the source domain and their corresponding labels into the viewing feature expansion module. The expanded features enter the multi-view fusion module. The entire network is pre-trained through the corresponding loss function to complete parameter optimization:

[0101] 4.1. Input the full-view extremely narrow pulse data and labels of the source domain into the view feature expansion module to obtain the multi-view feature F.

[0102] 4.2. Input the multi-view feature F into the multi-view fusion module to obtain the predicted label

[0103] 4.3. Update the parameters of the feature extractor E, discriminator D and generative network G through adversarial loss optimization, and update the parameters of the generative network G through reconstruction loss optimization.

[0104] 4.4. Update the feature extractor E, generate network G, feature fusion network F and classifier network C through cross entropy loss.

[0105] 4.5. Repeat steps 4.1, 4.2, 4.3, and 4.4 for pre-training until the network converges and obtains a generative multi-view recognition model that is well trained on the source domain dataset.

[0106] Step 5: Input the extremely narrow pulse echoes and their corresponding labels in the training set of the target domain into the view feature expansion module of the pre-trained network, and perform training optimization of the entire network again to complete the training of the model:

[0107] 5.1. Input the sparse viewpoint extremely narrow pulse data and corresponding category labels of the target domain training set into the viewpoint feature expansion module to obtain the multi-viewpoint feature F.

[0108] 5.2. Input the multi-view feature F into the multi-view fusion module to obtain the predicted label

[0109] 5.3. Update the feature extractor E, generate the network G and the feature fusion network F through cross entropy loss optimization.

[0110] 5.4. Repeat steps 501, 502, and 503 until the network converges, and obtain a generative multi-view recognition model that is optimized and fine-tuned using sparse view data.

[0111] Step 6: Input the extremely narrow pulse echoes and their corresponding labels in the target domain test set into the trained network. After view feature expansion and feature fusion, the recognition rate of the model for the target domain test set is finally obtained.

[0112] Example

[0113] This embodiment is combined with the measured data experiment to illustrate the effect of the present invention. In order to evaluate the performance of the proposed generative multi-view ultra-narrow pulse recognition model, an experiment was conducted using ultra-narrow pulse data obtained by inverting SAR images of the radar field public dataset MSTAR.

[0114] Dataset and parameter settings:

[0115] The MSTAR dataset contains synthetic aperture radar images of ten types of military targets. The radar used by MSTAR is a high-resolution spotlight synthetic aperture radar with a resolution of 0.3×0.3m, operating in the X-band, and using the HH polarization mode. By inverting the synthetic aperture radar images of the MSTAR dataset, the corresponding extremely narrow pulse radar echo data can be obtained.

[0116] The present invention uses extremely narrow pulses inverted by MSTAR to conduct experiments, and specifies 7 categories in MSTAR (BTR-60, 2S1, BRDM-2, D7, T-62, ZIL-131 and ZSU-23 / 4) as source domains for pre-training network parameters, and the remaining three categories (BMP2, BTR-70, T-72) are specified as non-cooperative targets to be recognized in the target domain, such as Figure 5 shown.

[0117] In practical applications, the data of non-cooperative targets can generally only obtain sparse perspective data, but in recognition, non-cooperative targets need to be identified from all angles. For the three types of non-cooperative targets BMP2, BTR-70, and T-72, the present invention selects all the data at a pitch angle of 15° as the test set of the target domain. For the data at a pitch angle of 17°, the present invention selects three different intervals of 10°, 20°, and 30° to form different sparse training sets of the target domain, as follows:

[0118] Table 1 Dataset details

[0119]

[0120] The source domain data is input into the network model for pre-training. After parameter migration, the target domain training set is input into the network model and trained again. Finally, the trained model is used to identify the test set of the target domain to obtain the recognition result.

[0121] Experimental results:

[0122] The following table shows the missing view recognition results of the present invention on the MSTAR dataset. By observation, the recognition results of the generative multi-view recognition model (GMVR) of the present invention show that compared with the baseline recognition method, the few-sample learning method (including the transfer learning method and the MAML model-independent meta-learning method), the classic method in the field of deep generative model DGM (including DCGAN, ReconAware and CVAE-GAN), and the data enhancement upper limit (DAUB, directly using real samples adjacent to the sparse view for data enhancement) method, the method of the present invention achieves the best recognition performance under three different sparse angle intervals, and even the performance at 20° and 30° pitch angles exceeds the data enhancement upper limit. The recognition results show an increasing trend with the number of multiple views. This reflects the effectiveness of the generated features and multi-view fusion algorithm in enhancing the recognition performance in sparse view scenes.

[0123] Table 2 Recognition results of different methods

[0124]

[0125] In order to explore the contribution of each module in the model proposed in the present invention, an ablation experiment was carried out. By comparing the recognition rates when different modules were removed, the present invention analyzed the importance of each module. Taking the generation of neighboring five-view features by the model as an example, the detailed results are shown in the table. Obviously, each module contributes to the recognition performance. The VFEM module effectively expands the view feature information of sparse view data, while the MVFM module adaptively fuses these features. When the two modules are integrated, the best recognition performance can be achieved.

[0126] Table 3 Ablation experiment results

[0127]

[0128] In summary, the above are only preferred embodiments of the present invention and are not intended to limit the protection scope of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A radar target HRRP recognition method based on sparse viewing feature generation and fusion, characterized in that: The steps include: Step 1: Construct an extremely narrow pulse echo dataset, divide it into a source domain training set, a target domain training set, and a test set, and perform normalization preprocessing; Step 2: Construct a view feature extension module (VFEM), including a feature extraction network E, a feature generation network G, and a feature discriminator D, and construct related adversarial losses and reconstruction losses; Step 3: Construct a multi-view fusion module (MVFM), including the fusion network F based on the attention mechanism and the fully connected network classifier C, and construct the relevant classification loss; Step 4: Input the extremely narrow pulse echoes of all viewing angles in the source domain and their corresponding labels into the viewing feature expansion module. The expanded features enter the multi-view fusion module. The entire network is pre-trained through the corresponding loss function to complete parameter optimization. Step 5: Input the extremely narrow pulse echoes and their corresponding labels in the training set of the target domain into the view feature expansion module of the pre-trained network, and perform training optimization of the entire network again to complete the training of the model; Step 6: Input the extremely narrow pulse echoes and their corresponding labels in the target domain test set into the trained network. After view feature expansion and feature fusion, the recognition rate of the model for the target domain test set is finally obtained.

2. The method according to claim 1, characterized in that The step 1 includes the following sub-steps: Step 1.1, obtain extremely narrow pulse echo data and corresponding category labels of M types of full-view targets as training data sets in the source domain; Step 1.2, obtain the extremely narrow pulse echo data and corresponding category labels of N types of full-view targets at different pitch angles, sample the azimuth angle of the data interval Δθ at a certain pitch angle to obtain the extremely narrow pulse echo data under sparse view angle and the corresponding category labels as the training set of the target domain, and use the all-round extremely narrow pulse data and corresponding category labels at another pitch angle as the test set of the target domain; Step 1.3: Perform maximum amplitude normalization on all data sets in the above source domain and target domain.

3. The method according to claim 1, characterized in that The step 2 includes the following sub-steps: Step 2.1: Construct a one-dimensional (1D) VGG network as a feature extractor E, which maps the input extremely narrow pulse data with an azimuth angle of θ into deep features f θ ; Step 2.2, construct the feature generation network G shown in Figure 2, G is based on a one-dimensional U-Net architecture, consisting of three layers of upsampling, three layers of downsampling and an intermediate layer; Step 2.3: Construct the discriminator network D and loss function.

4. The method according to claim 1, characterized in that The step 3 includes the following sub-steps: Step 3.1: Construct the attention-based fusion network F shown in Figure 3. The fusion network F consists of the channel attention block A C and spatial attention block A S Composition, input perspective expanded multi-view feature F, output fusion feature f us ; Step 3.2: Construct a classifier C, which consists of two fully connected layers. us Perform label prediction; Step 3.3: Construct a cross entropy loss function for parameter optimization of the multi-view fusion module.

5. The method according to claim 1, characterized in that The step 4 includes the following sub-steps: Step 4.1, input the full-view extremely narrow pulse data and labels of the source domain into the view feature expansion module to obtain the multi-view feature F; Step 4.2: Input the multi-view feature F into the multi-view fusion module to obtain the predicted label Step 4.3: Update the parameters of the feature extractor E, discriminator D and generator network G through adversarial loss optimization, and update the parameters of the generator network G through reconstruction loss optimization; Step 4.4, update the feature extractor E through cross entropy loss, generate network G, feature fusion network F and classifier network C; Step 4.5: Repeat steps 4.1, 4.2, 4.3, and 4.4 for pre-training until the network converges to obtain a generative multi-view recognition model that is well trained on the source domain dataset.

6. The method according to claim 1, characterized in that The step 5 includes the following sub-steps: Step 5.1, input the sparse view extremely narrow pulse data and corresponding category labels of the target domain training set into the view feature expansion module to obtain the multi-view feature F; Step 5.2: Input the multi-view feature F into the multi-view fusion module to obtain the predicted label Step 5.3, update the feature extractor E through cross entropy loss optimization, generate network G and feature fusion network F; Step 5.4: Repeat steps 5.1, 5.2, and 5.3 until the network converges to obtain a generative multi-view recognition model that is optimized and fine-tuned with sparse view data.

7. The method according to claim 1, characterized in that In step 2, the present invention introduces adversarial loss and reconstruction loss to guide the parameter update of the generation network G from the distribution level and pixel level respectively.

8. The method according to claim 1, characterized in that In step 2, the adversarial loss can be expressed as: L adv in For the generated network G, the optimization goal is to minimize the adversarial loss. For the discriminator network D, the optimization goal is to maximize the adversarial loss, and F represents the characteristics of the real multi-view extremely narrow pulse. represents F obeys p data The probability distribution of (F) and the mathematical expectation of the formula are calculated. Represents obedience fθ (f θ ) and find the mathematical expectation of the formula.

9. The method according to claim 1, characterized in that In step 2, the reconstruction loss can be expressed as: L recon L recon =min‖F-F‖1 Among them, min represents the minimum value, and ‖‖1 represents the L1 norm.

10. The method according to claim 1, characterized in that In step 3, the output fusion feature f us , the calculation process is as follows: in represents the channel-by-channel summation operation, c represents the number of channels, and N represents the number of multi-views generated.

11. The method according to claim 1, characterized in that In step 3, the cross entropy loss function is expressed as: Where M represents the number of categories, Indicates calculating the formula from 1 to M and accumulating, y i and represent the true label and the predicted label respectively.

Images