A method for generating multi-pose extremely narrow pulse echoes based on generative adversarial networks

By building a GAN with multi-pose extremely narrow pulse echo data generation capability, using attention decomposition and viewing angle embedding modules to generate extremely narrow pulse echoes of specified viewing angles, it solves the problem of generating clear viewing angles and posture sensitivity in the prior art. The problem of this has achieved the improvement of data generation efficiency in the study of electromagnetic characteristics of radar targets.

CN115308705BActive Publication Date: 2024-12-03BEIJING INST OF TECH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202210940267.3
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-08-05
Publication Date
2024-12-03
Estimated Expiration
2042-08-05

AI Technical Summary

Technical Problem

Existing GAN-based extremely narrow pulse echo generation methods cannot generate target extremely narrow pulse echoes with clear viewing angles, and the distribution fitting of input data cannot overcome the attitude sensitivity problem of extremely narrow pulse echoes.

Method used

By constructing a GAN with multi-pose extremely narrow pulse echo data generation capability, the factorization module of attention decouples the identity and attitude characteristics of the input extremely narrow pulse echo, and reorganizes the characteristics through the decoder module embedded in the continuous viewing angle , generates extremely narrow pulse echoes of the specified viewing angle.

Benefits of technology

It realizes the rapid acquisition of extremely narrow pulse echo data of unknown viewing angles, reduces the demand for actual measurement acquisition and electromagnetic simulation, and improves the data generation efficiency in the research on electromagnetic characteristics of radar targets and automatic identification.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN115308705B_ABST
    Figure CN115308705B_ABST
Patent Text Reader

Abstract

The present invention relates to the technical field of ultra-narrow pulse radar data generation, and in particular to a method for generating multi-attitude ultra-narrow pulse echoes based on a generative adversarial network (GAN). The method generates ultra-narrow pulse echoes of missing viewing angles of targets by constructing a GAN with the ability to generate multi-attitude ultra-narrow pulse echo data. This method can replace actual measurement acquisition, electromagnetic simulation and other methods to a certain extent to achieve rapid acquisition of ultra-narrow pulse echo data of unknown viewing angles, and solve the problem of ultra-narrow pulse echo attitude sensitivity faced in radar target electromagnetic characteristics and RATR research.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of ultra-narrow pulse radar data generation, and in particular to a method for generating multi-pose ultra-narrow pulse echoes based on a generative adversarial network (GAN). Background Art

[0002] Ultra-narrow pulse radar refers to a type of radar in which the width of a single echo pulse is much smaller than the target size after processing. For ultra-narrow pulse radar, the echo of the target contains multiple ultra-narrow pulses, which correspond to different scattering points on the target. Therefore, the target ultra-narrow pulse echo can represent the distribution of the target scattering points along the radar line of sight, which is also commonly referred to as the high-resolution range profile (HRRP) of the target.

[0003] Since the ultra-narrow pulse echo carries rich structural and geometric feature information of the target, it plays an important role in the research of electromagnetic characteristics of radar targets and radar automatic target recognition (RATR). Ultra-narrow pulse echoes can usually be obtained through actual measurement acquisition and electromagnetic simulation. However, the method of obtaining ultra-narrow pulse echoes through actual measurement acquisition requires a lot of manpower and material resources, and the method of obtaining ultra-narrow pulse echoes through electromagnetic simulation requires a high-precision three-dimensional model of the target and a lot of hardware resources and time resources, which are usually difficult to meet in practical applications. With the development of machine learning technology, relevant researchers have proposed a deep generative model based on GAN, which can quickly generate realistic data samples with low computational complexity and has been initially applied in the field of ultra-narrow pulse radar data generation.

[0004] However, the extremely narrow pulse echo data has high attitude sensitivity. When the relative attitude between the radar and the target changes slightly, the shape of the extremely narrow pulse echo will change significantly, which will have an adverse impact on the research of radar target electromagnetic characteristics and RATR. Obtaining the extremely narrow pulse echo of unknown perspective to complete the data set perspective is an effective way to overcome the attitude sensitivity problem of extremely narrow pulse echo. However, the existing GAN-based extremely narrow pulse echo generation method only performs distribution fitting on the input raw data and cannot generate extremely narrow pulse echoes of targets with clear perspectives. Summary of the invention

[0005] In view of the above problems, the present invention proposes a multi-posture ultra-narrow pulse echo generation method based on generative adversarial network. The method generates the ultra-narrow pulse echo of the target with missing perspective by constructing a GAN with the ability to generate multi-posture ultra-narrow pulse echo data. This method can replace the actual measurement acquisition, electromagnetic simulation and other methods to a certain extent to achieve the rapid acquisition of ultra-narrow pulse echo data with unknown perspective, and solve the problem of ultra-narrow pulse echo posture sensitivity faced in the electromagnetic characteristics of radar targets and RATR research.

[0006] The method of the present invention is achieved through the following technical solutions:

[0007] A method for generating multi-pose extremely narrow pulse echoes based on a generative adversarial network, wherein the method firstly realizes the decoupled representation of the input extremely narrow pulse echo based on an attention factorization module, and decomposes its identity feature and pose feature; then, based on a decoder module embedded with continuous viewpoints, the viewpoint label to be generated is converted into a new pose feature, and the pose feature is recombined with the previously decomposed identity feature, thereby realizing the generation of an extremely narrow pulse echo of a specified viewpoint; the steps of the method include:

[0008] Step 1, constructing an extremely narrow pulse echo data set, dividing the constructed extremely narrow pulse echo data set into a training data set and a data set to be completed, and performing normalization preprocessing on the divided training data set and the data set to be completed, respectively, to obtain a normalized preprocessed training data set and a normalized preprocessed data set to be completed;

[0009] Step 2, constructing a GAN generator network G, the constructed GAN generator network G includes an attention-based factorization module F and a continuous view embedding-based decoder module R;

[0010] Step 3: Construct the GAN discriminator network D and adversarial training loss;

[0011] Step 4, using the extremely narrow pulse echo data and the viewpoint label in the normalized preprocessed training data set in step 1 as the input of the GAN discriminator network D constructed in step 3, and optimizing the parameters of the GAN discriminator network D constructed in step 3 through the adversarial training loss constructed in step 3;

[0012] Step 5, using the extremely narrow pulse echo data and the viewpoint label in the normalized preprocessed training data set in step 1 as the input of the GAN generator network G constructed in step 2, and optimizing the parameters of the GAN generator network G constructed in step 2 through the GAN discriminator network D after parameter optimization in step 4 and the adversarial training loss constructed in step 3;

[0013] Step 6: The GAN generator network G with optimized parameters obtained in step 5 completes the missing perspectives of the dataset to be completed, and completes the generation of multi-pose ultra-narrow pulse echoes based on the generative adversarial network.

[0014] In step 1, the content of the extremely narrow pulse echo data set includes: Θ s ~Θ e The extremely narrow pulse echo data of the C-type target within the viewing angle range and the corresponding viewing angle label, where Θ s Represents the starting angle of the viewing range, Θ e Represents the end view angle of the view range.

[0015] In step 1, the method of dividing the constructed extremely narrow pulse echo data set into a training data set and a data set to be completed is:

[0016] The data angle labels of the extremely narrow pulse echo are counted. If the data angle label is a uniformly spaced full angle, N types of uniformly spaced Δθ full angle extremely narrow pulse echoes are obtained, which are recorded as the training data set X; otherwise, CN types of angles are obtained as θ1, θ2, ..., θ m The extremely narrow pulse echo of some viewing angles is recorded as the data set to be completed

[0017] In the step 1, the normalization preprocessing is performed by using the maximum amplitude normalization processing method, specifically:

[0018] The maximum amplitude normalization uses the maximum value of each frame of 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 is expressed as:

[0019] In step 2, the attention-based factorization module includes a feature extractor F, a feature decomposer T and a supervisory constraint network C. The attention-based factorization module is used to extract the identity feature f of the input extremely narrow pulse echo data x. i ,Right now:

[0020] f i =F(x)=T(F(x))

[0021] The extremely narrow pulse echo data x is a subset sampled from the training data set X;

[0022] The feature extractor F consists of a 4-layer convolutional neural network, which is used to extract the features of the extremely narrow pulse echo data x and output the mixed feature representation f mix ,Right now:

[0023] f mix=F(x)

[0024] The feature decomposer T consists of a 2-layer self-attention network and a 1-layer weighted decomposition network, and its input is the mixed feature representation f extracted by the feature extractor. mix , the output is the viewing feature f a and identity characteristics f i , where the self-attention network is used to obtain the weight factor λ of the mixed feature; the weighted decomposition network is implemented by f according to λ mix The decomposition of is:

[0025] [f a ,f i ]=T(f mix )

[0026]

[0027]

[0028] in, Represents matrix dot multiplication operation, 1 is a matrix of all 1s;

[0029] The supervised constraint network C is composed of the perspective fully connected regression network C a and identity fully connected classification network C i Composition, supervision constraint network for f mix The decomposition process is supervised and constrained;

[0030] The perspective fully connected regression network C a The input is f a , the output is the view regression result C a (f a ), the mean square error loss function is used for supervised constraint, and the perspective regression loss L is obtained a as follows:

[0031] L a =MSE(C a (f a )-y a )

[0032] Among them, y a is the true viewpoint label of the input extremely narrow pulse echo data, and MSE stands for the mean square error loss function;

[0033] Identity fully connected classification network C i Input is f i , the output is the identity classification result C i (f i ), the cross entropy loss function is used for supervised constraints to obtain the identity classification loss L i as follows:

[0034] L i =CE(C i (f i ),y i )

[0035] Among them, yi is the true identity label of the input extremely narrow pulse echo data, and CE represents the cross entropy loss function;

[0036] In step 2, the decoder module R based on continuous view embedding includes a view embedding representation E and a view reconstruction decoder R. The decoder module R based on continuous view embedding is based on the identity feature f i and the input perspective y to be generated a 'Reconstruct the extremely narrow pulse echo, denoted as x', that is:

[0037] x'=R(f i ,y a ')

[0038] The view embedding representation E consists of a 1-layer fully connected network to generate the view label y a 'To the viewing angle feature f a ' is embedded in:

[0039] f a '=E(y a ')

[0040] The view reconstruction decoder R consists of a 4-layer deconvolutional neural network to realize the generated view feature f a ' and identity characteristics f i The reorganization output is a reconstructed extremely narrow pulse echo, that is:

[0041] x'=R(f i ,f a ')=R(f i ,E(y a '));

[0042] In the step 3, the GAN discriminator network D is a discriminator composed of a 4-layer convolutional neural network, which is used to discriminate the authenticity of the input extremely narrow pulse echo;

[0043] The method to construct the adversarial training loss is:

[0044] (1) The generator training loss is constructed through the discriminator. The discriminator is responsible for distinguishing the extremely narrow pulse echo data x' generated by the generator as a "true" sample, that is, the output prediction label is True. The cross entropy loss function is used for supervised constraints to obtain the generator training loss L Gadv :

[0045] L Gadv=CE(D(x'),True)

[0046] (2) The discriminator training loss is constructed through the discriminator. When training the discriminator, the discriminator is responsible for distinguishing the extremely narrow pulse echo data x' generated by the generator as a "false" sample, that is, the output prediction label is False; and distinguishing the extremely narrow pulse echo data x sampled from the training data as a "true" sample, that is, the output prediction label is True. The cross entropy loss function is used for supervised constraints to obtain the discriminator training loss L Dadv :

[0047] L Dadv =CE(D(x'),False)+CE(D(x),True);

[0048] In step 4, the method for parameter optimization is: fix the parameters of the generator network G, train the discriminator, and use the discriminator to train the loss L Dadv , optimize the parameters of the discriminator network D:

[0049] L D =λ Dadv L Dadv

[0050] In step 5, the method for parameter optimization is:

[0051] Fix the parameters of the discriminator network D, train the generator, and use the perspective regression loss L a , identity classification loss L i And the generator training loss L Gadv , optimize the parameters of the generator network G, and the loss function used for generator training is:

[0052] L G =λ a L a +λ i L i +λ Gadv L Gadv

[0053] Perform cyclic parameter optimization until the network converges.

[0054] In step 6, the method for completing the missing perspectives of the data set to be completed is:

[0055] The existing data and full-angle labels of the dataset to be generated, namely Θ s ~Θ e The view labels are evenly spaced Δθ within the view range and input into the GAN generator network G trained in step 5. The output samples of the continuous view embedding encoder module are saved. This sample is the view completion sample.

[0056] Compared with the prior art, the method of the present invention has the following advantages:

[0057] (1) The existing GAN-based ultra-narrow pulse echo generation method cannot generate target ultra-narrow pulse echoes with a clear perspective. The present invention proposes a target ultra-narrow pulse echo generation network framework, which realizes the generation of multi-pose target ultra-narrow pulse echoes by reconstructing the perspective of the input ultra-narrow pulse echo.

[0058] (2) The existing GAN-based method for generating extremely narrow pulse echoes only performs distribution fitting on the input raw data, and there is a large amount of redundant information in the input data. The present invention utilizes an attention-based factorization module to realize the decoupled representation of the input extremely narrow pulse echoes, separates their identity features and posture features, realizes information refinement of the input data at the feature level, and improves the representation capability of the network.

[0059] (3) The present invention utilizes the “decomposition-recombination” method to realize the perspective reconstruction of extremely narrow pulse echoes. Based on the decoder module embedded with continuous perspective, the perspective label to be generated is converted into a new posture feature, which is then recombined with the identity feature decomposed from the input extremely narrow pulse echo, thereby realizing the generation of extremely narrow pulse echoes of multi-posture targets. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 It is the structure diagram of the generator network G;

[0061] Figure 2 It is the structure diagram of the discriminator network D;

[0062] Figure 3 The missing perspective data completion results for the GTRI dataset;

[0063] Figure 4 The result of feature evaluation. DETAILED DESCRIPTION

[0064] The present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0065] The present invention proposes a multi-view ultra-narrow pulse echo generation method based on GAN. GAN is a generative network model. Its basic idea is derived from the two-person zero-sum game in game theory. It is usually composed of a generator and a discriminator. It is trained by adversarial learning. The purpose is to estimate the potential distribution of data samples and generate new data samples that conform to the distribution. GAN has outstanding sample generation capabilities. It learns the sample distribution of the original data set through the fitting ability of deep neural networks and can generate samples with the same distribution as the original data set.

[0066] For the generation of multi-posture ultra-narrow pulse echoes, the present invention relies on the distribution fitting ability of GAN and designs the GAN network structure to implicitly model the changing law of multi-posture ultra-narrow pulse echoes, and realizes the generation of clear perspective targets through perspective reconstruction.

[0067] The technical solution for implementing the present invention is: firstly, the extremely narrow pulse echo data of the Class C target is obtained, the full-view data is divided into a training data set, the partial-view data is divided into a data set to be completed, and the maximum amplitude normalization is performed. Then, a designed multi-posture extremely narrow pulse echo generation GAN network is constructed, and the training data set samples and view labels are input into the network for training. Finally, the full-angle labels and the data in the data set to be completed are input into the trained GAN network to realize the generation of missing view data.

[0068] The specific steps include:

[0069] Step 1: construct an extremely narrow pulse echo data set, divide it into a training data set and a data set to be completed, and perform normalization preprocessing.

[0070] 101. Get Θ s ~Θ e The extremely narrow pulse echo data of Class C targets within the viewing angle range and the corresponding viewing angle labels.

[0071] 102. Count the angle labels of the extremely narrow pulse echo data to obtain N types of full-angle extremely narrow pulse echoes with uniform intervals of Δθ, which are recorded as training data set X.

[0072] 103. Count the remaining extremely narrow pulse echo data angle labels and get the CN angles as θ1, θ2, ..., θ m The extremely narrow pulse echo of some viewing angles is recorded as the data set to be completed

[0073] 104. Perform maximum amplitude normalization processing on the above training data set and the data set to be completed.

[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] Step 2: Construct the GAN generator network G, such as Figure 1 As shown, the GAN generator network includes an attention-based factorization module F and a continuous-view embedding-based decoder module R.

[0077] 201. Construct an attention-based factorization module F, including a feature extractor F, a feature decomposer T, and a supervision constraint network C, to extract the identity feature f of the input extremely narrow pulse echo data x i , that is: f i =F(x)=T(F(x))

[0078] The feature extractor F consists of a 4-layer convolutional neural network, which is used to extract the features of the extremely narrow pulse echo data x and can output a mixed feature representation f mix ,Right now:

[0079] f mix =F(x)

[0080] The feature decomposer T consists of a 2-layer self-attention network and a 1-layer weighted decomposition network, and its input is the mixed feature representation f extracted by the feature extractor. mix , the output is the viewing feature f a and identity characteristics f i The self-attention network is used to obtain the weight factor λ of the mixed feature; the weighted decomposition network is implemented by f according to λ mix The decomposition of is:

[0081] [f a ,f i ]=T(f mix )

[0082]

[0083]

[0084] in, Represents a matrix dot product operation, 1 is a matrix of all 1s.

[0085] The supervised constraint network C is composed of the perspective fully connected regression network C a and identity fully connected classification network C i Composition, for mix The decomposition process is supervised. The input of the perspective fully connected regression network is f a , the output is the view regression result C a (f a ), the mean square error loss function is used for supervised constraint, and the perspective regression loss L is obtained a as follows:

[0086] L a =MSE(C a (f a )-y a )

[0087] Among them, y aMSE is the true viewpoint label of the input extremely narrow pulse echo data, and MSE stands for the mean square error loss function.

[0088] The identity fully connected classification network input is f i , the output is the identity classification result C i (f i ), the cross entropy loss function is used for supervised constraints to obtain the identity classification loss L i as follows:

[0089] L i =CE(C i (f i ),y i )

[0090] Where yi is the true identity label of the input extremely narrow pulse echo data, and CE represents the cross entropy loss function. 202. Construct a decoder module R based on continuous view embedding, including a view embedding representation E and a view reconstruction decoder R. According to the identity feature f i and the input perspective y to be generated a 'Reconstruct the extremely narrow pulse echo, denoted as x', that is:

[0091] x'=R(f i ,y a ')

[0092] The view embedding representation E consists of a 1-layer fully connected network to realize the generated view label y a 'To the viewing angle feature f a ' is embedded in:

[0093] f a '=E(y a ')

[0094] The view reconstruction decoder R consists of a 4-layer deconvolutional neural network to realize the generated view feature f a ' and identity characteristics f i The reorganization output is a reconstructed extremely narrow pulse echo, that is:

[0095] x'=R(f i ,f a ')=R(f i ,E(y a '))

[0096] Step 3: Construct the GAN discriminator network D and adversarial training loss.

[0097] 301. Construct a discriminator consisting of a 4-layer convolutional neural network, such as Figure 2 As shown, it is used to determine the authenticity of the input extremely narrow pulse echo.

[0098] 302. Construct the generator training loss through the discriminator. When training the generator, the discriminator is responsible for distinguishing the extremely narrow pulse echo data x' generated by the generator as a "true" sample, that is, the output prediction label is True. Therefore, the cross entropy loss function is used for supervised constraints to obtain the generator training loss L Gadv :L Gadv =CE(D(x'),True)

[0099] 303. Construct the discriminator training loss through the discriminator. When training the discriminator, the discriminator is responsible for distinguishing the extremely narrow pulse echo data x' generated by the generator as a "false" sample, that is, the output prediction label is False; and distinguishing the extremely narrow pulse echo data x sampled from the real data as a "true" sample, that is, the output prediction label is True. Therefore, the cross entropy loss function is used for supervised constraints to obtain the discriminator training loss L Dadv :L Dadv =CE(D(x'),False)+CE(D(x),True)

[0100] Step 4: Use the extremely narrow pulse echo data and viewpoint labels in the normalized preprocessed training data set in step 1 as the input of the GAN discriminator network D constructed in step 3, and optimize the parameters of the GAN discriminator network D constructed in step 3 through the adversarial training loss constructed in step 3.

[0101] 401. Fix the parameters of the generator network G, train the discriminator, and use the discriminator to train the loss L Dadv , optimize the parameters of the discriminator network D:

[0102] L D =λ Dadv L Dadv

[0103] Step 5: Use the extremely narrow pulse echo data and viewpoint labels in the normalized preprocessed training data set in step 1 as the input of the GAN generator network G constructed in step 2, and optimize the parameters of the GAN generator network G constructed in step 2 through the GAN discriminator network D after parameter optimization in step 4 and the adversarial training loss constructed in step 3.

[0104] 501. Fix the parameters of the discriminator network D, train the generator, and use the perspective regression loss L a , identity classification loss L i And the generator training loss L Gadv , optimize the parameters of the generator network G, so the loss function used for generator training is:

[0105] L G =λ a La +λ i L i +λ Gadv L Gadv

[0106] 502. Return to step 401 and perform the next step of training until the network converges.

[0107] Step 6: Generate a GAN network based on the multi-pose ultra-narrow pulse echo with optimized parameters to complete the missing perspectives of the non-cooperative target data.

[0108] 601. The existing data and full-angle labels (i.e. Θ s ~Θ e The view labels are uniformly spaced Δθ within the view range and are input into the GAN generator network G trained in step 5.

[0109] 602. Save the output sample of the continuous view embedding encoder module, which is the view completion sample.

[0110] Example

[0111] This section describes the effect of the present invention in combination with the measured data experiment. In order to evaluate the performance of the proposed generative adversarial network, experiments were conducted using the radar field public dataset GTRI.

[0112] Dataset and parameter settings:

[0113] The GTRI dataset is a dataset of extremely narrow pulse echoes of T-72 tank targets measured by X-band radar. The radar center frequency is 9.6GHz and the synthesis bandwidth is 600MHz. Due to the limitation of the dataset, the extremely narrow pulse echoes at 10 different elevation angles are considered as 10 categories. The data of category 1 is recorded as the dataset to be generated, and the data of categories 2-10 are recorded as the training dataset. The angle interval of the training dataset is Δθ = 0.2°, and the viewing angle range is Θ s ~Θ e The angle of view of the dataset to be generated is 0.2°, 2°, and 4°. The details of the dataset are as follows:

[0114] Table 1 Dataset details

[0115]

[0116] The training data set set above is input into the GAN network for parameter optimization, and then the trained GAN is used to complete the missing perspectives of the generated data set.

[0117] Experimental results:

[0118] Figure 3The missing perspective data completion results of the present invention on the measured GTRI dataset are given. Through observation, the results show that compared with traditional interpolation methods and classic methods in the GAN field (including DRGAN method and PeaceGAN method), the present invention can generate sample data at all angles that is closer to the actual extremely narrow pulse echo data.

[0119] Figure 4 The similarity evaluation results of the scattering point distribution of the generated samples on the GTRI dataset are given. The KL divergence is used to evaluate the similarity between the scattering point distribution of the generated samples and the real samples. The smaller the KL divergence value, the closer the generated sample characteristics are to the real extremely narrow pulse echo characteristics. Therefore, the feature evaluation results show that compared with traditional interpolation methods and classic methods in the GAN field (including DRGAN method and PeaceGAN method), the extremely narrow pulse echo generated by the present invention is closest to the feature distribution of the real extremely narrow pulse echo data.

[0120] In order to explore the application prospects of the present invention, a data enhancement comparison experiment was carried out to verify the improvement of recognition performance when the present invention enhances the missing viewpoint data of the generated data set compared with the traditional interpolation method. The experimental settings are shown in the following table:

[0121] Table 2 Data augmentation experiment settings

[0122]

[0123]

[0124] In the basic experiment, the perspective of category 1 in the training set is set to be missing, the perspectives of categories 2-10 are set to be complete, and the perspectives of all targets in the test set are all in a complete state, which is used to verify the accuracy and recall rate of target recognition when the angle is missing. In the enhanced experiment, the missing perspectives of category 1 in the training set are supplemented by the present invention, the perspectives of categories 2-10 are set to be complete, and the perspectives of all targets in the test set are all in a complete state, which is used to verify the recognition performance improvement of the present invention for data enhancement. The present invention is used for data enhancement, and its recognition improvement is as follows:

[0125] Table 3 Data augmentation experiment settings

[0126] method Basic Experiments Enhanced Experiments Accuracy 65.3% 85.7% Recall 32.1% 73.2%

[0127] It can be seen that the present invention can effectively improve the accuracy and recall rate of target recognition by using data enhancement, and the improvement effect is better than traditional methods such as interpolation. The recognition accuracy reaches 85.7%, which is 20.4% higher than the basic experiment; the recall rate reaches 73.2%, which is 41.1% higher than the basic experiment.

[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 method for generating multi-pose ultra-narrow pulse echoes based on a generative adversarial network, characterized in that The steps of the method include: Step 1, constructing an extremely narrow pulse echo data set, dividing the constructed extremely narrow pulse echo data set into a training data set and a data set to be completed, and performing normalization preprocessing on the divided training data set and the data set to be completed, respectively, to obtain a normalized preprocessed training data set and a normalized preprocessed data set to be completed; Step 2, constructing a GAN generator network G, the constructed GAN generator network G includes an attention-based factorization module F and a continuous view embedding-based decoder module R; Step 3: Construct the GAN discriminator network D and adversarial training loss; Step 4, using the extremely narrow pulse echo data and the viewpoint label in the normalized preprocessed training data set in step 1 as the input of the GAN discriminator network D constructed in step 3, and optimizing the parameters of the GAN discriminator network D constructed in step 3 through the adversarial training loss constructed in step 3; Step 5, using the extremely narrow pulse echo data and the viewpoint label in the normalized preprocessed training data set in step 1 as the input of the GAN generator network G constructed in step 2, and optimizing the parameters of the GAN generator network G constructed in step 2 through the GAN discriminator network D after parameter optimization in step 4 and the adversarial training loss constructed in step 3; Step 6: The GAN generator network G with optimized parameters obtained in step 5 completes the missing perspectives of the data set to be completed, and completes the generation of multi-pose extremely narrow pulse echoes based on the generative adversarial network; In step 2, the decoder module R based on continuous view embedding includes a view embedding representation E and a view reconstruction decoder R. The decoder module R based on continuous view embedding is based on the identity feature f i and the input perspective y to be generated a 'Reconstruct the extremely narrow pulse echo, denoted as x', that is: x'=R(f i ,y a ') The view embedding representation E consists of a 1-layer fully connected network to generate the view label y a 'To the viewing angle feature f a ' is embedded in: f a '=E(y a ') The view reconstruction decoder R consists of a 4-layer deconvolutional neural network to realize the generated view feature f a ' and identity characteristics f i The reorganization output is a reconstructed extremely narrow pulse echo, that is: x'=R(f i ,f a ')=R(f i ,E(y a '))。 2. The method for generating multi-attitude ultra-narrow pulse echoes based on a generative adversarial network according to claim 1, characterized in that: In step 1, the content of the extremely narrow pulse echo data set includes: Θ s ~Θ e The extremely narrow pulse echo data of the C-type target within the viewing angle range and the corresponding viewing angle label, where Θ s Represents the starting angle of the viewing range, Θ e Represents the end view angle of the view range.

3. A method for generating multi-posture ultra-narrow pulse echoes based on a generative adversarial network according to claim 1 or 2, characterized in that: In step 1, the method of dividing the constructed extremely narrow pulse echo data set into a training data set and a data set to be completed is: The data angle labels of the extremely narrow pulse echo are counted. If the data angle label is a uniformly spaced full angle, N types of uniformly spaced Δθ full angle extremely narrow pulse echoes are obtained, which are recorded as the training data set X; otherwise, CN types of angles are obtained as θ1, θ2, ..., θ m The extremely narrow pulse echo of some viewing angles is recorded as the data set to be completed 4. The method for generating multi-posture ultra-narrow pulse echoes based on a generative adversarial network according to claim 3, characterized in that: In the step 1, the normalization preprocessing is performed by using the maximum amplitude normalization processing method, specifically: The maximum amplitude normalization uses the maximum value of each frame of 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 is expressed as:

5. The method for generating multi-posture ultra-narrow pulse echoes based on a generative adversarial network according to claim 4, characterized in that: In step 2, the attention-based factorization module includes a feature extractor F, a feature decomposer T and a supervisory constraint network C. The attention-based factorization module is used to extract the identity feature f of the input extremely narrow pulse echo data x. i ,Right now: f i =F(x)=T(F(x)); The extremely narrow pulse echo data x is a subset sampled from the training data set X; The feature extractor F consists of a 4-layer convolutional neural network, which is used to extract the features of the extremely narrow pulse echo data x and output the mixed feature representation f mix ,Right now: f mix =F(x) The feature decomposer T consists of a 2-layer self-attention network and a 1-layer weighted decomposition network, and its input is the mixed feature representation f extracted by the feature extractor. mix , the output is the viewing feature f a and identity characteristics f i , where the self-attention network is used to obtain the weight factor λ of the mixed feature; the weighted decomposition network is implemented by f according to λ mix The decomposition of is: [f a ,f i ]=T(f mix ) in, Represents a matrix dot product operation, 1 is a matrix of all 1s; The supervised constraint network C is composed of the perspective fully connected regression network C a and identity fully connected classification network C i Composition, supervision constraint network for f mix The decomposition process is supervised and constrained; The perspective fully connected regression network C a The input is f a , the output is the view regression result C a (f a ), the mean square error loss function is used for supervised constraint, and the perspective regression loss L is obtained a as follows: L a =MSE(C a (f a )-y a ) Among them, y a is the true viewpoint label of the input extremely narrow pulse echo data, and MSE stands for the mean square error loss function; Identity fully connected classification network C i Input is f i , the output is the identity classification result C i (f i ), the cross entropy loss function is used for supervised constraints to obtain the identity classification loss L i as follows: IT i =CE(C i (f i ),y i ) Among them, yi is the true identity label of the input extremely narrow pulse echo data, and CE represents the cross entropy loss function.

6. The method for generating multi-attitude ultra-narrow pulse echoes based on a generative adversarial network according to claim 1, characterized in that: In the step 3, the GAN discriminator network D is a discriminator composed of a 4-layer convolutional neural network, which is used to discriminate the authenticity of the input extremely narrow pulse echo; The method to construct the adversarial training loss is: (1) The generator training loss is constructed through the discriminator. The discriminator is responsible for distinguishing the extremely narrow pulse echo data x' generated by the generator as a "true" sample, that is, the output prediction label is True. The cross entropy loss function is used for supervised constraints to obtain the generator training loss L Gadv : L Gadv =CE(D(x'),True) (2) The discriminator training loss is constructed through the discriminator. When training the discriminator, the discriminator is responsible for distinguishing the extremely narrow pulse echo data x' generated by the generator as a "false" sample, that is, the output prediction label is False; and distinguishing the extremely narrow pulse echo data x sampled from the training data as a "true" sample, that is, the output prediction label is True. The cross entropy loss function is used for supervised constraints to obtain the discriminator training loss L Dadv : L Dadv =CE(D(x'),False)+CE(D(x),True)。 7. The method for generating multi-attitude ultra-narrow pulse echoes based on a generative adversarial network according to claim 1, characterized in that: In step 4, the method for parameter optimization is: fix the parameters of the generator network G, train the discriminator, and use the discriminator to train the loss L Dadv , optimize the parameters of the discriminator network D: L D =λ Dadv L Dadv 。 8. The method for generating multi-posture ultra-narrow pulse echoes based on a generative adversarial network according to claim 1, characterized in that: In step 5, the method for parameter optimization is: Fix the parameters of the discriminator network D, train the generator, and use the perspective regression loss L a , identity classification loss L i And the generator training loss L Gadv , optimize the parameters of the generator network G, and the loss function used for generator training is: L G =λ a L a +λ i L i +λ Gadv L Gadv Perform cyclic parameter optimization until the network converges.

9. The method for generating multi-attitude ultra-narrow pulse echoes based on a generative adversarial network according to claim 1, characterized in that: In step 6, the method for completing the missing perspectives of the data set to be completed is: The existing data and full-angle labels of the dataset to be generated, namely Θ s ~Θ e The view labels are evenly spaced Δθ within the view range and input into the GAN generator network G trained in step 5. The output samples of the continuous view embedding encoder module are saved. These samples are the view completion samples.