Sar target recognition method based on game against sample generation

Abstract

This invention discloses a SAR target recognition method based on game-theoretic adversarial sample generation. The method involves determining simulated and measured samples, constructing a recurrent generative adversarial network (RGAN) based on an encoder, converter, and decoder, and training the RGAN with the simulated and measured samples to obtain a generator and a discriminator. The simulated samples are then input into the generator and processed to obtain multiple virtual measured samples, forming a virtual measured sample set. The last layer of the discriminator is adjusted, and the virtual measured samples are input into the discriminator to train a target recognition network E. The measured samples are used to fine-tune or correct the target recognition network E to obtain the final target recognition network Q. The target recognition network Q is then used to test validation samples, which are measured data acquired by the radar at a preset or specified time. By utilizing prior target information, this method compensates for the lack of sufficient SAR measured samples, improving SAR target recognition performance under small sample conditions.

Description

Technical Field

[0001] This invention belongs to the technical field of automatic radar target recognition, and particularly relates to a SAR target recognition method based on game-theoretic adversarial sample generation. Background Technology

[0002] Synthetic Aperture Radar (SAR), as an active microwave sensor, can conduct all-weather, 24 / 7 reconnaissance of targets of interest regardless of natural conditions such as lighting and weather, and has a certain degree of penetration capability. As a result, it has become an indispensable military reconnaissance tool and has attracted widespread attention. Automatic target identification using SAR has also become a hot research topic both domestically and internationally.

[0003] Deep neural networks have been applied in SAR target recognition. However, existing applications have problems. The excellent recognition performance of deep learning requires the support of a large dataset. Due to the special nature and complexity of SAR images, they usually need to be labeled and analyzed by professionals. However, in complex environments, manual labeling is inefficient and the accuracy is difficult to guarantee. As a result, although a large amount of data has been accumulated in the process of equipment development, testing and use, very little data can actually be used for training. This causes the small sample problem, which seriously restricts the improvement of SAR target recognition performance and reduces the functionality of the radar.

[0004] In view of this, the present invention is hereby proposed. Summary of the Invention

[0005] The purpose of this invention is to provide a SAR target recognition method based on game-theoretic adversarial sample generation, solving the technical problem of reduced radar performance caused by existing methods. The technical solution of this invention has many beneficial effects, as described below:

[0006] A SAR target identification method based on game-theoretic adversarial sample generation is provided, the SAR target identification method comprising:

[0007] Simulation samples and measured samples are determined. The simulation samples form a simulation sample set X, and the measured samples form a measured sample set Y. The simulation samples are generated based on the target electromagnetic scattering model and target and environment characteristic analysis software. The measured samples are generated by processing radar echoes using imaging technology.

[0008] A recurrent generative adversarial network is built based on an encoder, a converter, and a decoder. The simulated samples and the measured samples are input into the recurrent generative adversarial network for training to obtain a generator and a discriminator.

[0009] The simulated samples are input into the generator, and after being processed to obtain multiple virtual measured samples. Multiple virtual measured samples form a virtual measured sample set.

[0010] Obtain the number of target categories C, modify the number of output nodes of the last layer of the discriminator to the number of target categories C, and input the virtual test sample set into the discriminator to determine the target recognition network E;

[0011] The target recognition network E is fine-tuned or corrected using the measured samples to obtain the final target recognition network Q.

[0012] The target recognition network Q is used to test the verification samples, which are the actual measured data of the samples acquired by the radar at a preset time or a specified time.

[0013] Compared with the prior art, the technical solution provided by the present invention has the following beneficial effects:

[0014] Based on the electromagnetic scattering model of the target, a large number of SAR simulation samples are generated using a rapid SAR target slice generation technique. A recurrent generative adversarial network is used to perform real-world processing on the simulation samples to enhance the similarity between the simulation samples and the real-world samples and improve the availability of the simulation samples. Finally, the SAR target recognition network is pre-trained using the real-world processed simulation data and fine-tuned using real-world SAR data to obtain the final SAR target recognition network. Transfer learning is used to solve the problem of low target recognition performance with a small amount of real-world SAR data. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art 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.

[0016] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0017] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, in the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0018] like Figure 1 A SAR target recognition method based on game-theoretic adversarial sample generation, the SAR target recognition method comprising:

[0019] The simulation samples and the measured samples are determined. The simulation samples are used to construct the simulation sample set X, and the measured samples are used to construct the measured sample set Y. The simulation samples are generated based on the target electromagnetic scattering model and target and environment characteristic analysis software; the measured samples are generated by processing radar echoes using imaging technology.

[0020] A recurrent generative adversarial network is built based on an encoder, a converter, and a decoder. Simulated samples and measured samples are input into the recurrent generative adversarial network for training to obtain a generator and a discriminator.

[0021] The simulated samples are input into the generator, and after being processed to resemble actual measurements, multiple virtual actual measurement samples are obtained. Multiple virtual measured samples form a virtual measured sample set.

[0022] Obtain the number of target categories C, modify the number of output nodes of the last layer of the discriminator to the number of target categories C, and input the virtual test sample set into the discriminator to determine the target recognition network E;

[0023] The target recognition network E is fine-tuned or corrected using actual test samples to obtain the final target recognition network Q.

[0024] The target recognition network Q is used to test the verification samples, which are the actual measured data of the samples acquired by the radar at a preset time or a specified time.

[0025] This method generates a large number of simulated target samples using a target electromagnetic scattering model, and then constructs a recurrent adversarial generative network to process the simulated samples to enhance their usability for assisting SAR target identification. By utilizing prior target information, it compensates for the lack of SAR measured samples, thus improving SAR target identification performance under small sample conditions.

[0026] As a specific implementation method provided in this case, a large number of target SAR simulation samples are generated using target and environmental characteristic analysis software to construct a simulation sample set. Where, x j Let I represent the j-th simulation sample. j ∈{0,...,C-1} represents the category label of the j-th simulation sample, C represents the total number of target categories, and M represents the number of simulation samples in the simulation sample set.

[0027] Furthermore, simulated samples and actual test samples are input into a recurrent generative adversarial network for training, resulting in a generator and a discriminator, including:

[0028] Simulated samples and actual test samples are input into a recurrent generative adversarial network (GAN) to train a generator and a discriminator, where:

[0029] Construct a generator G containing a convolutional layer P, a residual layer R, and a deconvolutional layer Q. X2Y and generator G Y2X And to build a discriminator D containing a convolutional layer L, a pooling layer S, a fully connected layer 1, and a Softmax layer for binary classification. X and D Y ;

[0030] Using generator G X2Y and G Y2X and the discriminator D X and D Y Constructing a recurrent generative adversarial network, specifically:

[0031] Determine the total cost function, where:

[0032] The relationship between simulated samples and measured samples is represented by the first loss function, which is expressed as follows: in, This represents the expected value of y;

[0033] The relationship between measured samples and simulated samples is represented by a second loss function, which is expressed as follows: in, This represents the expected value of x;

[0034] The cycle consistency between simulated and measured samples is represented by a third loss function, which is denoted as L. cvc (G X2Y G Y2X (X, Y) = E x～X [||G Y2X -x||1]+E y～Y [||G X2Y -y|1]

[0035] Where || ||1 is the norm constraint; Ex ~ x represents the expectation of the simulation sample set, and Ey ~ Y represents the expectation of the measured sample set;

[0036] Therefore, the total cost function L is:

[0037] L = L G (G X2Y D Y (X, Y) + L G (G Y2X D X (X, Y) + L cyc (G X2Y G Y2X (X, Y)

[0038] Objective: To make the simulated samples highly similar to the actual measured samples;

[0039] The simulated samples and measured samples are input into the total cost function to obtain the generator and discriminator, where:

[0040] Let the measured sample set be , where y j Let b represent the j-th measured sample. j Let {0, ..., C-1} represent the class label of the j-th measured sample, C represent the total number of target classes, and N represent the number of measured samples in the measured sample set. A training sample set is constructed using the simulated sample set X and the measured sample set Y. For example, 70% of the data from the training set is taken as training data, and 30% is taken as validation data. The remaining 70% of the data is input into the recurrent adversarial generative network (RBN), and the network parameters, such as weight updates, are updated using stochastic gradient descent.

[0041] A training sample set is constructed using a simulated sample set X and a measured sample set Y. The first proportion of data in the training sample set is used as training data, and the second proportion is used as validation data. This training data is then input into a recurrent adversarial generative network (RBN), and the network parameters are updated using stochastic gradient descent until the RBN converges, thus obtaining the generator G. X2Y and Discriminator D Y Network parameters include weights.

[0042] Unlike existing technologies that utilize simulation sample data, this method uses a recurrent generative adversarial network to perform real-world testing on the simulation sample data. This real-world testing improves the similarity between simulation data and real-world data, thereby enhancing the usability of the simulation samples.

[0043] Furthermore, the samples x in the simulation sample set X... j Input to generator G X2Y In the process, virtual measured samples are obtained. All virtual measured samples constitute the virtual measured sample set.

[0044] Furthermore, the number of target categories C is obtained, the number of output nodes of the last layer of the discriminator is modified to the number of target categories C, and the virtual test sample set is input into the discriminator to determine the target recognition network E, including:

[0045] Change the number of output nodes of the last layer of the discriminator to the number of target categories C, and use it as the target recognition network F;

[0046] virtual measured sample set The parameters of the target recognition network F are used as inputs to the target recognition network F and are updated using the stochastic gradient descent method until the target recognition network F converges. The final parameters are then used as the parameters of the target recognition network, and the converged network is called the target recognition network E.

[0047] Furthermore, the target recognition network E is fine-tuned or corrected using actual test samples to obtain the final target recognition network Q, including:

[0048] The measured sample set Y is used as the input to the target recognition network E. Stochastic gradient descent is used to learn the parameters, and this process is repeated until the target recognition network E converges. The converged network is then used as the target recognition network Q, and the final parameters are used as the parameters of the target recognition network Q. The aim is to use simulated samples to provide a better initial value (weights), reducing the need for real sample data. Therefore, higher recognition performance can be obtained with a smaller number of measured samples.

[0049] Furthermore, the target recognition network Q is used to test the verification samples, which are measured data acquired by the radar at a preset or specified time, including:

[0050] Let the validation sample set be ,in, Let c represent the j-th validation sample. j ∈{0,...,C-1} represents the measured class label of the j-th validation sample, C represents the total number of target classes, and O represents the number of test samples in the test sample set;

[0051] Test samples The samples are input into the target recognition network Q to obtain validation samples. Prediction category , The output of the target recognition network Q;

[0052] Actual test category label c j With network prediction category The comparison and verification show that if the two values ​​are the same, the sample is correctly classified; if the two values ​​are different, the sample is incorrectly classified.

[0053] Statistical validation sample set The final sample classification accuracy C is obtained by counting the number of correctly classified samples CP across all samples. p Among them, recognition rate

[0054] (1) Experimental conditions

[0055] Experimental platform: Anaconda3, Intel(R)Xeon(R)W-11955@2.60GHZ, 2.61GHz, NVIDIARTX A5000, Windows 10 Professional.

[0056] The data used in the experiment was the MSTAR measured SAR dataset jointly released by the U.S. Advanced Research Projects Agency (ARPA) and the Air Force Research Laboratory. This dataset contains SAR images of various military and civilian vehicle targets at different azimuth and elevation angles. The image resolution is 0.3m × 0.3m, and the image size ranges from a minimum of 128 × 128 to a maximum of 192 × 193. In the experiment, all SAR images were cropped to 128 × 128. The MSTAR dataset contains 10 types of ground vehicle targets, including BTR70, BMP2, and T72. Among them, BMP2 has three variants, namely SNC21, SN9566, and SN9563; T72 has 11 variants, namely SNS7, SN812, SN132, A04, A05, A07, A10, A32, A62, A63, and A64.

[0057] The MSTAR dataset includes measured data at elevation angles of 15° and 17°. Typically, the data at 17° is used as the training dataset, and the data at 15° is used as the validation dataset. In this experiment, the training set selected three different target types—BMP2, BTR70, and T72—at 17°, with only one model selected for each target type. The test set also selected three different target types—BMP2, BTR70, and T72—at 15°, but BMP2 and T72 each included three different variants. The specific data partitioning is shown in Table 1, which represents the MSTAR 3-class, 7-type data partitioning.

[0058]

[0059] (2) Experimental content and results analysis

[0060] Experiment 1: Experiments were conducted on MSTAR 3 type 7 data using the method of this invention.

[0061] Experiment 2: The target recognition network Q in the method of this invention was used to conduct experiments on MSTAR 3 types 7 data.

[0062] Experiment 3 converts SAR images from MSTAR 3 types 7 data into one-dimensional vectors, which are then used as inputs for Euclidean distance-constrained AEs for identification.

[0063] Experiment 4 converts SAR images from MSTAR 3 types 7 data into one-dimensional vectors, which are then used as inputs for sparse AEs (SAEs) for identification.

[0064] The recognition results of Experiments 1, 2, 3, and 4 are shown in Table 2:

[0065] Table 2. Recognition results of different recognition methods

[0066] Result 1 Result 2 Result 3 Result 4 Recognition rate 98.32％ 95.25％ 94.14％ 92.10％

[0067] As can be seen from Table 1, Result 1 represents the recognition rate of the method of the present invention, which is 98.32%. The others are the results of the comparative methods, which are 95.25%, 94.14%, and 92.10%. Result 1 achieves the highest recognition performance. By using simulated samples generated by electromagnetic scattering modeling to assist the actual test samples, the target recognition performance can be effectively improved when the actual test samples are insufficient.

[0068] The product provided by this invention has been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of this invention. The descriptions of the embodiments above are merely for the purpose of helping to understand the core ideas of this invention. It should be noted that those skilled in the art can make various improvements and modifications to the invention without departing from the principles of the invention, and these improvements and modifications also fall within the protection scope of the invention claims.

Claims

1. A SAR target recognition method based on game-theoretic adversarial sample generation, characterized in that, The SAR target identification method includes: Simulation samples and measured samples are determined. The simulation samples form a simulation sample set X, and the measured samples form a measured sample set Y. The simulation samples are generated based on the target electromagnetic scattering model and target and environment characteristic analysis software. The measured samples are generated by processing radar echoes using imaging technology. A recurrent generative adversarial network (RGAN) is constructed based on an encoder, a converter, and a decoder. Simulated and measured samples are input into the RGAN for training to obtain a generator and a discriminator. The generator includes a generator... and ; The simulated samples are input into the generator, and after being processed to obtain multiple virtual measured samples. Multiple virtual measured samples form a virtual measured sample set. Among them, the samples in the simulation sample set X Input to generator In the process, virtual measured samples are obtained. All the virtual measured samples constitute the virtual measured sample set. ; Obtain the number of target categories C, modify the number of output nodes of the last layer of the discriminator to the number of target categories C, and input the virtual test sample set into the discriminator to determine the target recognition network E. Specifically, the number of output nodes of the last layer of the discriminator is changed to the number of target categories C, and this becomes the target recognition network F. The virtual test sample set... The parameters of the target recognition network F are used as inputs and the stochastic gradient descent method is used to update the parameters of the target recognition network F until the target recognition network F converges. The final parameters are then used as the parameters of the target recognition network, and the converged network is used as the target recognition network E. The target recognition network E is fine-tuned or corrected using the measured samples to obtain the final target recognition network Q. The target recognition network Q is used to test the verification samples, which are the actual measured data of the samples acquired by the radar at a preset time or a specified time.

2. The SAR target identification method according to claim 1, characterized in that, The simulation samples and the measured samples are determined. The simulation samples are generated based on the target electromagnetic scattering model and target and environment characteristic analysis software. The measured samples are generated using imaging technology to process radar echoes, and include: A large number of target SAR simulation samples were generated using target and environment characteristic analysis software, and a simulation sample set was constructed. ,in, Let j represent the j-th simulation sample. Let C represent the category label of the j-th simulation sample, C represent the total number of target categories, and M represent the number of simulation samples in the simulation sample set.

3. The SAR target identification method according to claim 2, characterized in that, The simulated samples and measured samples are input into the recurrent generative adversarial network for training to obtain a generator and a discriminator, including: The simulated samples and the measured samples are input into the recurrent generative adversarial network (RGAN) to train the generator and discriminator, wherein: Establish a system containing convolutional layers Layer, residual layer Layers and deconvolution layers Layer generator and generator and establishing a system containing convolutional layers Pooling layer Discriminator consisting of a layer, a fully connected layer 1, and a Softmax layer that implements binary classification. and ; Using the generator and and the discriminator and Construct a recurrent generative adversarial network.

4. The SAR target identification method according to claim 3, characterized in that, Using the generator and and the discriminator and Generative adversarial networks (GANs) include: Determine the total cost function, where: The relationship between the simulated samples and the measured samples is represented by a first loss function, which is expressed as follows: = ,in, This represents the expected value of y; The relationship between measured samples and simulated samples is represented by a second loss function, which is expressed as follows: = ,in, This represents the expected value of x; The cycle consistency between simulated and measured samples is represented by a third loss function, which is expressed as follows: = in, For norm constraints; Ex~X represents the expectation of the simulation sample set, and Ey~Y represents the expectation of the measured sample set; Therefore, the total cost function L is: ； The simulated samples and measured samples are input into the total cost function to obtain the generator and discriminator, wherein: Let the measured sample set be ,in, Indicates the first One measured sample, Indicates the first The category labels of the measured samples, where C represents the total number of target categories and N represents the number of measured samples in the measured sample set; A training sample set is constructed using the simulated sample set X and the measured sample set Y. A first proportion of the data in the training sample set is used as training data, and a second proportion is used as validation data. This training data is then input into the recurrent adversarial generative network (ROG), and the network parameters are updated using stochastic gradient descent until the RPG converges, thus obtaining the generator. and discriminator The network parameters include weights.

5. The SAR target identification method according to claim 4, characterized in that, The target recognition network E is fine-tuned or corrected using the measured samples to obtain the final target recognition network Q, including: The measured sample set Y is used as the input of the target recognition network E, and the parameters are learned using the stochastic gradient descent method. This process is repeated until the target recognition network E converges. The converged network is then used as the target recognition network Q, and the final parameters are used as the parameters of the target recognition network Q.

6. The SAR target identification method according to claim 5, characterized in that, The target recognition network Q is used to test the verification samples, which are measured data acquired by the radar at a preset or specified time, including: Let the validation sample set be ,in, Let j represent the j-th verification sample. Let represent the measured class label of the j-th validation sample, C represent the total number of target classes, and O represent the number of test samples in the test sample set; Test samples The input is fed into the target recognition network Q to obtain validation samples. Prediction category , The output of the target recognition network Q; Actual category labels With network prediction category The comparison and verification show that if the two values ​​are the same, the sample is correctly classified; if the two values ​​are different, the sample is incorrectly classified. Statistical validation sample set The final sample classification accuracy C is obtained by counting the number of correctly classified samples CP across all samples. p Among them, the recognition rate C p = .

Images