A method and system for vortex electromagnetic wave inverse scattering based on generative adversarial networks

Abstract

This invention discloses a method and system for inverse scattering of vortex electromagnetic waves based on generative adversarial networks (GANs), relating to the radar field, for rapid and reliable inverse scattering inversion of vortex electromagnetic waves. The invention uses real targets and vortex electromagnetic wave parameters as real samples, and generated data as dummy samples, inputting both into a discriminator for training; conversely, it uses generated data as real samples input into the discriminator to train the generator. The trained GAN is then used to predict the radar cross section of the scattering target, obtaining the basic parameters of the target and the vortex electromagnetic wave. This invention fills a gap in vortex electromagnetic wave inverse scattering inversion, avoiding the iterative complexity of traditional electromagnetic inversion methods and improving inversion efficiency. Furthermore, this invention learns and reconstructs target features from real data, preserving detailed information from the real data, resulting in more reliable prediction results.

Description

Technical Field

[0001] This invention relates to the field of radar, and in particular to a method and system for vortex electromagnetic wave inverse scattering based on generative adversarial networks. Background Technology

[0002] Inverting the structure and size of a target based on its electromagnetic scattering characteristics is an important topic in the field of electromagnetic scattering. In recent years, with the development of deep learning technology, many scholars have incorporated deep learning techniques into solving this problem. Currently, research in this area is mostly based on plane electromagnetic waves, but vortex electromagnetic waves, due to their unique helical phase distribution and amplitude characteristics, have become a research hotspot for target detection in recent years. To widely apply vortex electromagnetic waves to radar detection, identification, and other fields, the inversion of vortex electromagnetic wave inverse scattering characteristics based on deep learning has developed into an important research topic.

[0003] Electromagnetic inverse scattering is a scientific problem of great significance, and its research began in the 1940s. During this period, with the development of radar, sonar, and other methods for detecting and locating objects, the question arose of using electromagnetic waves to detect more objects or backgrounds. In recent years, scholars both domestically and internationally have conducted extensive research on electromagnetic inverse scattering and have made some progress. Research on the inverse problem of electromagnetic fields generally falls into two categories: linear and nonlinear. Traditional inverse scattering theory can only solve linear patterns and their interrelationships, yielding only approximate solutions, such as the classic Born approximate iterative algorithm. Linear methods can only solve inverse scattering problems for relatively simple targets. Nonlinear methods, however, introduce the multiple scattering problem into nonlinear integral equations, effectively retrieving the physical properties of the scattering object and the medium. Nonlinear inverse scattering can be summarized into three types: 1) deterministic algorithms based on regularization, such as the improved Born iterative method; 2) stochastic algorithms with global search capabilities, such as genetic algorithms and particle swarm optimization; and 3) algorithms suitable for neural networks.

[0004] Two challenges have been encountered in solving general nonlinear inversion problems: First, as the dimension of the parameters to be inverted increases, conventional inversion algorithms struggle to handle the complex nonlinearities. Second, while some inverse scattering models perform well in specific situations, their robustness fails to meet broader application requirements, particularly their poor resistance to noise. With the rapid development of machine learning methods, new explorations have emerged in the study of electromagnetic field inversion problems. Machine learning has proven to have unique advantages in solving both of these problems, and datasets generated based on forward modeling can effectively overcome the shortcomings of existing research. For example, Bermani et al. used support vector machines to study the inversion of the position and electromagnetic parameters of two-dimensional objects. Wei et al. used a U-net structure to study the correspondence between scatterers and scattered fields. In the field of remote sensing, Zhang Qinghe et al. used SVM and BP neural networks to establish forward models and invert and reconstruct characteristic parameters such as sea surface wind speed and soil moisture. However, the nonlinear solution of electromagnetic wave inverse scattering increases the complexity of these machine learning algorithms, meaning that they are currently not fast enough in solving electromagnetic wave inverse scattering problems, and their reliability still needs to be further improved.

[0005] In general, from a mathematical perspective, electromagnetic inversion problems exhibit two main characteristics: ill-posedness and nonlinearity. Ill-posedness refers to the inability to accurately predict the spatial values ​​of all parameters when the dimension of the actual measurement is lower than the dimension of the reconstructed parameters. This is common in practical problems such as geological exploration and medical imaging. Methods to address ill-posedness include increasing the amount of observational data, introducing prior information, and employing regularization techniques; however, these methods often require modification based on the specific problem. On the other hand, electromagnetic inversion problems typically involve nonlinear relationships because the interaction between objects and electromagnetic waves is often complex and nonlinear. This renders traditional linear solution methods inapplicable, necessitating advanced numerical optimization and machine learning techniques. The nonlinear nature increases the problem's complexity, challenging the processing capabilities of traditional algorithms and making it difficult for traditional nonlinear methods to meet the speed requirements of electromagnetic inversion. Furthermore, traditional algorithms are mostly designed for plane electromagnetic waves, with limited application to vortex electromagnetic waves. Summary of the Invention

[0006] The purpose of this invention is to provide a vortex electromagnetic wave inverse scattering method and system based on generative adversarial networks to address all or part of the problems mentioned above, so as to quickly and reliably realize the inversion of vortex electromagnetic wave inverse scattering, which addresses the ill-posed and nonlinear problems of electromagnetic wave inverse scattering inversion.

[0007] The technical solution adopted in this invention is as follows:

[0008] A method for inverse scattering vortex electromagnetic waves based on generative adversarial networks, comprising:

[0009] The radar cross section of the scattering target is input into a trained generative adversarial network to generate target parameters and / or vortex electromagnetic wave parameters; where,

[0010] The generative adversarial network includes a generator and a discriminator;

[0011] The generator is configured to generate target parameters and / or vortex electromagnetic wave parameters based on the input radar cross section and random noise.

[0012] The discriminator is configured to: distinguish between true and false targets based on input training sample data or target parameters and / or vortex electromagnetic wave parameters generated by the generator, wherein the training sample data is a sample radar cross section and the corresponding true target parameters and / or vortex electromagnetic wave parameters.

[0013] The training process of the generative adversarial network includes:

[0014] Training the discriminator: Fix the network parameters of the generator, use the real target parameters and / or vortex electromagnetic wave parameters in the training sample data as real samples, and use the target parameters and / or vortex electromagnetic wave parameters generated by the generator with the sample radar cross section in the training sample data as input as false samples. Combine the real samples and false samples according to a predetermined ratio and input them into the discriminator for iterative training until the training end condition is met.

[0015] Training generator: Fix the network parameters of the discriminator, and use the target parameters and / or vortex electromagnetic wave parameters generated by the generator as real samples, input them into the discriminator for iterative training until the training termination condition is met.

[0016] Furthermore, when training the discriminator, the ratio of real samples to fake samples is 1:1.

[0017] Furthermore, in the training sample data, target parameters are obtained by measuring a given target, and the corresponding radar cross section is obtained by modeling and simulating a vortex electromagnetic wave with given vortex electromagnetic wave parameters incident on the given target.

[0018] Furthermore, during the training process of the generative adversarial network, stochastic gradient descent is used to update the network parameters.

[0019] Furthermore, the labeling of real and fake samples is achieved by binarizing the labels on the sample data.

[0020] This invention also provides a vortex electromagnetic wave inverse scattering system based on generative adversarial networks, which includes a data input module, a model training module, and a model prediction module, wherein:

[0021] The data input module is configured to receive the radar cross section of the scattering target and training sample data, wherein the training sample data is the sample radar cross section and the corresponding real target parameters and / or vortex electromagnetic wave parameters.

[0022] The model training module is configured to train a configured generative adversarial network, wherein the generative adversarial network includes a generator and a discriminator;

[0023] The generator is configured to generate target parameters and / or vortex electromagnetic wave parameters based on the input radar cross section and random noise.

[0024] The discriminator is configured to: distinguish between genuine and fake input training sample data or target parameters and / or vortex electromagnetic wave parameters generated by the generator;

[0025] The model training module trains the generative adversarial network according to the following configuration:

[0026] Training the discriminator: Fix the network parameters of the generator, use the real target parameters and / or vortex electromagnetic wave parameters in the training sample data as real samples, and use the target parameters and / or vortex electromagnetic wave parameters generated by the generator with the sample radar cross section in the training sample data as input as false samples. Combine the real samples and false samples according to a predetermined ratio and input them into the discriminator for iterative training until the training end condition is met.

[0027] Training generator: Fix the network parameters of the discriminator, and use the target parameters and / or vortex electromagnetic wave parameters generated by the generator as real samples, input them into the discriminator for iterative training until the training termination condition is met;

[0028] The model prediction module is configured to input the radar cross section of the scattering target received by the data input module into the generative adversarial network trained by the model training module, and predict the target parameters and / or vortex electromagnetic wave parameters.

[0029] Furthermore, when training the discriminator, the model training module configures the ratio of real samples to fake samples to be 1:1.

[0030] Furthermore, in the training sample data, target parameters are obtained by measuring a given target, and the corresponding radar cross section is obtained by modeling and simulating a vortex electromagnetic wave with given vortex electromagnetic wave parameters incident on the given target.

[0031] Furthermore, the model training module updates the network parameters using stochastic gradient descent during the training of the generative adversarial network.

[0032] Furthermore, the model training module achieves the labeling of real samples and fake samples by binarizing the labels of the sample data.

[0033] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0034] 1. Existing technologies mostly address electromagnetic inversion problems under plane electromagnetic wave illumination, and are not applicable to solving electromagnetic inverse scattering inversion problems of vortex electromagnetic waves. This invention fills the gap in vortex electromagnetic wave inverse scattering inversion, providing a foundation for the widespread application of vortex electromagnetic waves in radar detection, identification, and other fields.

[0035] 2. Compared with traditional electromagnetic inversion methods, this invention uses neural networks to quickly and effectively invert target parameters and / or vortex electromagnetic wave parameters (i.e., wave source information), avoiding the iterative complexity of traditional electromagnetic inversion methods. It can quickly and effectively establish the relationship between the target's radar cross section (i.e., scattering characteristics) and target parameters and vortex electromagnetic wave parameters, achieving target parameter inversion in a shorter time, improving inversion efficiency and reducing costs.

[0036] 3. By using generative adversarial networks, this invention can learn and reconstruct the features of a target from measurement data without directly relying on the traditional forward model, thus achieving more accurate prediction of unknown targets. This generative adversarial network-based method not only improves the solution efficiency of inverse scattering, but also better preserves the detailed information in the data, thereby producing more reliable reconstruction results. Attached Figure Description

[0037] The present invention will be described by way of example and with reference to the accompanying drawings, wherein:

[0038] Figure 1 This is an architecture diagram of a generative adversarial network model.

[0039] Figure 2 This is a flowchart of the training process for a generative adversarial network.

[0040] Figure 3 This is a comparison chart of the backscattering effects of vortex electromagnetic waves on a sphere dataset.

[0041] Figure 4 This is a comparison chart of the backscattering effects of vortex electromagnetic waves in a mixed dataset. Detailed Implementation

[0042] All features disclosed in this specification, or all steps in all disclosed methods or processes, may be combined in any way, except for mutually exclusive features and / or steps.

[0043] Any feature disclosed in this specification (including any appended claims and abstract) may be replaced by other equivalent or similar features, unless specifically stated otherwise. That is, unless specifically stated otherwise, each feature is merely one example of a series of equivalent or similar features.

[0044] This invention addresses the inherent nonlinearity and ill-posedness of traditional inverse scattering solutions, as well as the scarcity of solutions for vortex electromagnetic wave inverse scattering. It proposes a vortex electromagnetic wave inverse scattering method based on generative adversarial networks to solve the vortex electromagnetic wave inverse scattering inversion problem. This method primarily solves the following problems:

[0045] (1) Existing electromagnetic inverse scattering mainly focuses on the inversion of objects irradiated by plane electromagnetic waves, and rarely involves the case of vortex electromagnetic waves. This invention mainly uses a specific generative adversarial network to invert the scattering characteristics of vortex electromagnetic waves irradiating different targets. By training the network model, the parameters of the target and the wave source can be derived from the radar cross section (RCS) of the vortex electromagnetic waves.

[0046] (2) The electromagnetic inverse scattering problem involves inverting the acquired electromagnetic signal to obtain the target characteristics or wave source information of the measured object. Due to the inherent nonlinearity and ill-posedness of the inverse problem, linear approximation and iterative methods are generally required to solve it. These methods are time-consuming and labor-intensive, making it difficult to achieve high-performance solutions. This invention utilizes a generative adversarial network to infer the shape, size, and wavelength of the vortex electromagnetic wave from the radar cross section of the scattering target, thereby achieving the purpose of vortex electromagnetic wave inversion.

[0047] Example 1

[0048] This embodiment introduces a vortex electromagnetic wave inverse scattering method based on generative adversarial networks, which includes:

[0049] The radar cross section of the scattering target is input into a trained generative adversarial network to generate target parameters and / or vortex electromagnetic wave parameters. Target parameters include the target's size and shape, while vortex electromagnetic wave parameters include the vortex wave's topological charge, half-cone angle, polarization, and incident angle. The following descriptions in this embodiment will focus on predicting target parameters and vortex electromagnetic wave parameters.

[0050] like Figure 1 As shown, a generative adversarial network (GAN) consists of a generator G and a discriminator D. Wherein:

[0051] The input to generator G consists of two parts: random noise z and latent variable c. The latent variable c contains some useful information about the generated data, namely the radar cross section. The output of generator G is the generated data G(z,c). Generator G generates target parameters and vortex electromagnetic wave parameters based on the input latent variable c and random noise z.

[0052] Discriminator D is configured to distinguish between true and false data from the input training sample data x (real data) or the target parameters and vortex electromagnetic wave parameters (pseudo data) generated by generator G. Therefore, the output of discriminator D is a binary value: 1 for a true judgment and 0 for a false judgment. The training sample data x consists of sample radar cross sections (i.e., the radar cross sections obtained from actual measurements) and the corresponding real target parameters and vortex electromagnetic wave parameters. During the training phase, when a sample radar cross section is input into generator G, the target parameters and / or vortex electromagnetic wave parameters corresponding to that sample radar cross section are input into discriminator D.

[0053] The goal of the generator G is to generate generated data G(z,c) that is similar to the real training sample data x, such that the generated data G(z,c) has similar statistical characteristics to the training sample data x. The goal of the discriminator D is to distinguish the generated data G(z,c) from the real training sample data x as much as possible. Simultaneously, guided by the latent variable c, the generative adversarial network achieves fine-grained control over the generated data G(z,c), enabling the generator G to generate data with specified attributes.

[0054] like Figure 2 As shown, the training process of a generative adversarial network includes:

[0055] Training the discriminator D: The network parameters of the generator G are fixed. The real target parameters and vortex electromagnetic wave parameters in the training sample data x are used as real samples. The target parameters and vortex electromagnetic wave parameters generated by the generator G with the sample radar cross section in the training sample data x as input are used as fake samples. The real and fake samples are combined according to a predetermined ratio (usually a 1:1 sampling ratio) and input into the discriminator D for iterative training until the training termination condition is met. For the target parameters and vortex electromagnetic wave parameters in the training sample data x, their labels are set to 1, representing real samples, while the labels of the generated data G(z,c) are set to 0, representing fake samples. That is, the sample data is labeled using binary labeling to determine whether it is a real or fake sample.

[0056] Next, the generator G is trained: the network parameters of the discriminator D are fixed, and the target parameters and vortex electromagnetic wave parameters generated by the generator G are used as real samples and input into the discriminator D for iterative training until the training termination condition is met.

[0057] The training sample data x is mainly obtained through high-frequency modeling and simulation. Target parameters are obtained by measuring a given target through modeling and simulation. The radar cross section is obtained by modeling and simulating a vortex electromagnetic wave with given parameters incident on the given target. Thus, the corresponding target parameters, vortex electromagnetic wave parameters, and radar cross section are obtained. The radar cross section is used to train the generator G and the discriminator D, while the target parameters and vortex electromagnetic wave parameters are used to train the discriminator D.

[0058] Finally, by comparing the performance of the generated data G(z,c) and the training sample data x in the discriminator D, methods such as stochastic gradient descent are used to adjust the network parameters. This process is continuously iterated until an equilibrium is reached between the discriminator D and the generator G, completing the network training. The trained network then has the ability to invert different parameters of vortex electromagnetic waves and targets, and can quickly obtain the parameters of targets and vortex electromagnetic waves from the obtained scattering data.

[0059] Example 2

[0060] Corresponding to the method in Example 1, this example introduces a vortex electromagnetic wave inverse scattering system based on generative adversarial networks. The system includes a data input module, a model training module, and a model prediction module.

[0061] The data input module is configured to receive the radar cross section of a scattering target for predicting target parameters and / or vortex electromagnetic wave parameters. It also receives training sample data for training the generative adversarial network. The training sample data consists of sample radar cross sections and corresponding real target parameters and / or vortex electromagnetic wave parameters. In the training sample data, target parameters are obtained by measuring a given target, and the corresponding radar cross section is obtained by modeling and simulating a vortex electromagnetic wave with given parameters incident on the given target.

[0062] The model training module is configured to train the configured generative adversarial network.

[0063] like Figure 1 As shown, the generative adversarial network (GAN) includes a generator and a discriminator. The generator is configured to generate target parameters and / or vortex electromagnetic wave parameters based on the input radar cross section and random noise. The discriminator is configured to distinguish between genuine and false target parameters and / or vortex electromagnetic wave parameters from the input training sample data or the parameters generated by the generator.

[0064] The model training module trains the generative adversarial network according to the following configuration:

[0065] Training the discriminator: Fix the network parameters of the generator, use the real target parameters and / or vortex electromagnetic wave parameters in the training sample data as real samples (the real samples are labeled by marking the sample data with a label of 1), and use the target parameters and / or vortex electromagnetic wave parameters generated by the generator with the sample radar cross section in the training sample data as input as fake samples (the fake samples are labeled by marking the sample data with a label of 0). Combine the real samples and fake samples according to a predetermined ratio (the ratio of real samples to fake samples is usually 1:1) and input them into the discriminator for iterative training until the training termination condition is reached.

[0066] Training generator: Fix the network parameters of the discriminator, and use the target parameters and / or vortex electromagnetic wave parameters generated by the generator as real samples to input into the discriminator for iterative training until the training termination condition is met.

[0067] The model training module uses methods such as stochastic gradient descent to adjust the network parameters during the training of the generative adversarial network. This process is continuously iterated until an equilibrium is reached between the discriminator and the generator, completing the network training. The trained generative adversarial network can then be used to predict target parameters and / or vortex electromagnetic wave parameters.

[0068] The model prediction module is configured to input the radar cross section of the scattering target received by the data input module into the generative adversarial network trained by the model training module, and predict the target parameters and / or vortex electromagnetic wave parameters.

[0069] Example 3

[0070] This embodiment uses simple objects such as spheres, cylinders, cones, cylinders + cones, and double spheres as targets. It uses Matlab and CST co-simulation to generate scattering datasets with different topological charges and different parameters to obtain the target scattering RCS. Table 1 shows the distribution of the datasets, and Table 2 shows the distribution of the datasets for spheres.

[0071] Table 1 Dataset Distribution

[0072]

[0073] Table 2 Dataset of the Sphere

[0074]

[0075] The scattering RCS is used as the network input, and the topological charge of the vortex wave, the half-cone angle of the vortex wave, the polarization of the vortex wave, the size and shape of the target, and the incident angle are used as the output to train the adversarial neural network. A relatively good prediction effect is achieved by modifying the network parameters. Then, 2000 data points are randomly selected from the dataset for model validation. Figure 3This indicates the inversion effect for different parameters of the ball. Figure 4 The table shows the inversion effect for different parameters of the mixed dataset. The closer the inversion result is to the line y=1, the closer the actual data is to the generated data, indicating a better prediction result. Table 3 shows the error between the simulated and reconstructed data of the sphere dataset, and Table 4 shows the error between the simulated and reconstructed data of the mixed dataset. The correlation coefficient is used to quantitatively describe the actual and predicted results, verifying the reliability of the prediction method of this invention.

[0076] Table 3 Errors between Simulation and Reconstruction Data for the Sphere Dataset

[0077] Topological load Half cone angle polarization size Angle of incidence <![CDATA[R 2 (infoGAN)]]> 0.9497 0.9740 0.9341 0.9891 0.9688

[0078] Table 4 Errors between Simulated and Reconstructed Data in Hybrid Datasets

[0079] Topological load Half cone angle polarization Target shape Target size Angle of incidence <![CDATA[R 2 ]]> 0.9013 0.9282 0.8983 0.9712 0.9828 0.9134

[0080] This invention is not limited to the specific embodiments described above. The invention extends to any feature or objective proposed in this specification, or any combination thereof.

Claims

1. A method for inverse scattering vortex electromagnetic waves based on generative adversarial networks, characterized in that, include: The radar cross section of the scattering target is input into a trained generative adversarial network to generate target parameters and / or vortex electromagnetic wave parameters; where, The generative adversarial network includes a generator and a discriminator; The generator is configured to generate target parameters and / or vortex electromagnetic wave parameters based on the input radar cross section and random noise. The discriminator is configured to: distinguish between true and false targets based on input training sample data or target parameters and / or vortex electromagnetic wave parameters generated by the generator, wherein the training sample data is a sample radar cross section and the corresponding true target parameters and / or vortex electromagnetic wave parameters. The training process of the generative adversarial network includes: Training the discriminator: Fix the network parameters of the generator, use the real target parameters and / or vortex electromagnetic wave parameters in the training sample data as real samples, and use the target parameters and / or vortex electromagnetic wave parameters generated by the generator with the sample radar cross section in the training sample data as input as false samples. Combine the real samples and false samples according to a predetermined ratio and input them into the discriminator for iterative training until the training end condition is met. Training generator: Fix the network parameters of the discriminator, and use the target parameters and / or vortex electromagnetic wave parameters generated by the generator as real samples, input them into the discriminator for iterative training until the training termination condition is met.

2. The vortex electromagnetic wave inverse scattering method based on generative adversarial networks as described in claim 1, characterized in that, When training the discriminator, the ratio of real samples to fake samples is 1:

1.

3. The vortex electromagnetic wave inverse scattering method based on generative adversarial networks as described in claim 1, characterized in that, In the training sample data, target parameters are obtained by measuring a given target, and the corresponding radar cross section is obtained by modeling and simulating a vortex electromagnetic wave with given vortex electromagnetic wave parameters incident on the given target.

4. The vortex electromagnetic wave inverse scattering method based on generative adversarial networks as described in claim 1, characterized in that, During the training of the generative adversarial network, stochastic gradient descent is used to update the network parameters.

5. The vortex electromagnetic wave inverse scattering method based on generative adversarial networks as described in claim 1, characterized in that, The labeling of real and fake samples is achieved by binarizing the labels of the sample data.

6. A vortex electromagnetic wave inverse scattering system based on a generative adversarial network, characterized in that, It includes a data input module, a model training module, and a model prediction module, among which: The data input module is configured to receive the radar cross section of the scattering target and training sample data, wherein the training sample data is the sample radar cross section and the corresponding real target parameters and / or vortex electromagnetic wave parameters. The model training module is configured to train a configured generative adversarial network, wherein the generative adversarial network includes a generator and a discriminator; The generator is configured to generate target parameters and / or vortex electromagnetic wave parameters based on the input radar cross section and random noise. The discriminator is configured to: distinguish between genuine and fake input training sample data or target parameters and / or vortex electromagnetic wave parameters generated by the generator; The model training module trains the generative adversarial network according to the following configuration: Training the discriminator: Fix the network parameters of the generator, use the real target parameters and / or vortex electromagnetic wave parameters in the training sample data as real samples, and use the target parameters and / or vortex electromagnetic wave parameters generated by the generator with the sample radar cross section in the training sample data as input as false samples. Combine the real samples and false samples according to a predetermined ratio and input them into the discriminator for iterative training until the training end condition is met. Training generator: Fix the network parameters of the discriminator, and use the target parameters and / or vortex electromagnetic wave parameters generated by the generator as real samples, input them into the discriminator for iterative training until the training termination condition is met; The model prediction module is configured to input the radar cross section of the scattering target received by the data input module into the generative adversarial network trained by the model training module, and predict the target parameters and / or vortex electromagnetic wave parameters.

7. The vortex electromagnetic wave inverse scattering system based on generative adversarial networks as described in claim 6, characterized in that, When training the discriminator, the model training module configures a 1:1 ratio of real samples to fake samples.

8. The vortex electromagnetic wave inverse scattering system based on generative adversarial networks as described in claim 6, characterized in that, In the training sample data, target parameters are obtained by measuring a given target, and the corresponding radar cross section is obtained by modeling and simulating a vortex electromagnetic wave with given vortex electromagnetic wave parameters incident on the given target.

9. The vortex electromagnetic wave inverse scattering system based on generative adversarial networks as described in claim 6, characterized in that, The model training module uses stochastic gradient descent to update network parameters during the training of the generative adversarial network.

10. The vortex electromagnetic wave inverse scattering system based on generative adversarial networks as described in claim 6, characterized in that, The model training module distinguishes between real and fake samples by binarizing the labels of the sample data.

Images