ISAR image super-resolution method based on cycle generative adversarial network

Abstract

This application relates to an ISAR image super-resolution method based on a recurrent generative adversarial network (RBAN). It involves acquiring paired datasets of low-resolution and high-resolution ISAR images of targets; constructing training and testing sets for low-resolution and high-resolution ISAR images respectively; and employing a dynamic convolutional structure as the discriminator in the RBAN model construction. The receptive field size of the dynamic fully convolutional network is adjusted according to the target scattering points, ensuring that each receptive field can be bounded during convolution, thus solving the problem of increased misclassification rate in the discriminator module. By introducing the RBAN structure, end-to-end mapping from low-resolution ISAR images to high-resolution ISAR images is achieved. The trained network exhibits stronger generalization performance compared to existing ISAR image super-resolution networks, and its super-resolution performance is also superior to existing methods for ISAR images with resolutions different from the test data.

Description

Technical Field

[0001] This application relates to the field of inverse synthetic aperture radar (ISAR) imaging image processing technology, and in particular to an ISAR image super-resolution method based on cyclic generative adversarial networks. Background Technology

[0002] With the continuous advancement of radar system hardware and signal processing technology, existing broadband ISAR imaging radars already possess high resolution, capable of obtaining detailed information such as the shape, size, and material of moving targets like missiles and aircraft. To acquire clearer images of targets, the most direct method is to increase the radar's transmitted signal bandwidth to improve the range resolution of the ISAR image and to increase the coherent accumulation angle of the target relative to the radar to improve the azimuth resolution of the ISAR image. However, considering the high cost associated with increasing the transmitted signal bandwidth and the difficulty in guaranteeing a sufficient coherent accumulation angle for non-cooperative targets, obtaining higher-resolution ISAR images through super-resolution imaging methods is an effective approach.

[0003] In recent years, deep learning-based methods have been able to directly achieve super-resolution in the image domain, achieving good results in both optical and SAR images. Ledig C et al., in their paper "Photo-Realistic SingleImage Super-Resolution Using a Generative Adversarial Network" (IEEE ComputerSociety, 2016), designed a generative adversarial network structure to achieve super-resolution of a single optical image. Ao D et al., in their paper "Dialectical GAN ​​for SAR Image Translation: From Sentinel-1 to TerraSAR-X" (Remote Sensing, 2018, 10(10)), designed a dialectical generative adversarial network structure, which learns the mapping from Sentinel-1 SAR images to TerraSAR-X SAR images, thereby improving the resolution of Sentinel-1 SAR images. Similarly, for ISAR images, Dan Q et al. proposed a super-resolution method based on residual networks in their paper "ISAR Resolution Enhancement Using Residual Network" (IEEE 4th International Conference on Signal and Image Processing, 2019), which achieved some success. However, this method requires that the input test ISAR images must be preprocessed using the same method as the training ISAR images, which limits its practical application, especially when processing ISAR image data with different resolutions than the training data, where the super-resolution effect is not ideal. Summary of the Invention

[0004] Therefore, it is necessary to address the issue that the difference in resolution between test data and training data may lead to a decrease in super-resolution performance. To address this, a recurrent generative adversarial network-based ISAR image super-resolution method is needed to achieve end-to-end mapping from low-resolution ISAR images to high-resolution ISAR images, thereby obtaining ISAR images with higher target resolution.

[0005] A super-resolution ISAR image method based on recurrent generative adversarial networks, the method comprising:

[0006] Obtain paired target low-resolution ISAR image datasets and high-resolution ISAR image datasets; construct training and test sets for low-resolution ISAR images and high-resolution ISAR images respectively.

[0007] The low-resolution ISAR image training set and the high-resolution ISAR image training set are input into a preset recurrent generative adversarial network model for training to obtain a preliminary recurrent generative adversarial network model; the discriminator module in the preliminary recurrent generative adversarial network model is a dynamic fully convolutional network, and the receptive field size of the dynamic fully convolutional network is determined according to the target scattering point.

[0008] The low-resolution ISAR image test set is input into the preliminary recurrent generative adversarial network model for verification. The error between the model and the high-resolution ISAR image test set is calculated. When the error is large, the receptive field size of the discriminator module is adjusted, and the preliminary recurrent generative adversarial network model is retrained to obtain the recurrent generative adversarial network model.

[0009] In one embodiment, paired target low-resolution ISAR image datasets and high-resolution ISAR image datasets are obtained, and training and test sets for low-resolution and high-resolution ISAR images are constructed respectively, including:

[0010] By measuring in an anechoic chamber, the fully polarimetric radar echo data of the target in various X-band attitudes are obtained; according to the set imaging bandwidth and coherence accumulation angle, the echo data of the same attitude in the fully polarimetric radar echo data of each attitude are processed to obtain paired low-resolution ISAR images of the target in a specific attitude and high-resolution ISAR images corresponding to the low-resolution ISAR images; training sets and test sets of low-resolution ISAR images and high-resolution ISAR images are constructed respectively.

[0011] In one embodiment, a low-resolution ISAR image training set and a high-resolution ISAR image training set are input into a preset recurrent generative adversarial network (RGAN) model for training, resulting in a preliminary RGAN model, including:

[0012] The low-resolution ISAR image training set is divided into a low-resolution ISAR image training subset and a low-resolution ISAR image verification subset; the high-resolution ISAR image training set is divided into a high-resolution ISAR image training subset and a high-resolution ISAR image verification subset.

[0013] The low-resolution ISAR image training subset and the high-resolution ISAR image training subset are respectively input into the generator module to obtain the network loss function of the corresponding ISAR image, and the preset recurrent generative adversarial network model is trained based on the network loss function.

[0014] The preset cyclic generative adversarial network model is validated using the low-resolution ISAR image validation subset and the high-resolution ISAR image validation subset to obtain a preliminary cyclic generative adversarial network model.

[0015] In one embodiment, the network loss function includes: an adversarial loss function, a cyclic loss function, and a super-resolution loss function.

[0016] In one embodiment, the adversarial loss function includes: an adversarial loss function for low-resolution ISAR images and an adversarial loss function for high-resolution ISAR images.

[0017] The adversarial loss function for low-resolution ISAR images is expressed as follows:

[0018] ;

[0019] The adversarial loss function expression for high-resolution ISAR images is:

[0020] ;

[0021] in, This represents a subset of low-resolution images used for training. This represents a subset of high-resolution images used for training. From domain to Domain generator module, From domain to Domain generator module, , They are respectively Domain and Discriminator module on the domain, This represents the distribution of real samples in the training subset of low-resolution images. This represents the distribution of real samples in the training subset of high-resolution images.

[0022] In one embodiment, the cyclic loss function includes: a cyclic loss function for low-resolution ISAR images.

[0023] Its expression is:

[0024] .

[0025] In one embodiment, the super-resolution loss function includes: a super-resolution loss function for low-resolution ISAR images and a super-resolution loss function for high-resolution ISAR images.

[0026] The super-resolution loss function expression for low-resolution ISAR images is as follows:

[0027] ;

[0028] The super-resolution loss function expression for high-resolution ISAR images is:

[0029] .

[0030] In one embodiment, the network loss function includes: from domain to The generator module network loss function and from the domain domain to The network loss function of the domain generator module.

[0031] Among them, from domain to The network loss function of the domain generator module is expressed as:

[0032] ;

[0033] from domain to The network loss function of the domain generator module is expressed as:

[0034] .

[0035] In one embodiment, a low-resolution ISAR image test set is input into the preliminary recurrent generative adversarial network (RGAN) model for verification. The error between the model and the high-resolution ISAR image test set is calculated. When the error is large, the receptive field size of the discriminator module is adjusted, and the preliminary RGAN model is retrained to obtain the RGAN model, which includes:

[0036] The low-resolution ISAR image test set is input into the preliminary recurrent generative adversarial network model for verification. The absolute error between the output result and the high-resolution ISAR image test set is calculated. When the error is large, the receptive field size is adjusted by changing the discriminator convolution kernel size and the number of convolution layers. The preliminary recurrent generative adversarial network model is retrained. When the error does not decrease significantly, the recurrent generative adversarial network model is obtained. The absolute error refers to the sum and average of the absolute values ​​of the differences between each pixel in the two images.

[0037] In one embodiment, the generator module adopts a U-net structure; the discriminator module adopts a PatchGAN structure.

[0038] The aforementioned ISAR image super-resolution method based on recurrent generative adversarial networks (RBANs) acquires paired low-resolution and high-resolution ISAR image datasets. Training and testing sets for low-resolution and high-resolution ISAR images are constructed respectively. When constructing the RBAN model, a dynamic convolutional discriminator is used. The receptive field size of the dynamic fully convolutional network is adjusted according to the target scattering points, ensuring that any receptive field can contain the target scattering points. This ensures that each receptive field can be bounded during convolution, solving the problem that the receptive field bounding range is likely to be background rather than target during convolution, leading to an increased misclassification rate in the discriminator module. By introducing the RBAN structure, an end-to-end mapping from low-resolution to high-resolution ISAR images is achieved, resulting in ISAR images with higher target resolution. The trained network exhibits stronger generalization performance compared to existing ISAR image super-resolution networks, and its super-resolution performance is also superior to existing methods for ISAR images with resolutions different from the test data. Attached Figure Description

[0039] Figure 1 This is a flowchart of an ISAR image super-resolution method based on a recurrent generative adversarial network in one embodiment;

[0040] Figure 2 This is a flowchart illustrating the generation process of low-resolution ISAR images and high-resolution ISAR images in one embodiment.

[0041] Figure 3 This is a comparison diagram of the discriminator receptive field size before and after improvement in one embodiment;

[0042] Figure 4 The super-resolution result of a network trained on a dataset with a super-resolution ratio of 4 in one embodiment is compared with the corresponding ISAR image with an imaging bandwidth of 4GHz and a coherence accumulation angle of 20° after inputting an ISAR image with an imaging bandwidth of 4GHz and a coherence accumulation angle of 20°. Detailed Implementation

[0043] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0044] It is worth noting that the ISAR low-resolution image and high-resolution image mentioned in this invention are not obtained from the same image, but are obtained for the same target using ISAR with different transmit signal bandwidths and different coherent accumulation angles during turntable imaging.

[0045] The ISAR image super-resolution method based on recurrent generative adversarial networks provided in this application, such as Figure 1 The diagram shown is a flowchart of the present invention, which includes the following steps:

[0046] Step 1: Obtain paired target low-resolution ISAR image datasets and high-resolution ISAR image datasets; construct training and test sets for low-resolution ISAR images and high-resolution ISAR images respectively.

[0047] Specifically, the low-resolution ISAR image dataset is denoted as... The high-resolution ISAR image dataset is denoted as The low-resolution ISAR image training set is denoted as (Each image is 512×512), low-resolution ISAR image training set The range resolution of ISAR images includes Azimuth resolution The high-resolution ISAR image training set is denoted as... (Each image is 512×512), high-resolution ISAR image training set The range resolution of ISAR images includes Azimuth resolution .and , where n is the resolution ratio. The low-resolution ISAR image test set is denoted as . The high-resolution ISAR image test set is denoted as The ratio of range resolution to azimuth resolution between the two sets of measurement sets is n.

[0048] It is important to note that the low-resolution ISAR image dataset is relative to the high-resolution ISAR image dataset, and not an absolute concept. Before building the model, the dataset needs to be divided to prevent the training model from knowing the characteristics of the samples in the test set in advance, causing data snooping bias and making the model training results overly optimistic and deviating from reality. The low-resolution ISAR image dataset and the corresponding high-resolution ISAR image dataset are divided into training set, validation set, and test set. Since the imaging bandwidth and coherence accumulation angle when acquiring the ISAR images in the test set are different from those when acquiring the training / validation sets, the ISAR images obtained with different bandwidths and coherence accumulation angles are not included in the segmentation. In fact, only the training set and validation set with the same bandwidth and accumulation angle are segmented. In this embodiment of the invention, the segmentation ratio of the training set, validation set, and test set is 3:1:4, but different segmentation ratios can also be set according to the actual situation.

[0049] Step 2: Input the low-resolution ISAR image training set and the high-resolution ISAR image training set into the preset recurrent generative adversarial network model for training to obtain the preliminary recurrent generative adversarial network model; the discriminator module in the preliminary recurrent generative adversarial network model is a dynamic fully convolutional network, and the receptive field size of the dynamic fully convolutional network is determined according to the target scattering point.

[0050] It is worth noting that the low-resolution ISAR image training set includes a low-resolution ISAR image training subset and a low-resolution ISAR image training and validation subset, while the high-resolution ISAR image training set includes a high-resolution ISAR image training subset and a high-resolution ISAR image training and validation subset. A pre-defined recurrent generative adversarial network (GAN) model is trained using the training set, and then evaluated and validated using the validation set to find the optimal parameters.

[0051] In the recurrent generative adversarial network model of this invention, a dynamic fully convolutional network is used as the discriminator. Unlike conventional methods that select pixel images as targets, the target of this invention is a sparse ISAR image. When the receptive field size of the discriminator is too small, the area selected by its receptive field during convolution is likely to be the background rather than the target, leading to an increased misclassification rate. To solve this problem, the kernel size and the number of convolutional layers need to be increased according to the actual target size. Experimental results show that the super-resolution effect is better when the receptive field size of the discriminator can ensure that any receptive field can contain the target scattering points.

[0052] Step 3: Input the low-resolution ISAR image test set into the preliminary recurrent generative adversarial network model for verification, calculate the error with the high-resolution ISAR image test set, and adjust the receptive field size of the discriminator module when the error is large, and retrain the preliminary recurrent generative adversarial network model to obtain the recurrent generative adversarial network model.

[0053] It is worth noting that the preliminary recurrent generative adversarial network (RGAN) model obtained after training in step 2 has a certain generalization error. To verify this error, a low-resolution ISAR image test set is input into the preliminary RGAN model, and the absolute error between the output and the high-resolution ISAR image test set is calculated. Here, the absolute error refers to the average of the sum of the absolute values ​​of the differences between each pixel in the two images. When the error is large, the size of the discriminator's convolutional kernel and the number of convolutional layers are changed, thereby altering the receptive field size, and the network model is retrained. When the error does not decrease significantly, the desired RGAN model is obtained. The RGAN model trained using this method exhibits stronger generalization ability.

[0054] The aforementioned ISAR image super-resolution method based on recurrent generative adversarial networks (RBANs) acquires paired low-resolution and high-resolution ISAR image datasets. Training and testing sets for low-resolution and high-resolution ISAR images are constructed respectively. When constructing the RBAN model, a dynamic convolutional discriminator is used. The receptive field size of the dynamic fully convolutional network is adjusted according to the target scattering points, ensuring that any receptive field can contain the target scattering points. This ensures that each receptive field can be bounded during convolution, solving the problem that the receptive field bounding range is likely to be background rather than target during convolution, leading to an increased misclassification rate in the discriminator module. By introducing the RBAN structure, an end-to-end mapping from low-resolution to high-resolution ISAR images is achieved, resulting in ISAR images with higher target resolution. The trained network exhibits stronger generalization performance compared to existing ISAR image super-resolution networks, and its super-resolution performance is also superior to existing methods for ISAR images with resolutions different from the test data.

[0055] In one embodiment, paired target low-resolution ISAR image datasets and high-resolution ISAR image datasets are obtained, and training and test sets for low-resolution and high-resolution ISAR images are constructed respectively, including:

[0056] Through anechoic chamber measurements, fully polarimetric radar echo data of the target in various X-band attitudes were obtained. Based on the set imaging bandwidth and coherence accumulation angle, the echo data of the same attitude in the fully polarimetric radar echo data of each attitude were processed to obtain a pair of low-resolution ISAR image datasets of the target in a specific attitude and a high-resolution ISAR image dataset corresponding to the low-resolution ISAR images. Training sets and test sets of low-resolution ISAR images and high-resolution ISAR images were constructed respectively.

[0057] It is worth noting that the target acquired by this invention is radar echo data, such as... Figure 2 The diagram shows the process of generating low-resolution and high-resolution ISAR images. Echo data is obtained through anechoic chamber measurements. By setting different imaging bandwidths and coherence accumulation angles, data is truncated to obtain small-bandwidth echo data and corresponding large-bandwidth echo data. Low-resolution ISAR image datasets and high-resolution ISAR image datasets are then constructed respectively.

[0058] Specifically, echo data was obtained through anechoic chamber measurements. ISAR images with a center frequency of 10 GHz, an imaging coherence accumulation angle of 3°, and a bandwidth of 0.4 GHz were selected as the low-resolution ISAR image dataset; ISAR images with a bandwidth of 1.6 GHz and an imaging coherence accumulation angle of 12° were selected as the corresponding high-resolution ISAR image dataset. These were used to form the 0.4 GHz (3°) low-resolution ISAR image training set and the 1.6 GHz (12°) high-resolution ISAR image training set, aiming to train a network with a super-resolution ratio of 4. ISAR images with a bandwidth of 1 GHz and a coherence accumulation angle of 3° were used as the low-resolution ISAR image test set; ISAR images with a bandwidth of 4 GHz and a coherence accumulation angle of 12° at the corresponding attitude were used as the high-resolution ISAR image test set.

[0059] In one embodiment, a low-resolution ISAR image training set and a high-resolution ISAR image training set are input into a preset recurrent generative adversarial network (RGAN) model for training, resulting in a preliminary RGAN model, including:

[0060] The low-resolution ISAR image training set is divided into a low-resolution ISAR image training subset and a low-resolution ISAR image verification subset; the high-resolution ISAR image training set is divided into a high-resolution ISAR image training subset and a high-resolution ISAR image verification subset.

[0061] The low-resolution ISAR image training subset and the high-resolution ISAR image training subset are respectively input into the generator module to obtain the network loss function of the corresponding ISAR image. The preset recurrent generative adversarial network model is trained based on the network loss function.

[0062] The pre-defined cyclic generative adversarial network (CGAN) model was validated using a low-resolution ISAR image validation subset and a high-resolution ISAR image validation subset, resulting in a preliminary CGAN model.

[0063] Specifically, a pre-defined recurrent generative adversarial network model is constructed, including a generator module and a discriminator module.

[0064] The generator utilizes skip connections to ensure scattering point matching. Skip connections refer to the practice during deep network training of skipping certain layers of the neural network and using the output of one layer as the input of the next. The reason for using skip connections is that during deep network training, deeper layers cannot learn or even achieve identity mappings, leading to higher training errors. Neural networks with skip connections have a smoother loss surface and faster convergence speed. The discriminator is a dynamic fully convolutional network that maps input samples into a feature matrix through convolution. Each element of the feature matrix corresponds to a receptive field of the original image. The authenticity of the input sample is determined by averaging the entire feature matrix.

[0065] The 0.4 GHz (3°) low-resolution ISAR image training set was segmented to obtain a low-resolution ISAR image training subset and a low-resolution ISAR image validation subset; the 1.6 GHz (12°) high-resolution ISAR image training set was segmented to obtain a high-resolution ISAR image training subset and a high-resolution ISAR image validation subset. The ratio of training set to validation set was 3:1. Specifically, the data was arranged in ascending order of azimuth angle, and one data point was randomly selected from every four data points as the validation subset. The azimuth angles of the high-resolution and low-resolution subsets corresponded one-to-one.

[0066] The low-resolution ISAR image training subset and the high-resolution ISAR image training subset are respectively input into the generator module to obtain the network loss function of the corresponding ISAR image. The preset recurrent generative adversarial network model is trained based on the network loss function.

[0067] The pre-defined cyclic generative adversarial network (CGINN) model was validated using a low-resolution ISAR image validation subset and a high-resolution ISAR image validation subset, resulting in a preliminary CGINN model.

[0068] In one embodiment, the network loss function includes: an adversarial loss function, a cyclic loss function, and a super-resolution loss function.

[0069] Specifically, in a round of training, first there is Domain through Generate the corresponding , pass get Calculate the generator for updating the corresponding adversarial loss function, cyclic loss function, and super-resolution loss function. After that, regarding The data process in the domain is similar, except that the cyclic loss is removed during the process, and then the generator is updated. .

[0070] In one embodiment, the adversarial loss function includes: an adversarial loss function for low-resolution ISAR images and an adversarial loss function for high-resolution ISAR images;

[0071] The adversarial loss function for low-resolution ISAR images is expressed as follows:

[0072] ;

[0073] The adversarial loss function expression for high-resolution ISAR images is:

[0074] ;

[0075] in, This represents a subset of low-resolution images used for training. This represents a subset of high-resolution images used for training. From domain to Domain generator module, From domain to Domain generator module, , They are respectively Domain and Discriminator module on the domain, This represents the distribution of real samples in the training subset of low-resolution images. This represents the distribution of real samples in the training subset of high-resolution images.

[0076] In one embodiment, the cyclic loss function includes: a cyclic loss function for low-resolution ISAR images;

[0077] Its expression is:

[0078] .

[0079] In one embodiment, the super-resolution loss function includes: a super-resolution loss function for low-resolution ISAR images and a super-resolution loss function for high-resolution ISAR images.

[0080] The super-resolution loss function expression for low-resolution ISAR images is as follows:

[0081] ;

[0082] The super-resolution loss function expression for high-resolution ISAR images is:

[0083] .

[0084] In one embodiment, the network loss function includes: from domain to The generator module network loss function and from the domain domain to The network loss function of the domain generator module;

[0085] Among them, from domain to The network loss function of the domain generator module is expressed as:

[0086] ;

[0087] from domain to The network loss function of the domain generator module is expressed as:

[0088] .

[0089] In one embodiment, a low-resolution ISAR image test set is input into the preliminary recurrent generative adversarial network (RGAN) model for verification. The error between the model and the high-resolution ISAR image test set is calculated. When the error is large, the receptive field size of the discriminator module is adjusted, and the preliminary RGAN model is retrained to obtain the RGAN model, which includes:

[0090] The low-resolution ISAR image test set is input into the preliminary recurrent generative adversarial network model for verification. The absolute error between the output result and the high-resolution ISAR image test set is calculated. When the error is large, the receptive field size is adjusted by changing the discriminator convolution kernel size and the number of convolution layers. The preliminary recurrent generative adversarial network model is retrained. When the error does not decrease significantly, the recurrent generative adversarial network model is obtained. The absolute error refers to the sum and average of the absolute values ​​of the differences between each pixel in the two images.

[0091] In one embodiment, the generator module adopts the U-net structure; the discriminator module adopts the PatchGAN structure.

[0092] like Figure 3 The image shown is the PatchGAN structure for a corresponding image size of 512×512.

[0093] In one embodiment, the invention is verified, such as... Figure 4The image shows the super-resolution results after training the network on a dataset with a super-resolution ratio of 4, inputting an ISAR image with an imaging bandwidth of 1 GHz and a coherence accumulation angle of 5°, and comparing it with the corresponding ISAR image with an imaging bandwidth of 4 GHz and a coherence accumulation angle of 20°. It can be seen that the network still has good recovery capabilities for low-resolution ISAR images with larger bandwidths and coherence accumulation angles than those in the training data.

[0094] It should be understood that, although Figure 1 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order in which these steps are executed, and they can be performed in other orders. Figure 1 At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0095] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0096] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A super-resolution method for ISAR images based on recurrent generative adversarial networks, characterized in that, The method includes: Obtain paired target low-resolution ISAR image datasets and high-resolution ISAR image datasets; construct training and test sets for low-resolution ISAR images and high-resolution ISAR images respectively; The low-resolution ISAR image training set and the high-resolution ISAR image training set are input into a preset recurrent generative adversarial network model for training to obtain a preliminary recurrent generative adversarial network model; the discriminator module in the preliminary recurrent generative adversarial network model is a dynamic fully convolutional network, and the receptive field size of the dynamic fully convolutional network is determined according to the target scattering point. The low-resolution ISAR image test set is input into the preliminary recurrent generative adversarial network model for verification. The error with the high-resolution ISAR image test set is calculated. When the error is large, the receptive field size of the discriminator module is adjusted, and the preliminary recurrent generative adversarial network model is retrained to obtain the recurrent generative adversarial network model. The low-resolution ISAR image training set and the high-resolution ISAR image training set are input into a pre-defined recurrent generative adversarial network (RGAN) model for training, resulting in a preliminary RGAN model, including: The low-resolution ISAR image training set is divided into a low-resolution ISAR image training subset and a low-resolution ISAR image verification subset; the high-resolution ISAR image training set is divided into a high-resolution ISAR image training subset and a high-resolution ISAR image verification subset. The low-resolution ISAR image training subset and the high-resolution ISAR image training subset are respectively input into the generator module to obtain the network loss function of the corresponding ISAR image, and the preset recurrent generative adversarial network model is trained based on the network loss function. The preset cyclic generative adversarial network model is validated using the low-resolution ISAR image validation subset and the high-resolution ISAR image validation subset to obtain a preliminary cyclic generative adversarial network model. The network loss functions include: adversarial loss function, cyclic loss function, and super-resolution loss function.

2. The method according to claim 1, characterized in that, Obtain paired target low-resolution ISAR image datasets and high-resolution ISAR image datasets, and construct training and test sets for low-resolution ISAR images and high-resolution ISAR images respectively; including: By measuring in an anechoic chamber, the fully polarimetric radar echo data of the target in each X-band attitude is obtained; according to the set imaging bandwidth and coherence accumulation angle, the echo data of the same attitude in the fully polarimetric radar echo data of each attitude is processed to obtain a pair of low-resolution ISAR images of the target in a specific attitude and a high-resolution ISAR image corresponding to the low-resolution ISAR image. Training and testing sets for low-resolution and high-resolution ISAR images were constructed respectively.

3. The method according to claim 1, characterized in that, The adversarial loss function includes: an adversarial loss function for low-resolution ISAR images and an adversarial loss function for high-resolution ISAR images; The adversarial loss function for low-resolution ISAR images is expressed as follows: ； The adversarial loss function expression for high-resolution ISAR images is: ； in, This represents a subset of low-resolution images used for training. This represents a subset of high-resolution images used for training. From domain to Domain generator module, From domain to Domain generator module, , They are respectively Domain and Discriminator module on the domain, This represents the distribution of real samples in the training subset of low-resolution images. This represents the distribution of real samples in the training subset of high-resolution images.

4. The method according to claim 3, characterized in that, The cyclic loss function includes: a cyclic loss function for low-resolution ISAR images, the expression of which is: 。 5. The method according to claim 4, characterized in that, The super-resolution loss function includes: a super-resolution loss function for low-resolution ISAR images and a super-resolution loss function for high-resolution ISAR images. The super-resolution loss function expression for low-resolution ISAR images is as follows: ； The super-resolution loss function expression for high-resolution ISAR images is: 。 6. The method according to claim 5, characterized in that, The network loss function includes: from domain to The generator module network loss function and from the domain domain to The network loss function of the domain generator module; Among them, from domain to The network loss function of the domain generator module is expressed as: ； from domain to The network loss function of the domain generator module is expressed as: 。 7. The method according to claim 1, characterized in that, The low-resolution ISAR image test set is input into the preliminary recurrent generative adversarial network (RGAN) model for validation. The error with the high-resolution ISAR image test set is calculated. When the error is large, the receptive field size of the discriminator module is adjusted, and the preliminary RGAN model is retrained to obtain the RGAN model, which includes: The low-resolution ISAR image test set is input into the preliminary recurrent generative adversarial network model for verification. The absolute error between the output result and the high-resolution ISAR image test set is calculated. When the error is large, the receptive field size is adjusted by changing the discriminator convolution kernel size and the number of convolution layers. The preliminary recurrent generative adversarial network model is retrained. When the error does not decrease significantly, the recurrent generative adversarial network model is obtained. The absolute error refers to the sum and average of the absolute values ​​of the differences between each pixel in the two images.

8. The method according to any one of claims 1-7, characterized in that: The generator module adopts the U-net structure; the discriminator module adopts the PatchGAN structure.

Images