A multi-class, multi-azimuth SAR image generation method

Abstract

The present application relates to the technical field of image generation, and in particular to a multi-class and multi-azimuth angle SAR image generation method. The method first acquires an MSTAR data set; encodes the class label into a one-hot code, connects the one-hot vector and a latent code sampled from a uniform distribution in the channel dimension as conditional information, and inputs the conditional information and random noise into a generator to obtain a generated image; obtains a probability that the input image is consistent with the class label and is a real image, and a predicted latent code; alternately adjusts the parameters of the generator and the discriminator, so that the generator generates a fake image that conforms to the class information, and the latent code decouples the azimuth angle information; and controls the class label and the latent code to generate a multi-class and multi-azimuth angle SAR image. For different target image generation tasks, the present application does not need to use different target data sets to train the network multiple times, but generates a class and azimuth angle controllable SAR image through the class label and the latent code.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image generation, and particularly relates to a multi-class and multi-azimuth SAR image generation method based on a conditional generative adversarial network. BACKGROUND

[0002] Synthetic aperture radar (SAR) is an active microwave remote sensing technology, which has the advantages of being unaffected by weather and environment and being able to obtain ground information stably and continuously for a long time, and is widely used in military and civilian fields. Compared with optical imaging technology, SAR has greater difficulty in acquisition and higher equipment cost, resulting in a lack of complete SAR image data. Especially for a specific target, it is more difficult to collect a complete data set with ordered changes in azimuth due to the limitations of orbit shape and operating costs. Therefore, it is beneficial to further establish a complete SAR image target database by effectively expanding the SAR image target and generating SAR images with ordered changes in azimuth.

[0003] With the wide application of generative adversarial networks (GANs), researchers have used GANs to generate SAR images. For example, Guo [1] et al. proposed an end-to-end GAN model based on real images to simulate SAR images, and proposed a clutter normalization method to overcome the influence of clutter and speckle in SAR images on the training process, which helps to stabilize and accelerate the training process. Gao [2] et al. proposed a semi-supervised learning method based on DCGAN to solve the small sample target recognition task of SAR image labeled data set, improved the network structure of DCGAN, used two discriminators for joint training, generated false image expansion sample set similar to real images for data enhancement. Cui [3]Guo et al. generated SAR images using WGAN with gradient penalty, and designed a sample selection filter to segment and extract the contour of the generated SAR image, calculate the azimuth angle of the target, and select high-quality and specific azimuth samples in the generated image. However, the above algorithm needs to train the network with K real data training sets of K different categories of target images to obtain different target generation results, and the generation result has a large randomness. Although the generated image contains images with new azimuth angles, it cannot control the azimuth angle, such as generating images with ordered changes in azimuth angle within a certain angle range. At the same time, the network structure used by these algorithms is usually based on DCGAN, and DCGAN generates target images with limited size through simple stacking of convolutional layers, usually only suitable for generating 32*32 or 64*64 images, and the quality of the generated image is not high. In summary, the existing methods have the problems of being unable to generate different categories of target images through a single training, limited size and quality of the generated image, and random generation results.

[0004] [1]Guo J,Lei B,Ding C,et al.Synthetic aperture radar image synthesisby using generative adversarial nets[J].IEEE Geoscience and Remote Sensing Letters,2017,14(7):1111-1115.

[0005] [2]Gao F,Yang Y,Wang J,et al.A deep convolutional generativeadversarial networks(DCGANs)-based semi-supervised method for objectrecognition in synthetic aperture radar(SAR)images[J].Remote Sensing,2018,10(6):846.

[0006] [3]Cui Z,Zhang M,Cao Z,et al.Image data augmentation for SAR sensorvia generative adversarial nets[J].IEEE Access,2019,7:42255-42268. SUMMARY

[0007] In order to solve the problems existing in the prior art, the purpose of the present application is to provide a multi-class and multi-azimuth SAR image generation method based on conditional generative adversarial network, which does not need to train the network multiple times using different class target images, controls the generation result through the class label and the latent encoding, and strives to generate a SAR image with controllable class and azimuth, large size and high quality.

[0008] In order to achieve the above purpose, the present application adopts the following technical solutions:

[0009] The present application provides a conditional generative adversarial network, which comprises a generator, a discriminator and an auxiliary network.

[0010] The generator is mainly used for inputting the class label and the latent encoding as conditional information together with random noise into the generator to guide image generation; the generator comprises one linear layer and five neural network layers, each neural network layer comprises a residual module and an up-sampling layer, and the residual module and the up-sampling layer are alternately constructed; the residual module comprises two 3*3 convolution layers with the same number of output channels, and each convolution layer is followed by a batch normalization layer and a LeakyReLU activation function; and the up-sampling layer is used to continuously expand the size of the feature map.

[0011] The discriminator is mainly used for judging whether the input image is a real image consistent with the class label, and comprises two linear layers and five neural network layers; the first neural network layer is a convolution layer, and the last four neural network layers comprise a residual module and an average pooling layer, and the residual module and the average pooling layer are alternately constructed; the residual module comprises two 3*3 convolution layers with the same number of output channels, and each convolution layer is followed by a batch normalization layer and a LeakyReLU activation function; and the average pooling layer is used to continuously reduce the size of the feature map.

[0012] An auxiliary network is introduced at the end of the discriminator; the auxiliary network is mainly used for learning to predict the latent encoding of the input image; and the auxiliary network shares the remaining network structure with the discriminator except for the last linear layer.

[0013] The present application further provides a multi-class and multi-azimuth SAR image generation method using the conditional generative adversarial network, which comprises the following steps:

[0014] Step 1: obtaining an MSTAR data set, which comprises 10 classes of military target SAR images;

[0015] Step 2: converting each class label into a one-hot encoding vector with the same number of class labels, using N-bit 0 or 1 values to encode N different discrete features, and the one-hot encoding vector only has one valid bit at any time, indicating a certain state;

[0016] Step 3, latent code is obtained by uniformly distributed random sampling, and the one-hot vector representing the category is connected with the latent code in the channel dimension as a conditional vector;

[0017] Step 4, random noise is obtained by random sampling from the standard normal distribution, and the conditional vector and the random noise are input into the generator of the improved Resnet network structure to learn the conditional information and generate a high-resolution SAR image;

[0018] Step 5, the generated image or the real image and the one-hot encoding vector representing the category are input into the discriminator to obtain a feature map, which is unfolded into a column vector and output through a linear layer to obtain the probability that the input image conforms to the category label and is a real image;

[0019] Step 6, the column vector obtained in step 5 is input into the linear layer of the auxiliary network to obtain the predicted latent code of the input image;

[0020] Step 7, the conditional generative adversarial network is trained by an adversarial training strategy to adjust the parameters of the generator and the discriminator;

[0021] Step 8, after the training is completed, random noise is input, and a SAR image of a specific category and azimuth angle is generated by specifying the one-hot encoding vector and the latent code.

[0022] Further, the step 4 further comprises:

[0023] The random noise, the one-hot encoding vector and the latent code are combined into a vector of a certain length, which is input into the linear layer of the generator to obtain a vector of a certain length, and then the shape of the vector is adjusted according to the channel number*high*width, and then the vector is input into the residual module and the up-sampling layer of each neural network layer, and finally the vector channel number is converted to 3 through a 3*3 convolution layer with one output channel number in the 5th neural network layer to generate a high-resolution SAR image.

[0024] Further, the step 5 further comprises:

[0025] The one-hot encoding vector is input into a linear layer to obtain a column vector, and the size of the column vector is adjusted to be the same as that of the input SAR image, and then the column vector is connected with the input SAR image in the channel dimension, and finally a feature map is obtained through the 5-layer neural network layer, and then the feature map is unfolded into a column vector, and a linear layer is used to obtain the probability that the input image conforms to the category label and is a real image.

[0026] Further, the step 7 further comprises:

[0027] Step 7.1, fix the generator, train the discriminator, input the generated image and output 0, input the real image and output 1;

[0028] The loss function is as follows:

[0029]

[0030] Wherein: D is a discriminator, G is a generator, G * represents that the generator parameter is fixed, P z (z) is a standard normal distribution, z is random noise, c is a latent code, y is a category label, P data is a real data distribution, x is real data, G(z, c|y) is generated data, D(*) represents the output result of the discriminator, the probability that the output result conforms to the category label and is a real image, e represents the mathematical expectation, E x～p(x) [f(x)] = ∫p(x)f(x)dx;

[0031] Step 7.2, fix the discriminator, train the generator, and input the result generated by the generator into the discriminator, so that the output of the discriminator is 1, that is, the discriminator considers the generated image to be a real image, and the result of generating false images is generated.

[0032] The loss function is as follows:

[0033]

[0034] Wherein: D * represents that the discriminator parameter is fixed, and the target of the generator G is to minimize the above loss function V, that is, to maximize D * (G(z, c|y)), so that under the condition of category label constraint, D * (G(z, c|y))=1, that is, the output of the discriminator is 1 when the result generated by the generator is input into the discriminator.

[0035] Step 7.3, simultaneously train the generator and the discriminator, maximize the mutual information of the latent code predicted by the discriminator and the generated data, and decouple the obvious semantic information in the data, that is, the azimuth angle information.

[0036] The loss function is:

[0037] min G max p V I (D,G) = V(D,G)-λI(c;G(z,c|y))

[0038] Wherein: V(D, G) is a standard adversarial loss:

[0039]

[0040] Lambda is a hyperparameter, and I(c; G(z, c|y)) is the mutual information between the latent code and the generated data.

[0041] Compared with the prior art, the beneficial effects of the present application are:

[0042] 1) For different target SAR image generation tasks, different target data sets are not required to train the network respectively. By introducing class label information for the input of the generator, the generation network is guided to map the noise into different class domains, realizing one network generating different targets;

[0043] 2) To solve the problem of low explainability of GAN model generation results, increase the correlation constraint between input and generated image, decouple the azimuth angle information of SAR image by adding one-dimensional latent encoding to the input, and generate SAR images with azimuth angle sequence changes by linear interpolation of latent encoding.

[0044] 3) To solve the problem of limited image size generated by the network constructed by simply stacking convolutional layers, such as DCGAN, two relatively symmetric residual networks are designed using residual modules to generate large size images. BRIEF DESCRIPTION OF DRAWINGS

[0045] Figure 1 is the overall network diagram of conditional generative adversarial network.

[0046] Figure 2 is the corresponding display of the data set optical image and SAR image.

[0047] Figure 3 is the one-hot encoding of ten different targets.

[0048] Figure 4 is the network structure diagram of the residual module.

[0049] Figure 5 is the network structure diagram of the generator.

[0050] Figure 6 is the network structure diagram of the discriminator and auxiliary network.

[0051] Figure 7 is the generation result of the network with simply stacked convolutional layers, which generates target images with azimuth angle ordered changes of ten targets in order of encoding in units of lines. Figure 3

[0052] Figure 8 is the generation result of the improved residual network, which generates target images with azimuth angle ordered changes of ten targets in order of encoding in units of lines. Figure 3

[0053] Figures 9-11 are enlarged displays of the generation results of targets 2S1, BRDM_2 and T62 respectively, and the image size is 128 pixels*128 pixels. DETAILED DESCRIPTION

[0054] ​​The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0055] A schematic diagram of a conditional generative adversarial network is shown below. Figure 1 As shown, a conditional generative adversarial network consists of a generator, a discriminator, and an auxiliary network.

[0056] The generator is primarily used to input category labels and latent codes as conditional information, along with random noise, to guide image generation. The improved generator comprises one linear layer and five neural network layers. Each neural network layer includes a residual module and an upsampling layer, constructed alternately by the residual module and the upsampling layer. The residual module is as follows... Figure 4 As shown, it includes two 3x3 convolutional layers with the same number of output channels, each followed by a batch normalization layer and a LeakyReLU activation function. The residual block introduces skip connections, allowing the input information to bypass some layers and be added to the result. This input information is the feature map obtained by passing the generator's input through the network before the residual block. Skip connections reduce the loss of input information during inter-layer propagation. The input information is the feature map obtained by passing the generator's input through the network before the residual module. In this invention, the residual block is a module constituting the network, and it is used multiple times throughout the network structure. In the generator, for example... Figure 5 As shown, the entire network consists of upsampling layers, residual modules, and convolutional layers. The network preceding the residual modules is the network constructed from the remaining layers in the diagram. The upsampling layers are used to continuously expand the feature map size. The residual modules fuse features learned from higher-level networks through skip connections, giving the network structure stronger information encoding and generation capabilities, improving the quality of generated images while generating large-size images. The generator structure is as follows: Figure 5 As shown, conditional information is introduced into the generator, including a one-hot encoded vector and a latent encoding. The one-hot encoded vector provides category information, and the latent encoding provides azimuth information.

[0057] The discriminator, primarily used to determine whether the input image is a real image consistent with the category label, comprises two linear layers and five neural network layers. The first neural network layer is a convolutional layer, and the subsequent four neural network layers include residual modules and average pooling layers, constructed alternately. The residual module has the same structure as the generator's residual module; the residual module includes two 3*3 convolutional layers with the same number of output channels, each followed by a batch normalization layer and a LeakyReLU activation function; the average pooling layers are used to continuously reduce the feature map size. The discriminator structure is as follows: Figure 6As shown. In addition to the sample image, the input of the discriminator also includes a one-hot encoding vector representing the category information. The discriminator not only needs to determine whether the input sample image is real, but also needs to determine whether it is consistent with the object described by the category information. The discriminator only outputs 1 when the input sample image is a real image and meets the category information, so that the category label can guide the generator to generate images of a specific category.

[0058] An auxiliary network is introduced at the end of the discriminator. The auxiliary network is mainly used to learn to predict the latent encoding of the input image. The auxiliary network shares the network structure with the discriminator except for the last linear layer. The auxiliary network structure is as shown. Figure 6

[0059] The generator inputs the latent encoding and outputs the generated image, which hides some relationships of the latent encoding in the generated image. The auxiliary network inputs the generated image to learn the latent encoding, maximizes the mutual information between the latent encoding and the generated data, extracts these relationships from the image, and decouples the latent encoding that explicitly affects the generation result, so that the latent encoding can guide the generator to generate images with specific semantic information.

[0060] Through the adversarial training strategy, the generator and the discriminator are repeatedly iteratively trained, so that the generator gradually learns to generate images with specific categories and azimuth angles.

[0061] The multi-category and multi-azimuth SAR image generation method provided by the present application first acquires the MSTAR data set including 10 categories of military target SAR images as the training data set. Secondly, the category label is one-hot encoded, the one-hot vector and the latent encoding sampled from the uniform distribution are connected in the channel dimension as the conditional information, and the generated image is obtained by inputting the improved generator with random noise. Then, the image and the category label are input into the discriminator and the auxiliary network to obtain the probability that the input image is consistent with the category label and is a real image and the predicted latent encoding. Next, the network is trained by adopting the adversarial strategy, and the parameters of the generator and the discriminator are alternately adjusted, so that the generator generates images that meet the category information and are difficult to distinguish from real images, and the latent encoding decouples the obvious semantic information, i.e. azimuth angle information, in the generation distribution. Finally, after the training is completed, the multi-category and multi-azimuth SAR image generation is realized by controlling the category label and the latent encoding. For different target image generation tasks, the present application does not need to use different target data sets to train the network multiple times. The category and azimuth angle controllable SAR image is generated by the category label and the latent encoding. Specifically, the steps include:

[0062] ​Step 1, obtain the MSTAR dataset, which includes target images of multiple vehicle targets obtained at different azimuth angles, the targets including 2S1 (self-propelled howitzer), BRDM2 (armored reconnaissance vehicle), BTR60 (armored transport vehicle), D7 (bulldozer), T62 (tank), ZIL131 (cargo truck), ZSU234 (self-propelled anti-aircraft gun), T72 (tank), BTR70 (armored transport vehicle), and BMP2 (infantry fighting vehicle). The optical image corresponds to the SAR image as shown in Figure 2 .

[0063] Step 2, convert each class label into a one-hot encoding vector equal to the number of class labels. One-hot encoding is suitable for discrete features with no size significance. N different discrete features are encoded using N-bit 0 or 1 values. The one-hot encoding vector has only one valid bit at any time, representing a certain state. Taking the MSTAR dataset as an example, the correspondence between different categories and one-hot encoding is shown in Figure 3 .

[0064] Step 3, obtain latent encoding by uniformly distributing random sampling from [-1, 1], and concatenate the latent encoding and the one-hot encoding vector representing the class label in the channel dimension as the condition vector.

[0065] Step 4, obtain random noise by random sampling from a standard normal distribution, and input the condition vector and random noise into the generator of the improved Resnet network structure.

[0066] As shown in Figure 5 , the random noise (length 64), one-hot encoding (length 10), and latent encoding (length 1) are combined into a vector of length 75, which is processed through a linear layer to obtain a vector of length 1024. The vector is adjusted to have a shape of 64*4*4 (channel number*height*width), and its output size is Figure 5 the label under each layer of the network. At the end of the 5th network, a convolutional layer is used to convert the channel number from 64 to 3 (RGB image 3 channels), generating a high-resolution SAR image.

[0067] Step 5, input the generated image or real image, and the one-hot encoding vector representing the class, into the improved discriminator, whose output is the probability that the input image conforms to the class label and is a real image, to obtain the true or false judgment result.

[0068] The input image size is 3*128*128, and the one-hot encoded vector is a column vector of length 10. The one-hot encoded vector is input into the discriminator's linear layer and transformed into a vector of length 16384. The shape is then adjusted to 1*128*128 and concatenated with the input SAR image in the channel dimension, resulting in a size of 4*128*128. After passing through the first convolutional layer, a feature map of size 64*128*128 is obtained. After passing through the residual module and average pooling layer of the next four neural network layers, a feature map of size 64*7*7 is obtained. This feature map is unfolded into a column vector of length 3136 and passed through a linear layer to obtain the result. The result represents the probability D(x|y) that matches the class label and is a real image.

[0069] Step 6: The auxiliary network and the discriminator share the network structure except for the last linear layer. The column vector of length 3136 from Step 5 is passed through the linear layer of the auxiliary network to obtain the latent code of the predicted input image.

[0070] Step 7: Train the Conditional Generative Adversarial Network (CGN) using an adversarial training strategy, adjusting the parameters of the generator and discriminator. The Adam optimizer is used with a learning rate of 0.0002 and parameters β1 = 0.5 and β2 = 0.999. The specific training steps are as follows:

[0071] Step 7.1: Fix the generator, train the discriminator, input the generated image and output 0, input the real image and output 1, the loss function is as follows:

[0072]

[0073] Where: D is the discriminator, G is the generator, and Gi is the generator. * This indicates that the generator parameters are fixed, P z (z) represents a standard normal distribution, z is random noise, c is the latent code, y is the class label, and P data Let x be the real data distribution, G(z, c|y) be the generated data, and D(*) represent the output of the discriminator, which is the probability that the image matches the label and is a real image. The goal of the discriminator D is to maximize the loss function V, i.e., maximize D(x|y) and minimize D(G(z, c|y)). * (z, c|y)) such that, under the category label constraint, D(x|t) = 1, D(G * (z, c|y)) = 0, meaning the output is 0 when the input is the generated image and 1 when the input is the real image.

[0074] Step 7.2: Fix the discriminator and train the generator so that when the generator's output is input into the discriminator, the discriminator's output is 1, meaning the discriminator considers the generated image to be a real image, thus generating a highly realistic result. The loss function is as follows:

[0075]

[0076] where: D * denotes the discriminator parameters are fixed, and the rest are defined as in step 7.1. The goal of the generator G is to minimize the above loss function V, i.e. maximize D * (G(z, c|y)) under the constraint of the category label, i.e. D * (G(z, c|y)) = 1, i.e. the output of the discriminator is 1 when the generator generates the result and inputs it to the discriminator;

[0077] Step 7.3, the generator and the discriminator are trained simultaneously, a regularization method in information theory is adopted to maximize the mutual information between the latent code and the generation distribution.

[0078] In information theory, the mutual information I(X; Y) between X and Y measures the amount of information learned from the random variable Y about the random variable X, and the mutual information can be expressed as the difference between the two entropies:

[0079] I(X; Y) = H(X) - H(X|Y) = H(Y) - H(Y|X)

[0080] If X and Y are independent, I(X; Y) = 0, i.e. knowing one variable does not reveal the other variable. In contrast, if X and Y are associated by a deterministic reversible function, the maximum mutual information is obtained. Maximizing the mutual information between the latent code and the generation distribution means that the information of the latent code is not lost in the generation process. By constraining the latent code through mutual information, it can be understood as a process of self-encoding, i.e. inputting the latent code into the generator to obtain the generated image, and then learning the predicted latent code through the auxiliary network. It is the inverse process of encoding-decoding, if the latent code input into the generator can have a clear impact on the generated result, i.e. the correlation between the two is very high, the predicted code output by the auxiliary network should be consistent with the latent code input into the generator as much as possible. By training, the latent code can significantly affect the generated result, and the semantic information in the data distribution is decoupled. Since the training data set contains SAR images of different angles, the decoupled semantic information is the azimuth information. On the basis of the standard adversarial loss, a mutual information loss is added to achieve it, and the total loss function is:

[0081] min G max D V I (D, G) = V(D, G) - λI(c; G(z, c|y))

[0082] where: V(D, G) is the standard adversarial loss:

[0083]

[0084] where λ is a hyper-parameter and I(c; G(z, c|y)) is the mutual information between the latent code and the generated data. λ is set to 0.1 and the whole training process is iterated 500 times.

[0085] Step 8, after the training, input random noise, generate SAR images of specific class and azimuth angle by specifying one-hot encoding vector and latent code. The generation results are shown in Figures 7-11 Figure 7 The generation results of the network with simple stacked convolutional layers are shown in Figure 8 The generation results of the improved residual network are shown in, where each row is a SAR image generated by 10 linear interpolations of the latent code for a specific class and azimuth angle. Figure 8 Compared with Figure 7 The generated images have higher quality and more texture features of the targets in real images, and the edges of the intermediate targets are clearer. Figures 9-11 The zoomed-in display of the generation results of targets 2S1, BRDM_2 and T62 are shown in Figures 9-11 It can be seen that the network can decouple the azimuth information of the SAR image, and linear interpolation in the dimension where the azimuth information is located can achieve controllable generation of the azimuth of the SAR image.​

Claims

1. A multi-class, multi-azimuth SAR image generation method of a conditional generative adversarial network, characterized in that, The method comprises the following steps: Step 1, obtaining a MSTAR data set, which comprises 10 categories of military target SAR images; Step 2, converting each category label into a one-hot encoding vector equal in number to the category label, using N-bit 0 or 1 values to encode N different discrete features, and the one-hot encoding vector being valid at any time only one bit, representing a certain state; Step 3, obtaining latent encoding from uniform distribution random sampling, connecting the latent encoding and the one-hot vector representing the category in the channel dimension as a conditional vector; Step 4, obtaining random noise from a standard normal distribution, taking the conditional vector and the random noise as inputs, learning the conditional information through a generator of an improved Resnet network structure, and generating a high-resolution SAR image; Step 5, inputting the generated image or the real image and the one-hot encoding vector representing the category into a discriminator to obtain a feature map, expanding the feature map into a column vector, and outputting the probability that the input image conforms to the category label and is a real image through a linear layer; Step 6, obtaining the column vector in step 5 through a linear layer of an auxiliary network to obtain the predicted output encoding of the input image; Step 7, training the conditional generative adversarial network through an adversarial training strategy to adjust the parameters of the generator and the discriminator; Step 8, after the training is completed, inputting random noise to generate a SAR image of a specific category and azimuth angle through a specified one-hot encoding vector and latent encoding; The conditional generative adversarial network comprises a generator, a discriminator and an auxiliary network; The generator is mainly used for inputting the category label and the latent code as the condition information together with random noise into the generator to guide image generation; the generator comprises one linear layer and five neural network layers, each neural network layer comprises a residual module and an up-sampling layer, and the residual module and the up-sampling layer are alternately constructed; the residual module comprises two 3 3 convolutional layers, each convolutional layer being followed by a batch normalization layer and a LeakyReLU activation function; and the up-sampling layer is used to continuously expand the size of a feature map; The discriminator is mainly used for judging whether the input image is a real image consistent with the category label, and includes two linear layers and five neural network layers, the first neural network layer is a convolution layer, and the last four neural network layers include residual modules and average pooling layers, and are alternately constructed by the residual modules and the average pooling layers; the residual module includes two 3 3 convolution layers, each of which is followed by a batch normalization layer and a LeakyReLU activation function; and the average pooling layer is used to continuously reduce the size of the feature map. An auxiliary network is introduced at the end of the discriminator; the auxiliary network is used to learn the latent encoding of the input image; and the auxiliary network shares the remaining network structure with the discriminator except for a last linear layer.

2. The multi-class, multi-azimuth SAR image generation method of claim 1, wherein, The step 4 further comprises: The random noise, the one-hot encoding vector and the latent encoding are merged into a vector of a certain length, and a linear layer of the generator is used to obtain a vector of a certain length, and the number of channels High After adjusting the shape of the vector, the residual module and the up-sampling layer of each neural network layer are used, and the output channel number of the 5th neural network layer is 3 3 convolutional layers are used to convert the vector channel number to 3, and a high-resolution SAR image is generated.

3. The multi-class, multi-azimuth SAR image generation method of claim 1, wherein, The step 5 further comprises: The one-hot encoding vector is inputted into a linear layer to obtain a column vector, the size of the column vector is adjusted to be the same as that of the input SAR image, the column vector is connected with the input SAR image in the channel dimension, a feature map is obtained through 5 neural network layers, the feature map is expanded into a column vector, and a linear layer is used to obtain the probability that the input image conforms to the category label and is a real image.

4. The multi-class, multi-azimuth SAR image generation method of claim 1 or 3, wherein, The step 7 further comprises: Step 7.1, fixing the generator, training the discriminator, inputting the generated image to output 0, and inputting the real image to output 1; The loss function is as follows: wherein: is a discriminator, is a generator, denotes that the generator parameters are fixed, is a standard normal distribution, is a random noise, is a latent code, is a class label, is a real data distribution, is real data, G is generated data, denotes the output of the discriminator, the probability that the generated data is a real image with the class label, denotes the mathematical expectation, ; Step 7.2, fixing the discriminator, training the generator, so that the output of the discriminator is 1 when the generated result is inputted into the discriminator, that is, the discriminator considers the generated image as a real image, and the result of generating a false image is generated; The loss function is as follows: wherein: represents that the discriminator parameter is fixed, and the target of the generator G is to minimize the above loss function V, that is, to maximize so that under the condition of the category label constraint, = 1, that is, even if the generated result of the generator is input to the discriminator, the output of the discriminator is 1; Step 7.3, training the generator and the discriminator at the same time, maximizing the mutual information between the latent encoding predicted by the discriminator and the generated data, and decoupling the obvious semantic information in the data, that is, the azimuth angle information; The loss function is as follows: wherein: is the standard contrastive loss: is an upper bound on the mutual information between the latent code and the generated data, is an upper bound on the mutual information between the latent code and the generated data.

Images