A method for batch synthesis of specific angle SAR images by using a generative adversarial network

Abstract

The application discloses a method for batch synthesizing specific-angle SAR images by using a generative adversarial network, and comprises the following steps: S1, setting a generative network G, and mapping a random n-dimensional vector Z obtained by random sampling from a hidden space Z into SAR pictures I of different scales, so that all the SAR pictures form a space J; S2, setting a multi-scale discriminative network J->R for the SAR pictures, and outputting a probability that an input picture is a real picture; S3, training in the image space by using the generative adversarial network, so that the generative network can simulate real SAR pictures; S4, decoupling the generative network, and obtaining a direction vector n and a boundary vector Z related to the semantic of the azimuth angle of the SAR picture; and S5, batch synthesizing a SAR image sequence of a specific azimuth angle by using the generative network G, the direction vector n and the boundary vector Z, wherein the method is characterized in that a novel generative adversarial network structure and an efficient SAR target image simulation algorithm are carefully designed according to the characteristics of the SAR picture, and a high-resolution target can be generated according to a given azimuth angle. inf inf ​​

Description

Technical Field

[0001] This invention relates to the field of image processing technology, specifically a method for batch synthesizing SAR images at a specific angle using generative adversarial networks. Background Technology

[0002] Generative Adversarial Networks (GANs) consist of a generator network and a discriminator network. The generator network fits the target data distribution, while the discriminator network measures the difference between the generated data distribution and the real data distribution. Currently, researchers have completed numerous studies on GANs in areas such as super-resolution and text-to-image conversion. A core application of GANs in computer vision is generating image datasets, including objects such as numbers, signatures, faces, and outdoor scenes. Furthermore, GANs have excellent applications in cutting-edge fields such as image restoration, image dehazing and raindrop effects, and natural language processing.

[0003] Generative Adversarial Networks (GANs) have become a new research hotspot in computer vision. Compared to traditional methods, GANs do not rely on labeled data but optimize parameters through internal adversarial processes within the model itself. However, achieving Nash equilibrium during GAN training is difficult, making the training process challenging. Yau's team trained and optimized only the discriminator network in a low-dimensional case, deriving an explicit solution for the generator network and demonstrating that the training method of WGAN did not yield the optimal solution. To address this issue, Rodfird proposed DCGAN, experimentally enumerating the generator and discriminator network frameworks to obtain a better set of network structures, but this did not fundamentally solve the problem. Martín Arjovsky et al. proposed WGAN, solving the problem of unstable GAN training and largely resolving issues such as pattern collapse and training instability in the original GAN ​​models. Subsequently, Ishaan Gulrajani et al. proposed WGAN-GP, which addresses the potential gradient vanishing or exploding problems in WGAN.

[0004] To further improve the quality of generated images, Karras proposed StyleGAN, a GAN model that incorporates style transfer. StyleGAN can stably generate high-quality, high-resolution images. However, when testing and optimizing SAR image recognition software, it is necessary to simulate images with angles inconsistent with existing images, and the inherent speckle noise of SAR images increases the difficulty of this problem.

[0005] Before the era of deep learning, most SAR image simulation techniques were based on ray tracing or rasterization methods. To simulate high-quality images, these methods typically required specialized expertise to manually construct accurate computer-aided drawing (CAD) models for each target, increasing time and financial costs. With the development of deep learning algorithms, generative models have gradually been used to simulate SAR images, obtaining target SAR images through support vector machines and principal component analysis. Additionally, conditional GANs (DCGANs) are used to simulate SAR images. However, these methods have two main drawbacks: 1) the structure of the GAN used is too trivial for SAR images, making it difficult to produce high-quality simulation results; 2) they cannot accurately simulate specific SAR images based on the required angle. Summary of the Invention

[0006] The purpose of this invention is to propose a method for batch synthesis of SAR images at specific angles using generative adversarial networks. This method carefully designs a novel generative adversarial network (GAN) structure and an efficient SAR target image simulation algorithm based on the characteristics of SAR images, which can generate high-resolution targets according to a given azimuth angle.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] A method for batch synthesis of SAR images at a specific angle using generative adversarial networks includes the following steps:

[0009] S1. Set up a generator network G, and map the random n-dimensional vector Z obtained by random sampling from the latent space Z to SAR images I at different scales. All SAR images constitute space J.

[0010] S2. Set up a multi-scale discrimination network J→R for SAR images, and output the probability that the input image is a real image;

[0011] S3. Generative adversarial networks are trained in the image space to enable the network to simulate real SAR images.

[0012] S4. Decouple the generator network to obtain the direction vector n and boundary vector Z, which are semantically related to the azimuth angle of the SAR image. inf ;

[0013] S5. Utilizing the generator network G, direction vector n, and boundary vector Z inf SAR image sequences with specific azimuth angles are synthesized in batches.

[0014] Preferably, step S1 is achieved through the following steps:

[0015] S11. A multi-scale generative network structure is adopted, and the input noise of the first layer is removed;

[0016] S12. Perform an affine transformation on the input latent vector Z. The formula for the affine transformation is as follows:

[0017] G1(Z) = Vz + b

[0018] Where V is a matrix, b is a t-dimensional bias vector, and G1: R k →R t As a feature-based unentanglement transformation, the k-dimensional latent space Z is projected onto the t-dimensional feature space. Where χ is the semantic space, X∈χ is the semantic vector, and each dimension of χ independently controls a semantic of the generated image;

[0019] S13. The semantic vector X is sent to the generation module. Generation modules of different scales are sequentially combined to form the entire generation network. All inputs to the generation modules contain the same semantic vector, and the first generation module synthesizes an image with the lowest resolution m0×m0 from the semantic vector. The intermediate generation modules receive the image generated by the previous generation module and combine it with the semantic vector X to synthesize an image with a higher resolution. The last generation module simulates an image with the target resolution, using the following formula:

[0020]

[0021]

[0022] Where i represents the sequence number of the generation module, The image with the lowest resolution m0 is referred to as B0(S), where S is the semantic vector and B0(S) is the initial generation module and its input semantic vector S. Refers to a resolution of 2 i The image of m0 Refers to the i-th generating module and its input semantic vector S and image.

[0023] S14. The semantic vector X is generated through the key submodule Adain in each generation module. i Normalization, and then through feature y i The corresponding scalars are scaled and biased, and semantic vectors and feature maps are fused through the key submodule Adain to control the style of the generated image. The key submodule Adain is:

[0024]

[0025] Among them, AdaIN(X) i ,y) refers to the AdaIN module, y s,i The variance of the input semantic vector at layer i, y b,i X refers to the mean of the input semantic vector of the i-th layer.i The input feature map of the i-th layer, μ refers to the mean calculation, and α refers to the variance calculation.

[0026] Preferably, the multi-scale discriminant network in step S2 is symmetrical to the generator network and is a multi-scale classifier composed of several discriminant modules; the discriminant module consists of a representation structure DMAM for SAR images and several convolutional layers; the representation structure DMAM is described by the following formula:

[0027] M i =S i (A(F 7×7 (I)))

[0028]

[0029] Among them, F 7×7 It is a 2D convolution with a kernel size of 7×7, A is a self-attention module, S i It is a deformable convolution with a kernel size of 2i+1; using F, A, S i Three feature maps with different spatial information M1 to M3 were obtained; The tensor connection operator is represented by P, which is the ASPP module, and U is the final feature map, i.e., the output of DMAM. DMAM is represented by the uppercase letter H, and the above two functions are used to fuse spatial information.

[0030] Based on the representation structure DMAM, the discriminant modules at different scales are represented as follows:

[0031] Y1=K(H(I1)),i=1

[0032] Y i =K(αH(I) i )+(1-α)R i (Y i-1 ),i>1

[0033] y i =L(Y i ), i≥1

[0034] Where Y1 refers to the probability that the lowest resolution input image is a real image, K refers to a convolutional layer of size 3, i refers to the image sequence at each resolution, and Y i Refers to a resolution of 2 i The image of m0, α refers to the decreasing weighting factor, R i It is a linear interpolation function used to interpolate Y i-1 The resolution is doubled; L is a fully connected layer with a sigmoid activation function, providing the final output; scalar y i Indicates input I i It is the probability of a real SAR image.

[0035] Preferably, the generative adversarial network in step S3 is trained using the following formula:

[0036]

[0037]

[0038] Where G and D represent the generator network and the discriminator network, respectively, I is the input image, and z is the input vector.

[0039] Preferably, step S4 is implemented through the following steps:

[0040] S41. For the linear transformation matrix V in the equation G1(Z) = Vz + b in step S12, find V T The V feature vector is obtained using the following formula:

[0041]

[0042] The eigenvector corresponding to its largest eigenvalue is the direction vector n;

[0043] Where, λ i Refers to the i-th largest eigenvalue, n i λ i The corresponding feature vector, where c refers to the number of selected features;

[0044] S42. Given a random latent vector Z and a step size α, generate an image sequence {I} according to the following formula. i}:

[0045] I i =G(Z-iαn)

[0046] Among them, I i Let G refer to the i-th image, G to the generator network, i to the index, n to the semantic direction vector, and the sequence {I} be the image. i In the image, as the number i increases, the semantic azimuth angle gradually decreases. The image with the smallest azimuth angle of 0 degrees is recorded as having a corresponding boundary vector Z. inf .

[0047] Preferably, step S5 is implemented through the following steps: based on the generator network G and its semantic direction vector n, boundary vector Z inf A SAR image sequence with azimuth angles in the interval [a, b] and linearly increasing azimuth angles is generated using the following formula:

[0048] {I i}=G(Z inf +(iα+α-θ)n),i=0,1,...(ba) / α

[0049] Where i refers to the generated image number, α refers to the step size, θ refers to the orientation angle of the specific SAR image to be synthesized, and n refers to the semantic direction vector.

[0050] By adopting the above technical solution, the present invention has the following advantages compared with the prior art:

[0051] 1. Based on the characteristics of SAR images, this invention has carefully designed a novel generative adversarial network (GAN) structure and an efficient SAR target image simulation algorithm, which can generate high-resolution targets according to a given azimuth angle.

[0052] 2. The Generative Adversarial Network (GAN) used in this invention can be trained end-to-end without the need for specialized knowledge, making it simple and convenient to use.

[0053] 3. This invention addresses the problem of insufficient training samples caused by insufficient source data in SAR identification tasks. By generating adversarial networks to simulate specific SAR images, the simulated data samples can be directly used as training samples. This invention addresses the problem of inconsistent features between modeled and simulated images and key technologies such as transfer learning of feature-based generative adversarial networks. It constructs an operable SAR feature transformation, improves the intelligent and automated processing level of SAR data, provides a solid theoretical foundation for subsequent military target detection, classification and identification, and realizes efficient application in military reconnaissance and combat. Attached Figure Description

[0054] Figure 1 The network structure diagrams of the generative adversarial network (a) and image simulation algorithm (b) of this invention are shown below;

[0055] Figure 2 This is a network structure diagram of the discrimination network (a) and the noise reduction morphological attention module (b) of the present invention;

[0056] Figure 3 This is a network structure diagram of the affine module (a) and the generator network (b) of the present invention;

[0057] Figure 4 The images are SAR images simulated by four schemes: DCGAN, ProGAN, StyleGAN, and the present invention.

[0058] Figure 5 This is a qualitative comparison chart of the first set of real SAR images and synthetic SAR images;

[0059] Figure 6 This is a qualitative comparison chart of the second set of real SAR images and synthetic SAR images. Detailed Implementation

[0060] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0061] This invention mainly relates to a method for batch synthesizing SAR images at a specific angle using generative adversarial networks. The following is a detailed description of this embodiment with reference to the accompanying drawings.

[0062] Please see Figures 1 to 6 This invention discloses a method for batch synthesis of SAR images at a specific angle using a generative adversarial network, comprising the following steps:

[0063] S1. Set up a generator network G, and map the random n-dimensional vector Z obtained by random sampling from the latent space Z to SAR images I at different scales. All SAR images constitute space J.

[0064] Step S1 is achieved through the following steps:

[0065] S11. A multi-scale generative network structure is adopted, and the input noise of the first layer is removed;

[0066] S12. Perform an affine transformation on the input latent vector Z. The formula for the affine transformation is as follows:

[0067] G1(Z) = Vz + b

[0068] Where V is a matrix, b is a t-dimensional bias vector, and G1: R k →R t As a feature-based unentanglement transformation, the k-dimensional latent space Z is projected onto the t-dimensional feature space. Where χ is the semantic space, X∈χ is the semantic vector, and each dimension of χ independently controls a semantic of the generated image;

[0069] S13. The semantic vector X is sent to the generation module. Generation modules of different scales are sequentially combined to form the entire generation network. All inputs to the generation modules contain the same semantic vector, and the first generation module synthesizes an image with the lowest resolution m0×m0 from the semantic vector. The intermediate generation modules receive the image generated by the previous generation module and combine it with the semantic vector X to synthesize an image with a higher resolution. The last generation module simulates an image with the target resolution, using the following formula:

[0070]

[0071]

[0072] Where i represents the sequence number of the generation module, The image with the lowest resolution m0 is referred to as B0(S), where S is the semantic vector and B0(S) is the initial generation module and its input semantic vector S. Refers to a resolution of 2 i The image of m0 Refers to the i-th generating module and its input semantic vector S and image.

[0073] S14. The semantic vector X is generated through the key submodule Adain in each generation module. i Normalization, and then through feature y i The corresponding scalars are scaled and biased, and semantic vectors and feature maps are fused through the key submodule Adain to control the style of the generated image. The key submodule Adain is:

[0074]

[0075] Among them, AdaIN(X) i ,y) refers to the AdaIN module, y s,i The variance of the input semantic vector at layer i, y b,i X refers to the mean of the input semantic vector of the i-th layer. i The i-th layer input feature map, μ refers to the mean calculation, and α refers to the variance calculation;

[0076] S2. Set up a multi-scale discrimination network J→R for SAR images, and output the probability that the input image is a real image;

[0077] The multi-scale discriminant network described in step S2 is symmetrical to the generator network and is a multi-scale classifier composed of several discriminant modules. Each discriminant module consists of a representation structure DMAM for SAR images and several convolutional layers. The representation structure DMAM is described by the following formula:

[0078] M i =S i (A(F 7×7 (I)))

[0079]

[0080] Among them, F 7×7 It is a 2D convolution with a kernel size of 7×7, A is a self-attention module, S i It is a deformable convolution with a kernel size of 2i+1; using F, A, S i Three feature maps with different spatial information M1 to M3 were obtained; The tensor connection operator is represented by P, which is the ASPP module, and U is the final feature map, i.e., the output of DMAM. DMAM is represented by the uppercase letter H, and the above two functions are used to fuse spatial information.

[0081] Based on the representation structure DMAM, the discriminant modules at different scales are represented as follows:

[0082] Y1=K(H(I1)),i=1

[0083] Y i =K(αH(I) i )+(1-α)R i (Y i-1 ),i>1

[0084] y i =L(Y i ), i≥1

[0085] Where Y1 refers to the probability that the lowest resolution input image is a real image, K refers to a convolutional layer of size 3, i refers to the image sequence at each resolution, and Y i Refers to a resolution of 2 i The image of m0, α refers to the decreasing weighting factor, R i It is a linear interpolation function used to interpolate Y i-1 The resolution is doubled; L is a fully connected layer with a sigmoid activation function, providing the final output; scalar y i Indicates input I i It is the probability of a true SAR image;

[0086] S3. Generative adversarial networks are trained in the image space to enable the network to simulate real SAR images.

[0087] The generative adversarial network described in step S3 is trained using the following formula:

[0088]

[0089]

[0090] Where G and D represent the generator network and the discriminator network, respectively, I is the input image, and z is the input vector.

[0091] S4. Decouple the generator network to obtain the direction vector n and boundary vector Z, which are semantically related to the azimuth angle of the SAR image. inf ;

[0092] Step S4 is achieved through the following steps:

[0093] S41. For the linear transformation matrix V in the equation G1(Z) = Vz + b in step S12, find V T The V feature vector is obtained using the following formula:

[0094]

[0095] The eigenvector corresponding to its largest eigenvalue is the direction vector n;

[0096] Where, λ i Refers to the i-th largest eigenvalue, n i λ i The corresponding feature vector, where c refers to the number of selected features;

[0097] S42. Given a random latent vector Z and a step size α, generate an image sequence {I} according to the following formula. i}:

[0098] I i =G(Z-iαn)

[0099] Among them, I i Let G refer to the i-th image, G to the generator network, i to the index, n to the semantic direction vector, and the sequence {I} be the image. i In the image, as the number i increases, the semantic azimuth angle gradually decreases. The image with the smallest azimuth angle of 0 degrees is recorded as having a corresponding boundary vector Z. inf ;

[0100] S5. Utilizing the generator network G, direction vector n, and boundary vector Z inf Batch synthesis of SAR image sequences at specific azimuth angles;

[0101] Step S5 is achieved through the following steps: based on the generative network G and its semantic direction vector n, boundary vector Z inf A SAR image sequence with azimuth angles in the interval [a, b] and linearly increasing azimuth angles is generated using the following formula:

[0102] {I i}=G(Z inf +(iα+α-θ)n),i=0,1,...(ba) / α

[0103] Where i refers to the generated image number, α refers to the step size, θ refers to the orientation angle of the specific SAR image to be synthesized, and n refers to the semantic direction vector.

[0104] To further confirm the effectiveness of the invention, 232 SAR images were collected. Images with even-numbered angles (112 in total) were used as the training set for the model, and the remaining images were used as the test set to evaluate the algorithm's performance. This embodiment uses Fréchet Inception Distance (FID) (Heusel et al. 2017) and SSIM (Z. Wang et al. 2004) to measure the overall similarity of images and the similarity between each pair of synthesized images and real images at specific angles, respectively.

[0105] Table 1. Quantitative evaluation results of FID for different schemes

[0106] 0-90 90-180 180-270 270-360 DCGAN with WGAN-GP loss 73.8743 72.3742 70.8921 74.8421 ProGAN 22.3611 21.7037 22.1784 22.7119 StyleGAN 18.7791 17.1869 17.5780 18.4513 Ours 14.8482 15.1413 14.9335 15.1307

[0107] Table 2. Quantitative evaluation results of different SSIM schemes

[0108] 0-90 90-180 180-270 270-360 DCGAN with WGAN-GP loss 0.3221 0.3189 0.3673 0.3685 ProGAN 0.5694 0.5423 0.5814 0.5893 StyleGAN 0.6174 0.6073 0.5989 0.6003 Ours 0.6478 0.6389 0.6435 0.6265

[0109] As shown in Table 1, the present invention significantly outperforms DCGAN in overall image quality. Compared with other multi-scale GAN designs, the present invention also has significant advantages. Table 2 shows the average SSIM values ​​for each pair of real / synthetic images in the test set. As can be seen from Table 2, the present invention also has significant advantages.

[0110] Figure 4 The images randomly generated by various schemes are shown: (a) DCGAN: low simulation quality and limited patterns; (b) ProGAN: relatively high simulation quality but slightly limited patterns; (c) StyleGAN: a balance between simulation quality and patterns; (d) Ours: high simulation quality and rich patterns. Figure 4 It can be seen that traditional GAN ​​models can only generate SAR images with limited patterns and poor quality. Multi-scale GANs (b), (c), and (d) show highly synthetic qualitative results, while the SAR images simulated by the GAN used in invention (d) have the highest image quality and rich image patterns. This confirms the effectiveness of the GAN model designed in this invention.

[0111] Figure 5 The first set of images shows pairs of real SAR images with azimuth angles ranging from 0 to 90 degrees and synthetic SAR images synthesized using the present invention. The first row shows real SAR images and the second row shows synthetic SAR images. Figure 6 The second set of images, showing pairs of real SAR images with azimuth angles ranging from 0 to 90 degrees, and synthetic SAR images synthesized using this invention, is displayed. The first row contains real SAR images, and the second row contains synthetic SAR images. Figure 5 and Figure 6 It can be seen that the present invention can effectively simulate SAR images at specific angles.

[0112] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for batch synthesizing SAR images at a specific angle using a generative adversarial network, characterized in that, Includes the following steps: S1. Set up a generator network G, and map the random n-dimensional vector Z obtained by random sampling from the latent space Z to SAR images I at different scales. All SAR images constitute space J. Step S1 is achieved through the following steps: S11. A multi-scale generative network structure is adopted, and the input noise of the first layer is removed; S12. Perform an affine transformation on the input latent vector Z; S13. The semantic vector X is sent to the generation module. Generation modules of different scales are sequentially formed into the entire generation network. All inputs of the generation module contain the same semantic vector. The first generation module synthesizes an image with the lowest resolution m0×m0 from the semantic vector. The middle generation module receives the image generated by the previous generation module and combines it with the semantic vector X to synthesize an image with a higher resolution. The last generation module simulates an image with the target resolution. S14. The feature map X is generated through the key sub-module Adain in each generation module. i Normalization, and then through semantic vector y i The corresponding scalars are scaled and biased, and semantic vectors and feature maps are fused through the key submodule Adain to control the style of the generated image; S2. Set up a multi-scale discrimination network J→R for SAR images, and output the probability that the input image is a real image; S3. Generative adversarial networks are trained in the image space to enable the network to simulate real SAR images. S4. Decouple the generator network to obtain the direction vector n and boundary vector Z, which are semantically related to the azimuth angle of the SAR image. inf ; Step S4 is achieved through the following steps: S41. For the linear transformation matrix V in the equation G1(Z) = Vz + b in step S12, find V T The V feature vector is obtained using the following formula: The eigenvector corresponding to its largest eigenvalue is the direction vector n; Where, λ i Refers to the i-th largest eigenvalue, n i λ i The corresponding feature vector, where c refers to the number of selected features; S42. Given a random latent vector Z and a step size α, generate an image sequence {I} according to the following formula. i }: I i =G(Z-iαn) Among them, I i Let G refer to the i-th image, G to the generator network, i to the index, n to the semantic direction vector, and the sequence {I} be the image. i In the image, as the number i increases, the semantic azimuth angle gradually decreases. The image with the smallest azimuth angle of 0 degrees is recorded as having a corresponding boundary vector Z. inf ; S5. Utilizing the generator network G, direction vector n, and boundary vector Z inf Batch synthesis of SAR image sequences at specific azimuth angles; Step S5 is achieved through the following steps: based on the generative network G and its semantic direction vector n, boundary vector Z inf A SAR image sequence with azimuth angles in the interval [a, b] and linearly increasing azimuth angles is generated using the following formula: {I i }=G(Z inf +(iα+α-θ)n),i=0,1,…(b-a) / α Where i refers to the generated image number, α refers to the step size, θ refers to the orientation angle of the specific SAR image to be synthesized, and n refers to the semantic direction vector.

2. The method for batch synthesis of SAR images at a specific angle using a generative adversarial network as described in claim 1, characterized in that: In step S12, the formula for the affine transformation is as follows: G1(Z) = Vz + b Where V is a matrix, b is a t-dimensional bias vector, and G1: R k →R t As a feature-based unentanglement transformation, the k-dimensional latent space Z is projected onto the t-dimensional feature space. Where χ is the semantic space, X∈χ is the semantic vector, and each dimension of χ independently controls a semantic of the generated image; The formula used in step S13 is as follows: Where i represents the sequence number of the generation module, The image with the lowest resolution m0 is referred to as B0(S), where S is the semantic vector and B0(S) is the initial generation module and its input semantic vector S. Refers to a resolution of 2 i The image of m0 Refers to the i-th generating module and its input semantic vector S and image. In step S14, the key submodule Adain is: Among them, AdaIN(X) i ,y) refers to the AdaIN module, y s,i The variance of the input semantic vector at layer i, y b,i X refers to the mean of the input semantic vector of the i-th layer. i The input feature map of the i-th layer, μ refers to the mean calculation, and α refers to the variance calculation.

3. The method for batch synthesis of SAR images at a specific angle using a generative adversarial network as described in claim 2, characterized in that, The multi-scale discriminant network described in step S2 is symmetrical to the generator network and is a multi-scale classifier composed of several discriminant modules. Each discriminant module consists of a representation structure DMAM for SAR images and several convolutional layers. The representation structure DMAM is described by the following formula: M i =S i (A(F 7×7 (I))) Among them, F 7×7 It is a 2D convolution with a kernel size of 7×7, A is a self-attention module, S i It is a deformable convolution with a kernel size of 2i+1; using F, A, S i Three feature maps with different spatial information M1 to M3 were obtained; The tensor connection operator is represented by P, which is the ASPP module, and U is the final feature map, i.e., the output of DMAM. DMAM is represented by the uppercase letter H, and the above two functions are used to fuse spatial information. Based on the representation structure DMAM, the discriminant modules at different scales are represented as follows: Y1=K(H(I1)),i=1 Y i =K(αH(I i )+(1-α)R i (Y i-1 )),i>1 and i =L(Y i ),i≥1 Where Y1 refers to the probability that the lowest resolution input image is a real image, K refers to a convolutional layer of size 3, i refers to the image sequence at each resolution, and Y i Refers to a resolution of 2 i The image of m0, α refers to the decreasing weighting factor, R i It is a linear interpolation function used to interpolate Y i-1 The resolution is doubled; L is a fully connected layer with a sigmoid activation function, providing the final output; scalar y i Indicates input I i It is the probability of a real SAR image.

4. The method for batch synthesis of SAR images at a specific angle using a generative adversarial network as described in claim 3, characterized in that, The generative adversarial network described in step S3 is trained using the following formula: Where G and D represent the generator network and the discriminator network, respectively, I is the input image, and z is the input vector.

Images