Synthetic aperture radar image directional generation method based on image statistical characteristics

Abstract

A synthetic aperture radar image orientation generation method based on image statistical characteristics comprises image preprocessing, dividing the image attitude angle intervals, fitting the image statistical distribution, and sampling and fitting G... 0 The invention consists of the following steps: distribution, construction of a statistical generative adversarial network, training of the statistical generative adversarial network, and targeted image generation. Because this invention employs G... 0 This invention uses a distribution fitting method to obtain the statistical distribution of real synthetic aperture radar (SAR) images, capturing their statistical characteristics. During image generation, random noise conforming to the statistical characteristics of SAR images is sampled as input to a generative adversarial network (GAN). The generated SAR images possess statistical characteristics similar to those of real SAR images. Compared with existing methods, this invention produces higher quality images, exhibiting advantages such as realistic image generation and the ability to generate SAR images within a specific attitude angle range. It can be used for SAR image generation.

Description

Technical Field

[0001] This invention belongs to the field of image generation technology, specifically relating to synthetic aperture radar. Background Technology

[0002] Synthetic Aperture Radar (SAR) is an active microwave imaging sensor that can operate in all weather conditions and is widely used in both civilian and military fields. Deep learning-based automatic target recognition of SAR images aims to build a recognition network to identify the category of an input SAR image.

[0003] Due to the limited number of realistic synthetic aperture radar (SAR) images, automatic target recognition based on deep learning-based SAR images cannot achieve optimal performance. Generative adversarial networks (GANs) generate images to address the technical challenge of the limited number of realistic SAR images.

[0004] Existing synthetic aperture radar (SAR) image generation methods based on generative adversarial networks (GANs) do not incorporate the statistical characteristics of SAR images during generation, resulting in significant differences in statistical properties between the generated images and the real images. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to overcome the shortcomings of the prior art and provide a synthetic aperture radar image directional generation method based on image statistical characteristics that generates high-quality and realistic images.

[0006] The technical solution adopted to solve the above technical problems consists of the following steps:

[0007] (1) Preprocessed image

[0008] The radar imaging data at an elevation angle of 17° in the MSTAR dataset is used as the training set. A 64×64 pixel sub-image is cropped from the center of the training set. Pixel values ​​greater than 1 in the sub-image are changed to 1, and pixel values ​​equal to 0 are changed to the second minimum value. The pixel values ​​of the sub-image are normalized to the interval [-1,1].

[0009] (2) Divide the image attitude angle interval

[0010] Starting from S°, where S°∈[0°,360°], rotate clockwise for one revolution. Each interval of W° is used to divide the image attitude angle into one interval. The value of W° is a number less than 360° and divisible by 360°. This gives us the image attitude angle intervals.

[0011] (3) Statistical distribution of fitted image

[0012] The pixel values ​​of the training set contained within the image pose angle range are extracted and merged into a vector ξ. Then, G is applied based on the Mellin transform. 0Distributed parameter estimation methods perform G on the vector ξ 0 Distribution fitting yields the fitted G. 0 distributed.

[0013] (4) G of sampling fitting 0 distributed

[0014] The fitted G is obtained by using the acceptance-rejection sampling method. 0 Sampling is performed on the distribution to obtain G 0 Random noise distribution.

[0015] (5) Constructing a statistical generative adversarial network

[0016] Statistical generative adversarial networks consist of a generator and a discriminator connected in series.

[0017] The generator is composed of generator module 1, generator module 2, generator module 3, generator module 4, and generator module 5 connected in series; the discriminator is composed of discriminator module 1, discriminator module 2, discriminator module 3, discriminator module 4, and discriminator module 5 connected in series.

[0018] The generator module 1 is composed of generator residual module 1 and generator residual module 2 connected in parallel; the structures of generator module 2, generator module 3, generator module 4 and generator module 5 are the same as those of generator module 1.

[0019] The discriminator module 1 is composed of discriminator residual module 1 and discriminator residual module 2 connected in parallel; the structures of discriminator module 2, discriminator module 3, discriminator module 4 and discriminator module 5 are the same as those of discriminator module 1.

[0020] (6) Training statistical generative adversarial networks

[0021] 1) Construct the generator loss function L G

[0022] The generator loss function L is constructed as follows: G .

[0023]

[0024] in, This indicates that for D(G(z) c The expectation calculation of c), where D is the discriminator, G is the generator, and c represents the image pose angle interval number. G is the image fitted within the image attitude angle interval c. 0 Distribution, z c From Random noise in the sampled data.

[0025] 2) Construct the discriminator loss function

[0026] Construct the discriminator loss function L using the following formula D :

[0027]

[0028]

[0029] in, This represents the expectation calculation for D(x,c). Indicates to The expected value is calculated, where x is the true image within the image attitude angle interval c, and P c It represents the distribution of the true image within the image pose angle interval c, where λ is a constraint parameter with a finite number of positive integers, and |||² is the L2 norm. For the gradient operator, ε follows a uniform distribution on (0,1). for The data distribution it follows.

[0030] 3) Training statistical generative adversarial networks

[0031] The training set, the image pose angle intervals corresponding to the images in the training set are numbered c and z. c The data is input into a statistical generative adversarial network for training. The training batch size is 64. The discriminator is trained 5 times and the generator is trained 1 time in each batch. Adam is used as the optimizer for both the generator and the discriminator. The initial learning rate is 0.0002 and the decay rate is 0.5. The learning rate decays to 0.999 times the previous value each time. The training is conducted for 1000 rounds.

[0032] (7) Directional image generation

[0033] The image attitude angle interval number c and random noise z are used. c The input is fed into a trained statistical generative adversarial network to generate synthetic aperture radar images with statistical characteristics of the image within the image attitude angle range c.

[0034] In step (5) of this invention, which constructs a statistical generative adversarial network, the generator residual module 1 is composed of a conditional batch normalization layer 1, a convolutional layer 1, a conditional batch normalization layer 2, a deconvolutional layer 1, and a convolutional layer 2 connected in series; the generator residual module 2 is composed of a deconvolutional layer 2; and the generator module outputs X. GM :

[0035] X GM =X GR1 +X GR2

[0036] XGR1 =C2(D1(σ1(N) CB2 (c,C1(σ1(N CB1 (c,X GI )))))))

[0037] X GR2 =D2(X GI )

[0038] Among them, X GR1 This is the output of generator residual module 1, X GR2 This is the output of generator residual module 2, C2 is convolutional layer 2, D1 is deconvolutional layer 1, σ1 is the ReLU activation function, and N... CB2 This is conditional batch normalization 2, c is the image pose angle interval number, C1 is convolutional layer 1, and N... CB1 It is conditional batch normalization 1, X GI D2 is the input to the generator module, and D2 is the deconvolution layer 2.

[0039] In this invention, the convolutional kernel of convolutional layer 1 is 3×3 and the padding size is 1; the convolutional kernel of deconvolutional layer 1 is 4×4; the convolutional kernel of convolutional layer 2 is 3×3 and the padding size is 1; and the convolutional kernel of deconvolutional layer 2 is 4×4.

[0040] In step (5) of this invention, which constructs a statistical generative adversarial network, the discriminator residual module 1 is composed of a conditional batch normalization layer 3 connected in series with a convolutional layer 3, a conditional batch normalization layer 4, a convolutional layer 4, and a convolutional layer 5; the discriminator residual module 2 is composed of a convolutional layer 6. The discriminator module outputs X. DM :

[0041] X DM =X DR1 +X DR2

[0042] X DR1 =C5(C4(σ3(N) CB4 (c,C3(σ3(N CB3 (c,X DI )))))))

[0043] X DR2 =C6(X DI )

[0044] Among them, X DR1 It is the output of the discriminator residual module 1, X DR2 It is the output of the discriminator residual module 2, C k It is a convolutional layer k, k∈{3,4,5,6}, σ3 is the LeakyReLU activation function, N CB4 It is a conditional batch normalization layer 4, NCB3 It is conditional batch normalization layer 3, X DI This is the input to the discriminator module.

[0045] In this invention, the convolutional kernel of convolutional layer 3 is 3×3 and the padding size is 1; the convolutional kernel of convolutional layer 4 is 4×4; the convolutional kernel of convolutional layer 5 is 3×3 and the padding size is 1; and the convolutional kernel of convolutional layer 6 is 4×4.

[0046] In step (6) of the present invention, training the statistical generative adversarial network according to equation (2), the stated This represents the expectation calculation for D(x,c). Indicates to The expected value is calculated, where x is the true image within the image attitude angle interval c, and P c It represents the distribution of the true image within the image pose angle interval c, where λ is a constraint parameter with a value ranging from 5 to 15, and |||² is the L2 norm. For the gradient operator, ε follows a uniform distribution on (0,1). for The data distribution it follows.

[0047] Compared with the prior art, the present invention has the following technical advantages:

[0048] Because this invention uses G 0 This invention uses a distribution fitting method to fit the statistical distribution of real synthetic aperture radar (SAR) images, obtaining their statistical characteristics. During image generation, random noise conforming to the statistical characteristics of SAR images is sampled as input to the generative adversarial network (GAN). The generated SAR images possess statistical characteristics similar to real SAR images, resulting in higher image quality. Compared with existing methods, the FID values ​​of the images generated by this invention are 42.964, 23.713, and 1.049 lower than those of methods 1, 2, and 3, respectively, indicating that the generated images have the best quality. This invention has the advantages of generating realistic images and being able to directionally generate SAR images within the attitude angle range, making it suitable for SAR image generation. Attached Figure Description

[0049] Figure 1 This is a flowchart of Embodiment 1 of the present invention.

[0050] Figure 2 This is a schematic diagram of the statistical generative adversarial network structure of Embodiment 1 of the present invention.

[0051] Figure 3 yes Figure 2 A schematic diagram of the structure of Generator Module 1.

[0052] Figure 4 yes Figure 2 A schematic diagram of the structure of the discriminator module 1.

[0053] Figure 5 This is a synthetic aperture radar image generated in Embodiment 1 of the present invention.

[0054] Figure 6 It is a real image of a synthetic aperture radar image. Detailed Implementation

[0055] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments, but the present invention is not limited to the following embodiments.

[0056] Example 1

[0057] The synthetic aperture radar image orientation generation method based on image statistical characteristics in this embodiment consists of the following steps (see...). Figure 1 ):

[0058] (1) Preprocessed image

[0059] The radar imaging data at an elevation angle of 17° in the MSTAR dataset is used as the training set. A 64×64 pixel sub-image is cropped from the center of the training set. Pixel values ​​greater than 1 in the sub-image are changed to 1, and pixel values ​​equal to 0 are changed to the second minimum value. The pixel values ​​of the sub-image are normalized to the interval [-1,1].

[0060] (2) Divide the image attitude angle interval

[0061] Using S° as the starting point for dividing the image attitude angle interval, S°∈[0°,360°], in this embodiment, S° is 0°. Rotating clockwise, each interval W° is divided into one image attitude angle interval. W° is a number less than 360° and divisible by 360°. In this embodiment, W° is 5°, thus obtaining the image attitude angle interval.

[0062] (3) Statistical distribution of fitted image

[0063] The pixel values ​​of the training set contained within the image pose angle range are extracted and merged into a vector ξ. Then, G is applied based on the Mellin transform. 0 Distributed parameter estimation methods perform G on the vector ξ 0 Distribution fitting yields the fitted G. 0 distributed.

[0064] (4) G of sampling fitting 0 distributed

[0065] The fitted G is obtained by using the acceptance-rejection sampling method. 0 Sampling is performed on the distribution to obtain G 0 Random noise distribution.

[0066] (5) Constructing a statistical generative adversarial network

[0067] The statistical generative adversarial network in this embodiment consists of a generator and a discriminator connected in series.

[0068] Figure 2 A schematic diagram of the generator in this embodiment is provided. Figure 2 In this embodiment, the generator is composed of generator module 1 connected in series with generator module 2, generator module 3, generator module 4, and generator module 5. Similarly, the discriminator is composed of discriminator module 1 connected in series with discriminator module 2, discriminator module 3, discriminator module 4, and discriminator module 5.

[0069] Figure 3 A schematic diagram of the generator module 1 in this embodiment is provided. Figure 3 In this embodiment, generator module 1 is composed of generator residual module 1 and generator residual module 2 connected in parallel. The structures of generator module 2, generator module 3, generator module 4, and generator module 5 are the same as those of generator module 1.

[0070] In this embodiment, the generator residual module 1 is composed of a conditional batch normalization layer 1, a convolutional layer 1, a conditional batch normalization layer 2, a deconvolutional layer 1, and a convolutional layer 2 connected in series. In this embodiment, the convolutional kernel of convolutional layer 1 is 3×3 with a padding size of 1, the convolutional kernel of deconvolutional layer 1 is 4×4, the convolutional kernel of convolutional layer 2 is 3×3 with a padding size of 1, and the convolutional kernel of deconvolutional layer 2 is 4×4. The generator residual module 2 in this embodiment is composed of a deconvolutional layer 2. The generator module outputs X. GM :

[0071] X GM =X GR1 +X GR2

[0072] X GR1 =C2(D1(σ1(N) CB2 (c,C1(σ1(N CB1 (c,X GI )))))))

[0073] X GR2 =D2(X GI )

[0074] Among them, X GR1 This is the output of generator residual module 1, X GR2 This is the output of generator residual module 2, C2 is convolutional layer 2, D1 is deconvolutional layer 1, σ1 is the ReLU activation function, and N... CB2 This is conditional batch normalization 2, c is the image pose angle interval number, C1 is convolutional layer 1, and N... CB1It is conditional batch normalization 1, X GI D2 is the input to the generator module, and D2 is the deconvolution layer 2.

[0075] Figure 4 A schematic diagram of the discriminator module 1 in this embodiment is provided. Figure 4 In this embodiment, the discriminator module 1 is composed of discriminator residual module 1 and discriminator residual module 2 connected in parallel. The structures of discriminator module 2, discriminator module 3, discriminator module 4, and discriminator module 5 are the same as those of discriminator module 1.

[0076] In this embodiment, the discriminator residual module 1 is composed of a conditional batch normalization layer 3, a convolutional layer 3, a conditional batch normalization layer 4, a convolutional layer 4, and a convolutional layer 5 connected in series. In this embodiment, the convolutional kernel of convolutional layer 3 is 3×3 with a padding size of 1, the convolutional kernel of convolutional layer 4 is 4×4, the convolutional kernel of convolutional layer 5 is 3×3 with a padding size of 1, and the convolutional kernel of convolutional layer 6 is 4×4. The discriminator residual module 2 in this embodiment is composed of convolutional layer 6; the discriminator module outputs X. DM :

[0077] X DM =X DR1 +X DR2

[0078] X DR1 =C5(C4(σ3(N) CB4 (c,C3(σ3(N CB3 (c,X DI )))))))

[0079] X DR2 =C6(X DI )

[0080] Among them, X DR1 It is the output of the discriminator residual module 1, X DR2 It is the output of the discriminator residual module 2, C k It is a convolutional layer k, k∈{3,4,5,6}, σ3 is the LeakyReLU activation function, N CB4 It is a conditional batch normalization layer 4, N CB3 It is conditional batch normalization layer 3, X DI This is the input to the discriminator module.

[0081] (6) Training statistical generative adversarial networks

[0082] 1) Construct the generator loss function L G

[0083] The generator loss function L is constructed as follows: G :

[0084]

[0085] in, This indicates that for D(G(z) c The expectation calculation of c), where D is the discriminator, G is the generator, and c represents the image pose angle interval number. G is the image fitted within the image attitude angle interval c. 0 Distribution, z c From Random noise in the mid-sample;

[0086] 2) Construct the discriminator loss function:

[0087] Construct the discriminator loss function L using the following formula D :

[0088]

[0089]

[0090] in, This represents the expectation calculation for D(x,c). Indicates to The expected value is calculated, where x is the true image within the image attitude angle interval c, and P c λ is the distribution of the true image within the image pose angle interval c, λ is a constraint parameter, λ takes a value of 5 to 15, and in this embodiment λ takes a value of 10, |||2 is the L2 norm. For the gradient operator, ε follows a uniform distribution on (0,1). for The data distribution it follows;

[0091] 3) Training statistical generative adversarial networks

[0092] The training set, the image pose angle intervals corresponding to the images in the training set are numbered c and z. c The data is input into a statistical generative adversarial network for training. The training batch size is 64. The discriminator is trained 5 times and the generator is trained once per batch. Adam is used as the optimizer for both the generator and the discriminator. The initial learning rate is 0.0002 and the decay rate is 0.5. The learning rate decays to 0.999 times the previous value each time. The training is conducted for 1000 rounds.

[0093] (7) Directional image generation

[0094] The image attitude angle interval number c and random noise z are used. c The input is fed into a trained statistical generative adversarial network to generate synthetic aperture radar images with statistical characteristics of the image within the image attitude angle range c.

[0095] A method for directional generation of synthetic aperture radar images based on image statistical properties was developed.

[0096] Example 2

[0097] The synthetic aperture radar image orientation generation method based on image statistical characteristics in this embodiment consists of the following steps:

[0098] (1) Preprocessed image

[0099] The steps are the same as in Example 1.

[0100] (2) Divide the image attitude angle interval

[0101] Using S° as the starting point for dividing the image attitude angle interval, S°∈[0°,360°], in this embodiment, S° is 180°. Rotating clockwise, each interval W° is divided into one image attitude angle interval. W° is a number less than 360° and divisible by 360°. In this embodiment, W° is 1°, thus obtaining the image attitude angle interval.

[0102] (3) Statistical distribution of fitted image

[0103] The steps are the same as in Example 1.

[0104] (4) G of sampling fitting 0 distributed

[0105] The steps are the same as in Example 1.

[0106] (5) Constructing a statistical generative adversarial network

[0107] The steps are the same as in Example 1.

[0108] (6) Training statistical generative adversarial networks

[0109] 1) Construct the generator loss function L G

[0110] The steps are the same as in Example 1.

[0111] 2) Construct the discriminator loss function:

[0112] Construct the discriminator loss function L using the following formula D :

[0113]

[0114]

[0115] in, This represents the expectation calculation for D(x,c). Indicates to The expected value is calculated, where x is the true image within the image attitude angle interval c, and P c λ is the distribution of the true image within the image pose angle interval c, λ is a constraint parameter, λ takes a value of 5 to 15, and in this embodiment λ takes a value of 5, |||2 is the L2 norm. For the gradient operator, ε follows a uniform distribution on (0,1). for The data distribution it follows.

[0116] 3) Training statistical generative adversarial networks

[0117] The steps are the same as in Example 1.

[0118] The other steps are the same as in Example 1. This completes the method for directional generation of synthetic aperture radar images based on image statistical characteristics.

[0119] Example 3

[0120] The synthetic aperture radar image orientation generation method based on image statistical characteristics in this embodiment consists of the following steps:

[0121] (1) Preprocessed image

[0122] The steps are the same as in Example 1.

[0123] (2) Divide the image attitude angle interval

[0124] Using S° as the starting point for dividing the image attitude angle interval, S°∈[0°,360°], in this embodiment, S° is 360°. Rotating clockwise, each interval W° is divided into one image attitude angle interval. W° is a number less than 360° and divisible by 360°. In this embodiment, W° is 180°, thus obtaining the image attitude angle interval.

[0125] (3) Statistical distribution of fitted image

[0126] The steps are the same as in Example 1.

[0127] (4) G of sampling fitting 0 distributed

[0128] The steps are the same as in Example 1.

[0129] (5) Constructing a statistical generative adversarial network

[0130] The steps are the same as in Example 1.

[0131] (6) Training statistical generative adversarial networks

[0132] 1) Construct the generator loss function L G

[0133] The steps are the same as in Example 1.

[0134] 2) Construct the discriminator loss function:

[0135] Construct the discriminator loss function L using the following formula D :

[0136]

[0137]

[0138] in, This represents the expectation calculation for D(x,c). Indicates to The expected value is calculated, where x is the true image within the image attitude angle interval c, and P c λ is the distribution of the true image within the image pose angle interval c, λ is a constraint parameter, λ takes a value of 5 to 15, and in this embodiment λ takes a value of 15, |||2 is the L2 norm. For the gradient operator, ε follows a uniform distribution on (0,1). for The data distribution it follows.

[0139] 3) Training statistical generative adversarial networks

[0140] The steps are the same as in Example 1.

[0141] The other steps are the same as in Example 1. This completes the method for directional generation of synthetic aperture radar images based on image statistical characteristics.

[0142] To verify the beneficial effects of the present invention, the synthetic aperture radar image directional generation method based on image statistical characteristics according to Embodiment 1 of the present invention was used for the following experiment:

[0143] (1) Experimental study on the beneficial effects of the present invention

[0144] The synthetic aperture radar image generated using the method of Embodiment 1 of this invention is shown below. Figure 5 , Figure 6 It is a real image from synthetic aperture radar, produced by Figure 5 and Figure 6 By comparison, visual contrast in the images shows that the images generated by this invention are of high quality and closely resemble real images. Observation Figure 5 It has been found that the synthetic aperture radar images generated by this invention are diverse and can be generated directionally within the image attitude angle range.

[0145] (2) Comparative simulation experiment of the present invention

[0146] Image quality comparison experiments were conducted using the method of Example 1 with Conditional Generative Adversarial Network (hereinafter referred to as Comparison Method 1), Conditional Wasserstein Generative Adversarial Network (hereinafter referred to as Comparison Method 2), and Statistical Generative Adversarial Network with normally distributed random noise as input (hereinafter referred to as Comparison Method 3).

[0147] The image quality was evaluated using the Fréchet Inception Distance (FID) metric in the comparative simulation experiments. The experimental results are shown in Table 1.

[0148] Table 1. Comparison of FID values ​​for synthetic aperture radar images generated by different methods.

[0149]

[0150] As shown in Table 1, the FID values ​​of the images generated by Embodiment 1 of the present invention are 42.964, 23.713, and 1.049 lower than those of Comparison Method 1, Comparison Method 2, and Comparison Method 3, respectively, indicating that the images generated by the present invention have the best quality.

[0151] Through the above-mentioned comparative experiments between the present invention and other comparative methods, the present invention uses the noise of the statistical characteristics of synthetic aperture radar images as the input of generative adversarial networks to generate synthetic aperture radar images with image statistical characteristics within the image attitude angle range, resulting in higher image quality.

Claims

1. A method for directional generation of synthetic aperture radar images based on image statistical characteristics, characterized in that... It consists of the following steps: (1) Preprocessed image The radar imaging data at an elevation angle of 17° from the MSTAR dataset was used as the training set. A 64×64 pixel sub-image was cropped from the center of the training set. Pixel values ​​greater than 1 in the sub-image were changed to 1, and pixel values ​​equal to 0 were changed to the second minimum value. The pixel values ​​of the sub-image were then normalized to the interval [range missing]. Inside; (2) Divide the image attitude angle interval by º is the starting point for dividing the image attitude angle interval, Sº [0º, 360º], rotate clockwise one revolution, at each interval The image is divided into one pose angle interval. The value of º is a number that is less than 360º and divisible by 360º, which is used to obtain the image attitude angle range; (3) Statistical distribution of the fitted image Extract the pixel values ​​of the training set contained within the image pose angle range and merge them into a vector. Using Mellin transform-based Distributed parameter estimation methods for vectors conduct Distribution fitting, obtaining the fitted value distributed; (4) Sampling fit distributed The fit was achieved using the accept-rejection sampling method. Sampling is performed on the distribution to obtain Distributed random noise; (5) Constructing a statistical generative adversarial network Statistical generative adversarial networks consist of a generator and a discriminator connected in series. The generator is composed of generator module 1 connected in series with generator module 2, generator module 3, generator module 4, and generator module 5; the discriminator is composed of discriminator module 1 connected in series with discriminator module 2, discriminator module 3, discriminator module 4, and discriminator module 5. The generator module 1 is composed of generator residual module 1 and generator residual module 2 connected in parallel; the structures of generator module 2, generator module 3, generator module 4, and generator module 5 are the same as those of generator module 1. The discriminator module 1 is composed of discriminator residual module 1 and discriminator residual module 2 connected in parallel; the structures of discriminator module 2, discriminator module 3, discriminator module 4 and discriminator module 5 are the same as those of discriminator module 1. (6) Training statistical generative adversarial networks 1) Construct the generator loss function Construct the generator loss function as follows : （1） in, Indicates to Expectation calculation, For the discriminator, For generator, Indicates the image attitude angle interval number. , Image attitude angle range Intra-image fitting distributed, From Random noise in the mid-sample; 2) Construct the discriminator loss function Construct the discriminator loss function as follows : （2） in, Indicates to Expectation calculation, Indicates to Expectation calculation, Image attitude angle range Real images inside, Image attitude angle range The distribution that the real images within the image follow. For constraint parameters, The value can be a finite number of positive integers. It is the L2 norm. For gradient operators, It follows a uniform distribution on (0,1). for The data distribution it follows; 3) Training statistical generative adversarial networks Number the image pose angle intervals corresponding to the images in the training set. as well as The data is input into a statistical generative adversarial network for training. The training batch size is 64. The discriminator is trained 5 times and the generator is trained once per batch. Adam is used as the optimizer for both the generator and the discriminator. The initial learning rate is 0.0002 and the decay rate is 0.

5. The learning rate decays to 0.999 times the previous value each time. The training is conducted for 1000 rounds. (7) Directional image generation Number the image attitude angle intervals and random noise The input is fed into a trained statistical generative adversarial network to generate images with pose angle ranges. Synthetic aperture radar images with statistical properties of internal images.

2. The synthetic aperture radar image orientation generation method based on image statistical characteristics according to claim 1, characterized in that: In step (5), when constructing the statistical generative adversarial network, the generator residual module 1 is composed of a conditional batch normalization layer 1, a convolutional layer 1, a conditional batch normalization layer 2, a deconvolutional layer 1, and a convolutional layer 2 connected in series; the generator residual module 2 is composed of a deconvolutional layer 2; the generator module outputs... : in, This is the output of generator residual module 1. This is the output of generator residual module 2. It is convolutional layer 2. It is deconvolution layer 1. It is a ReLU activation function. It is conditional batch normalization 2. It is the image attitude angle interval number. It is convolutional layer 1. It is conditional batch normalization 1. As input to the generator module, It is deconvolution layer 2.

3. The synthetic aperture radar image orientation generation method based on image statistical characteristics according to claim 2, characterized in that: The convolutional layer 1 has a 3×3 kernel and a padding size of 1, the deconvolutional layer 1 has a 4×4 kernel, the convolutional layer 2 has a 3×3 kernel and a padding size of 1, and the deconvolutional layer 2 has a 4×4 kernel.

4. The synthetic aperture radar image orientation generation method based on image statistical characteristics according to claim 1, characterized in that: In step (5), when constructing the statistical generative adversarial network, the discriminator residual module 1 is composed of a conditional batch normalization layer 3 and a convolutional layer 3, a conditional batch normalization layer 4, a convolutional layer 4, and a convolutional layer 5 connected in series; the discriminator residual module 2 is composed of a convolutional layer 6; the discriminator module output... : in, This is the output of the discriminator residual module 1. This is the output of the discriminator residual module 2. It is a convolutional layer , , It is the LeakyReLU activation function. It is conditional batch normalization layer 4. It is conditional batch normalization layer 3. This is the input to the discriminator module.

5. The synthetic aperture radar image orientation generation method based on image statistical characteristics according to claim 4, characterized in that: The convolutional layer 3 has a 3×3 kernel and a padding size of 1, the convolutional layer 4 has a 4×4 kernel, the convolutional layer 5 has a 3×3 kernel and a padding size of 1, and the convolutional layer 6 has a 4×4 kernel.

6. The synthetic aperture radar image orientation generation method based on image statistical characteristics according to claim 1, characterized in that: In step (6), training the statistical generative adversarial network using equation (2), the stated... Indicates to Expectation calculation, Indicates to Expectation calculation, Image attitude angle range Real images inside, Image attitude angle range The distribution that the real images within the image follow. For constraint parameters, The value ranges from 5 to 15. It is the L2 norm. For gradient operators, It follows a uniform distribution on (0,1). for The data distribution it follows.

Images