Intelligent Generation Method of SAR Target Data by Fusion of Geometric Information

Abstract

The present invention discloses an intelligent generation method of SAR target data integrating geometric information, extracts the contour of the target in the optical image, obtains the geometric structure diagram and calculates the dominant scattering boundary of the target under different radar incident directions; constructs a SAR image generation network, including a generator, a discriminator and a classifier that introduce a multi-layer attention mechanism; inputs the dominant scattering boundary, the geometric structure diagram and the category label into the generator to obtain the generated SAR image; inputs the generated SAR image or the real SAR image into the discriminator to determine whether the image belongs to a real sample or a generated sample; inputs the generated SAR image or the real SAR image and the corresponding category label into the classifier to determine the category of the image; the discriminator and the classifier both feed back the judgment result to the generator. The SAR image generated by the present invention has an obvious dominant scattering boundary, which effectively improves the problem that the generated target data does not conform to the physical law of SAR imaging.

Description

Technical Field

[0001] The present invention relates to the technical field of image generation, and in particular to a method for intelligently generating SAR target data by fusing geometric information. Background Art

[0002] Synthetic Aperture Radar (SAR) is an active remote sensing technology that has the ability to observe and image over a wide range at all times of the day and in all weather conditions. Due to its advantages in extreme environments, research in SAR-related fields continues to receive a lot of attention. In recent years, with the further development of deep learning, deep neural networks have also been widely used in SAR-related fields. A large number of applications involve SAR target image processing, such as SAR automatic target recognition and target classification. The MSTAR (Moving and Stationary Target Acquisition and Recognition) dataset is currently the most widely used multi-angle public dataset for target image processing. However, this dataset is difficult to support deep learning model training in terms of target type richness and data scale. It is costly to rely on real experiments to obtain real SAR data of multiple angles and multiple target types. Therefore, some scholars have proposed to use SAR image generation methods to expand existing datasets.

[0003] At present, the commonly used SAR image generation methods are mainly divided into two categories. The first category is to generate corresponding SAR data based on target SAR data, which can be specifically divided into two methods. The first is electromagnetic computational simulation, which commonly includes physical optics method, integral equation method, etc. These methods usually rely on the physical model of the target to perform electromagnetic calculations to generate SAR images. The second is to generate SAR target data based on the Generative Adversarial Network (GAN). GAN consists of a generator module and a discriminator module. The generator module uses the input information provided to capture the data distribution and generate image data. The discriminator module determines whether the sample belongs to a training sample or a generated sample. The two achieve sample generation that meets the real data distribution through adversarial training based on game theory. The second category generates corresponding SAR data based on target optical data, takes relevant technologies in the field of computer vision as prototypes, and realizes SAR data generation through style transfer.

[0004] There are the following disadvantages in using electromagnetic computing simulation to generate SAR images: 1) Electromagnetic simulation takes little consideration of image features, and the generated SAR target data is different from the real data, which is easy to interfere with the deep learning model; 2) This generation method is time-consuming and costly, and the target data generation efficiency is low. The method of generating SAR target data based on GAN also has certain limitations: 1) When traditional GAN ​​is applied to SAR target data generation, random noise is used as input, resulting in the generated data having no clear physical meaning; 2) Improved GAN, such as Deep Convolutional GAN ​​(DCGAN) and Conditional GAN ​​(CGAN), does not fully consider the complex imaging mechanism of SAR, the generated data does not conform to the physical imaging law, and the model training stability is low. Similarly, the method of generating corresponding SAR data based on target optical data also has some problems: the imaging geometric features of the target are not reasonably utilized, and although the generated SAR image shows similar grayscale statistical characteristics to the real image, it still does not meet the imaging physical law of SAR targets. Summary of the invention

[0005] In order to solve the above technical problems, the present invention proposes a SAR target data intelligent generation method integrating geometric information. The generated SAR image has an obvious dominant scattering boundary, which effectively improves the problem that the generated target data does not conform to the physical laws of SAR imaging.

[0006] In order to achieve the above object, the technical solution of the present invention is as follows:

[0007] The method for intelligently generating SAR target data by integrating geometric information comprises the following steps:

[0008] Extract the contour of the target in the optical image to obtain the geometric structure diagram of the target, and use the geometric structure diagram to calculate the dominant scattering boundary of the target under different radar incident directions; manually mark the target in the optical image to obtain the category label of the target;

[0009] A SAR image generation network is constructed based on a generative adversarial network, wherein the SAR image generation network includes a generator, a discriminator, and a classifier that introduce a multi-layer attention mechanism;

[0010] The dominant scattering boundary, geometric structure diagram and category label of the target at different radar incident azimuths are input into the generator to obtain the generated SAR image;

[0011] The generated SAR image or the real SAR image is input into the discriminator to determine whether the image belongs to a real sample or a generated sample; the generated SAR image or the real SAR image and the corresponding category label are input into the classifier to determine the category of the image; the discriminator and the classifier both feed back the judgment results to the generator that introduces the multi-layer attention mechanism.

[0012] Preferably, the generator introducing the multi-layer attention mechanism comprises an encoder, a connection unit and a decoder connected in sequence, wherein:

[0013] The encoder is composed of an attention unit and four downsampling modules, each of which includes a convolutional layer, an InstanceNorm layer and a ReLU activation function connected in sequence; the attention unit includes a first attention module, a second attention module and a third attention module, the first attention module includes a convolutional layer and a Sigmoid function layer; the second attention module and the third attention module both include a convolutional layer, a ReLU activation function and a convolutional attention mechanism module;

[0014] The connection unit includes five residual modules connected in sequence, each residual module includes a convolution layer, an InstanceNorm layer, a ReLU activation function, an InstanceNorm layer and a ReLU activation function connected in sequence;

[0015] The decoder includes three upsampling modules and an output module connected in sequence; each upsampling module includes a deconvolution layer, an InstanceNorm layer and a ReLU activation function connected in sequence; the output module includes a convolution layer and a Tanh function connected in sequence.

[0016] Preferably, the processing of the encoder comprises the following steps:

[0017] The dominant scattering boundary is input into the first attention module to obtain a weight map, and the weight map is element-wise multiplied with the dominant scattering boundary to obtain an adjusted feature map; the geometric structure map, the category label and the adjusted feature map are concatenated and input into the first downsampling module;

[0018] The dominant scattering boundary is input into the second attention module to perform feature attention weighting in the channel and spatial dimensions in turn. The output of the second attention module is spliced ​​with the output of the first downsampling module as the input of the second downsampling module; the dominant scattering boundary passes through the convolution layer and ReLU activation function in the second attention module and is input into the third attention module to perform feature attention weighting in the channel and spatial dimensions in turn. The output of the third attention module is spliced ​​with the output of the second downsampling module as the input of the third downsampling module.

[0019] Preferably, the convolution kernel size of the convolution layer in the first attention module is 7×7;

[0020] The convolution kernel sizes of the convolution layers in the four downsampling modules are all 3×3, and except for the convolution layer step size of the first downsampling module being set to 1, the convolution layer step sizes of the remaining downsampling modules are all set to 2;

[0021] The convolution kernel size of the deconvolution layer in the three upsampling modules is 3×3, and the step size is set to 2;

[0022] The convolution kernel size of the convolution layer in the output module is 7×7.

[0023] Preferably, the discriminator includes a first convolution module, a second convolution module, a third convolution module, a fourth convolution module and a fifth convolution module connected in sequence; the classifier includes a first convolution module, a second convolution module, a third convolution module, a first fully connected module and a second fully connected module connected in sequence, and the discriminator and the classifier share the first convolution module, the second convolution module and the third convolution module.

[0024] Preferably, the first convolution module is composed of a convolution layer and a LeakyReLU activation function, the second convolution module, the third convolution module and the fourth convolution module all include a convolution layer, an InstanceNorm layer and a LeakyReLU activation function connected in sequence, and the fifth convolution module is composed of a convolution layer and a Sigmoid activation function; the convolution kernel size of the convolution layer in the first convolution module, the second convolution module, the third convolution module, the fourth convolution module and the fifth convolution module is 3×3, and the step size is 2.

[0025] Preferably, the first fully connected module uses LeakyReLU as the activation function; the second fully connected module uses Softmax as the activation function.

[0026] Based on the above technical solution, the beneficial effects of the present invention are:

[0027] 1) The present invention makes full use of the target SAR imaging geometric information, and the generated SAR target image meets the physical laws of SAR imaging in terms of image style and electromagnetic scattering;

[0028] 2) The present invention designs a generator based on a multi-layer attention mechanism, which emphasizes the dominant scattering boundaries in the image by adjusting the attention mechanism, and incorporates these adjusted attention weights into the generator to guide the network to better understand and learn the imaging geometry of the image and make full use of the dominant scattering boundary information;

[0029] 3) The present invention designs a discriminant module based on multi-task learning, which enhances the generalization performance of the model by adding a classifier and jointly optimizing it with the discriminator. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 is a schematic diagram of a flow chart of a method for intelligently generating SAR target data by fusing geometric information in an embodiment;

[0031] Figure 2 The figure is a schematic diagram of the process of obtaining the dominant scattering boundary in the intelligent generation method of SAR target data integrating geometric information in an embodiment, wherein: Figure 2 (a) is a schematic diagram of four radar incident directions; Figure 2 (b) Schematic diagram of the target dominant scattering boundary calculation process corresponding to a radar incident angle of 0°;

[0032] Figure 3 is a schematic diagram of a network structure of a generator in a method for intelligently generating SAR target data that integrates geometric information in an embodiment;

[0033] Figure 4 is a schematic diagram of the structure of a convolutional attention mechanism module in a method for intelligently generating SAR target data that integrates geometric information in an embodiment;

[0034] Figure 5 is a schematic diagram of the structure of a residual module in a method for intelligently generating SAR target data that integrates geometric information in an embodiment;

[0035] Figure 6 The diagram is a schematic diagram of the structure of a discriminator module in a method for intelligently generating SAR target data by fusing geometric information in an embodiment. DETAILED DESCRIPTION

[0036] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0037] like Figure 1 As shown, this embodiment provides a method for intelligently generating SAR target data by fusing geometric information, including the following steps:

[0038] Step 1: extract the contour of the target in the optical image to obtain the geometric structure diagram of the target, and use the geometric structure diagram to calculate the dominant scattering boundary of the target under different radar incident directions; manually mark the target in the optical image to obtain the category label of the target.

[0039] In this embodiment, the position of the target's dominant scattering boundary at different radar incident angles is calculated using the target contour structure information of the optical image. Figure 2 (a) shows the four radar incident directions. Figure 2(b) shows the calculation process of the target dominant scattering boundary when the radar is at an incident angle of 0°, where the dominant scattering boundary is the L-shaped scattering boundary of the vehicle target in the SAR data, such as Figure 2 First, the contour of the optical image is extracted to obtain its geometric structure information. The dominant scattering boundary is usually part of the outer contour of the target. Therefore, the target geometric contour map is traversed pixel by pixel from the radar incident direction (the direction corresponding to 0° in this case) to obtain the outermost contour boundary in this direction. This boundary can be considered as the dominant scattering boundary of the target at this radar incident angle.

[0040] Step 2: construct a SAR image generation network based on the adversarial generative network, wherein the SAR image generation network includes a generator, a discriminator, and a classifier that introduce a multi-layer attention mechanism.

[0041] In this embodiment, in order to provide constraints from both the target geometry and the target category, the generator takes the target geometry diagram and the category label as input. In addition, the dominant scattering boundary is injected into the generator as prior information to guide the weight distribution of the network so that the generated SAR vehicle target data has a dominant scattering boundary that conforms to the laws of physics. Multi-task learning is used in the discriminator module to construct a discriminator and an auxiliary classifier for joint optimization to improve the generalization ability of the model. Among them, the discriminator takes the generated or real SAR image as input to determine whether the input data belongs to real data or generated data. In addition to the generated or real data, the auxiliary classifier input information also includes the category label of the data, which is used to determine whether the image belongs to the corresponding target category.

[0042] Specifically, the main structure of the generator consists of three modules: an encoder, a decoder, and a connection unit between the two. The encoder consists of an attention unit and four downsampling modules, and the decoder module consists of three upsampling modules and an output module. In order to enable the generated network to learn more recognizable information, this embodiment adds five residual modules between the encoder and the decoder. The details of each part are as follows:

[0043] Encoder: The encoder contains an attention unit and four downsampling modules, which are used to sample the input image features into feature maps of size [512, 16, 16]. Each downsampling module of the encoder consists of a convolutional layer, an instance normalization (InstanceNorm) layer, and a rectified linear unit (ReLU) activation function. The convolution kernel size of the four convolutional layers is 3×3. Except for the convolution layer stride of the first downsampling module, which is set to 1, the convolution layer stride of the remaining downsampling modules is 2.

[0044] In addition to the optical structure map and target category features, the generator input also includes the target dominant scattering boundary map, which is used to guide the network to learn the target imaging geometric information. This embodiment uses a physical injection network to inject the dominant scattering boundary maps of different scales obtained by the convolutional network into the generator network. The physical injection network injects the physical information obtained from other modules into the traditional deep neural network, and uses additional injected features to provide rich prior knowledge. In order to make the model pay more attention to the area where the dominant scattering boundary is distributed, the dominant scattering boundary map is first adjusted before feature injection, and the corresponding attention mechanism is selected according to the scale of the feature map. Through the first attention module in the attention unit, the dominant scattering boundary feature map that is first adjusted for attention is a single-channel image, so only a simplified spatial attention mechanism is used for it. The simplified spatial attention mechanism module first performs a convolution operation with a convolution kernel of 7×7 on the input feature map and activates it using the S-type (Sigmoid) function to obtain a weight map. The weight map obtained is then multiplied by the original input feature map to obtain the adjusted feature map. In the second and third attention modules, a convolutional block attention module (CBAM) that mixes spatial and channel attention is used.

[0045] The specific structure of the convolutional attention mechanism module is as follows Figure 4 As shown in the figure, it contains two sub-modules: Channel Attention Mechanism (CAM) and Spatial Attention Mechanism (SAM), which are used to perform feature attention weighting in channel and spatial dimensions in turn.

[0046] Channel attention module: The input feature map (size is H×W×C) is passed through a global maximum pooling layer and a global average pooling layer based on one layer to obtain two feature maps of size 1×1×C. The two generated feature maps are input into a shared multilayer perceptron (MLP) respectively. The output of the MLP is added element by element and activated by the Sigmoid function to generate the channel attention map. Finally, the channel attention map and the input feature map are multiplied element by element to generate the input features required by the spatial attention module. The process of generating features in the channel attention module can be expressed as:

[0047] M CAM (F)=σ(MLP(AvgPool(F))+MLP(MaxPool(F))) (1)

[0048] Where σ represents the Sigmoid activation function, AvgPool represents global average pooling, MaxPool represents global maximum pooling, MLP represents multi-layer perceptron, and F represents the input feature map.

[0049] Spatial attention module: The spatial attention module has the ability to mine feature spatial relationships and generate spatial attention maps. The spatial attention module uses the feature map output by the channel attention module as input features. First, the feature map is input into the maximum pooling layer and the average pooling layer respectively to obtain two features of size H×W×1. The two feature maps are then concatenated at the channel level and input into the convolution layer. The output of the convolution layer is activated by the Sigmoid function to output a spatial attention map of size H×W×1. Finally, the spatial attention map is element-wise multiplied with the input feature to obtain the final generated feature. The overall process of generating features by the spatial attention module can be expressed as:

[0050] M SAM (F)=σ(f([AvgPool(F));MLP(MaxPool(F)])), (2)

[0051] Where σ represents the Sigmoid activation function, f(·) represents the convolution operation, [·;·] represents the concatenation of feature maps, AvgPool represents global average pooling, MaxPool represents global maximum pooling, MLP represents multi-layer perceptron, and F represents the input feature map.

[0052] The attention-adjusted feature maps are concatenated and fused with the input features of the lower layers of the encoder, and the new features obtained are used as the input of the next sampling module for subsequent operations.

[0053] In order to enable the generative network to learn more recognizable information, this embodiment adds five residual modules between the encoder and the decoder. The residual module structure is as follows: Figure 5 shown.

[0054] Specifically, the decoder contains three upsampling modules and an output module, which is used to restore the feature map of size [512, 16, 16] to the output image of size [1, 128, 128]. Figure 3 As shown in the figure. Each upsampling module consists of a deconvolution layer, an InstanceNorm layer, and a ReLU activation function. The convolution kernel size of the three deconvolution layers is 3×3, and the stride is 2. The output feature map size after the upsampling module is [64, 128, 128]. The output module consists of a convolution layer with a convolution kernel size of 7×7 and a hyperbolic tangent (Tanh) activation function, which adjusts the feature map to the final output size [1, 128, 128].

[0055] Specifically, the multi-task learning idea is used to construct the discriminator module. The discriminator module structure is as follows: Figure 6 shown.

[0056] The input of the discriminator module is a real SAR image (or generated SAR image) and its category label. Similar to the generator input, the input SAR image is a single-channel image of size [128, 128], and the target category label is encoded as an N-dimensional one-hot vector (N represents the number of categories). The discriminator module contains a discriminator for determining whether the image belongs to a real sample or a generated sample and a classifier for determining the image category, which can effectively improve the generalization ability of the model.

[0057] The discriminator and the classifier share three convolution modules (the first convolution module, the second convolution module, and the third convolution module) in the network structure. Among them, the first convolution module consists of a convolution layer and a leaky linear rectifier (Leaky Rectified Linear Unit, LeakyReLU) activation function. Compared with the first convolution module, the other two modules (the second convolution module and the third convolution module) have an additional InstanceNorm layer between the convolution layer and the LeakyReLU activation function. The convolution kernel size of the convolution layer of the above three modules is 3×3, and the step size is 2. The output feature map size after the three convolution modules is [256, 16, 16], and the output feature map continues to be input into the remaining modules of the discriminator and classifier respectively. The remaining modules of the discriminator are the fourth convolution module and the fifth convolution module structure. Among them, the fourth convolution module has the same structure as the second and third convolution modules. The fifth convolution module consists of a convolution layer and a Sigmoid activation function. The convolution kernel size of the convolution layer in the fourth and fifth convolution modules is 3×3, and the step size is 2.

[0058] In order to improve the training speed, the input image is divided into multiple regions, and each region is classified as true or false. The final output size of the discriminator is [1, 14, 14], where each element represents the discriminator's judgment on the probability that the block of image belongs to the real image.

[0059] The remaining modules of the classifier include two fully connected modules (the first fully connected module and the second fully connected module). The first fully connected module uses LeakyReLU as the activation function, which is mainly used for feature conversion and compression to capture multiple feature patterns of input features. The second fully connected module uses the normalized exponential function (Soft Maximum, Softmax) as the activation function to normalize the output probability to [0, 1] and predict the probability that the input image belongs to each category.

[0060] Step 3: Input the dominant scattering boundary, geometric structure diagram and category label of the target under different radar incident directions into the generator to obtain the generated SAR image.

[0061] Step 4: Input the generated SAR image or the real SAR image into the discriminator to determine whether the image belongs to a real sample or a generated sample; input the generated SAR image or the real SAR image and the corresponding category label into the classifier to determine the category of the image; the discriminator and the classifier both feed back the judgment results to the generator that introduces the multi-layer attention mechanism.

[0062] In an embodiment, a method for intelligently generating SAR target data by fusing geometric information is also provided, and a processing process of an encoder is also provided, including the following steps:

[0063] The dominant scattering boundary is input into the first attention module to obtain a weight map, and the weight map is element-wise multiplied with the dominant scattering boundary to obtain an adjusted feature map; the geometric structure map, the category label and the adjusted feature map are concatenated and input into the first downsampling module;

[0064] The dominant scattering boundary is input into the second attention module to perform feature attention weighting in the channel and spatial dimensions in turn. The output of the second attention module is spliced ​​with the output of the first downsampling module as the input of the second downsampling module; the dominant scattering boundary passes through the convolution layer and ReLU activation function in the second attention module and is input into the third attention module to perform feature attention weighting in the channel and spatial dimensions in turn. The output of the third attention module is spliced ​​with the output of the second downsampling module as the input of the third downsampling module.

[0065] It should be understood that, although the various steps in the above-mentioned flow chart are displayed in sequence according to the indication of the arrows, these steps are not necessarily executed in sequence according to the order indicated by the arrows. Unless there is a clear explanation in this article, the execution of these steps is not strictly limited in order, and these steps can be executed in other orders. Moreover, at least a part of the steps in the above-mentioned flow chart may include a plurality of sub-steps or a plurality of stages, and these sub-steps or stages are not necessarily executed at the same time, but can be executed at different times, and the execution order of these sub-steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a part of the sub-steps or stages of other steps.

[0066] The above is only a preferred implementation of the method for intelligently generating SAR target data integrating geometric information disclosed in the present invention, and is not intended to limit the protection scope of the embodiments of this specification. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the embodiments of this specification shall be included in the protection scope of the embodiments of this specification.

Claims

1. A method for intelligently generating SAR target data by integrating geometric information, characterized in that: The steps include: Extract the contour of the target in the optical image, obtain the geometric structure diagram of the target, and use the geometric structure diagram to calculate the dominant scattering boundary of the target under different radar incident azimuths; Manually mark the targets in the optical image to obtain the target category labels; A SAR image generation network is constructed based on a generative adversarial network, wherein the SAR image generation network includes a generator, a discriminator, and a classifier that introduce a multi-layer attention mechanism; The dominant scattering boundary, geometric structure diagram and category label of the target at different radar incident azimuths are input into the generator to obtain the generated SAR image; The generated SAR image or the real SAR image is input into the discriminator to determine whether the image belongs to a real sample or a generated sample; The generated SAR image or the real SAR image and the corresponding category label are input into the classifier to determine the category of the image; the discriminator and the classifier both feed back the judgment results to the generator that introduces the multi-layer attention mechanism.

2. The method for intelligently generating SAR target data by integrating geometric information according to claim 1, characterized in that: The generator introducing the multi-layer attention mechanism includes an encoder, a connection unit and a decoder connected in sequence, wherein: The encoder is composed of an attention unit and four downsampling modules, each of which includes a convolutional layer, an InstanceNorm layer and a ReLU activation function connected in sequence; the attention unit includes a first attention module, a second attention module and a third attention module, the first attention module includes a convolutional layer and a Sigmoid function layer; the second attention module and the third attention module both include a convolutional layer, a ReLU activation function and a convolutional attention mechanism module; The connection unit includes five residual modules connected in sequence, each residual module includes a convolution layer, an InstanceNorm layer, a ReLU activation function, an InstanceNorm layer and a ReLU activation function connected in sequence; The decoder includes three upsampling modules and an output module connected in sequence; each upsampling module includes a deconvolution layer, an InstanceNorm layer and a ReLU activation function connected in sequence; the output module includes a convolution layer and a Tanh function connected in sequence.

3. The method for intelligently generating SAR target data by integrating geometric information according to claim 2, characterized in that: The processing process of the encoder includes the following steps: The dominant scattering boundary is input into the first attention module to obtain a weight map, and the weight map is element-wise multiplied with the dominant scattering boundary to obtain an adjusted feature map; the geometric structure map, the category label and the adjusted feature map are concatenated and input into the first downsampling module; The dominant scattering boundary is input into the second attention module to perform feature attention weighting in the channel and spatial dimensions in turn. The output of the second attention module is spliced ​​with the output of the first downsampling module as the input of the second downsampling module; the dominant scattering boundary passes through the convolution layer and ReLU activation function in the second attention module and is input into the third attention module to perform feature attention weighting in the channel and spatial dimensions in turn. The output of the third attention module is spliced ​​with the output of the second downsampling module as the input of the third downsampling module.

4. The method for intelligently generating SAR target data by integrating geometric information according to claim 2, characterized in that: The convolution kernel size of the convolution layer in the first attention module is 7×7; The convolution kernel sizes of the convolution layers in the four downsampling modules are all 3×3, and except for the convolution layer step size of the first downsampling module being set to 1, the convolution layer step sizes of the remaining downsampling modules are all set to 2; The convolution kernel size of the deconvolution layer in the three upsampling modules is 3×3, and the step size is set to 2; The convolution kernel size of the convolution layer in the output module is 7×7.

5. The method for intelligently generating SAR target data by integrating geometric information according to claim 1, characterized in that: The discriminator includes a first convolution module, a second convolution module, a third convolution module, a fourth convolution module and a fifth convolution module connected in sequence; the classifier includes a first convolution module, a second convolution module, a third convolution module, a first fully connected module and a second fully connected module connected in sequence, and the discriminator and the classifier share the first convolution module, the second convolution module and the third convolution module.

6. The method for intelligently generating SAR target data by integrating geometric information according to claim 5, characterized in that: The first convolution module is composed of a convolution layer and a LeakyReLU activation function, the second convolution module, the third convolution module and the fourth convolution module all include a convolution layer, an InstanceNorm layer and a LeakyReLU activation function connected in sequence, and the fifth convolution module is composed of a convolution layer and a Sigmoid activation function; the convolution kernel size of the convolution layer in the first convolution module, the second convolution module, the third convolution module, the fourth convolution module and the fifth convolution module is 3×3, and the step size is 2.

7. The method for intelligently generating SAR target data by integrating geometric information according to claim 5, characterized in that: The first fully connected module uses LeakyReLU as the activation function; the second fully connected module uses Softmax as the activation function.

Images