SAR (Synthetic Aperture Radar) image generation method of feature-guided multi-stage generative adversarial network, network, equipment and medium

Abstract

The invention provides an SAR image generation method of a feature-guided multi-stage generative adversarial network, a network, equipment and a medium, and the method comprises the following steps: in a background stage, carrying out the fitting of the statistical characteristics of an SAR image through G0 distribution, carrying out the random sampling, and generating a background noise image, thereby precisely capturing the statistical characteristics of the SAR image, and providing a real background environment for a subsequent stage; in the shape stage, an azimuth angle label of an SAR image is used as input to ensure that the similarity between the generated image and a real SAR image in shape is higher, and the accuracy and reality sense of the generated image in geometric structure are improved; in the target stage, the scattering center parameter of the SAR image is extracted to act on the generative network in the target stage, so that the detail performance of the target is enhanced, and the identifiability and definition of the target area are improved. According to the method, the statistical characteristics, azimuth angle characteristics and scattering center characteristics of the SAR image are utilized, image generation is guided in different stages, the generated image is highly similar to a real image, the detail fidelity is better, and better diversity is achieved.

Description

Technical Field

[0001] The present invention belongs to the field of image processing technology, and in particular relates to a SAR image generation method, network, device and medium based on a feature-guided multi-stage generative adversarial network. Background Art

[0002] Synthetic Aperture Radar (SAR) images have important applications in remote sensing, military reconnaissance, and environmental monitoring. However, the acquisition of SAR images is often affected by weather, equipment performance, and imaging environment, resulting in data scarcity and unstable quality.

[0003] CN119850970A discloses a SAR image generation model, method, device, system, and program product. The SAR image generation model is a generative adversarial network (GAC-GAN) with controlled generation angles, which is used to generate SAR images at a specified azimuth angle. The SAR image generation model includes a generator and a discriminator. The generator includes a feature extraction layer, a feature fusion layer, and an image generation layer. The generator extracts SAR image features from a first SAR input image and a second SAR input image at different azimuth angles and generates a SAR target image at a specified azimuth angle based on an azimuth angle control variable. The discriminator includes a similarity discriminator and an azimuth angle discriminator. The discriminator distinguishes the SAR target image generated by the generator from the corresponding real SAR label image, calculates the azimuth angle information of the generated image, and provides optimization direction for the generator based on the discrimination results. By introducing an azimuth angle control variable into the generative adversarial network, the generator is trained to generate SAR images at a specified azimuth angle. In this method, the generator performs only a single-stage generation process, without performing staged generation to address different generation requirements (such as background modeling, geometric structure, and detail enhancement). This approach is not conducive to optimizing background and target features separately, and may result in poor quality or inconsistency in some generated areas. Feature extraction utilizes a convolution kernel modulated by attributed scattering centers, but this still relies heavily on the image representation itself and does not explicitly incorporate higher-level target physical modeling, such as scattering center parameters, resulting in limited reliance on scattering feature extraction.

[0004] Existing networks for generating SAR images pay little attention to the inherent characteristics of SAR images. While network models based on generative adversarial networks have shown improvements in training stability and generated image quality, the generated images suffer from limited target details, such as high noise, geometric distortion, and texture loss. Therefore, there is an urgent need for an optimized generative adversarial network model to improve the quality of SAR image generation. Summary of the Invention

[0005] In order to solve the problem in the prior art that the network model based on the generative adversarial network for generating SAR images has poor image quality due to neglect of the inherent characteristics of SAR images, the present invention provides a SAR image generation method, network, device and medium based on a feature-guided multi-stage generative adversarial network. High-quality SAR images are generated through three stages, so that the generated SAR images perform well in terms of authenticity, detail fidelity and diversity, and have a higher similarity with real SAR images.

[0006] To achieve the above object, the present invention provides the following technical solutions: A SAR image generation method using a feature-guided multi-stage generative adversarial network includes the following steps: Step 1: Generate corresponding SAR background image, SAR shape image and SAR target image according to the background stage, shape stage and target stage; Among them: The generation process of SAR background image is: according to G 0 The distribution model is used to fit the statistical characteristics of the SAR image to obtain background noise that conforms to the statistical characteristics of the real SAR image; the background noise is used to generate a SAR background image that conforms to the statistical characteristics of the real SAR image; The generation process of SAR shape image is as follows: the SAR image azimuth is partitioned and re-encoded to reduce the dimension, and the azimuth code after dimension reduction is used as the category label input to generate the SAR shape image in the shape stage; The generation process of SAR target image is as follows: extract the attribute scattering center parameters of SAR image, take the extracted attribute scattering center parameters of SAR image as input, add them to the generation of SAR image target through attention mechanism, and obtain SAR target image; Step 2: Synthesize the SAR background image, SAR shape image, and SAR attribute image to obtain the final SAR generated image.

[0007] Preferably, in step 1, according to G 0 The distribution model fits the statistical characteristics of the SAR image and randomly samples noise to generate the SAR background image, including the following steps: Using G 0 The statistical model is constructed based on the statistical characteristics of the real SAR image, and random sampling is performed according to the statistical model to generate noise samples with corresponding statistical characteristics; The noise samples are input into the generative network of the background stage to generate the SAR background image.

[0008] Preferably, in step 1, the specific steps of partitioning and re-encoding are as follows: dividing the azimuth angle labels of the SAR image from 0° to 360° into intervals of 5°, and encoding the SAR image azimuth angles using N-Progressive encoding.

[0009] Preferably, in step 1, the process of extracting the attribute scattering center parameters of the SAR image is as follows: given the original SAR image , the estimated parameter set ; The attribute scattering center model is as follows:

[0010] in, is the estimated parameter set, Indicates the Parameters of the scattering centers; is the sampling frequency of radar reflected waves; is the radar center frequency; is the azimuth, and its value range is ,in represents the maximum imaging observation angle; Indicates the speed of radio waves in vacuum; is the position of the scattering center in range and azimuth; is an imaginary unit; is the complex amplitude; represents frequency dependence; when When , it corresponds to the local scattering center; represents the azimuthal dependence of the scattering center, when , corresponding to the distribution scattering center, its dependence on the azimuth is determined by the physical length of the target Tilt angle representation; Parameter set obtained from attribute scattering center model The maximum likelihood estimate of for:

[0011] In the formula, + represents the summation of positive terms.

[0012] Preferably, the specific steps of generating the SAR target image are as follows: First, the coordinates of the scattering center are extracted from the attribute scattering center parameters of the SAR image and the corresponding attribute values , and mapped to the corresponding position of the feature map to form a two-dimensional attribute heat map; The average pooling, maximum pooling, and attribute heatmap obtained after the convolution layer in the shape stage are channel-concatenated to obtain an input feature map that integrates global statistical information, local saliency information, and scattering center attribute information related to spatial position; The convolution layer is applied to map the spliced ​​input image to a single-channel output, and the final attention map is generated through the Sigmoid activation function. The attention map is multiplied element-by-element with the original input feature map to obtain the weighted output feature map, which is the SAR target image.

[0013] Preferably, in step 2, during synthesis, the SAR shape image is first superimposed on the SAR background image through a mask to obtain an intermediate image; then the SAR target image is superimposed on the intermediate image through a mask to obtain a final SAR generated image containing the background, shape and target.

[0014] A SAR image generation network based on a feature-guided multi-stage generative adversarial network, including background stage network structure, shape stage network structure, target stage network structure and synthesis network structure; The background stage network structure is used according to G 0 The distribution model is used to fit the statistical characteristics of the SAR image to obtain background noise that conforms to the statistical characteristics of the real SAR image; the background noise is used to generate a SAR background image that conforms to the statistical characteristics of the real SAR image; The shape stage network structure is used to reduce the SAR image azimuth by partitioning and re-encoding, and use the azimuth encoding after dimensionality reduction as the category label input to generate the SAR shape image in the shape stage; The target phase network structure is used to extract the attribute scattering center parameters of the SAR image, and then add the extracted attribute scattering center parameters of the SAR image to the generation of the SAR image target through the attention mechanism to obtain the SAR target image; The synthesis network structure is used to synthesize the SAR background image, SAR shape image and SAR attribute image to obtain a complete SAR generated image.

[0015] Preferably, it further includes an identification module, which includes a background identifier, a shape identifier and a target identifier; The background discriminator is used to calculate the background truth and the background-foreground category probability; The shape discriminator is used to ensure shape encoding You can control the foreground and mask images in the shape stage The generation of The object discriminator is used to output category scores and judge whether the image is real or fake.

[0016] A computer device comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the steps of the method are implemented when the processor executes the computer program.

[0017] A computer-readable storage medium stores a computer program, wherein the computer program implements the steps of the method when executed by a processor.

[0018] Compared with the prior art, the present invention has the following beneficial technical effects: The present invention proposes a SAR image generation method based on feature-guided multi-stage generative adversarial network to generate high-quality SAR images through three stages. Specifically, first, in the background generation stage, G 0 The distribution model fits the statistical characteristics of the SAR image, and a background noise image is generated through random sampling. This ensures that the generated background image accurately reflects the statistical characteristics of the real SAR image, providing a realistic background environment for subsequent processing. Secondly, in the shape generation stage, the SAR image's azimuth label is used as a conditional input to ensure that the generated image is highly consistent in geometry with the real SAR image, improving the image's geometric accuracy and realism. Finally, in the target generation stage, the scattering center parameters of the SAR image are extracted and used as the core input of the attention mechanism, guiding the generative network to extract and fuse features, thereby enhancing the expressiveness of target features and improving the detail and clarity of the target area.

[0019] The present invention utilizes the statistical characteristics, azimuth characteristics and scattering center characteristics of SAR images to guide image generation at different stages. The generated images are not only highly similar to real images, but also have better detail fidelity and greater diversity. The generated SAR images can be used to expand the target recognition dataset, thereby improving the accuracy of SAR target recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 This is a flow chart of a SAR image generation method based on a feature-guided multi-stage generative adversarial network according to the present invention; Figure 2 This is the background stage G of the present invention 0 Schematic diagram of the network structure for fitting SAR statistical characteristics with distribution models and image generation; Figure 3 This is a schematic diagram of the network structure of the shape stage azimuth label encoding method and image generation in the present invention; Figure 4 Schematic diagram of the network structure for attribute scattering center parameter extraction and image generation in the target stage of the present invention; Figure 5 These are images of each stage generated using the present invention. DETAILED DESCRIPTION

[0021] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of the present invention.

[0022] It should be noted that the terms "first", "second", etc. in the description and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that the numbers used in this way are interchangeable where appropriate so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices. The present invention is described in detail below in conjunction with the accompanying drawings.

[0023] The present invention discloses a SAR image generation method based on a feature-guided multi-stage generative adversarial network. Figure 1 , including the following steps: Step 1: Generate corresponding SAR background image, SAR shape image and SAR target image according to the background stage, shape stage and target stage; Among them: The generation process of SAR background image is: according to G 0 The distribution model is fitted with the statistical characteristics of the SAR image to obtain background phase noise with SAR statistical characteristics; the background phase noise is used to generate a SAR background phase image to generate a SAR background part image with SAR statistical characteristics; The generation process of SAR shape images is as follows: the SAR image azimuth is partitioned and encoded, and the azimuth encoding is used as the category label input to guide the generation of the SAR image shape stage, so that the generated image has azimuth controllability; The SAR target image generation process is as follows: extracting the attribute scattering center parameters of the SAR image; adding the extracted SAR image attribute scattering center parameters to the image generation of the SAR image target stage through the attention mechanism, guiding the generated SAR image to have more realistic details; In step 2, the SAR images of different stages (SAR background image, SAR shape image, and SAR target image) generated in step 1 are stitched together to obtain a complete SAR generated image.

[0024] In the present invention, Figure 2 Background stage G 0 Schematic diagram of the network structure for fitting SAR statistical characteristics with the distribution model and image generation. The generation of the SAR background image in the background stage includes the following steps: Give the generator Input noise and background coding Concatenate to form a feature matrix, where the background code is the category label of the SAR image, noise From G 0 The noise is randomly sampled after distribution fitting of the real SAR image dataset; Use five upsampling layers to gradually enlarge the feature map, and finally map the feature map to the pixel space to generate the background image; In the background identification stage, the background discriminator is divided into true and false discriminators and background classification discriminator , where the feature extraction part includes three 4×4 convolutional layers to extract background features, the true and false discriminator part includes a convolutional layer and a Sigmoid activation function to calculate the authenticity; the background classification discriminator also includes a convolutional layer and a Sigmoid activation function to calculate the probability that the image is in the background stage; Background stage true and false discriminator The loss function for:

[0025] in, For real samples, background classification discriminator The loss function for:

[0026] Representation based on real images The loss term is the expected value of the real image under the discriminator output, Represents noise and tags The expected loss obtained under , reflects the consistency between the generated image and the corresponding label.

[0027] In the present invention, Figure 3 This is a diagram of the network structure for encoding the azimuth label and generating images in the shape phase. Image generation in the shape phase includes the following steps: The shape stage network structure includes a generator and the discriminator , to the generator Input noise and azimuth labels The feature map formed by splicing is gradually enlarged using five upsampling layers, and the obtained feature map is then combined with the azimuth label The shape stage feature map is obtained by combining the convolution layer and the residual block. The feature map is mapped to the pixel space through a 3×3 convolution layer. The foreground image of the generated shape stage and generate the shape stage mask image Composition. The feature matrix extracted from the shape stage is used as the input of the target stage. The target stage continues to extract features based on this, and the azimuth label Labels after azimuth repartition encoding of SAR images; In the shape identification stage, it also includes feature extraction and category prediction parts. The feature extraction part includes multiple convolutional layers, batch normalization and LeakyReLU Activation function. The category prediction part outputs the final prediction value through a 4×4 convolution; Shape stage loss for:

[0028] in, The azimuth label input to the generator, They are The output foreground image and Output mask image, ensuring shape encoding Can control foreground and mask images The generation of Represents noise and tags The expected loss obtained under , reflects the consistency between the generated image and the corresponding label.

[0029] In the present invention, Figure 4 The network structure diagram for the target phase attribute scattering center parameter extraction and image generation is shown in Figure 1. The extraction of SAR image attribute scattering center parameters includes the following steps: Image segmentation, model category selection, parameter initialization selection and parameter optimization, etc. The details of model selection and parameter initialization and optimization are as follows: The specific form of the attribute scattering center model is as follows:

[0030] in, is the estimated parameter set, Indicates the Parameters of the scattering centers; is the sampling frequency of radar reflected waves; is the radar center frequency; is the azimuth, and its value range is ,in represents the maximum imaging observation angle; Indicates the speed of radio waves in vacuum; is the position of the scattering center in range and azimuth; is an imaginary unit; is the complex amplitude; Indicates frequency dependence. When , it corresponds to the local scattering center; represents the azimuthal dependence of the scattering center, when , corresponding to the distribution scattering center, its dependence on the azimuth is determined by the physical length of the target Tilt angle representation.

[0031] The problem of estimating the center parameters of target scattering features based on SAR images can be summarized as follows: given a SAR image , the estimated parameter set , so that the SAR image simulated by this parameter set can best fit ; Assume that flat( 、 They are direction), the frequency domain resampled data of the target A sum factor of the scattering response that exactly matches the model With Gaussian white noise Composed of:

[0032] The above formula is 2-dimensional Matrix representation method, since the parameter initialization and optimization part in the scattering center extraction requires vector form, in order to facilitate subsequent calculations, it is converted into dimensional column vectors and are denoted by lowercase letters:

[0033] at this time, is a normally distributed random vector, After the imaging process, the relationship between the image domain data becomes:

[0034] Noise Factor , is a Gaussian color noise vector. A Gaussian color noise vector refers to a noise vector that obeys Gaussian distribution but has correlation. dimensional image vector The probability density function of is:

[0035] in, represents the covariance matrix of the noise, -1 represents the generalized inverse, represents the complex conjugate transpose, is a dimensional image vector, the parameter set can be obtained by the attribute scattering center model The maximum likelihood estimate of for:

[0036] In the formula, + represents the sum of positive terms; The parameters of the attribute scattering center are used as the input of the target stage to guide the generation network to perform feature extraction and fusion.

[0037] The attention mechanism is used to control the generation of SAR image targets using parameters, including the following steps: After extracting the attribute scattering center parameters of the SAR image, an attribute heat map is generated for each sample. Specifically, the coordinates of the scattering center and the corresponding attribute values The attributes are mapped to the corresponding positions of the feature map to form a two-dimensional attribute map. The attribute values ​​are normalized to ensure that they are within the valid range and mapped to the heat map in the following way:

[0038]

[0039] here represents the index of the sample, is the position of the scattering center; The average pooling, maximum pooling and attribute heat map generated after the convolution layer are channel-joined to obtain a fused input feature map containing global statistical information, local saliency information and scattering center attribute information related to spatial position; The fused input feature map is processed through convolution operations. Specifically, the present invention applies a convolution layer to map the spliced ​​input image to a single-channel output, and generates a final attention map through a Sigmoid activation function. The attention map is element-wise multiplied with the original input feature map to obtain a weighted output feature map. Target Phase Coding is the SAR image category label, which is used as the conditional input vector and the feature matrix extracted in the previous stage Perform splicing as input to the target stage.

[0040] The target generator consists of a concatenated convolutional layer, a residual layer, a same-size convolutional layer, and an attention mechanism module. The attention mechanism module pools the feature map and fuses the attribute heat map of the scattering center parameters to enhance the feature expression of the scattering center area. The target stage network structure includes the generator and the category discriminator and the true and false discriminator . By generating the foreground image of the target stage and generate the target stage mask image composition; Target stage discriminators include true and false discriminators and category discriminator , which includes five layers of convolution and LeakyReLU activation function. Category discriminator The identification module is a 4×4 convolutional layer and a Sigmoid activation function, which outputs the category score. The identification module is a 4×4 convolutional layer, which is used to judge whether the image is real or fake; Among them, the true and false discriminator in the target stage The loss function for:

[0041] Category Discriminator The loss function for:

[0042] in, For real samples, It is the final generated image after the three-stage generated images are stitched together. encoding the target phase of the generator input, They are The output foreground image and the output Mask image to ensure target stage encoding Can control foreground and mask images The generation of Representation based on real images The loss term is the expected value of the real image under the discriminator output, Representation based on tags The loss term reflects the similarity and consistency between the generated image and the corresponding label.

[0043] The present invention provides a SAR image generation method based on a feature-guided multi-stage generative adversarial network. In the background stage, G 0 The statistical characteristics of the SAR image are distributed and fitted, and background noise images are randomly sampled to accurately capture the statistical characteristics of the SAR image and provide a realistic background environment for subsequent stages. In the shape stage, the azimuth label of the SAR image is used as input to ensure that the generated image is more similar to the real SAR image in shape, improving the accuracy and realism of the generated image in terms of geometric structure. In the target stage, the scattering center parameters of the SAR image are extracted and applied to the target stage generation network to enhance the detail representation of the target and improve the recognizability and clarity of the target area. The present invention utilizes the statistical characteristics, azimuth characteristics, and scattering center characteristics of SAR images to guide image generation at different stages. The generated images are not only highly similar to the real images, but also have better detail fidelity and greater diversity.

[0044] Compared with the method in CN119850970A, the present invention introduces a multi-stage generation structure to refine the task objectives. 0 Distributed simulated noise is independently modeled to enhance background realism. Azimuth labels are used to generate geometric structures during the shape phase. During the target phase, scattering center parameters are used to enhance detail, ensuring accurate and clear key areas, improving the control accuracy and realism of various features. Furthermore, the scattering center parameters in this invention are not implicitly learned but are explicitly extracted and applied to the generative network, making their physical meaning more explicit.

[0045] In order to verify the effect of the present invention, the following simulation experiments are carried out: The MSTAR 10-category dataset was used as experimental data. The dataset contains SAR images of 10 different ground vehicles. These SAR images, with a resolution of 0.3 m, are all X-band and 128×128 pixels. The images in this dataset are at two elevation angles: 15° and 17°. During the experiment, SAR images at a 17° elevation angle were selected as training data, and those at a 15° elevation angle were selected as testing data. The azimuth angle range is 0°–360°. The number of SAR images of different categories in the 10-category dataset is shown in Table 1. Preprocessing was performed to select only the central 64×64 images from the original 128×128 images as training images. The SAR images were normalized before use.

[0046] Table 1 Number of SAR images of different categories

[0047] In the present invention, the training round Set to 1500, train the discriminator in each training round =5 times. Training batch Set to 64, learning rate All are set to 0.0002. The optimizer uses Adam. and are the parameters of Adam.

[0048] In the present invention, Figure 5 These are images generated at various stages in the example disclosed herein.

[0049] In order to verify the beneficial effects of the present invention, the present invention is used together with the InfoGAN algorithm (Chen X, Duan Y, Houthooft R, et al. "InfoGAN: Interpretable representation learning byinformation maximizing generative adversarial nets," Advances in neural information processing systems, vol. 29, pp. 2180-2188, 2016.), the ACGAN algorithm (Odena A, Olah C, Shlens J, "Conditional image synthesis with auxiliary classifierGANs," International conference on machine learning, pp. 2642-2651, 2017.), and the WGAN-GP algorithm (Jozdani S, Chen D, Pouliot D, et al. "A review and meta-analysis ofgenerative adversarial networks and their applications in remote sensing," International Journal of Applied Earth Observation and Geoinformation, vol. 108, 102734, 2022, doi: 10.1016 / j.jag.2022.102734.) conducted comparative experiments by using SAR image generation algorithms of different models for image generation and comparing the FID values ​​of different generated images. The experimental results are shown in Table 2. The recognition accuracy of different recognition networks after the generated images were used to expand the original dataset was compared. The recognition networks were Resnet18 and Vggnet11, respectively. The experimental results are shown in Table 3.

[0050] Table 2 FID values ​​of SAR generation algorithms for different models

[0051] As can be seen from Table 2, the images generated by the present invention have smaller FID values ​​than those generated by other algorithms, indicating that the SAR images generated by the present invention perform better in quality and diversity than the other three comparative experiments.

[0052] Table 3 Recognition experiment results

[0053] As shown in Table 3, using the ResNet18 recognition network, the recognition accuracy of the image augmentation dataset generated using the present invention is improved compared to both the case without data augmentation and the image augmentation datasets generated using other algorithms. Using the VggNet11 recognition network, the recognition rate is also improved when the image augmentation dataset is generated using the present invention, compared to both the case without data augmentation and the image augmentation datasets generated using other algorithms. In both recognition networks, the recognition rate of the SAR images generated using the present invention's augmentation datasets is improved, demonstrating that the present method can improve the recognition rate of SAR datasets.

[0054] The present invention also discloses a SAR image generation network of a feature-guided multi-stage generative adversarial network, including a background stage network structure, a shape stage network structure, a target stage network structure and a synthesis network structure; The background stage network structure is used according to G 0 The distribution model fits the statistical characteristics of the SAR image to obtain background noise that conforms to the statistical characteristics of the real SAR image; the background noise is used to generate a SAR background image that conforms to the statistical characteristics of the real SAR image; the shape stage network structure is used to reduce the SAR image azimuth by partitioning and re-encoding, and the azimuth encoding after dimensionality reduction is used as the category label input to generate the SAR shape image of the shape stage; the target stage network structure is used to extract the attribute scattering center parameters of the SAR image, and the extracted SAR image attribute scattering center parameters are added to the generation of the SAR image target through the attention mechanism to obtain the SAR target image; the synthesis network structure is used to synthesize the SAR background image, SAR shape image and SAR attribute image to obtain a complete SAR generated image.

[0055] Also included is an identification module, the identification module including a background identifier, a shape identifier, and a target identifier; The background discriminator is used to calculate the background truth and the background foreground category probability; the shape discriminator is used to ensure the shape encoding You can control the foreground and mask images in the shape stage The target discriminator is used to output category scores and judge whether the image is real or fake.

[0056] The present invention also discloses a computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the method when executing the computer program. The processor may be a central processing unit (CPU), or may be other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing core and control core of the terminal, and is suitable for implementing one or more instructions, specifically for loading and executing one or more instructions in a computer storage medium to implement the corresponding method flow or corresponding function; the processor described in the embodiment of the present invention can implement the operation of the SAR image generation method of the feature-guided multi-stage generative adversarial network.

[0057] The present invention also discloses a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the method. The computer-readable storage medium is a memory device in a computer device, used to store programs and data. It is understood that the computer-readable storage medium herein may include both built-in storage media in the computer device and, of course, extended storage media supported by the computer device. The computer-readable storage medium provides storage space that stores the terminal's operating system. Furthermore, the storage space also stores one or more instructions suitable for being loaded and executed by the processor. These instructions may be one or more computer programs (including program code). It should be noted that the computer-readable storage medium herein may be a high-speed RAM memory or a non-volatile memory, such as at least one disk storage device. The processor may load and execute the one or more instructions stored in the computer-readable storage medium to implement the corresponding steps of the SAR image generation method using a feature-guided multi-stage generative adversarial network described in the above-mentioned embodiment.

[0058] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0059] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0060] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0061] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0062] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A SAR image generation method using a feature-guided multi-stage generative adversarial network, comprising the following steps: Step 1: Generate corresponding SAR background image, SAR shape image and SAR target image according to the background stage, shape stage and target stage; Among them: The generation process of SAR background image is: according to G 0 The distribution model is used to fit the statistical characteristics of the SAR image to obtain background noise that conforms to the statistical characteristics of the real SAR image; the background noise is used to generate a SAR background image that conforms to the statistical characteristics of the real SAR image; The generation process of SAR shape image is as follows: the SAR image azimuth is partitioned and re-encoded to reduce the dimension, and the azimuth code after dimension reduction is used as the category label input to generate the SAR shape image in the shape stage; The generation process of SAR target image is as follows: extract the attribute scattering center parameters of SAR image, take the extracted attribute scattering center parameters of SAR image as input, add them to the generation of SAR image target through attention mechanism, and obtain SAR target image; Step 2: Synthesize the SAR background image, SAR shape image, and SAR attribute image to obtain the final SAR generated image.

2. The SAR image generation method using a feature-guided multi-stage generative adversarial network according to claim 1, characterized in that: In step 1, according to G 0 The distribution model fits the statistical characteristics of the SAR image and randomly samples noise to generate the SAR background image, including the following steps: Using G 0 The statistical model is constructed based on the statistical characteristics of the real SAR image, and random sampling is performed according to the statistical model to generate noise samples with corresponding statistical characteristics; The noise samples are input into the generative network of the background stage to generate the SAR background image.

3. The SAR image generation method using a feature-guided multi-stage generative adversarial network according to claim 1, characterized in that: In step 1, the specific steps of partitioning and re-encoding are as follows: the azimuth labels of the SAR image from 0° to 360° are divided into intervals of 5°, and the azimuth angles of the SAR image are encoded using N-Progressive encoding.

4. The SAR image generation method using a feature-guided multi-stage generative adversarial network according to claim 1, characterized in that: In step 1, The process of extracting the attribute scattering center parameters of the SAR image is as follows: given the original SAR image , the estimated parameter set ; The attribute scattering center model is as follows: in, is the estimated parameter set, Indicates the Parameters of the scattering centers; is the sampling frequency of radar reflected waves; is the radar center frequency; is the azimuth, and its value range is ,in represents the maximum imaging observation angle; Indicates the speed of radio waves in vacuum; is the position of the scattering center in range and azimuth; is an imaginary unit; is the complex amplitude; represents frequency dependence; when When , it corresponds to the local scattering center; represents the azimuthal dependence of the scattering center, when , corresponding to the distribution scattering center, its dependence on the azimuth is determined by the physical length of the target Tilt angle representation; Parameter set obtained from attribute scattering center model The maximum likelihood estimate of for: In the formula, + represents the summation of positive terms.

5. The SAR image generation method using a feature-guided multi-stage generative adversarial network according to claim 1, characterized in that: The specific steps of SAR target image generation are as follows: First, the coordinates of the scattering center are extracted from the attribute scattering center parameters of the SAR image and the corresponding attribute values , and mapped to the corresponding position of the feature map to form a two-dimensional attribute heat map; The average pooling, maximum pooling, and attribute heatmap obtained after the convolution layer in the shape stage are channel-concatenated to obtain an input feature map that integrates global statistical information, local saliency information, and scattering center attribute information related to spatial position; The convolution layer is applied to map the spliced ​​input image to a single-channel output, and the final attention map is generated through the Sigmoid activation function. The attention map is multiplied element-by-element with the original input feature map to obtain the weighted output feature map, which is the SAR target image.

6. The SAR image generation method using a feature-guided multi-stage generative adversarial network according to claim 1, characterized in that: In step 2, during synthesis, the SAR shape image is first superimposed on the SAR background image through a mask to obtain an intermediate image; then the SAR target image is superimposed on the intermediate image through a mask to obtain the final SAR generated image containing the background, shape, and target.

7. A SAR image generation network based on a feature-guided multi-stage generative adversarial network, characterized in that: Including background stage network structure, shape stage network structure, target stage network structure and synthesis network structure; The background stage network structure is used according to G 0 The distribution model fits the statistical characteristics of the SAR image to obtain the background noise that conforms to the statistical characteristics of the real SAR image; Use background noise to generate SAR background image, and generate SAR background image that conforms to the statistical characteristics of real SAR image; The shape stage network structure is used to reduce the SAR image azimuth by partitioning and re-encoding, and the azimuth encoding after dimensionality reduction is used as the category label input to generate the SAR shape image in the shape stage; The target phase network structure is used to extract the attribute scattering center parameters of the SAR image, and then add the extracted attribute scattering center parameters of the SAR image to the generation of the SAR image target through the attention mechanism to obtain the SAR target image; The synthesis network structure is used to synthesize the SAR background image, SAR shape image and SAR attribute image to obtain a complete SAR generated image.

8. The SAR image generation network of the feature-guided multi-stage generative adversarial network according to claim 7, characterized in that: Also included is an identification module, the identification module including a background identifier, a shape identifier, and a target identifier; The background discriminator is used to calculate the background truth and the background-foreground category probability; The shape discriminator is used to ensure shape encoding You can control the foreground and mask images in the shape stage The generation of The object discriminator is used to output category scores and judge whether the image is real or fake.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 6 are implemented.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

Images