A method and device for three-dimensional reconstruction of synthetic aperture radar images

Abstract

The application provides a method and device for three-dimensional reconstruction of synthetic aperture radar images, which comprises: converting single-view complex data collected based on a small number of observation angles into an initial three-dimensional point cloud covering the target contour; parameterizing the point cloud into an initial three-dimensional Gaussian scene representation; using a pre-trained two-dimensional diffusion model to optimize the representation through multi-view consistency optimization by score distillation sampling to complete geometric and texture details; and finally extracting a target three-dimensional point cloud with enhanced details from the optimized Gaussian representation. The data acquisition requirement is significantly reduced, high-quality, high-consistency and physically interpretable three-dimensional reconstruction is achieved, and the reconstruction efficiency and practicability are improved.

Description

Technical Field

[0001] This application relates to the field of image processing technology, and more specifically, to a method and apparatus for three-dimensional reconstruction of synthetic aperture radar images. Background Technology

[0002] Synthetic Aperture Radar (SAR), as an active microwave imaging sensor, possesses all-weather, all-day Earth observation capabilities. SAR images acquired by SAR are of significant value in surveying, reconnaissance, and disaster monitoring. Recovering the three-dimensional structure of targets from SAR images—a process known as SAR 3D reconstruction—is a crucial step in enhancing its application value.

[0003] Traditional SAR 3D reconstruction methods are mainly divided into two categories: one is tomographic SAR (TomoSAR) technology based on the multi-baseline tomography principle, and the other is a data-driven method based on deep learning. TomoSAR technology constructs a synthetic aperture in the elevation direction using multi-track, multi-baseline SAR observation data, and uses spectral estimation or sparse reconstruction algorithms to invert the vertical structure of the target to achieve 3D imaging. This method heavily relies on a large amount of precisely registered multi-baseline data, resulting in high data acquisition costs and long cycles. It is also sensitive to orbital errors and system noise, easily producing tomographic artifacts, and the reconstruction quality degrades when data is insufficient or noise is high.

[0004] Data-driven methods, such as deep learning, learn the 3D structure mapping relationship directly from SAR data by training neural networks. These methods avoid complex physical model inversion, but often require large-scale, high-quality labeled data for training, and the models have poor interpretability, making it difficult to incorporate the electromagnetic scattering physics mechanism unique to SAR into the reconstruction results, resulting in limited generalization ability. Summary of the Invention

[0005] In view of this, the purpose of this application is to provide a method and apparatus for three-dimensional reconstruction of synthetic aperture radar images, so as to overcome the problems in the prior art.

[0006] In a first aspect, embodiments of this application provide a method for three-dimensional reconstruction of synthetic aperture radar images, the method comprising: The initial single-view complex data collected based on the first number of observation angles is transformed into an initial three-dimensional point cloud that can cover the outline of the observed object. The initial 3D point cloud is initialized with Gaussian to obtain a parameterized initial 3D Gaussian scene representation; Using a pre-trained two-dimensional diffusion model, the initial three-dimensional Gaussian scene representation is optimized for multi-view consistency through score distillation sampling to complete the geometric and texture details of the observed object, thereby obtaining the optimized target parameterized three-dimensional Gaussian scene representation. The detailed-enhanced 3D point cloud of the target is extracted from the parameterized 3D Gaussian scene representation of the target.

[0007] In some technical solutions of this application, the above-mentioned initial single-view complex data acquired based on a first number of observation angles is transformed into an initial three-dimensional point cloud capable of covering the outline of the observed object, including: The initial single-view complex data acquired based on the first number of observation angles is processed into a grayscale image under multiple viewpoints. The initial three-dimensional point cloud is generated by synthesizing three-dimensional information based on the grayscale image and extracting the main scattering structures.

[0008] In some technical solutions of this application, the above-mentioned processing of the initial single-view complex data acquired based on a first number of observation angles into a grayscale image under multiple views includes: Phase error estimation and compensation are performed on the initial single-view complex data at each observation angle to obtain the target single-view complex data. The target single-view complex data is resampled on the frequency axis so that the sampling interval meets the preset imaging requirements, and uniformly sampled frequency domain data is obtained. The uniformly sampled frequency domain data is transformed from the frequency domain to the spatial domain to generate grayscale images corresponding to each viewpoint.

[0009] In some technical solutions of this application, the above-mentioned synthesis of three-dimensional information based on the grayscale image and extraction of the main scattering structure to generate the initial three-dimensional point cloud includes: The grayscale images from the multiple perspectives are projected onto a unified three-dimensional spatial coordinate system, and incoherent energy accumulation is performed in the three-dimensional voxel grid to generate a three-dimensional scattering intensity distribution. The three-dimensional scattering intensity distribution is segmented by a threshold, and voxels with energy higher than a preset threshold are extracted; The spatial locations of the extracted voxels are used as points to generate the initial three-dimensional point cloud.

[0010] In some technical solutions of this application, the above-mentioned initialization of the initial 3D point cloud into a Gaussian representation to obtain a parameterized initial 3D Gaussian scene representation includes: The spatial coordinates of the initial three-dimensional point cloud are mapped to the center position parameters of a three-dimensional Gaussian distribution. Based on the spatial distribution characteristics of the initial 3D point cloud, the scale parameters of the 3D Gaussian distribution are determined. The electromagnetic scattering intensity information of the initial three-dimensional point cloud is mapped to the opacity parameter of a three-dimensional Gaussian distribution. The geometric orientation information of the initial 3D point cloud is mapped to color parameters of a 3D Gaussian distribution.

[0011] In some technical solutions of this application, the above-mentioned pre-trained two-dimensional diffusion model is used to perform multi-view consistency optimization on the initial three-dimensional Gaussian scene representation through score distillation sampling to complete the geometric and texture details of the observed object, thereby obtaining the optimized target parameterized three-dimensional Gaussian scene representation, including: The initial three-dimensional Gaussian scene representation is rendered from multiple different perspectives to obtain a two-dimensional rendered image from each perspective. The noise-added 2D rendered images are input into the pre-trained 2D diffusion model to calculate the score distillation sampling gradient for each viewpoint. The scoring distillation sampling gradients from multiple different perspectives are fused to obtain the fused optimized gradient; Based on the fused optimized gradient, the parameters of the initial 3D Gaussian scene representation are iteratively updated until the target parameterized 3D Gaussian scene representation is obtained.

[0012] In some technical solutions of this application, the extraction of detail-enhanced target 3D point clouds from the target parameterized 3D Gaussian scene representation includes: Obtain the geometric center position parameters and opacity parameters of each three-dimensional Gaussian distribution from the target parameterized three-dimensional Gaussian scene representation; Based on preset opacity filtering conditions, target distributions that meet the conditions are selected from each of the three-dimensional Gaussian distributions. The geometric center of the selected target distribution is determined as a point in the point cloud, and the points are aggregated to form the detailed enhanced target 3D point cloud.

[0013] Secondly, embodiments of this application provide an apparatus for three-dimensional reconstruction of synthetic aperture radar images, the apparatus comprising: The data conversion module is used to convert the initial single-view complex data collected based on a first number of observation angles into an initial three-dimensional point cloud that can cover the outline of the observed object. The Gaussian modeling module is used to initialize the initial 3D point cloud with Gaussian to obtain a parameterized initial 3D Gaussian scene representation; The diffusion optimization module is used to optimize the initial three-dimensional Gaussian scene representation by multi-view consistency through scoring distillation sampling using a pre-trained two-dimensional diffusion model, so as to complete the geometric and texture details of the observed object and obtain the optimized target parameterized three-dimensional Gaussian scene representation. The point cloud extraction module is used to extract the target 3D point cloud with enhanced details from the target parameterized 3D Gaussian scene representation.

[0014] Thirdly, embodiments of this application provide an electronic device, a processor, a memory, and a bus. The memory stores machine instructions executed by the processor. When the electronic device is running, the processor communicates with the memory via the bus. When the machine instructions are executed by the processor, the steps of the above-described method for three-dimensional reconstruction of synthetic aperture radar images are performed.

[0015] Fourthly, embodiments of this application provide a computer storage medium storing a computer program, which, when run by a processor, executes the steps of the above-described method for three-dimensional reconstruction of synthetic aperture radar images.

[0016] The technical solutions provided by the embodiments of this application may include the following beneficial effects: The method of this application includes: acquiring initial single-view complex data based on a first number of observation angles and converting it into an initial 3D point cloud capable of covering the outline of the observed object; initializing the initial 3D point cloud with Gaussian to obtain a parameterized initial 3D Gaussian scene representation; using a pre-trained 2D diffusion model, performing multi-view consistency optimization on the initial 3D Gaussian scene representation through score distillation sampling to complete the geometric and texture details of the observed object, thereby obtaining an optimized target parameterized 3D Gaussian scene representation; and extracting the detail-enhanced target 3D point cloud from the target parameterized 3D Gaussian scene representation.

[0017] This method requires only a small amount of observation data to generate an initial 3D point cloud, significantly reducing data acquisition costs and time. By converting the point cloud into a finely tuned 3D Gaussian scene representation and utilizing a pre-trained diffusion model for multi-view optimization, it can intelligently complete geometric details, enhance textures, and ensure consistent reconstruction across different viewpoints. The entire process incorporates the physical properties of SAR data during optimization, resulting in results that combine high accuracy with good physical interpretability, thus improving overall processing efficiency.

[0018] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1A schematic flowchart of a method for three-dimensional reconstruction of synthetic aperture radar images provided in an embodiment of this application is shown; Figure 2 A schematic diagram of a prior art method provided in an embodiment of this application is shown; Figure 3 A schematic diagram of another prior art method provided by an embodiment of this application is shown; Figure 4 A schematic diagram of a pre-imaging stage flowchart provided in an embodiment of this application is shown; Figure 5 A schematic diagram of a completion stage provided by an embodiment of this application is shown; Figure 6 A schematic diagram of an apparatus for three-dimensional reconstruction of synthetic aperture radar images provided in an embodiment of this application is shown; Figure 7 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. It should be understood that the accompanying drawings in this application are for illustrative and descriptive purposes only and are not intended to limit the scope of protection of this application. Furthermore, it should be understood that the schematic drawings are not drawn to scale. The flowcharts used in this application illustrate operations implemented according to some embodiments of this application. It should be understood that the operations in the flowcharts may not be implemented in sequence, and steps without logical contextual relationships may be reversed or implemented simultaneously. In addition, those skilled in the art, guided by the content of this application, may add one or more other operations to the flowcharts, or remove one or more operations from the flowcharts.

[0022] Furthermore, the described embodiments are merely some, not all, of the embodiments of this application. The components of the embodiments of this application described and illustrated herein can typically be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of the application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0023] It should be noted that the term "comprising" will be used in the embodiments of this application to indicate the presence of the features declared thereafter, but does not exclude the addition of other features.

[0024] Existing SAR (Synthetic Aperture Radar) image reconstruction methods mainly include multi-baseline sparse reconstruction methods and deep learning-based reconstruction methods. Multi-baseline sparse reconstruction methods, represented by TomoSAR, are techniques that, based on traditional SAR two-dimensional imaging, reconstruct the scattering structure in the third dimension (vertical / elevation direction) through multi-view observations. It can be used to separate multiple targets from a single resolution cell, solving problems such as overlapping and volume scattering aliasing in traditional SAR images. Its core idea is to conduct multi-track, multi-baseline, and multi-temporal radar observations of the target area from different angles to obtain radar scattering information of the same ground object at different incident angles. By constructing a tomographic imaging model, the scattering distribution in the elevation direction is reconstructed, essentially extending two-dimensional SAR imaging to three-dimensional volume imaging. A diagram of the TomoSAR technical scheme is shown below. Figure 2 As shown.

[0025] TomoSAR performs radar imaging of the same ground target from multiple perspectives, obtaining multi-track, multi-baseline SAR observation data, and synthesizes the aperture in the elevation direction. It then performs unified processing on the multi-view SAR images, including geometric correction, phase registration, and coherence filtering, to improve the accuracy of subsequent modeling and ensure geometric and phase consistency among the observation data. After preprocessing, the multi-view SAR echo signals are modeled as a linear observation system.

[0026] Where y: echo vectors acquired from N viewpoints; A: observation matrix composed of incident geometry and system parameters from each viewpoint; x: elevation direction scattering coefficient distribution to be solved; n: noise and modeling error terms.

[0027] The above model is solved using a high-resolution spectral estimation algorithm or a sparse inversion method to recover the volume scattering structure along the elevation direction from the observation data, achieving pixel-level vertical separation and structural decoupling. The final reconstructed model outputs a three-dimensional scattering point cloud map.

[0028] This method heavily relies on a large amount of accurately registered multi-baseline data, which is costly and time-consuming to acquire data. It is also sensitive to orbital errors and system noise, and is prone to tomographic artifacts. When data is insufficient or noise is strong, the reconstruction quality deteriorates.

[0029] With the widespread application of deep neural networks in the field of image processing, researchers have gradually introduced them into the reconstruction task of SAR images to solve the problems of insufficient image quality, noise resistance, and recovery ability in scenarios with incomplete data in traditional imaging methods.

[0030] Deep learning-based SAR image reconstruction methods typically take SAR echo signals, undersampled images, or low-resolution images as input, and train and optimize them using convolutional neural networks (CNNs), generative adversarial networks (GANs), or Transformer architectures to generate high-quality, high-resolution reconstructed images. A diagram of a deep learning-based SAR image reconstruction technique is shown below. Figure 3 As shown.

[0031] By utilizing satellite or airborne platforms to image the same ground area multiple times at different times or orbital locations, rich information on perspective and temporal differences is obtained, resulting in multi-view, multi-temporal SAR images. Geometric alignment and phase and amplitude corrections are then performed on the multi-temporal images to eliminate systematic biases such as orbital errors, temporal drift, and radiometric inconsistencies. The calibrated SAR images are input into a depth network to extract multi-scale, cross-view spatial features, providing rich contextual information for depth estimation. Based on the features extracted by the network, depth estimation predicts the 3D depth corresponding to each pixel, and regression optimization is performed in conjunction with the SAR imaging geometric model. Finally, the depth estimation results are mapped to 3D coordinate space, and the preliminary point cloud is denoised, filtered, geometrically corrected, and reconstructed to obtain the reconstructed result.

[0032] These methods avoid complex physical model inversion, but often require large-scale, high-quality labeled data for training, and the models are not very interpretable. The reconstruction results are difficult to incorporate the electromagnetic scattering physical mechanism unique to SAR, and the generalization ability is limited.

[0033] Based on this, the present application provides a method and apparatus for three-dimensional reconstruction of synthetic aperture radar images, which will be described below through embodiments.

[0034] Figure 1 The diagram illustrates a flowchart of a method for three-dimensional reconstruction of a synthetic aperture radar image according to an embodiment of this application, wherein the method includes steps S101-S104; specifically: S101. The initial single-view complex data collected based on the first number of observation angles is converted into an initial three-dimensional point cloud that can cover the outline of the observed object. S102. Initialize the initial 3D point cloud with Gaussian to obtain a parameterized initial 3D Gaussian scene representation; S103. Using a pre-trained two-dimensional diffusion model, the initial three-dimensional Gaussian scene representation is optimized for multi-view consistency through score distillation sampling to complete the geometric and texture details of the observed object, thereby obtaining the optimized target parameterized three-dimensional Gaussian scene representation. S104. Extract the detailed enhanced target 3D point cloud from the target parameterized 3D Gaussian scene representation.

[0035] This method requires only a small amount of observation data to generate an initial 3D point cloud, significantly reducing data acquisition costs and time. By converting the point cloud into a finely tuned 3D Gaussian scene representation and utilizing a pre-trained diffusion model for multi-view optimization, it can intelligently complete geometric details, enhance textures, and ensure consistent reconstruction across different viewpoints. The entire process incorporates the physical properties of SAR data during optimization, resulting in results that combine high accuracy with good physical interpretability, thus improving overall processing efficiency.

[0036] The following describes some embodiments of this application in detail. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0037] This application's method acquires a first number of synthetic aperture radar (SAR) single-look complex data (SAR SLC echo data) collected from different observation angles using an airborne or spaceborne SAR platform. These different observation angles characterize the differences in the relative spatial geometry between the radar platform and the target when imaging the same target area, with the core being the diversity of spatial baselines. Traditional methods require a very dense and uniformly distributed second number of spatial baselines (typically dozens or even hundreds) to obtain sufficient elevation resolution, necessitating long-term, multi-orbit observation accumulation, resulting in high costs and long cycles. The first number in this application's method has the core advantage of being significantly less than the aforementioned second number. This method can fully extract and reconstruct the complete three-dimensional structure of the target from these limited, potentially sparse or unevenly distributed observation angles, thereby significantly reducing the requirements for data acquisition scale and cycle time.

[0038] In practice, an airborne platform equipped with synthetic aperture radar performs multiple echo scans of the same target area at different azimuth and elevation angles. Each scan corresponds to a unique observation angle, thus obtaining an independent single-view complex data file. Changes in azimuth mean that the radar observes the target from different horizontal directions, such as the side, front, and back; while changes in elevation mean that the radar observes the target from different tilt angles, such as looking down or sideways. This multi-angle observation design aims to capture the target's structural information from multiple spatial dimensions.

[0039] After obtaining the initial single-view complex data, it needs to be transformed into an initial 3D point cloud that can cover the outline of the observed object. The initial 3D point cloud that can cover the outline of the observed object represents a sparse but structurally faithful 3D geometric skeleton. It is not a dense sampling of the complete surface of the target, but a set of key spatial feature points that are quickly and robustly extracted from limited observation data using this method.

[0040] like Figure 4As shown, the initial single-view complex data acquired based on a first number of observation angles is transformed into an initial three-dimensional point cloud that can cover the outline of the observed object. This process includes two core sub-stages: first, the initial single-view complex data is converted into grayscale images that can be used for geometric analysis; and second, three-dimensional structural information is extracted from these grayscale images.

[0041] The first stage aims to process the initial single-view complex data into grayscale images from multiple viewpoints. As mentioned earlier, each observation angle corresponds to an independent set of single-view complex data. Although these data contain the target's phase and amplitude information, they are in the original radar coordinate domain and cannot be directly used for 3D geometric analysis. Therefore, each set of single-view complex data needs to undergo imaging processing. This processing typically involves a series of standard synthetic aperture radar (SAR) signal processing steps, such as motion compensation, range migration correction, and final 2D focusing. The result after focusing is a series of SAR grayscale images. The pixel values ​​of each image mainly reflect the backscattering intensity (i.e., amplitude information) of the target at that specific radar viewpoint, while the implicit, precise phase information used for 3D reconstruction has been transformed into constraints on the geometric relationships between images. This results in a set of 2D intensity images of the same scene from multiple different viewpoints, i.e., "grayscale images from multiple viewpoints."

[0042] The second stage aims to synthesize 3D information from grayscale images and extract the main scattering structures to generate an initial 3D point cloud. Since a single 2D image only provides the outline of a target in a certain projection direction, information from all viewpoints must be fused to recover its 3D shape. This process is called 3D information synthesis. In practice, each pixel in each 2D grayscale image is typically back-projected into the 3D space of its possible source, based on the geometric model of radar imaging. The projection lines from all viewpoints converge and enhance near the actual 3D scattering point location. By accumulating and superimposing the projection intensities from multiple viewpoints on a regular grid (i.e., a voxel grid) in 3D space, a 3D scattering intensity distribution volume can be constructed. The data values ​​in this 3D volume characterize the combined scattering response intensity from all radar viewpoints at that spatial location. Subsequently, by extracting the main scattering structures, such as setting an adaptive intensity threshold, the 3D intensity volume is segmented, retaining only those voxels with significantly higher intensities than the background noise. These retained voxels represent the most dominant and stable scattering centers in the scene. Finally, the center coordinates of these voxels in 3D space are output, forming the initial 3D point cloud. Although this point cloud is sparse and contains only the strongest scattering points, it is still able to effectively delineate the main framework and external outline of the observed object (such as buildings, vehicles, etc.) in space, providing crucial initial geometric constraints and spatial support for subsequent refined 3D reconstruction.

[0043] In practice, the process of converting multi-angle SLC data into an initial three-dimensional point cloud can be systematically divided into two core stages that are connected one after the other: the first stage is sub-aperture imaging, and the second stage is incoherent stacking.

[0044] Phase 1: Sub-aperture imaging processing.

[0045] The sub-aperture imaging processing in this application aims to rapidly and accurately convert the raw SLC single-look complex data acquired at each observation angle into SAR grayscale images that can be used for 3D information extraction. This processing flow mainly includes three sequentially executed steps: First, phase calibration is performed. To eliminate phase errors introduced by radar platform trajectory deviations or atmospheric disturbances and to ensure strict geometric consistency among multiple images, this step models the phase history of each echo line (i.e., each observation angle data). Specifically, a quadratic polynomial phase error model is used for fitting and compensation:

[0046] The model coefficients a2, a1, and a0 are automatically estimated through optimization algorithms (such as minimizing the peak width of cross-correlation between complex data in each channel), thereby achieving high-precision trajectory error correction.

[0047] Secondly, interpolation resampling is performed. The sampling points on the frequency axis of the phase-calibrated data may be non-uniform. To meet the Fast Fourier Transform's requirement for equally spaced sampling, this step uniformly resamples the SLC echo data onto a preset equally spaced grid on the frequency axis. In practice, the amplitude and phase information of each echo are processed using a cubic spline interpolation algorithm, with the key frequency interval parameter set to Δf = 0.8MHz to balance imaging resolution and computational complexity.

[0048] Finally, inverse Fourier transform imaging is performed. This is a crucial step in generating the final grayscale image. This application uses techniques such as Stolt mapping to map the equally spaced frequency domain data obtained after interpolation and resampling to a three-dimensional wavenumber domain consisting of range wavenumber kx, azimuth wavenumber ky, and elevation wavenumber kz.

[0049] In this process, the bilinear interpolation algorithm is used to resample from a non-uniform wavenumber grid to a uniform grid. Then, through inverse Fourier transform, the data is finally transformed from the wavenumber domain to the two-dimensional image spatial domain to generate a single SAR amplitude grayscale image.

[0050] This sub-aperture imaging pipeline is highly optimized and boasts excellent computational efficiency; for example, it can generate a high-quality 2048×2048 pixel grayscale image within a delay of approximately 0.5 seconds. The multiple SAR grayscale images obtained after processing all observation angles serve as input for the next stage of incoherent overlay.

[0051] Second stage: Incoherent superposition.

[0052] After obtaining SAR grayscale images from multiple perspectives, this application uses an incoherent overlay process to fuse and invert these two-dimensional image information into a preliminary three-dimensional spatial structure, which is then extracted as an initial three-dimensional point cloud. First, 3D projection and accumulation are performed. Based on the precise imaging geometry model, the grayscale images from all viewpoints generated during the sub-aperture imaging stage are back-projected into a unified 3D Cartesian coordinate system using voxelization projection technology, filling a discrete 3D voxel mesh V(x, y, z). For intensity (amplitude) values ​​projected onto the same voxel from different viewpoints, the values ​​are accumulated using a "sum of squared amplitudes" method.

[0053] This incoherent accumulation strategy effectively suppresses noise and highlights the strong scattering signal that is stable across multiple viewpoints, thereby rapidly synthesizing a three-dimensional scattering intensity volume that characterizes the three-dimensional spatial scattering energy distribution.

[0054] Subsequently, point cloud extraction and optimization are performed. From the generated 3D scattering intensity volume, points representing the main scattering structures need to be automatically identified and extracted. This application first automatically calculates an adaptive threshold τ to distinguish strong scatterers from the background. Voxels with intensity values ​​higher than the threshold τ are retained, initially screening out high-energy scattering regions. To further improve the point cloud quality, isolated noise points are filtered out. 3D connected component analysis is performed on the retained voxels, and only the largest connected component is retained to ensure that the point cloud corresponds to the most important target structure. Finally, the 3D center coordinates (xi, yi, zi) of each remaining voxel after the above screening are extracted, and the resulting set of points constitutes the preprocessed result—the initial 3D point cloud Pin.

[0055] This method achieves a good balance between the efficiency of sub-aperture imaging and the robustness of incoherent superposition, and can quickly convert multi-angle SAR SLC echo data into a sparse and coarse but three-dimensional consistent point cloud skeleton that can effectively carry the structural information of the target. This reliable initial three-dimensional geometric representation provides crucial spatial constraints and an iterative starting point for subsequent fine-grained completion and optimization.

[0056] After obtaining an initial 3D point cloud that covers the target contour, the process moves to the completion stage of the initial 3D point cloud, such as... Figure 5 As shown. The first step in this stage is to perform Gaussian initialization on the aforementioned sparse point cloud (initial 3D point cloud) to construct a continuous and differentiable 3D scene representation that can be efficiently processed by subsequent optimization algorithms. This process is not a simple encapsulation, but rather transforms the geometric and physical properties of the point cloud into a set of parameterized 3D Gaussian distributions, thereby forming the initial 3D Gaussian scene representation. The specific implementation includes the mapping of the following four key parameters: First, spatial location mapping is performed. Each point in the initial 3D point cloud represents a precise scattering center in 3D space. Therefore, the most direct mapping is to directly assign the 3D coordinates of each point to the center position parameters of the corresponding 3D Gaussian distribution. This ensures that the Gaussian distribution is precisely "placed" in the spatial location of the original scattering point, faithfully inheriting the geometric skeleton of the initial point cloud.

[0057] Secondly, the scaling parameter needs to be determined. To initialize the shape and coverage of each Gaussian distribution (typically represented by the covariance matrix), calculations are required based on the spatial distribution characteristics of the initial point cloud. One efficient and effective implementation is to first calculate the average nearest neighbor distance for all points in the point cloud, which reflects the approximate sparsity of the point cloud in space. Then, using this average distance as a reference, an isotropic scaling parameter (i.e., initialized as a sphere) is initialized for all Gaussian distributions. This provides a reasonable starting point for adapting the Gaussian shape to complex geometries in subsequent optimizations.

[0058] Secondly, the opacity parameter is mapped. The initial 3D point cloud is not purely geometric points; its generation process has assigned each point an intensity value reflecting the strength of electromagnetic scattering from the target. This crucial physical property is then incorporated into the Gaussian representation. Specifically, the electromagnetic scattering intensity information associated with each point is normalized to fall within a suitable range (e.g., between 0 and 1), and then mapped to the corresponding Gaussian distribution's opacity parameter. Thus, points with stronger scattering intensity have a more opaque Gaussian distribution during rendering, making them more visually prominent and preserving the unique physical scattering characteristics of SAR data in the 3D model.

[0059] Finally, color parameters are mapped. Since SAR data itself does not contain optically meaningful color information, directly assigning colors to Gaussian distributions lacks basis. Therefore, geometric information from the point cloud is used to generate visual guidance. By calculating the normal vector of the initial 3D point cloud (e.g., using principal component analysis or a dedicated point cloud library), the geometric orientation information of the local surface where each point is located is obtained. Then, the three components of the normal vector are normalized to the standard RGB color space range (e.g., the [0,1] interval) through a linear transformation. The pseudo-color generated in this way is mapped to the color parameters of the corresponding Gaussian distribution (e.g., represented by low-order spherical harmonic function coefficients). This mapping not only provides rich visual features for the scene to assist subsequent optimization, but more importantly, it encodes the geometric surface orientation information of the target into optimizable parameters.

[0060] By completing the mapping process described above four steps, the initial, discrete, and sparse 3D point cloud is successfully transformed into a parameterized initial 3D Gaussian scene representation composed of tens of thousands or even millions of adjustable Gaussian primitives. This representation not only fully preserves the geometric and physical skeleton of the original point cloud but also possesses the key characteristics of continuity, differentiability, and support for efficient rendering, laying a solid foundation for subsequent intelligent completion and global optimization using a diffusion model.

[0061] In practical implementation, in order to map the scattered point cloud output from the pre-imaging stage to a set of parameterized Gaussian kernels, the correspondence of each attribute and the calculation method are explained as follows.

[0062] Location: The three-dimensional coordinates of the point cloud are the center of the Gaussian curve.

[0063] Scale (covariance matrix): To balance memory usage and rendering efficiency, isotropic spheres are used during initialization, with the uniform radius estimated by the average nearest neighbor distance.

[0064] Opacity: The intensity normalization of the electromagnetic point cloud reflects the contribution of points to the radar echo, controlling the intensity of the Gaussian effect on pixel accumulation.

[0065] Color: Since SAR point clouds lack natural RGB color labels and Gaussian points cannot be directly mapped to color information suitable for image rendering, a point cloud normal mapping representation is introduced to generate pseudo-color information to enhance visualization. The Open3D tool is used to calculate the 3D normal vector of the input point cloud P, and the normal vector is linearly normalized and mapped to the [0,1] interval to form a pseudo-color with RGB three channels.

[0066] After constructing an initial 3D Gaussian scene representation, a pre-trained 2D diffusion model is used to intelligently and globally optimize the scene representation for multi-view consistency through a score distillation sampling technique. The purpose of this process is to guide the coarse initial geometry towards a high-fidelity 3D structure with rich detail, visual plausibility, and consistency across all viewpoints, ultimately obtaining the optimized target parameterized 3D Gaussian scene representation. The specific implementation process includes the following four steps: First, multi-view rendering is performed. To evaluate the quality of the current 3D Gaussian scene representation and obtain optimization signals, it needs to be rendered differently from multiple different, pre-planned viewpoints. Using GPU-accelerated Gaussian spray rasterization, corresponding 2D rendered images can be generated in real time from each selected viewpoint. These images typically contain RGB color channels and may also contain depth channels. This step connects the 3D optimization problem with the 2D visual quality assessment.

[0067] Next, gradient calculation is performed. This is the core of score distillation sampling. The diffusion model is trained to understand "what kind of image is realistic and reasonable." To utilize this knowledge, the optimization algorithm intentionally adds a specific level of random Gaussian noise to the rendered images obtained from each viewpoint in the previous step, resulting in a set of noisy images. These noisy images, along with textual prompts or reference images that may serve as conditional information, are then input into a pre-trained 2D diffusion model (such as Stable Diffusion). Based on its powerful denoising capabilities, the diffusion model predicts a "better" clean image that conforms to the natural image distribution. By comparing the model's prediction with the original noisy input, a special gradient can be calculated: the score distillation sampling gradient. This gradient indicates in which direction the pixels of the current rendered image should change to better approximate the "high-quality image" prior known to the diffusion model. The gradient field is calculated independently for each viewpoint.

[0068] Next, gradient fusion is implemented. Since the goal is to obtain a consistent and reasonable 3D model across all viewpoints, a single viewpoint gradient cannot be simply used. Therefore, it is necessary to integrate the score distillation sampled gradients from all selected viewpoints. An effective approach is to perform weighted average fusion, where the weights can be set considering the rendering quality, occlusion level, or importance of each viewpoint. The fused optimized gradient integrates the visual consistency requirements from multiple directions, providing a more comprehensive and stable guiding signal for updating the 3D scene.

[0069] Finally, the parameter iteration and update are completed. The optimized gradient obtained by merging is backpropagated through the chain rule of the differentiable rendering pipeline. The gradient information is passed to every learnable parameter of the 3D Gaussian scene representation, including but not limited to: the position of the Gaussian center, the covariance matrix (controlling scale and rotation), the coefficients of the spherical harmonic function (controlling color and appearance), and the opacity. Based on the gradient direction, these parameters are finely adjusted using optimizers such as Adam. The above process of "rendering-calculating gradient-merging-updating parameters" constitutes an iterative loop. After hundreds to thousands of iterations, under the continuous guidance of the powerful visual priors provided by the diffusion model, the geometry of the 3D Gaussian scene representation is gradually completed (e.g., generating missing back faces or internal structures), the surface texture becomes more detailed and rich, and the rendering results from any new perspective are naturally consistent. When the optimization converges or reaches the preset number of iterations, a target parameterized 3D Gaussian scene representation with significantly enhanced details and greatly improved visual quality is obtained. This representation is the core result of the completion stage, preparing for the final high-precision point cloud output.

[0070] In practical implementation, after 3DGS initialization, the Chamfer distance is first used as an optimization metric to filter viewpoints from a large pool of candidates, determining the optimal reference viewpoint V with minimal self-occlusion. Then, a depth map and an RGBA pseudo-color image are rendered from this viewpoint. After determining the reference viewpoint, a CUDA-based splatter rasterizer is used to generate two images in real-time at a resolution of 1024×1024: a depth map and an RGBA pseudo-color image. The depth map records the nearest depth value projected onto the point cloud for each pixel at that viewpoint, used for structure verification. Subsequent algorithms perform depth difference analysis with the original point cloud to quickly and objectively measure whether the 3D structure has distorted. The RGBA image, which fuses the pseudo-color (RGB) and opacity (A) outputs of the normal mapping, is used to guide SDS. As the conditional image Ir for SDS, it provides consistent semantic anchors, reducing the gradient amplification effect of the Diffusion model on the blob texture.

[0071] By employing a viewpoint selection strategy in conjunction with efficient CUDA rasterization rendering, a set of reference images with minimal geometric occlusion and sufficient semantic information was generated. This viewpoint-image pair provides a solid reference foundation for subsequent SDS-Diffusion iterations.

[0072] View-Guided SDS Optimization: In the 3D reconstruction process using Gaussian modeling, a view-guided SDS optimization method is proposed to improve the geometric consistency and multi-view robustness of point cloud structure completion. This method introduces zero-mean Gaussian noise to perturb the initial point cloud, generating a noisy version to better represent the geometric space surrounding the point cloud. Simultaneously, to prevent the initial geometric structure from being destroyed during the noisy point cloud completion process, structural constraints are introduced to ensure that the geometric structure is as consistent as possible under the reference viewpoint. Based on the previously obtained reference viewpoint, a reference image is rendered from the current Gaussian distribution. The rendered image is input into the SDS module and compared with the image representation generated by the 2D Diffusion model. Through backpropagation of the loss gradient in the image space, the Gaussian attributes (position, scale, color, opacity) are guided towards a better structure through iteration. The gradient calculation method of SDS is as follows:

[0073] Based on a viewpoint evaluation strategy, k high-quality viewpoints with low occlusion and uniform distribution are selected. A CUDA-based splatter rasterizer is used to render the corresponding depth map and RGBA pseudo-color map in parallel for each viewpoint. Then, the rendered image Ik of each viewpoint Vk is input into a finely tuned 2D Diffusion model to calculate the SDS gradient of each pixel. Finally, the SDS gradients of all k viewpoints are weighted and fused.

[0074] The goal is to integrate structural completion objectives from multiple directions to guide the Gaussian field to consistently optimize its property parameters from multiple perspectives.

[0075] After completing the multi-view consistency optimization guided by the diffusion model and obtaining a detailed parameterized 3D Gaussian scene representation of the target, the final result generation stage begins: extracting discrete, directly applicable 3D point clouds of the target from this continuous volume representation, thus completing the final output of the 3D reconstruction. This extraction process is not a simple location recording, but a refined screening and reconstruction step based on the physical properties of the optimization results. The specific implementation includes the following three key steps: First, parameter acquisition. From the optimized and converged parameterized 3D Gaussian scene representation, its constituent elements are systematically extracted, including the descriptive parameters of each 3D Gaussian distribution. Among them, the geometric center position parameter directly defines the coordinates of the Gaussian primitive in 3D space and is a potential candidate point cloud location. The opacity parameter is a crucial attribute. After previous optimization iterations, it is no longer an initial simple mapping value, but a comprehensive indicator that integrates multi-view consistency constraints and diffusion model priors, reflecting the degree of contribution of the location to the final rendered image and its confidence as an "entity" structure. High opacity values ​​usually correspond to enhanced, visually significant, and stable structural parts in the model.

[0076] Secondly, opacity-based filtering is crucial. To extract the truly representative, detailed structure from hundreds of thousands or even millions of Gaussian pixels, a filtering process based on preset opacity criteria is necessary. The most common criterion is setting a global threshold δ. This threshold can be adaptively determined using statistical methods (such as taking the median or a certain quantile of all opacity values) or set empirically. Subsequently, all Gaussian distributions are iterated, and only distributions with opacity parameters higher than the threshold are selected. This operation has clear physical and geometric significance: in the context of SAR 3D reconstruction, high-opaque regions correspond to strong and stable electromagnetic scattering centers, i.e., the target's bulk structure; simultaneously, within the visual optimization framework, high opacity also means that this region contributes significantly to forming a high-quality, reasonable rendered image, representing the reconstructed "entity." Through this filtering, low-opaque Gaussian distributions representing background noise, weak scattering regions, or uncertain surfaces generated during the optimization process can be effectively removed, thereby significantly improving the purity and structural clarity of the output point cloud.

[0077] After the above screening, the retained "target distributions" represent the enhanced 3D structures that, after completion and optimization, are determined to belong to the observed object. The geometric center position parameters of each such target distribution are directly extracted as a 3D point. By combining all the selected points, the final enhanced 3D point cloud of the target is formed.

[0078] In practical implementation, after completing Gaussian field modeling and viewpoint guidance optimization, a high-precision 3D point cloud extraction method based on transparency screening mechanism is provided. The aim is to effectively extract target point clouds with clear structure and boundaries from the optimized 3D Gaussian representation, reduce scattering noise interference, and improve the accuracy of 3D reconstruction.

[0079] The center position of each Gaussian element in the optimized 3D Gaussian volume is used as a potential 3D point cloud candidate to construct an initial point set. This operation preserves the spatial geometric structure information embodied by the optimized Gaussian distribution, providing a basic framework for point cloud reconstruction.

[0080] A transparency filtering mechanism is then introduced. Points with higher transparency values ​​typically correspond to regions with significant backscattering in SAR images, representing structural boundaries or the outline of the main target. A Gaussian transparency threshold δ is set accordingly. By retaining foreground Gaussian points with transparency > δ, unstructured points in background clutter and weakly scattering regions commonly found in SAR imaging are eliminated, thus more effectively extracting the boundary structure of the main target.

[0081] Figure 6 This invention provides a schematic diagram of a device for three-dimensional reconstruction of synthetic aperture radar images, comprising: The data conversion module is used to convert the initial single-view complex data collected based on a first number of observation angles into an initial three-dimensional point cloud that can cover the outline of the observed object. The Gaussian modeling module is used to initialize the initial 3D point cloud with Gaussian to obtain a parameterized initial 3D Gaussian scene representation; The diffusion optimization module is used to optimize the initial three-dimensional Gaussian scene representation by multi-view consistency through scoring distillation sampling using a pre-trained two-dimensional diffusion model, so as to complete the geometric and texture details of the observed object and obtain the optimized target parameterized three-dimensional Gaussian scene representation. The point cloud extraction module is used to extract the target 3D point cloud with enhanced details from the target parameterized 3D Gaussian scene representation.

[0082] The process of converting initial single-view complex data acquired based on a first number of observation angles into an initial 3D point cloud capable of covering the outline of the observed object includes: The initial single-view complex data acquired based on the first number of observation angles is processed into a grayscale image under multiple viewpoints. The initial three-dimensional point cloud is generated by synthesizing three-dimensional information based on the grayscale image and extracting the main scattering structures.

[0083] The step of processing the initial single-view complex data acquired based on the first number of observation angles into a grayscale image under multiple views includes: Phase error estimation and compensation are performed on the initial single-view complex data at each observation angle to obtain the target single-view complex data. The target single-view complex data is resampled on the frequency axis so that the sampling interval meets the preset imaging requirements, and uniformly sampled frequency domain data is obtained. The uniformly sampled frequency domain data is transformed from the frequency domain to the spatial domain to generate grayscale images corresponding to each viewpoint.

[0084] The step of synthesizing three-dimensional information based on the grayscale image and extracting the main scattering structures to generate the initial three-dimensional point cloud includes: The grayscale images from the multiple perspectives are projected onto a unified three-dimensional spatial coordinate system, and incoherent energy accumulation is performed in the three-dimensional voxel grid to generate a three-dimensional scattering intensity distribution. The three-dimensional scattering intensity distribution is segmented by a threshold, and voxels with energy higher than a preset threshold are extracted; The spatial locations of the extracted voxels are used as points to generate the initial three-dimensional point cloud.

[0085] The initialization of the initial 3D point cloud into a Gaussian representation to obtain a parameterized initial 3D Gaussian scene representation includes: The spatial coordinates of the initial three-dimensional point cloud are mapped to the center position parameters of a three-dimensional Gaussian distribution. Based on the spatial distribution characteristics of the initial 3D point cloud, the scale parameters of the 3D Gaussian distribution are determined. The electromagnetic scattering intensity information of the initial three-dimensional point cloud is mapped to the opacity parameter of a three-dimensional Gaussian distribution. The geometric orientation information of the initial 3D point cloud is mapped to color parameters of a 3D Gaussian distribution.

[0086] The method utilizes a pre-trained two-dimensional diffusion model to perform multi-view consistency optimization on the initial three-dimensional Gaussian scene representation through score distillation sampling, in order to complete the geometric and texture details of the observed object, and obtain the optimized target parameterized three-dimensional Gaussian scene representation, including: The initial three-dimensional Gaussian scene representation is rendered from multiple different perspectives to obtain a two-dimensional rendered image from each perspective. The noise-added 2D rendered images are input into the pre-trained 2D diffusion model to calculate the score distillation sampling gradient for each viewpoint. The scoring distillation sampling gradients from multiple different perspectives are fused to obtain the fused optimized gradient; Based on the fused optimized gradient, the parameters of the initial 3D Gaussian scene representation are iteratively updated until the target parameterized 3D Gaussian scene representation is obtained.

[0087] The step of extracting the enhanced-detail target 3D point cloud from the target parameterized 3D Gaussian scene representation includes: Obtain the geometric center position parameters and opacity parameters of each three-dimensional Gaussian distribution from the target parameterized three-dimensional Gaussian scene representation; Based on preset opacity filtering conditions, target distributions that meet the conditions are selected from each of the three-dimensional Gaussian distributions. The geometric center of the selected target distribution is determined as a point in the point cloud, and the points are aggregated to form the detailed enhanced target 3D point cloud.

[0088] like Figure 7 As shown, this application provides an electronic device for performing the synthetic aperture radar image three-dimensional reconstruction method of this application. The device includes a memory, a processor, a bus, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the synthetic aperture radar image three-dimensional reconstruction method.

[0089] Specifically, the aforementioned memory and processor can be general-purpose memory and processor, without any specific limitations. When the processor runs the computer program stored in the memory, it can execute the aforementioned method for three-dimensional reconstruction of synthetic aperture radar images.

[0090] Corresponding to the method for three-dimensional reconstruction of synthetic aperture radar images in this application, this application embodiment also provides a computer storage medium storing a computer program, which is executed by a processor to perform the steps of the above-described method for three-dimensional reconstruction of synthetic aperture radar images.

[0091] Specifically, the storage medium can be a general-purpose storage medium, such as a portable disk or hard disk. When the computer program on the storage medium is run, it can execute the above-mentioned method for three-dimensional reconstruction of synthetic aperture radar images.

[0092] In the embodiments provided in this application, it should be understood that the disclosed systems and methods can be implemented in other ways. The system embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and there may be other division methods in actual implementation. Furthermore, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Additionally, the coupling or direct coupling or communication connection shown or discussed may be through some communication interface; the indirect coupling or communication connection between systems or units may be electrical, mechanical, or other forms.

[0093] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0094] In addition, the functional units in the embodiments provided in this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0095] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0096] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. In addition, the terms "first", "second", "third", etc. are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0097] Finally, it should be noted that the above-described embodiments are merely specific implementations of this application, used to illustrate the technical solutions of this application, and not to limit them. The protection scope of this application is not limited thereto. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some of the technical features, within the scope of the technology disclosed in this application; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application. All should be covered within the protection scope of this application. Therefore, the protection scope of this application should be determined by the protection scope of the claims.

Claims

1. A method for three-dimensional reconstruction of synthetic aperture radar images, characterized in that, The method includes: The initial single-view complex data collected based on the first number of observation angles is transformed into an initial three-dimensional point cloud that can cover the outline of the observed object. The initial 3D point cloud is initialized with Gaussian to obtain a parameterized initial 3D Gaussian scene representation; Using a pre-trained two-dimensional diffusion model, the initial three-dimensional Gaussian scene representation is optimized for multi-view consistency through score distillation sampling to complete the geometric and texture details of the observed object, thereby obtaining the optimized target parameterized three-dimensional Gaussian scene representation. The detailed-enhanced 3D point cloud of the target is extracted from the parameterized 3D Gaussian scene representation of the target.

2. The method according to claim 1, characterized in that, The process of converting initial single-view complex data acquired based on a first number of observation angles into an initial 3D point cloud capable of covering the outline of the observed object includes: The initial single-view complex data acquired based on the first number of observation angles is processed into a grayscale image under multiple viewpoints. The initial three-dimensional point cloud is generated by synthesizing three-dimensional information based on the grayscale image and extracting the main scattering structures.

3. The method according to claim 2, characterized in that, The step of processing the initial single-view complex data acquired based on the first number of observation angles into a grayscale image under multiple views includes: Phase error estimation and compensation are performed on the initial single-view complex data at each observation angle to obtain the target single-view complex data. The target single-view complex data is resampled on the frequency axis so that the sampling interval meets the preset imaging requirements, and uniformly sampled frequency domain data is obtained. The uniformly sampled frequency domain data is transformed from the frequency domain to the spatial domain to generate grayscale images corresponding to each viewpoint.

4. The method according to claim 2, characterized in that, The step of synthesizing three-dimensional information based on the grayscale image and extracting the main scattering structures to generate the initial three-dimensional point cloud includes: The grayscale images from the multiple perspectives are projected onto a unified three-dimensional spatial coordinate system, and incoherent energy accumulation is performed in the three-dimensional voxel grid to generate a three-dimensional scattering intensity distribution. The three-dimensional scattering intensity distribution is segmented by a threshold, and voxels with energy higher than a preset threshold are extracted; The spatial locations of the extracted voxels are used as points to generate the initial three-dimensional point cloud.

5. The method according to claim 1, characterized in that, The initialization of the initial 3D point cloud into a Gaussian representation to obtain a parameterized initial 3D Gaussian scene representation includes: The spatial coordinates of the initial three-dimensional point cloud are mapped to the center position parameters of a three-dimensional Gaussian distribution. Based on the spatial distribution characteristics of the initial 3D point cloud, the scale parameters of the 3D Gaussian distribution are determined; The electromagnetic scattering intensity information of the initial three-dimensional point cloud is mapped to the opacity parameter of a three-dimensional Gaussian distribution. The geometric orientation information of the initial 3D point cloud is mapped to color parameters of a 3D Gaussian distribution.

6. The method according to claim 1, characterized in that, The method utilizes a pre-trained two-dimensional diffusion model to perform multi-view consistency optimization on the initial three-dimensional Gaussian scene representation through score distillation sampling, in order to complete the geometric and texture details of the observed object, and obtain the optimized target parameterized three-dimensional Gaussian scene representation, including: The initial three-dimensional Gaussian scene representation is rendered from multiple different perspectives to obtain a two-dimensional rendered image from each perspective. The noise-added 2D rendered images are input into the pre-trained 2D diffusion model to calculate the score distillation sampling gradient for each viewpoint. The scoring distillation sampling gradients from multiple different perspectives are fused to obtain the fused optimized gradient; Based on the fused optimized gradient, the parameters of the initial 3D Gaussian scene representation are iteratively updated until the target parameterized 3D Gaussian scene representation is obtained.

7. The method according to claim 1, characterized in that, The step of extracting the enhanced-detail target 3D point cloud from the target parameterized 3D Gaussian scene representation includes: Obtain the geometric center position parameters and opacity parameters of each three-dimensional Gaussian distribution from the target parameterized three-dimensional Gaussian scene representation; Based on preset opacity filtering conditions, target distributions that meet the conditions are selected from each of the three-dimensional Gaussian distributions. The geometric center of the selected target distribution is determined as a point in the point cloud, and the points are aggregated to form the detailed enhanced target 3D point cloud.

8. A device for three-dimensional reconstruction of synthetic aperture radar images, characterized in that, The device includes: The data conversion module is used to convert the initial single-view complex data collected based on a first number of observation angles into an initial three-dimensional point cloud that can cover the outline of the observed object. The Gaussian modeling module is used to initialize the initial 3D point cloud with Gaussian to obtain a parameterized initial 3D Gaussian scene representation; The diffusion optimization module is used to optimize the initial three-dimensional Gaussian scene representation by multi-view consistency through scoring distillation sampling using a pre-trained two-dimensional diffusion model, so as to complete the geometric and texture details of the observed object and obtain the optimized target parameterized three-dimensional Gaussian scene representation. The point cloud extraction module is used to extract the target 3D point cloud with enhanced details from the target parameterized 3D Gaussian scene representation.

9. An electronic device, characterized in that, include: The device includes a processor, a memory, and a bus. The memory stores machine instructions that the processor executes. When the electronic device is running, the processor communicates with the memory via the bus. When the machine instructions are executed by the processor, they perform the steps of the method for three-dimensional reconstruction of synthetic aperture radar images as described in any one of claims 1 to 7.

10. A computer storage medium, characterized in that, The computer storage medium stores a computer program that, when executed by a processor, performs the steps of the method for three-dimensional reconstruction of synthetic aperture radar images as described in any one of claims 1 to 7.

Images