Target reconstruction method based on SAR Gaussian sputtering

Abstract

The invention discloses a target reconstruction method based on SAR Gaussian sputtering, and the method comprises the following steps: obtaining original SAR images of different targets, including multi-azimuth and multi-pitch-angle SAR images; registering and preprocessing the original SAR image to obtain radar parameters, azimuth angle information and pitch angle information; constructing a training data set and a test data set, and constructing an SAR Gaussian scattering model by using the initial point cloud; training an SAR Gaussian scattering model, rendering the SAR image under the visual angle of the training data set, and obtaining a second SAR Gaussian scattering model; and calling the second SAR Gaussian model, and testing a multi-view generation result and a target reconstruction result. According to the method, a mapping projection algorithm and a Gaussian sputtering method are combined, a Gaussian ellipsoid is used for representing a target geometric structure and background modeling, the Gaussian ellipsoid is endowed with scattering characteristics, and multi-view generation and geometric reconstruction are carried out on a target by learning target scattering characteristics of a multi-view SAR image.

Description

Technical Field

[0001] The present invention relates to the fields of radar image processing and target reconstruction, and in particular to a target reconstruction method based on SAR Gaussian sputtering. Background Art

[0002] The imaging process of synthetic aperture radar (SAR) involves complex electromagnetic scattering mechanisms, making SAR images difficult to interpret. Three-dimensional reconstruction based on SAR images is a key method for SAR image interpretation. Traditional SAR 3D reconstruction techniques include interferometric synthetic aperture radar (InSAR), tomographic synthetic aperture radar (TomoSAR), polarimetric synthetic aperture radar (PoISAR), and SAR image-based 3D reconstruction. Compared to the first three, image-based 3D reconstruction techniques have lower data requirements and offer faster reconstruction speeds.

[0003] In recent years, with the advancement of deep learning technology, data-driven neural networks have achieved remarkable results in 3D reconstruction. NeRF (Neural Radiance Fields), a representative method, represents scenes as continuous volumetric density and radiance fields, enabling highly accurate synthesis of images from any perspective, achieving realistic 3D scene reconstruction. Furthermore, 3D-GS (3D Gaussian Splatting) has further advanced the development of 3D generative models. By leveraging techniques such as generative adversarial networks, it not only improves the detail and realism of reconstructions but also enhances the model's generalization capabilities in complex scenes. These technological advances have not only expanded the application scope of 3D reconstruction, such as virtual reality, augmented reality, and digital twins, but also provided new insights and methods for related research. With continued algorithm optimization and increased computing resources, neural network-based 3D reconstruction technology is expected to achieve even greater breakthroughs in the future.

[0004] Three-dimensional reconstruction of SAR images is a fundamental problem in SAR image interpretation. It can be used for target classification, target geometry analysis, and target material modeling. However, current research on 3D reconstruction based on SAR images is limited. Image-based 3D reconstruction requires precisely registered images, appropriate image normalization methods, and SAR image data from multiple elevation and azimuth angles, making it difficult to research. Summary of the Invention

[0005] According to an embodiment of the present invention, a target reconstruction method based on SAR Gaussian sputtering is provided, comprising the following steps:

[0006] Acquire original SAR images of different targets, including multi-azimuth and multi-elevation SAR images;

[0007] Perform registration and preprocessing on the original SAR image to obtain radar parameters, azimuth information, and pitch angle information;

[0008] According to radar parameters, azimuth information, and elevation angle information, a training data set and a test data set are constructed, and the SAR Gaussian scattering model is constructed using the initial point cloud.

[0009] Train the SAR Gaussian scattering model, render the SAR image under the perspective of the training dataset, and obtain the second SAR Gaussian scattering model;

[0010] Retrieve the second SAR Gaussian model and test the multi-view generation results and target reconstruction results.

[0011] Preferably, acquiring multi-azimuth and multi-elevation SAR images of different targets comprises the following steps:

[0012] Generate SAR images based on the echo data of the airborne SAR front-side scanning;

[0013] The target in the SAR image is calibrated and registered, and the corresponding measured wave data target slice is extracted.

[0014] Preferably, registering and preprocessing the original SAR image to obtain radar parameters, azimuth information, and elevation angle information includes the following steps:

[0015] Perform radiation correction on the original SAR image;

[0016] Perform amplitude normalization on the original SAR image after radiation correction to eliminate the amplitude differences between different images;

[0017] Perform logarithmic scale conversion on the original SAR image after amplitude normalization. The logarithmic scale conversion formula is: .

[0018] Preferably, performing radiation correction on the original SAR image comprises the following steps:

[0019] Perform sensor calibration, atmospheric correction, and geometric correction;

[0020] Convert radar return signals into physical quantities.

[0021] Preferably, converting the radar return signal into a physical quantity includes: normalizing the scattering cross-sectional area.

[0022] Preferably, Min-Max normalization or Z-score normalization is used for amplitude normalization.

[0023] Preferably, Min-Max normalization linearly maps the amplitude values ​​to a preset range: .

[0024] Preferably, Z-score normalization transforms the amplitude values ​​into a distribution with mean 0 and standard deviation 1: , is the image mean, is the standard deviation.

[0025] Preferably, constructing the SAR Gaussian scattering model comprises the following steps:

[0026] Construct the transformation matrix from the world coordinate system to the radar coordinate system based on the height difference, pitch angle and azimuth angle between the radar and the target. , is the rotation matrix, , is the offset vector, ;

[0027] The projection matrix J is determined based on the resolution and image size of the SAR image, where the projection matrix of the projection plane is calculated as , the projection matrix of the imaging plane is , is the number of distance pixels, is the range resolution, is the number of pixels in azimuth, and the radar parameters of each SAR image are obtained;

[0028] Based on the radar parameters of each SAR image, a training dataset and a test dataset are constructed respectively;

[0029] Construct the initial scattered point cloud, use the plane to initialize the background for the target with background, use the hemispherical shell to initialize the target, and assign the initial covariance matrix, backscattering coefficient and phase function to each point cloud object as the initial representation of the scene.

[0030] Preferably, the data in the test dataset needs to be ensured to have the same resolution and image size as the data in the training dataset.

[0031] Preferably, obtaining the second SAR Gaussian scattering model includes the following steps:

[0032] Get the scattering intensity of the Gaussian element in the scene;

[0033] The Gaussian element is projected onto the imaging plane. The difference between the two-dimensional covariance matrix, the plane coordinates, and the pixel center is used as the contribution weight of the Gaussian element scattering intensity to the pixel scattering energy. The pixel scattering energy is used as the pixel intensity to obtain the rendered SAR image.

[0034] Obtain loss values ​​of the rendered SAR image and the true SAR image, and backpropagate the gradient to optimize the parameters of the Gaussian element.

[0035] Preferably, the scattering intensity of each Gaussian element is:

[0036] , yes Rays in the The backward energy decay of the Gaussian element under depth, 、 and are the phase function, covariance and weighting of the ray center and the backscattering coefficient of the Gaussian element respectively.

[0037] Preferably, the distance to pixel The scattered energy is: ,in is the distance to pixel Next The influence weight of each Gaussian element on it, is the scattering intensity of the Gaussian element.

[0038] Preferably, the loss function between the rendered SAR image obtained by rendering the SAR image under the perspective of the training dataset and the true SAR image tends to be stable and close to 0, and the loss function is: ,in .

[0039] According to an embodiment of the present invention, a target reconstruction method based on SAR Gaussian sputtering is proposed. The Marine Predators Algorithm (MPA) is combined with the Gaussian sputtering method. A Gaussian ellipsoid is used to characterize the target geometry and background modeling. The Gaussian ellipsoid is then given a scattering feature. By learning the target scattering features from multi-view SAR images, multi-view generation and geometric reconstruction of the target are performed.

[0040] It is to be understood that both the foregoing general description and the following detailed description are exemplary, and are intended to provide further explanation of the technology as claimed. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 Flowchart of a target reconstruction method based on SAR Gaussian sputtering according to an embodiment of the present invention;

[0042] Figure 2 A diagram showing the conversion relationship between the world coordinate system and the radar coordinate system according to an embodiment of the present invention;

[0043] Figure 3 Schematic diagram of forward rendering according to an embodiment of the present invention;

[0044] Figure 4 This is the backward conduction gradient map of the SAR differentiable Gaussian grating device. DETAILED DESCRIPTION

[0045] The preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings to further illustrate the present invention.

[0046] First, combine Figures 1 to 4 A target reconstruction method based on SAR Gaussian sputtering according to an embodiment of the present invention is described, which is used in the fields of radar image processing and target reconstruction and has a wide range of application scenarios.

[0047] like Figure 1 As shown, the target reconstruction method based on SAR Gaussian sputtering according to an embodiment of the present invention has the following steps:

[0048] In S1, as Figure 1 As shown in FIG, obtaining original SAR images of different targets, which include multi-azimuth and multi-elevation SAR images, includes the following sub-steps:

[0049] In S11, SAR imaging is generated based on the echo data of the airborne SAR front-side scanning. The specific embodiment of the SAR flight parameters for collecting echo signals is shown in the following table:

[0050] Table 1 MiniSAR flight experiment parameters

[0051] Parameter Quantity Actual surface resolution 6.25cm Downward perspective 60° Mop Corner 30° frequency 15GHz Flight altitude 475m Waveform FMCW

[0052] In S12, the target in the SAR image is calibrated and then registered, that is, regional registration is performed with the overall scene image of the collected data. The horizontal and vertical coordinates of the SAR image are the azimuth and range directions, respectively, which are used to determine the coordinates of the two to locate the target area and extract the corresponding measured wave data target slice.

[0053] In S2, the original SAR image is registered and preprocessed to obtain radar parameters, azimuth information, and pitch angle information. This includes the following sub-steps:

[0054] In S21, the original SAR image is subjected to radiation correction, thereby ensuring that the SAR image in the data set is converted into scattering parameters with clear physical meaning, which is convenient for subsequent analysis and processing.

[0055] In S211 , sensor calibration, atmospheric correction, and geometric correction are performed to ensure the geometric accuracy of the image so that images acquired from different viewing angles and at different times can be accurately aligned.

[0056] In S212 , the radar return signal is converted into a physical quantity, such as a normalized radar cross section (NRCS).

[0057] In S22, amplitude normalization is performed on the original SAR image after radiation correction to eliminate amplitude differences between different images caused by sensor gain or other system factors. In this embodiment, Min-Max normalization or Z-score normalization is used for amplitude normalization.

[0058] Among them, Min-Max normalization linearly maps the amplitude value to a preset range (usually [0, 1]): .

[0059] Z-score normalization transforms the magnitude values ​​into a distribution with a mean of 0 and a standard deviation of 1: , is the image mean, is the standard deviation.

[0060] In S23, the original SAR image after amplitude normalization is logarithmically scaled. The logarithmic scale conversion formula is: , where 1 is added to avoid the negative infinity problem in logarithmic operations.

[0061] In S3, based on radar parameters, azimuth information, and elevation information, training and test datasets are constructed. The SAR Gaussian scattering model is constructed using the initial point cloud. This includes the following sub-steps:

[0062] In S31, if Figure 2 As shown, according to the height difference between the radar and the target , pitch angle , azimuth Construct the transformation matrix from the world coordinate system to the radar coordinate system, , is the rotation matrix, , is the offset vector, .

[0063] In S32, the projection matrix J is determined based on the resolution and image size of the SAR image, wherein the projection matrix of the projection plane is calculated as , the projection matrix of the imaging plane is , is the number of distance pixels, is the range resolution, is the number of pixels in azimuth, and the radar parameters of each SAR image are obtained.

[0064] In S33, based on the radar parameters of each SAR image, a training dataset and a test dataset are constructed. In this embodiment, the data of the test dataset must have the same resolution and image size as the data of the training dataset.

[0065] In S34, an initial scattered point cloud is constructed. For a target with a background, a plane is used to initialize the background, and a hemispherical shell is used to initialize the target. An initial covariance matrix, backscatter coefficient, and phase function are assigned to each point cloud object as an initial representation of the scene.

[0066] In S4, as Figure 3 As shown, the SAR Gaussian scattering model is trained, the SAR image under the perspective of the training data set is rendered, and the second SAR Gaussian scattering model is obtained.

[0067] In S41, if Figure 3 As shown, the scattering intensity of the Gaussian element in the scene is calculated, and it is emitted from the calculation plane. A radar beam passes through the Gaussian element and is attenuated. The scattering intensity of each Gaussian element is:

[0068]

[0069] in yes Rays in the The backward energy decay of the Gaussian element under depth, 、 and are the phase function, covariance, ray center weighting and backscattering coefficient of the Gaussian element respectively.

[0070] In S42, if Figure 3 As shown in Figure 1, the Gaussian element is projected onto the imaging plane, and the difference between the two-dimensional covariance matrix, the plane coordinates, and the pixel center is used as the contribution weight of the Gaussian element scattering intensity to the pixel scattering energy. The pixel scattering energy is used as the pixel intensity to render the SAR image. The scattered energy is:

[0071]

[0072] in is the distance to pixel Next The influence weight of each Gaussian element on it, is the scattering intensity of the Gaussian element.

[0073] In S43, if Figure 3 、 4As shown, the loss between the rendered SAR image and the ground-truth SAR image is calculated, and the gradient is back-propagated to optimize the parameters of the Gaussian element. In this embodiment, the loss function of the rendered SAR image and the ground-truth SAR image obtained by rendering the SAR image under the perspective of the training dataset tends to be stable and close to 0. The loss function is: ,in .

[0074] In S5, the second SAR Gaussian model is retrieved to test the multi-view generation and target reconstruction results. Specifically, the trained SAR Gaussian model is tested by retrieving the radar parameters in the test set and rendering the SAR image. The image performance, such as SSIM and PSNR, is compared with the ground-truth SAR image.

[0075] In the above embodiment, the Gaussian splashing method is combined with the MPA algorithm, the target and scene are characterized by a Gaussian scattering ellipsoid, the scattering intensity of the Gaussian ellipsoid is calculated by simulating the attenuation of the radar beam passing through the ellipsoid, and then the ellipsoid is projected onto the imaging plane, and the covariance matrix of the ellipsoid and the difference between the plane coordinates and the pixel center are used as mixing weights to affect the influence of the Gaussian ellipsoid on the distance-to-pixel scattering energy.

[0076] The positions of the Gaussian ellipsoids in the scene are independent of each other, and the scene distribution is fitted using a Gaussian random distribution. Since accurate representation of the scene using a fixed-point Gaussian distribution is difficult, during training, threshold Gaussian densification is applied to appropriately increase or decrease the number of Gaussian elements to achieve a better fit to the scene and target. Densification involves culling, splitting, and duplication. Culling refers to removing Gaussian elements with too small backscatter coefficients in the scene, while splitting and duplication are densification operations performed on Gaussian elements when a single Gaussian element is difficult to represent.

[0077] Above, refer to Figures 1 to 4 This paper describes a target reconstruction method based on SAR Gaussian sputtering according to an embodiment of the present invention. The method combines the Mapping and Projection Algorithm (MPA) with the Gaussian sputtering method, uses a Gaussian ellipsoid to characterize the target geometry and background modeling, and assigns the Gaussian ellipsoid a scattering feature. By learning the target scattering features of multi-view SAR images, multi-view generation and geometric reconstruction of the target are performed.

[0078] It should be noted that, in this specification, the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, article, or apparatus comprising a series of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus. In the absence of further limitations, the elements defined by the phrase "comprising..." do not exclude the presence of other identical elements in the process, method, article, or apparatus comprising the elements.

[0079] Although the present invention has been described in detail through the above preferred embodiments, it should be understood that the above description is not intended to limit the present invention. After reading the above description, various modifications and substitutions of the present invention will become apparent to those skilled in the art. Therefore, the scope of protection of the present invention should be defined by the appended claims.

Claims

1. A target reconstruction method based on SAR Gaussian sputtering, characterized in that: The following steps are included: Acquire original SAR images of different targets, wherein the original SAR images include multi-azimuth and multi-elevation SAR images; Performing registration and preprocessing on the original SAR image to obtain radar parameters, azimuth information, and pitch angle information; Constructing a training data set and a test data set according to the radar parameters, the azimuth information, and the pitch angle information, and constructing a SAR Gaussian scattering model using the initial point cloud; Training the SAR Gaussian scattering model, rendering the SAR image under the perspective of the training data set, and obtaining the target SAR Gaussian scattering model; Retrieve the target SAR Gaussian model and test the multi-view generation results and target reconstruction results.

2. The target reconstruction method based on SAR Gaussian sputtering according to claim 1, characterized in that: Acquiring multi-azimuth and multi-elevation SAR images of different targets involves the following steps: Generate SAR images based on the echo data of the airborne SAR front-side scanning; The target in the SAR image is calibrated and registered, and the corresponding measured wave data target slice is extracted.

3. The target reconstruction method based on SAR Gaussian sputtering according to claim 2, characterized in that: Performing registration and preprocessing on the original SAR image to obtain radar parameters, azimuth information, and elevation angle information includes the following steps: Performing radiation correction on the original SAR image; performing amplitude normalization processing on the original SAR image after radiation correction to eliminate amplitude differences between different images; The original SAR image after amplitude normalization is subjected to logarithmic scale conversion, and the logarithmic scale conversion formula is: .

4. The target reconstruction method based on SAR Gaussian sputtering according to claim 3, characterized in that: Performing radiation correction on the original SAR image includes the following steps: Perform sensor calibration, atmospheric correction, and geometric correction; Convert radar return signals into physical quantities.

5. The target reconstruction method based on SAR Gaussian sputtering according to claim 4, characterized in that: Converting the radar return signal into physical quantities includes: normalized scattering cross-sectional area.

6. The target reconstruction method based on SAR Gaussian sputtering according to claim 3, characterized in that: Amplitude normalization was performed using Min-Max normalization or Z-score normalization.

7. The target reconstruction method based on SAR Gaussian sputtering according to claim 6, characterized in that: The Min-Max normalization linearly maps the amplitude value to a preset range: .

8. The target reconstruction method based on SAR Gaussian sputtering according to claim 6, characterized in that: The Z-score normalization transforms the amplitude values ​​into a distribution with a mean of 0 and a standard deviation of 1: , is the image mean, is the standard deviation.

9. The target reconstruction method based on SAR Gaussian sputtering according to claim 1, characterized in that: Building a SAR Gaussian scattering model includes the following steps: Construct the transformation matrix from the world coordinate system to the radar coordinate system based on the height difference, pitch angle and azimuth angle between the radar and the target. , is the rotation matrix, , is the offset vector, ; The projection matrix J is determined based on the resolution and image size of the SAR image, where the projection matrix of the projection plane is calculated as , the projection matrix of the imaging plane is , is the number of distance pixels, is the range resolution, is the number of pixels in the azimuth direction, Obtain radar parameters of each SAR image for azimuth resolution; Based on the radar parameters of each SAR image, a training data set and a test data set are constructed respectively; Construct the initial scattered point cloud, use the plane to initialize the background for the target with background, use the hemispherical shell to initialize the target, and assign the initial covariance matrix, backscattering coefficient and phase function to each point cloud object as the initial representation of the scene.

10. The target reconstruction method based on SAR Gaussian sputtering according to claim 1 or 9, characterized in that: The data in the test dataset must have the same resolution and image size as the data in the training dataset.

11. The target reconstruction method based on SAR Gaussian sputtering according to claim 1, characterized in that: Obtaining the second SAR Gaussian scattering model includes the following steps: Get the scattering intensity of the Gaussian element in the scene; The Gaussian element is projected onto the imaging plane. The difference between the two-dimensional covariance matrix, the plane coordinates, and the pixel center is used as the contribution weight of the Gaussian element scattering intensity to the pixel scattering energy. The pixel scattering energy is used as the pixel intensity to obtain the rendered SAR image. Obtain loss values ​​of the rendered SAR image and the true SAR image, and backpropagate the gradient to optimize the parameters of the Gaussian element.

12. The target reconstruction method based on SAR Gaussian sputtering according to claim 11, characterized in that: The scattering intensity of each Gaussian element is: , yes Rays in the The backward energy decay of the Gaussian element under depth, 、 and are the phase function, covariance and weighting of the ray center and the backscattering coefficient of the Gaussian element respectively.

13. The target reconstruction method based on SAR Gaussian sputtering according to claim 11, characterized in that: Distance to Pixels The scattered energy is: ,in is the distance to pixel Next The influence weight of each Gaussian element on it, is the scattering intensity of the Gaussian element.

14. The target reconstruction method based on SAR Gaussian sputtering according to claim 11, characterized in that: The loss function between the rendered SAR image obtained by rendering the SAR image under the perspective of the training dataset and the ground-truth SAR image tends to be stable and close to 0. The loss function is: ,in .

Images