Multi-view sar and isar self-supervised three-dimensional imaging method for arbitrary observation geometry

Abstract

A self-supervised 3D imaging method for multi-view SAR and ISAR with arbitrary observation geometry is proposed, comprising: acquiring observation geometric parameters and 2D imaging results from multiple views; constructing 2D projection relationships for each view based on the observation geometric parameters; acquiring the scattering parameter set of sampling points in the 3D sampling space corresponding to each view based on the implicit scattering field network constrained by geometric conditions and the 2D projection relationships; performing differentiable forward imaging based on the scattering parameter set and the 2D projection relationships to generate 2D predicted images for each view; constructing a multi-view consistency loss based on the 2D imaging results and 2D predicted images for each view; performing self-supervised iterative optimization of the implicit scattering field network based on the multi-view consistency loss until a preset iteration condition is met, and performing 3D scattering reconstruction based on the scattering parameter set to obtain the 3D scattering field reconstruction result. This method eliminates the need for 3D ground truth annotation, achieves stable 3D imaging, and improves adaptability.

Description

Technical Field

[0001] This application relates to the field of radar imaging and 3D reconstruction technology, specifically to a multi-view SAR and ISAR self-supervised 3D imaging method for arbitrary observation geometry. Background Technology

[0002] Existing radar 3D imaging technologies generally rely on coherent observations and geometric diversity to acquire 3D information, forming an overall technical route of "2D imaging processing link + 3D extended inversion". Currently, relatively mature 2D imaging and 3D reconstruction schemes have been developed for two systems: Synthetic Aperture Radar (SAR) and Inverse Synthetic Aperture Radar (ISAR). SAR 3D imaging primarily introduces height diversity information through multi-baseline / multi-view observations, combining it with tomographic inversion (TomoSAR), sparse reconstruction, or regularized inversion to recover the 3D scattering distribution. ISAR 3D imaging acquires range-Doppler 2D ISAR images with motion compensation. 3D ISAR schemes generally introduce third-dimensional information and output 3D point-like scattering distributions or point cloud structures through multi-channel, interferometric (Interferometric Inverse Synthetic Aperture Radar (InISAR)), array, or multi-station observations.

[0003] In complex application scenarios such as low-altitude integrated sensing, there is a need for mobile platforms (e.g., UAVs) to observe static targets / scenes under irregular flight paths and multiple altitude changes, generating multi-view 2D synthetic aperture radar (SAR) imaging results. Conversely, there is a need for fixed base stations to observe moving targets, forming an equivalent aperture through target motion and obtaining multi-view 2D inverse synthetic aperture radar (ISAR) imaging results. Existing 3D imaging methods often heavily rely on observation geometry assumptions, trajectory regularity, or multi-channel configurations. When faced with non-uniform viewpoint distribution, significant changes in geometric parameters, and the coexistence of static and moving targets, it is difficult to simultaneously achieve stable 3D reconstruction and 3D imaging output within the same imaging framework. Furthermore, existing solutions mainly revolve around traditional imaging links, tomographic reconstruction frameworks, or interferometric angle measurement processes under specific systems, and still have shortcomings in intelligent modeling capabilities, observation geometry adaptability, and unified processing capabilities for SAR and ISAR, making it difficult to meet the diverse, flexible, and highly adaptable 3D imaging requirements of low-altitude integrated sensing scenarios. Summary of the Invention

[0004] At least one embodiment of this application provides a multi-view SAR and ISAR self-supervised 3D imaging method for arbitrary observation geometry, which is used to solve the problem that the existing technology is difficult to meet the diverse, flexible and highly adaptable 3D imaging requirements in low-altitude integrated sensing scenarios.

[0005] To solve the above-mentioned technical problems, this application is implemented as follows:

[0006] One embodiment of this application provides a self-supervised 3D imaging method for multi-view SAR and ISAR oriented to arbitrary observation geometry, including:

[0007] The observation geometric parameters and two-dimensional imaging results are obtained from multiple perspectives. The two-dimensional imaging results include two-dimensional SAR images of stationary targets obtained by a mobile platform and two-dimensional ISAR images of moving targets obtained by a fixed base station.

[0008] Based on the observation geometric parameters, construct the two-dimensional projection relationship corresponding to each viewpoint and form an observation geometric model;

[0009] Based on the implicit scattering field network constrained by geometric conditions and the two-dimensional projection relationship, the scattering parameter set of the sampling points in the three-dimensional sampling space corresponding to each viewpoint is obtained. The three-dimensional sampling space is constructed according to the observation geometric parameters and the projection relationship corresponding to each viewpoint.

[0010] Differentiable forward imaging is performed based on the set of scattering parameters and the two-dimensional projection relationship to generate two-dimensional predicted images corresponding to each viewpoint.

[0011] Based on the two-dimensional imaging results and the two-dimensional predicted images corresponding to each viewpoint, a multi-view consistency loss is constructed.

[0012] The implicit scattering field network is self-supervised iteratively optimized based on the multi-view consistency loss until the preset iteration conditions are met, and the three-dimensional scattering reconstruction result is obtained by performing three-dimensional scattering reconstruction based on the scattering parameter set.

[0013] Specifically, the self-supervised 3D imaging method for multi-view SAR and ISAR oriented to arbitrary observation geometry, as described above, involves constructing the two-dimensional projection relationship corresponding to each viewpoint based on the acquired observation geometric parameters and two-dimensional imaging results from multiple views, and forming an observation geometric model, including:

[0014] Construct corresponding two-dimensional imaging plane bases based on the observation geometric parameters corresponding to each viewpoint;

[0015] The relative distance change rate between the three-dimensional scattering point and the reference point, as well as the unit direction vector of the reference point pointing to the radar, are obtained according to preset parameter information. The preset parameter information includes: radar position, reference point position, position of the three-dimensional scattering point relative to the reference point, rotation matrix, equivalent rotation vector, and target angular velocity.

[0016] Based on the range and azimuth unit bases in the two-dimensional imaging plane base, the equivalent rotation vector, and the relative distance change rate, the range and azimuth coordinates of the three-dimensional scattering point in the two-dimensional imaging result are obtained.

[0017] Based on the three-dimensional coordinates corresponding to each of the three-dimensional scattering points, as well as the two-dimensional range coordinates and azimuth coordinates, the two-dimensional projection relationship is determined, and the observation geometric model is formed.

[0018] Specifically, the self-supervised 3D imaging method for multi-view SAR and ISAR oriented to arbitrary observation geometry, as described above, involves obtaining the set of scattering parameters of sampling points in the 3D sampling space corresponding to each viewpoint based on the implicit scattering field network constrained by geometric conditions and the 2D projection relationship.

[0019] Based on the multi-view radar coding network, the two-dimensional imaging results corresponding to each view are processed to obtain the two-dimensional feature map corresponding to each view.

[0020] Based on the preprocessed sampling points, the observed geometric parameters, and the two-dimensional projection relationship, the sampling points are projected onto the two-dimensional feature map, and bilinear sampling is performed on the projected features on the two-dimensional feature map to obtain a set of point-level features aligned to multiple views.

[0021] Based on statistical enhancement and attention fusion, the point-level feature set is subjected to cross-view adaptive weighting to obtain fused features;

[0022] Based on the fusion features, the coordinates of the sampling points, and the observation geometric parameters, a three-dimensional scattering field is rendered to obtain the scattering parameter set corresponding to each sampling point. The scattering parameter set includes the complex scattering response of each sampling point at each viewpoint and preset parameters related to the scattering intensity or attenuation characteristics of the sampling point.

[0023] Specifically, the self-supervised 3D imaging method for multi-view SAR and ISAR oriented to arbitrary observation geometry, as described above, further includes obtaining the set of scattering parameters of sampling points in the 3D sampling space corresponding to each viewpoint based on the implicit scattering field network constrained by geometric conditions and the 2D projection relationship, as well as:

[0024] The sampling point coordinates and the observation geometric parameters corresponding to each viewpoint are preprocessed to obtain the preprocessed sampling points and observation geometric parameters.

[0025] Specifically, the self-supervised 3D imaging method for multi-view SAR and ISAR oriented to arbitrary observation geometry, as described above, involves performing differentiable forward imaging based on the scattering parameter set and the two-dimensional projection relationship to generate two-dimensional predicted images corresponding to each viewpoint, including:

[0026] Based on the viewpoint type of each viewpoint, a differentiable mapping strategy is determined, wherein the viewpoint type includes SAR viewpoint and ISAR viewpoint, and the differentiable mapping strategy includes the range Doppler imaging mechanism corresponding to the SAR viewpoint and the line integral projection mechanism along the projection direction or its discretized expression corresponding to the ISAR viewpoint.

[0027] Based on the differentiable mapping strategy corresponding to each viewpoint, differentiable forward imaging is performed based on the scattering parameter set and the two-dimensional projection relationship to generate the two-dimensional predicted image corresponding to each viewpoint.

[0028] Another embodiment of this application provides a control device, including:

[0029] The first processing module is used to acquire observation geometric parameters and two-dimensional imaging results from multiple perspectives. The two-dimensional imaging results include two-dimensional SAR images of stationary targets acquired by a mobile platform from multiple perspectives and two-dimensional ISAR images of moving targets acquired by a fixed base station from multiple perspectives.

[0030] The second processing module is used to construct the two-dimensional projection relationship corresponding to each viewpoint based on the observation geometric parameters, the obtained observation geometric parameters under multiple views, and the two-dimensional imaging results, and form an observation geometric model.

[0031] The third processing module is used to obtain the scattering parameter set of sampling points in the three-dimensional sampling space corresponding to each viewpoint based on the implicit scattering field network constrained by geometric conditions and the two-dimensional projection relationship. The three-dimensional sampling space is constructed based on the observation geometric parameters and the projection relationship corresponding to each viewpoint.

[0032] The fourth processing module is used to perform differentiable forward imaging based on the scattering parameter set and the two-dimensional projection relationship to generate two-dimensional prediction images corresponding to each viewpoint.

[0033] The fifth processing module is used to construct a multi-view consistency loss based on the two-dimensional imaging results and the two-dimensional predicted image corresponding to each viewpoint.

[0034] The sixth processing module is used to perform self-supervised iterative optimization of the implicit scattering field network based on the multi-view consistency loss until the preset iteration conditions are met, and to perform three-dimensional scattering reconstruction based on the scattering parameter set to obtain the three-dimensional scattering field reconstruction result.

[0035] Another embodiment of this application provides an electronic device, including: a processor, a memory, and a program stored in the memory and executable on the processor, wherein when the program is executed by the processor, it implements the steps of the multi-view SAR and ISAR self-supervised three-dimensional imaging method for arbitrary observation geometry as described above.

[0036] Another embodiment of this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the self-supervised three-dimensional imaging method for multi-view SAR and ISAR oriented to arbitrary observation geometry as described above.

[0037] Another embodiment of this application provides a computer program product, including computer instructions that, when executed by a processor, implement the steps of the self-supervised 3D imaging method for multi-view SAR and ISAR oriented to arbitrary observation geometry as described above.

[0038] Compared with existing technologies, the self-supervised 3D imaging method for multi-view SAR and ISAR oriented to arbitrary observation geometry provided in this application, by uniformly organizing multi-view 2D SAR / ISAR imaging results and their corresponding observation geometric parameters, and establishing the projection relationship between 2D observation and 3D space, can adapt to irregular trajectories, non-uniform view distribution, and randomly changing observation conditions. It supports SAR imaging of static targets and ISAR imaging of moving targets, improves the versatility and transferability of 3D imaging, and has good adaptability to arbitrary random observation geometry, making it more suitable for complex application scenarios such as low-altitude platform mobile observation and distributed collaborative sensing. Moreover, it does not rely on strict signal-level phase synchronization between the original echoes of multiple stations, thus making it more suitable for distributed collaborative observation in low-altitude integrated sensing scenarios. Through a differentiable forward imaging process, the predicted 3D scattering field is used to generate a 2D image under the corresponding view, and the consistency error between the predicted image and the real observation image is used for back-side optimization, thereby constructing a self-supervised training mechanism for 3D imaging. It does not require 3D ground truth labeling, nor does it rely on traditional multi-step analytical processing chains, which can effectively reduce the impact of error accumulation in intermediate steps on the final 3D imaging result. An implicit scattering field network is employed to model the 3D scattering field of a target in continuous space. This directly learns the scattering distribution of the target in continuous 3D space, enabling a description of the target's 3D structure from a continuous spatial perspective. This is more conducive to expressing the spatial continuity and detailed changes of complex targets. Furthermore, by directly using the continuous 3D scattering field as the optimization object, 3D scattering field reconstruction is completed without the need for 3D ground truth annotation through multi-view consistency constraints. This reduces the reliance on explicit scattering point matching, complex preprocessing, and manual parameter tuning. It exhibits better robustness and practical application value even under conditions of sparse viewpoints, large geometric changes, and limited sample conditions. Attached Figure Description

[0039] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of this application. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0040] Figure 1 This is one of the flowcharts illustrating the multi-view SAR and ISAR self-supervised 3D imaging method for arbitrary observation geometry proposed in this application.

[0041] Figure 2 This is the second flowchart illustrating the self-supervised 3D imaging method for multi-view SAR and ISAR oriented to arbitrary observation geometry proposed in this application.

[0042] Figure 3 This is the third flowchart illustrating the self-supervised 3D imaging method for multi-view SAR and ISAR oriented to arbitrary observation geometry proposed in this application.

[0043] Figure 4 This is the fourth flowchart illustrating the self-supervised 3D imaging method for multi-view SAR and ISAR oriented to arbitrary observation geometry proposed in this application.

[0044] Figure 5 This is a schematic diagram of the control device of this application;

[0045] Figure 6 This is a schematic diagram of the projection mapping from the scattering point to the two-dimensional imaging plane under the observation geometry model of this application. Detailed Implementation

[0046] Exemplary embodiments of the present application will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present application are shown in the drawings, it should be understood that the present application may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this application will be thorough and complete, and will fully convey the scope of the present application to those skilled in the art.

[0047] The terms “first,” “second,” etc., used in the specification and claims of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented, for example, in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus. The terms “and / or” in the specification and claims indicate at least one of the connected objects.

[0048] Please refer to Figure 1 This application provides a self-supervised 3D imaging method for multi-view SAR and ISAR with arbitrary observation geometry, comprising:

[0049] Step S101: Obtain observation geometric parameters and two-dimensional imaging results from multiple perspectives. The two-dimensional imaging results include two-dimensional SAR images of a stationary target obtained by a mobile platform from multiple perspectives and two-dimensional ISAR images of a moving target obtained by a fixed base station from multiple perspectives.

[0050] Step S102: Based on the observation geometric parameters, construct the two-dimensional projection relationship corresponding to each viewpoint and form an observation geometric model;

[0051] Step S103: Based on the implicit scattering field network constrained by geometric conditions and the two-dimensional projection relationship, obtain the scattering parameter set of the sampling points in the three-dimensional sampling space corresponding to each viewpoint. The three-dimensional sampling space is constructed according to the observation geometric parameters and the projection relationship corresponding to each viewpoint.

[0052] Step S104: Perform differentiable forward imaging based on the scattering parameter set and the two-dimensional projection relationship to generate two-dimensional prediction images corresponding to each viewpoint.

[0053] Step S105: Construct a multi-view consistency loss based on the two-dimensional imaging results and the two-dimensional predicted image corresponding to each viewpoint;

[0054] Step S106: Perform self-supervised iterative optimization on the implicit scattering field network according to the multi-view consistency loss until the preset iteration conditions are met, and perform three-dimensional scattering reconstruction according to the scattering parameter set to obtain the three-dimensional scattering field reconstruction result.

[0055] In this embodiment, by acquiring observation geometric parameters and 2D imaging results from multiple perspectives as input, SAR and ISAR are unified into an input format of "2D observation results + observation geometric parameters," and used for 3D reconstruction under the same implicit scattering field network and the same optimization framework. This enables simultaneous support for SAR imaging of stationary targets and ISAR imaging of moving targets, improving the versatility and transferability of 3D imaging. The observation geometric parameters include at least line-of-sight parameters and imaging plane basis parameters / projection direction parameters. Furthermore, the observation geometric parameters are allowed to be sparsely distributed, non-uniformly distributed, or irregularly distributed.

[0056] Then, based on the observation geometric parameters under each viewpoint, the two-dimensional projection relationship corresponding to each viewpoint can be constructed to describe the projection relationship from the three-dimensional scattering point to the two-dimensional image coordinates, which facilitates subsequent steps such as obtaining scattering parameters based on the implicit scattering field network and generating two-dimensional prediction images through differentiable forward imaging.

[0057] Furthermore, based on the implicit scattering field network constrained by geometric conditions and the aforementioned two-dimensional projection relationship, the set of scattering parameters for sampling points in the three-dimensional sampling space corresponding to each viewpoint can be obtained. This implicit scattering field network is then used to parameterize the three-dimensional scattering distribution of the target. Instead of directly solving for the scattering value of each voxel independently on a fixed discrete voxel, a differentiable mapping model is constructed to directly predict the scattering parameters at the sampling point, thus forming a continuous three-dimensional scattering field representation. This method can express the spatial scattering structure of complex targets with a finite parameter scale and provides a unified implementation form for optimization solutions under consistency constraints of multi-view two-dimensional images.

[0058] After obtaining the set of scattering parameters, differentiable forward imaging can be performed on each viewpoint according to the set of scattering parameters and the two-dimensional projection relationship to generate two-dimensional predicted images corresponding to each viewpoint. Since the two-dimensional predicted image is generated by differentiable forward imaging (forward projection) from the output of the implicit scattering field network, it is differentiable with respect to the network parameters of the implicit scattering field network. Based on this, a multi-view consistency loss can be obtained from the two-dimensional predicted image and the two-dimensional imaging result under the corresponding viewpoint, so as to facilitate self-supervised training of the implicit scattering field network. That is, the error between the observed two-dimensional image and the predicted two-dimensional image is directly used to back-optimize the network parameters. It does not require the provision of three-dimensional real annotation of the target, nor does it rely on the traditional multi-step analytical processing chain. It can effectively reduce the impact of the accumulation of errors in intermediate steps on the final three-dimensional imaging result. The multi-view consistency loss includes at least the rendering consistency difference term between the predicted two-dimensional image and the observed two-dimensional image of each viewpoint, and can optionally include scattering field sparsity constraints, spatial smoothness constraints, or energy constraints.

[0059] Finally, after the iterative training meets the preset iterative conditions, the three-dimensional scattering reconstruction is performed based on the scattering parameter set to obtain the three-dimensional scattering field reconstruction result, which helps to ensure the accuracy of the three-dimensional scattering field reconstruction.

[0060] In summary, this application, by unifying the organization of multi-view 2D SAR / ISAR imaging results and their corresponding observation geometric parameters, and establishing the projection relationship between 2D observation and 3D space, can adapt to irregular trajectories, non-uniform viewpoint distributions, and randomly changing observation conditions. It supports both static target SAR imaging and moving target ISAR imaging, improving the versatility and transferability of 3D imaging. It exhibits good adaptability to any random observation geometry and is more suitable for complex application scenarios such as low-altitude platform mobile observation and distributed collaborative sensing. Furthermore, it does not rely on strict signal-level phase synchronization between the original echoes from multiple stations, making it more suitable for distributed collaborative observation in low-altitude integrated sensing scenarios. Through a differentiable forward imaging process, the obtained 3D scattering field generates a 2D predicted image from the corresponding viewpoint. Backward optimization is performed using the multi-view consistency loss between the predicted image and the real observation image, thereby constructing a self-supervised training mechanism for 3D imaging. This eliminates the need for 3D ground truth labeling and traditional multi-step analytical processing chains, effectively reducing the impact of intermediate step error accumulation on the final 3D imaging result. An implicit scattering field network is employed to model the 3D scattering field of a target in continuous space. This directly learns the scattering distribution of the target in continuous 3D space, enabling a description of the target's 3D structure from a continuous spatial perspective. This is more conducive to expressing the spatial continuity and detailed changes of complex targets. Furthermore, by directly using the continuous 3D scattering field as the optimization object, 3D scattering field reconstruction is completed without the need for 3D ground truth annotation through multi-view consistency loss constraints. This reduces the reliance on explicit scattering point matching, complex preprocessing, and manual parameter tuning. It exhibits better robustness and practical application value even under conditions of sparse viewpoints, large geometric changes, and limited sample conditions.

[0061] It should be noted that the final 3D scattering field reconstruction result of the target can be further represented in various forms, such as 3D voxel intensity distribution, 3D point cloud, spatial slice map, isosurface or volume rendering image, depending on the application requirements, to achieve visualization of the target's 3D structure and facilitate subsequent analysis. Alternatively, 3D imaging results can be generated based on the 3D scattering field reconstruction result. The 3D imaging result includes at least SAR 3D imaging output for static targets and ISAR 3D imaging output for moving targets, thereby improving the system's 3D imaging capabilities for both static and dynamic targets.

[0062] In another embodiment of this application, after the iterative training meets the preset iteration conditions (e.g., reaching a preset number of iterations or the multi-view consistency loss is less than a preset value), more dense sampling points can be selected in the three-dimensional sampling space, and point-by-point queries can be performed based on the optimized implicit scattering field network to obtain the corresponding set of scattering parameters (scattering parameter distribution), thereby obtaining the three-dimensional scattering field reconstruction result of the target. It should be noted that this iterative training refers to returning to step S103 if the preset iteration conditions are not met.

[0063] In one embodiment, the multi-view consistency loss uses the multi-view amplitude consistency error, which can be written in the form of:

[0064]

[0065] in, This represents the operation of complex image magnitude. This represents the summation or averaging of absolute values ​​pixel by pixel. Type error measure, Indicates the first The actual two-dimensional imaging results from each perspective Indicates the first Two-dimensional predicted images from multiple perspectives. By minimizing the loss function shown in the above equation, the three-dimensional scattering field represented by the network can gradually approximate the real two-dimensional SAR / ISAR image after forward projection from each perspective, thereby achieving three-dimensional scattering structure learning under multi-view joint constraints.

[0066] See Figure 2 Specifically, the self-supervised 3D imaging method for multi-view SAR and ISAR oriented to arbitrary observation geometry, as described above, involves constructing the two-dimensional projection relationship corresponding to each viewpoint based on the acquired observation geometric parameters and two-dimensional imaging results from multiple views, and forming an observation geometric model, including:

[0067] Step S201: Construct corresponding two-dimensional imaging plane bases according to the observation geometric parameters corresponding to each viewpoint;

[0068] Step S202: Obtain the relative distance change rate between the three-dimensional scattering point and the reference point and the unit direction vector of the reference point pointing to the radar according to the preset parameter information. The preset parameter information includes: radar position, reference point position, position of the three-dimensional scattering point relative to the reference point, rotation matrix, equivalent rotation vector, and target angular velocity.

[0069] Step S203: Based on the range unit basis and azimuth unit basis in the two-dimensional imaging plane basis, as well as the equivalent rotation vector and the relative distance change rate, obtain the range coordinates and azimuth coordinates of the three-dimensional scattering point in the two-dimensional imaging result;

[0070] Step S204: Based on the three-dimensional coordinates corresponding to each of the three-dimensional scattering points, as well as the two-dimensional range coordinates and azimuth coordinates, determine the two-dimensional projection relationship and form the observation geometric model.

[0071] In this embodiment, the steps for obtaining the two-dimensional projection relationship corresponding to each viewpoint are illustrated. To facilitate understanding by those skilled in the art, in a specific embodiment, the following observation geometric model is established in the inertial reference coordinate system, and the steps for obtaining the two-dimensional projection relationship are explained based on the observation geometric model.

[0072] Assume the radar is at time... Spatial location is denoted as The reference point selected on the target at time Spatial location is denoted as Any scattering point on the target The fixed coordinates of this reference point in the target fixed coordinate system are as follows: The rotational relationship of the target attitude relative to the inertial frame is denoted by a matrix. Then the instantaneous position of the scattering point in the inertial frame can be written as:

[0073]

[0074] in, Let be an orthogonal rotation matrix, satisfying: ,in, Represents the identity matrix.

[0075] Record the scattering point The instantaneous distance to the radar is: ;

[0076] Meanwhile, the instantaneous distance from the target reference point to the radar is defined as: .

[0077] Furthermore, the unit direction vector of the target reference point pointing to the radar from the current perspective, i.e., the line-of-sight unit vector, is:

[0078]

[0079] in, This represents the unit direction vector pointing from the target reference point to the radar.

[0080] Under far-field conditions, the target size is negligible relative to the observation distance. A first-order expansion can be performed near the reference point. Let the instantaneous relative position of the scattering point with respect to the reference point in the inertial frame be:

[0081]

[0082] The distance to the scattering point can then be approximated as:

[0083]

[0084] This equation shows that the additional distance of the scattering point relative to the reference point is approximately equal to the projection of the scattering point onto the current line of sight.

[0085] In actual imaging processing, it is also necessary to first compensate for the translational component of the reference point, that is, to eliminate the effect of... The resulting common phase term. After translational compensation, the equivalent distance term corresponding to the scattering point can be written as:

[0086]

[0087] This equation gives the equivalent geometric quantity of the scattering point in the range direction after compensation. Taking its derivative with respect to time, we get:

[0088]

[0089] The sources of the two terms in the above formula will be analyzed below.

[0090] First, define the relative velocity of the radar with respect to the target reference point as: ,in, The derivative of the radar's spatial position with respect to time. The derivative of the spatial position of the reference point with respect to time;

[0091] From the definition of the unit line-of-sight vector, its rate of change over time satisfies:

[0092]

[0093] This equation shows that the change in the line-of-sight direction is caused only by the component of the relative velocity perpendicular to the line-of-sight direction. This change can be further expressed in angular velocity form:

[0094]

[0095] in,

[0096]

[0097] This represents the line-of-sight rotation angular velocity caused by the relative motion between the radar and the target reference point.

[0098] On the other hand, the target's own attitude also changes over time, so its angular velocity in the inertial frame is denoted as . Then the position change of the scattering point relative to the reference point satisfies:

[0099]

[0100] Will and Substitution We can obtain:

[0101]

[0102] Using vector identity can The second item in the middle is rewritten as:

[0103]

[0104] therefore, It can be uniformly organized as follows:

[0105]

[0106] For ease of subsequent explanation, a unified equivalent rotation vector is defined: Then, the rate of change of distance after scattering point compensation can be further written as: .

[0107] This equation is a key relationship in the observation geometry model of this application. It shows that after completing the translational compensation of the reference point, the rate of change of the distance of the scattering point is determined by two parts: one part comes from the rotation of the observation line of sight, and the other part comes from the rotation of the target itself; the two can be unified and combined into an equivalent rotation vector. Therefore, whether it is SAR or ISAR, the azimuth information in a two-dimensional image can essentially be regarded as the three-dimensional scattering point in a vector... Projection in the direction.

[0108] Within a relatively short coherent accumulation time, the target's turning angle can usually be approximated as small, and and At the center moment The surrounding area is changing slowly. Therefore, the following approach is appropriate: And the relative positions of the scattering points are approximated as constant. At this point, two unit basis vectors can be defined in the two-dimensional image plane.

[0109] The first basis vector is defined as the locative unit basis: ;

[0110] The second basis vector is defined as the azimuth-oriented unit basis: ;

[0111] Therefore, the range coordinates of the scattering point in a two-dimensional SAR / ISAR image can be expressed as: ;

[0112] The corresponding normalized azimuth coordinates can be expressed as: ;

[0113] Therefore, the mapping from the three-dimensional coordinates of the scattering point to the two-dimensional image coordinates can be uniformly expressed as the following linear mapping form:

[0114]

[0115] Let the projection matrix be:

[0116]

[0117] Then we have:

[0118]

[0119] This formula is the two-dimensional projection model used in this application to uniformly describe the two-dimensional projection relationship between SAR and ISAR. The model shows that, after translational compensation, the position of the three-dimensional scattering point in the two-dimensional radar image is determined by a set of unified two-dimensional basis vectors. The first basis vector is the current line-of-sight direction, and the second basis vector is the azimuth direction obtained by normalizing the cross product of the equivalent rotation vector and the line-of-sight direction. Therefore, SAR and ISAR have a unified three-dimensional to two-dimensional mapping mechanism in mathematical form; the main difference between them lies only in the equivalent rotation vector. The physical sources of rotation differ: for SAR, the target is usually approximately stationary, and the equivalent rotation mainly comes from the line-of-sight rotation caused by platform motion; for ISAR, the radar position is usually approximately fixed, and the equivalent rotation mainly comes from the target's own motion. Despite the different physical sources, the mapping between two-dimensional image coordinates and three-dimensional scattering position is consistent.

[0120] Furthermore, due to Only the projected coordinates of the three-dimensional scattering point on the two-dimensional imaging plane are given; therefore, single-view observation will still lose information along the normal direction of this imaging plane. Let this normal direction be denoted as: ;

[0121] Therefore, a two-dimensional image essentially only preserves the image of the scatterer in the form of... and The projected structure on the plane, and along Depth information in direction cannot be uniquely recovered under single-view conditions. For example... Figure 6 As shown, under the observation geometry model, a three-dimensional scattering point can be projected onto a two-dimensional imaging plane spanned by the range basis and the azimuth basis, and its two-dimensional coordinates are determined by the projection of the point onto the two plane basis vectors.

[0122] See Figure 3 Specifically, the self-supervised 3D imaging method for multi-view SAR and ISAR oriented to arbitrary observation geometry, as described above, involves obtaining the set of scattering parameters of sampling points in the 3D sampling space corresponding to each viewpoint based on the implicit scattering field network constrained by geometric conditions and the 2D projection relationship. This includes:

[0123] Step S301: Based on the multi-view radar coding network, process the two-dimensional imaging results corresponding to each view to obtain the two-dimensional feature map corresponding to each view.

[0124] Step S302: Based on the preprocessed sampling points, the observation geometric parameters, and the two-dimensional projection relationship, the sampling points are projected onto the two-dimensional feature map, and bilinear sampling is performed on the projected features on the two-dimensional feature map to obtain a multi-view aligned point-level feature set.

[0125] Step S303: Based on statistical enhancement and attention fusion, the point-level feature set is subjected to cross-view adaptive weighting to obtain fused features;

[0126] Step S304: Render the three-dimensional scattering field according to the fusion features, the sampling point coordinates and the observation geometric parameters to obtain the scattering parameter set corresponding to each sampling point. The scattering parameter set includes the complex scattering response of each sampling point at each viewpoint and preset parameters related to the scattering intensity or attenuation characteristics of the sampling point.

[0127] This embodiment illustrates the specific steps of obtaining the scattering parameter set using the implicit scattering field network described above. This implicit scattering field network can be composed of a multi-view radar coding network, a cross-view fusion network, and a three-dimensional scattering field rendering network. Specifically, the observed geometric parameters are first preprocessed through position encoding or harmonic embedding before participating in fusion and prediction to obtain view-dependent scattering prediction capabilities.

[0128] The multi-view radar coding network takes the two-dimensional imaging results of each view as input and outputs two-dimensional feature maps of each view. The cross-view fusion network uses the sampling points as indexes, projects the sampling points onto the two-dimensional feature maps according to the observation geometric parameters of each view, and performs bilinear sampling on the features at the projection points on the two-dimensional feature maps to obtain a set of point-level features aligned to multiple views. The point-level feature sets can be further weighted across views through statistical enhancement and attention fusion to form fused features. The three-dimensional scattering field rendering network takes the fused features, the coordinates of the sampling points, and the observation geometric parameters as input and outputs a set of scattering parameters corresponding to each sampling point.

[0129] This embodiment illustrates the steps for obtaining the scattering parameter set described above. Specifically, this application employs a geometrically constrained implicit scattering field network to parametrically model the three-dimensional scattering distribution of the target. This network takes the coordinates of the three-dimensional sampling points and the corresponding observation geometric parameters as input, and outputs the scattering field parameters of each sampling point after implicit mapping, thereby forming a continuous three-dimensional scattering field representation.

[0130] For the For each observation viewpoint, a local observation coordinate system is first constructed based on its corresponding line-of-sight and projection directions, and a regular three-dimensional sampling mesh is generated within this coordinate system. Let the current viewpoint be the first... The spatial coordinates of each sampling point are: All sampling points together constitute the sampling point set corresponding to the current viewpoint. Since the generation process of these sampling points is directly related to the current observation geometry, the sampling space under different viewpoints can adaptively match the imaging direction of their respective two-dimensional images, thus providing a geometric basis for subsequently establishing the correspondence between the three-dimensional scattering field and the two-dimensional image.

[0131] Based on this, the geometrically constrained implicit scattering field network constructed in this application is denoted as:

[0132]

[0133] in, Indicates by parameters Implicit mapping model of control Indicates the first From the perspective of the first Coordinates of each sampling point This indicates the observation geometry parameters corresponding to the current viewpoint, used to characterize parameters such as the line-of-sight direction and projection direction. This represents the set of scattering parameters output by the network at this sampling point.

[0134] In one specific embodiment of this application, the scattering field parameters include preset parameters related to the volume and scattering intensity or attenuation characteristics, as well as the real and imaginary parts of the complex scattering response, namely:

[0135]

[0136] in, This represents a preset parameter related to the scattering intensity or attenuation characteristics of the sampling point. and Let represent the real and imaginary parts of the complex scattering response at this sampling point, respectively. Therefore, the complex scattering response at this sampling point can be constructed as:

[0137]

[0138] in, It is the imaginary unit.

[0139] It should also be noted that, as shown in the aforementioned observation geometry model, at the m-th observation viewpoint, the position of any spatial point in the target within the 2D SAR / ISAR image can be determined by the 2D imaging basis vector corresponding to that viewpoint. Therefore, if the target is generalized to a continuous three-dimensional scattering field, the 2D image can be considered as the projection result of this three-dimensional scattering field under the current observation geometry. Based on this, a forward model from the three-dimensional scattering field to the 2D image can be further established, and a mathematical expression for multi-view joint 3D reconstruction can be given accordingly.

[0140] Let the three-dimensional scattering distribution of the target be denoted as: ,in, Indicates spatial location The scattering response at point . For the ... Each observation perspective, based on the local imaging basis defined in the previous section. , and normal base Any point in space can be written as: ,in, and These represent the range and azimuth coordinates of the point within the current two-dimensional imaging plane, respectively. This represents the depth coordinates along the normal direction.

[0141] Therefore, the first Two-dimensional images from different perspectives at the pixel level The response at that point can be expressed as the cumulative result of the three-dimensional scattered field along the normal direction of that viewpoint, that is:

[0142]

[0143] The above equation shows that a single 2D SAR / ISAR image is essentially a projection of the 3D scattering field onto the current observation viewpoint. For ease of numerical calculation, the normal direction can be discretized to obtain the discrete form:

[0144]

[0145] in, Indicates the first Each depth sampling location This represents the corresponding projection weight. Therefore, a two-dimensional image from a single viewpoint only reflects the result of the three-dimensional scattering field being compressed along a certain normal direction, and thus cannot uniquely recover the complete three-dimensional structure.

[0146] When the same goal is achieved Two-dimensional images from different perspectives At that time, each viewpoint corresponds to a different local imaging base, thus forming multiple different projections onto the same three-dimensional scattering field. The first... The forward imaging process from each perspective is uniformly denoted as: ,

[0147] in, Indicates the first The 3D-to-2D projection operator corresponding to each viewpoint. Therefore, the multi-view joint observation model can be written as: ,

[0148] in, This is the complete set of two-dimensional observation images. It is a set of multi-view joint projection operators.

[0149] Based on the above relationships, the 3D reconstruction problem can be formulated as follows: Given multi-view 2D SAR / ISAR images and corresponding observation geometric parameters, find a 3D scattering field that, after projection from each viewpoint, is as consistent as possible with the observed images, i.e.:

[0150]

[0151] in, This represents a measure of the difference between the predicted image and the observed image. This expression represents the estimation result of the three-dimensional scattering field that satisfies the consistency of multi-view observations. It provides the basic mathematical form of the three-dimensional imaging problem in this application. Since the three-dimensional scattering field to be solved is a high-dimensional continuous unknown, and the number of multi-view two-dimensional images is limited, this problem is essentially an underdetermined inverse problem. To improve representational capability and enhance solution stability, this application employs a parameterized approach to model the three-dimensional scattering field in subsequent steps and constructs a differentiable forward projection process, enabling multi-view two-dimensional observations to jointly constrain and optimize the three-dimensional scattering field parameters. Based on this, a three-dimensional imaging implementation method based on an implicit scattering field network is further presented.

[0152] Specifically, the self-supervised 3D imaging method for multi-view SAR and ISAR oriented to arbitrary observation geometry, as described above, further includes obtaining the set of scattering parameters of sampling points in the 3D sampling space corresponding to each viewpoint based on the implicit scattering field network constrained by geometric conditions and the 2D projection relationship, as well as:

[0153] The sampling point coordinates and the observation geometric parameters corresponding to each viewpoint are preprocessed to obtain the preprocessed sampling points and observation geometric parameters.

[0154] In this embodiment, to enhance the network's ability to represent changes in spatial position and geometric direction, the coordinates of the three-dimensional sampling points and the observed geometric parameters can be preprocessed, for example, by position encoding or harmonic expansion, before being input into the implicit scattering field network for mapping. After network calculation, a set of scattering field parameters is output for each sampling point, thus obtaining the scattering parameter set corresponding to each sampling point.

[0155] See Figure 4 Specifically, the self-supervised 3D imaging method for multi-view SAR and ISAR oriented to arbitrary observation geometry, as described above, includes the following: Differentiable forward imaging is performed based on the scattering parameter set and the 2D projection relationship to generate 2D predicted images corresponding to each viewpoint.

[0156] Step S401: Determine a differentiable mapping strategy based on the viewpoint type of each viewpoint, wherein the viewpoint type includes SAR viewpoint and ISAR viewpoint, and the differentiable mapping strategy includes the range Doppler imaging mechanism corresponding to the SAR viewpoint and the line integral projection mechanism along the projection direction or its discretized expression corresponding to the ISAR viewpoint.

[0157] Step S402: Based on the differentiable mapping strategy corresponding to each viewpoint, perform differentiable forward imaging based on the scattering parameter set and the two-dimensional projection relationship to generate the two-dimensional prediction image corresponding to each viewpoint.

[0158] In this embodiment, the steps of differentiable forward imaging are illustrated, wherein, after obtaining the scattering parameters of each sampling point, a forward generation process for a two-dimensional image is further constructed based on the aforementioned three-dimensional to two-dimensional projection relationship. For the first... The position of any pixel in the two-dimensional image from a single observation perspective. This corresponds to a set of depth sampling points arranged along the normal direction in the current viewpoint sampling space. Let the set of indices corresponding to these sampling points be denoted as... Then the predicted complex response at that pixel location can be written as:

[0159]

[0160] in, Indicates the first From the perspective of the first The weight of each depth sampling point in the current pixel generation process is determined by the volume scattering parameters of the corresponding sampling point. In one implementation of this application, the weight can be obtained by monotonically mapping scattering intensity related parameters, thereby reflecting the contribution of different depth sampling points to the two-dimensional image response.

[0161] As can be seen from the formula, this application achieves differentiable forward imaging from a three-dimensional scattering field to a two-dimensional image by weighted summation of the multi-depth scattering responses corresponding to the same two-dimensional pixel.

[0162] Applying the above process to each observation viewpoint yields a set of two-dimensional predicted images corresponding to all viewpoints:

[0163]

[0164] Since the two-dimensional predicted image is generated by forward projection from the output of the implicit scattering field network, it is sensitive to the network parameters. Differentiable. Based on this, this application employs multi-view image consistency constraints to perform self-supervised training of the network, that is, without providing the three-dimensional true annotation of the target, it directly uses the error between the observed two-dimensional image and the predicted two-dimensional image to optimize the network parameters in reverse.

[0165] Furthermore, since the principles of differentiable mapping differ depending on the viewpoint type, to ensure the accuracy of differentiable mapping under different viewpoint types, a differentiable mapping strategy is first determined based on the viewpoint type of each input viewpoint. The viewpoint types include SAR viewpoints and ISAR viewpoints, and the differentiable mapping strategies include the range Doppler imaging mechanism corresponding to the SAR viewpoint and the line integral projection mechanism along the projection direction or its discretized expression corresponding to the ISAR viewpoint. Further, based on the differentiable mapping strategy corresponding to each viewpoint, differentiable forward imaging is performed based on the scattering parameter set and the two-dimensional projection relationship to generate the two-dimensional predicted image corresponding to each viewpoint.

[0166] It should be noted that the above only provides reference differentiable mapping strategies for SAR and ISAR perspectives. In practical applications, strategies that can achieve the same effect should also fall within the scope of protection of this application.

[0167] It should also be noted that the above-described differentiable forward imaging / rendering processes all support gradient backpropagation of the network output and network parameters in their numerical implementation, thus enabling end-to-end self-supervised optimization. The two-dimensional predicted image can be further used to obtain a two-dimensional predicted amplitude image, which is used to construct a rendering consistency loss with the observed two-dimensional image. Various methods of the embodiments of this application have been described above. Apparatus for implementing the above methods will be further provided below.

[0168] Please refer to Figure 5 This application also provides an electronic device, including:

[0169] Another embodiment of this application provides a control device, including:

[0170] The first processing module 501 is used to acquire observation geometric parameters and two-dimensional imaging results from multiple perspectives. The two-dimensional imaging results include two-dimensional SAR images of a stationary target acquired by a mobile platform from multiple perspectives and two-dimensional ISAR images of a moving target acquired by a fixed base station from multiple perspectives.

[0171] The second processing module 502 is used to construct the two-dimensional projection relationship corresponding to each viewpoint based on the observation geometric parameters, the obtained observation geometric parameters under multiple views and the two-dimensional imaging results, and form an observation geometric model.

[0172] The third processing module 503 is used to obtain the scattering parameter set of sampling points in the three-dimensional sampling space corresponding to each viewpoint based on the implicit scattering field network constrained by geometric conditions and the two-dimensional projection relationship. The three-dimensional sampling space is constructed based on the observation geometric parameters and the projection relationship corresponding to each viewpoint.

[0173] The fourth processing module 504 is used to perform differentiable forward imaging based on the scattering parameter set and the two-dimensional projection relationship to generate two-dimensional prediction images corresponding to each viewpoint.

[0174] The fifth processing module 505 is used to construct a multi-view consistency loss based on the two-dimensional imaging results and the two-dimensional predicted image corresponding to each viewpoint.

[0175] The sixth processing module 506 is used to perform self-supervised iterative optimization of the implicit scattering field network based on the multi-view consistency loss until the preset iteration conditions are met, and to perform three-dimensional scattering reconstruction based on the scattering parameter set to obtain the three-dimensional scattering field reconstruction result.

[0176] Specifically, in the control device described above, the second processing module includes:

[0177] The first processing unit is used to construct corresponding two-dimensional imaging plane bases according to the observation geometric parameters corresponding to each viewpoint;

[0178] The second processing unit is used to obtain the relative distance change rate between the three-dimensional scattering point and the reference point and the unit direction vector of the reference point pointing to the radar according to preset parameter information. The preset parameter information includes: radar position, reference point position, position of the three-dimensional scattering point relative to the reference point, rotation matrix, equivalent rotation vector, and target angular velocity.

[0179] The third processing unit is used to obtain the range coordinates and azimuth coordinates of the three-dimensional scattering point in the two-dimensional imaging result based on the range unit basis and azimuth unit basis in the two-dimensional imaging plane basis, the equivalent rotation vector, and the relative distance change rate.

[0180] The fourth processing unit is used to determine the two-dimensional projection relationship and form an observation geometric model based on the three-dimensional coordinates corresponding to each of the three-dimensional scattering points, as well as the two-dimensional range coordinates and azimuth coordinates.

[0181] Specifically, the control device described above, the third processing module, includes:

[0182] The fifth processing unit is used to process the two-dimensional imaging results corresponding to each view according to the multi-view radar coding network to obtain the two-dimensional feature map corresponding to each view.

[0183] The sixth processing unit is used to project the sampling points onto the two-dimensional feature map based on the preprocessed sampling points, the observation geometric parameters, and the two-dimensional projection relationship, and to perform bilinear sampling on the projected features on the two-dimensional feature map to obtain a multi-view aligned point-level feature set.

[0184] The seventh processing unit is used to perform cross-view adaptive weighting on the point-level feature set based on statistical enhancement and attention fusion to obtain fused features;

[0185] The eighth processing unit is used to perform three-dimensional scattering field rendering based on the fused body features, the sampling point coordinates and the observation geometric parameters, to obtain the scattering parameter set corresponding to each sampling point.

[0186] Specifically, in the control device described above, the third processing module further includes:

[0187] The ninth processing unit is used to preprocess the sampling point coordinates of the sampling points and the observation geometric parameters corresponding to each viewpoint to obtain the preprocessed sampling points and the observation geometric parameters.

[0188] Specifically, in the control device described above, the fourth processing module includes:

[0189] The tenth processing unit is used to determine a differentiable mapping strategy based on the viewpoint type of each viewpoint, wherein the viewpoint type includes SAR viewpoint and ISAR viewpoint, and the differentiable mapping strategy includes the range Doppler imaging mechanism corresponding to the SAR viewpoint and the line integral projection mechanism along the projection direction corresponding to the ISAR viewpoint or its discretized expression.

[0190] The eleventh processing unit is used to perform differentiable forward imaging based on the scattering parameter set and the two-dimensional projection relationship according to the differentiable mapping strategy corresponding to each viewpoint, and generate the two-dimensional prediction image corresponding to each viewpoint.

[0191] The apparatus embodiments of this application are apparatuses corresponding to the embodiments of the methods described above. All implementation means in the method embodiments described above are applicable to the apparatus embodiments and can achieve the same technical effects. The apparatus provided in this application embodiments can implement all the method steps implemented in the method embodiments described above and can achieve the same technical effects. Therefore, the parts and beneficial effects that are the same as those in the method embodiments in this embodiment will not be described in detail here.

[0192] Another embodiment of this application provides an electronic device, including: a processor, a memory, and a program stored in the memory and executable on the processor. When the program is executed by the processor, it implements the steps of the self-supervised three-dimensional imaging method of multi-view SAR and ISAR for arbitrary observation geometry as described above, and can achieve the same technical effect. To avoid repetition, it will not be described again here.

[0193] Another embodiment of this application provides a computer-readable storage medium storing a computer program. When executed by a processor, the computer program implements the steps of the self-supervised 3D imaging method for multi-view SAR and ISAR oriented to arbitrary observation geometry as described above, and achieves the same technical effect. To avoid repetition, it will not be described again here. The computer-readable storage medium may be a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk, etc.

[0194] Another embodiment of this application provides a computer program product, including computer instructions. When executed by a processor, these computer instructions implement the steps of the self-supervised three-dimensional imaging method for multi-view SAR and ISAR oriented to arbitrary observation geometry as described above, and can achieve the same technical effect. To avoid repetition, they will not be described again here.

[0195] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0196] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0197] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.

Claims

1. A self-supervised 3D imaging method for multi-view SAR and ISAR oriented to arbitrary observation geometry, characterized in that, include: The observation geometric parameters and two-dimensional imaging results are obtained from multiple perspectives. The two-dimensional imaging results include two-dimensional SAR images of stationary targets obtained by a mobile platform and two-dimensional ISAR images of moving targets obtained by a fixed base station. Based on the observation geometric parameters, construct the two-dimensional projection relationship corresponding to each viewpoint and form an observation geometric model; Based on the implicit scattering field network constrained by geometric conditions and the two-dimensional projection relationship, the scattering parameter set of the sampling points in the three-dimensional sampling space corresponding to each viewpoint is obtained. The three-dimensional sampling space is constructed according to the observation geometric parameters and the projection relationship corresponding to each viewpoint. Differentiable forward imaging is performed based on the set of scattering parameters and the two-dimensional projection relationship to generate two-dimensional predicted images corresponding to each viewpoint. Based on the two-dimensional imaging results and the two-dimensional predicted image corresponding to each viewpoint, a multi-view consistency loss is constructed. The implicit scattering field network is self-supervised iteratively optimized based on the multi-view consistency loss until the preset iteration conditions are met, and the three-dimensional scattering reconstruction result is obtained by performing three-dimensional scattering reconstruction based on the scattering parameter set.

2. The self-supervised 3D imaging method for multi-view SAR and ISAR oriented to arbitrary observation geometry as described in claim 1, characterized in that, The step of constructing two-dimensional projection relationships for each viewpoint based on the observed geometric parameters and forming an observed geometric model includes: Construct corresponding two-dimensional imaging plane bases based on the observation geometric parameters corresponding to each viewpoint; The relative distance change rate between the three-dimensional scattering point and the reference point, as well as the unit direction vector of the reference point pointing to the radar, are obtained based on preset parameter information. The preset parameter information includes: radar position, reference point position, position of the three-dimensional scattering point relative to the reference point, rotation matrix, equivalent rotation vector, and target angular velocity. Based on the range and azimuth unit bases in the two-dimensional imaging plane base, the equivalent rotation vector, and the relative distance change rate, the range and azimuth coordinates of the three-dimensional scattering point in the two-dimensional imaging result are obtained. Based on the three-dimensional coordinates corresponding to each of the three-dimensional scattering points, as well as the two-dimensional range coordinates and azimuth coordinates, the two-dimensional projection relationship is determined, and the observation geometric model is formed.

3. The self-supervised 3D imaging method for multi-view SAR and ISAR oriented to arbitrary observation geometry as described in claim 1, characterized in that, The process of obtaining the scattering parameter set of sampling points in the three-dimensional sampling space corresponding to each viewpoint, based on the implicit scattering field network constrained by geometric conditions and the two-dimensional projection relationship, includes: Based on the multi-view radar coding network, the two-dimensional imaging results corresponding to each view are processed to obtain the two-dimensional feature map corresponding to each view. Based on the preprocessed sampling points, the observed geometric parameters, and the two-dimensional projection relationship, the sampling points are projected onto the two-dimensional feature map, and bilinear sampling is performed on the projected features on the two-dimensional feature map to obtain a set of point-level features aligned to multiple views. Based on statistical enhancement and attention fusion, the point-level feature set is subjected to cross-view adaptive weighting to obtain fused features; Based on the fusion features, the coordinates of the sampling points, and the observation geometric parameters, a three-dimensional scattering field is rendered to obtain the scattering parameter set corresponding to each sampling point. The scattering parameter set includes the complex scattering response of each sampling point at each viewpoint and preset parameters related to the scattering intensity or attenuation characteristics of the sampling point.

4. The self-supervised 3D imaging method for multi-view SAR and ISAR oriented to arbitrary observation geometry as described in claim 3, characterized in that, The method of obtaining the scattering parameter set of sampling points in the three-dimensional sampling space corresponding to each viewpoint based on the implicit scattering field network constrained by geometric conditions and the two-dimensional projection relationship further includes: The sampling point coordinates and the observation geometric parameters corresponding to each viewpoint are preprocessed to obtain the preprocessed sampling points and observation geometric parameters.

5. The self-supervised 3D imaging method for multi-view SAR and ISAR oriented to arbitrary observation geometry as described in claim 1, characterized in that, The step of performing differentiable forward imaging based on the set of scattering parameters and the two-dimensional projection relationship to generate two-dimensional predicted images corresponding to each viewpoint includes: Based on the viewpoint type of each viewpoint, a differentiable mapping strategy is determined, wherein the viewpoint type includes SAR viewpoint and ISAR viewpoint, and the differentiable mapping strategy includes the range Doppler imaging mechanism corresponding to the SAR viewpoint and the line integral projection mechanism along the projection direction or its discretized expression corresponding to the ISAR viewpoint. Based on the differentiable mapping strategy corresponding to each viewpoint, differentiable forward imaging is performed based on the scattering parameter set and the two-dimensional projection relationship to generate the two-dimensional predicted image corresponding to each viewpoint.

6. A control device, characterized in that, include: The first processing module is used to acquire observation geometric parameters and two-dimensional imaging results from multiple perspectives. The two-dimensional imaging results include two-dimensional SAR images of stationary targets acquired by a mobile platform from multiple perspectives and two-dimensional ISAR images of moving targets acquired by a fixed base station from multiple perspectives. The second processing module is used to construct the two-dimensional projection relationship corresponding to each viewpoint based on the observation geometric parameters, the obtained observation geometric parameters under multiple views, and the two-dimensional imaging results, and form an observation geometric model. The third processing module is used to obtain the scattering parameter set of sampling points in the three-dimensional sampling space corresponding to each viewpoint based on the implicit scattering field network constrained by geometric conditions and the two-dimensional projection relationship. The three-dimensional sampling space is constructed based on the observation geometric parameters and the projection relationship corresponding to each viewpoint. The fourth processing module is used to perform differentiable forward imaging based on the set of scattering parameters and the two-dimensional projection relationship, respectively, to generate two-dimensional prediction images corresponding to each viewpoint; The fifth processing module is used to construct a multi-view consistency loss based on the two-dimensional imaging results and the two-dimensional predicted image corresponding to each viewpoint. The sixth processing module is used to perform self-supervised iterative optimization of the implicit scattering field network based on the multi-view consistency loss until the preset iteration conditions are met, and to perform three-dimensional scattering reconstruction based on the scattering parameter set to obtain the three-dimensional scattering field reconstruction result.

7. An electronic device, characterized in that, include: A processor, a memory, and a program stored in the memory and executable on the processor, wherein the program, when executed by the processor, implements the steps of the self-supervised three-dimensional imaging method for multi-view SAR and ISAR oriented to arbitrary observation geometry as described in any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the self-supervised three-dimensional imaging method for multi-view SAR and ISAR oriented to arbitrary observation geometry as described in any one of claims 1 to 5.

9. A computer program product, characterized in that, It includes computer instructions that, when executed by a processor, implement the steps of the multi-view SAR and ISAR self-supervised three-dimensional imaging method for arbitrary observation geometry as described in any one of claims 1 to 5.

Images