An electric tower layer chromatography three-dimensional reconstruction method, device, terminal equipment and computer readable storage medium

Abstract

The application discloses a kind of electric tower chromatography three-dimensional reconstruction method, device, terminal equipment and computer readable storage medium, belong to power grid intelligent inspection technical field.Method includes: obtaining the SAR original image data of target electric tower to be reconstructed;It is input into the super-resolution reconstruction network of pre-training, and initial time sequence high-resolution SAR image data is obtained;Repeating execution joint optimization operation to generate final electric tower three-dimensional point cloud data and construct three-dimensional reconstruction model;Joint optimization operation includes: according to current high image data, generate three-dimensional point cloud, if not meet accuracy, it is projected to two-dimensional image domain to generate simulation data, calculate residual error and update network parameters, and original image is input into updated network again to generate the data required for next cycle.The application solves the problem that the existing method lacks three-dimensional to two-dimensional dynamic feedback verification, leading to error accumulation and difficulty in achieving high-precision modeling.

Description

Technical Field

[0001] This invention relates to the field of intelligent power grid inspection technology, and in particular to a method, apparatus, terminal equipment, and computer-readable storage medium for three-dimensional reconstruction of power towers by tomography. Background Technology

[0002] With the continuous expansion of my country's power grid, the safe and stable operation of transmission towers is crucial to ensuring the reliability of the entire power system. In recent years, Synthetic Aperture Radar (SAR) has demonstrated great potential in high-risk areas under complex weather conditions due to its all-weather, all-time, and cloud-penetrating advantages. In particular, TomoSAR, a tomographic imaging technique based on multi-baseline SAR imagery, can reconstruct the three-dimensional point cloud of power towers by inverting the elevation information of the target, becoming an important research direction in the field of power inspection. However, existing SAR tomographic tower reconstruction techniques still have many limitations: most existing conventional reconstruction methods are one-way open-loop processes of front-end image enhancement and back-end tomographic inversion. Due to the lack of an effective dynamic feedback verification mechanism from three-dimensional to two-dimensional, errors introduced in the front-end image processing stage will be uncontrollably propagated downstream and accumulated in the final three-dimensional reconstruction result. This error divergence problem caused by the lack of a closed-loop optimization mechanism seriously restricts the further improvement of the accuracy of power tower reconstruction and makes it difficult to meet the actual needs of high-precision engineering modeling. Summary of the Invention

[0003] This invention provides a method for three-dimensional reconstruction of power towers using tomography, which can solve the problem that conventional power tower reconstruction methods in the prior art cannot meet the requirements of high-precision engineering modeling.

[0004] An embodiment of the present invention provides a method for three-dimensional reconstruction of power towers by tomography, comprising: Acquire raw SAR image data of the target power tower to be reconstructed; The raw SAR image data is input into a pre-trained super-resolution reconstruction network for forward propagation to obtain initial temporal high-resolution SAR image data. Based on the initial time-series high-resolution SAR image data, the joint optimization operation is repeatedly performed to generate the final three-dimensional point cloud data of the power tower. Based on the final three-dimensional point cloud data of the power tower, the structural parameters of the target power tower to be reconstructed are extracted to obtain the three-dimensional reconstruction model of the target power tower to be reconstructed. The joint optimization operation includes: Based on the current temporal high-resolution SAR image data, generate the current three-dimensional point cloud data of the power tower; wherein, the temporal high-resolution SAR image data when the joint optimization operation is first performed is the initial temporal high-resolution SAR image data. Based on the current 3D point cloud data of the power tower, determine whether the super-resolution reconstruction network meets the preset accuracy requirements; If so, the current 3D point cloud data of the power tower is used as the final 3D point cloud data of the power tower; otherwise, the current 3D point cloud data of the power tower is projected onto the 2D SAR image domain to generate simulation data, the residual between the simulation data and the current time-series high-resolution SAR image data is calculated, and backpropagation is performed based on the residual to update the parameters of the super-resolution reconstruction network. The raw SAR image data is then input again into the updated super-resolution reconstruction network to generate the temporal high-resolution SAR image data required for the next joint optimization operation.

[0005] Furthermore, the super-resolution reconstruction network includes a low-resolution phase information feature processing branch and a low-resolution amplitude information feature processing branch; the low-resolution phase information feature processing branch includes: a recurrent convolutional layer, a gated recurrent unit, and a phase unwrapping layer connected in sequence; the low-resolution amplitude information feature processing branch includes: a multi-level stride convolutional layer and a deconvolutional layer connected in sequence. The step of inputting the raw SAR image data into a pre-trained super-resolution reconstruction network for forward propagation to obtain initial temporal high-resolution SAR image data includes: processing the phase information in the raw SAR image data using a recurrent convolutional layer based on the low-resolution phase information feature processing branch to obtain spatial long-range dependent phase features; processing the spatial long-range dependent phase features using a gated recurrent unit to obtain temporal phase change features; unwrapping and reconstructing the temporal phase change features using a phase unwrapping layer to generate a high-resolution phase image; compressing the amplitude information in the raw SAR image data using a multi-level stride convolutional layer based on the low-resolution amplitude information feature processing branch to obtain spatial abstract features; upsampling the spatial abstract features using a deconvolutional layer to generate a high-resolution amplitude image; and combining the high-resolution phase image and the high-resolution amplitude image to obtain the initial temporal high-resolution SAR image data.

[0006] Furthermore, the loss function used by the super-resolution reconstruction network during the pre-training phase is specifically as follows: ; in, For the total loss function, These are the regularization coefficients for each term; The The basic amplitude loss is as follows: Where T is the total number of time-series nodes, and t is the current time-series node; For high-resolution amplitude information obtained from network prediction, For true high-resolution amplitude information; The The phase continuity loss term is as follows: ,in, To obtain the high-resolution phase for prediction, This is the high-resolution phase predicted at the previous moment; The The phase consistency loss term is as follows: Where LPF(·) is the low-pass filter operation, The low-resolution phase is the input; The For symmetric regularization, specifically: ;in, (·) indicates a vertical mirror flip operation.

[0007] Further, generating the current 3D point cloud data of the power tower based on the current temporal high-resolution SAR image data includes: extracting a binary mask of the target power tower based on the current temporal high-resolution SAR image data; forcing the 3D scattering tensor of the background region to zero based on the binary mask, and mapping the current temporal high-resolution SAR image data to the 3D scattering tensor of the target region; solving the 3D scattering tensor of the target region under joint horizontal and vertical sparsity constraints to obtain the current 3D point cloud data of the power tower.

[0008] Furthermore, the step of solving the three-dimensional scattering tensor of the target region under the combined horizontal and vertical sparsity constraints to obtain the current three-dimensional point cloud data of the power tower specifically includes: constructing an optimization objective function with the three-dimensional scattering tensor as the solution variable; obtaining the optimal three-dimensional complex scattering coefficient distribution by solving the optimization objective function, and using the three-dimensional complex scattering coefficient distribution as the current three-dimensional point cloud data of the power tower; The specific optimization objective function is as follows: Y is a SAR three-dimensional observation matrix containing the current temporal high-resolution SAR image data. For the observation imaging operator, * denotes the mode product of the tensor and the matrix. It is the F-norm. Let S be the regularization coefficient, S be the three-dimensional scattering tensor to be solved, and M(x,y) be the extracted binary mask. It is a mixed norm; The mixing norm The specific calculation formula is as follows: ; Where x and y are the pixel indices in the azimuth and range directions, respectively. , , representing the total number of pixels in the azimuth and range directions, respectively; z is the elevation resolution meta-index. Let S be the total resolution element in the elevation direction, and |S(x,y,z)| represent taking the modulus of the three-dimensional scattering tensor at the spatial location (x,y,z).

[0009] Further, the step of projecting the current 3D point cloud data of the power tower onto the 2D SAR image domain to generate simulation data, calculating the residual between the simulation data and the current temporal high-resolution SAR image data, and performing backpropagation based on the residual to update the parameters of the super-resolution reconstruction network specifically includes: multiplying the 3D complex scattering coefficients at each spatial location in the current 3D point cloud data of the power tower with the baseline phase factor of the corresponding elevation dimension, summing and integrating the multiplication result along the elevation dimension to obtain the pixel value at the corresponding pixel location in the 2D SAR image domain, and constructing the simulation data from all the pixel values ​​at the pixel locations; calculating the degree of difference between the pixel values ​​at each pixel location in the current temporal high-resolution SAR image data and the pixel values ​​at the corresponding pixel locations in the simulation data, and obtaining the residual in the 2D SAR image domain; using the residual in the 2D SAR image domain as a supervision signal for physical consistency constraints, calculating the gradient of the residual with respect to the parameters of the super-resolution reconstruction network through the backpropagation algorithm, and updating the parameters of the super-resolution reconstruction network based on the gradient.

[0010] Further, the step of constructing a three-dimensional reconstruction model of the target power tower to be reconstructed based on the final three-dimensional point cloud data of the power tower includes: extracting point clouds of dense vertical regions from the final three-dimensional point cloud data of the power tower, fitting a cylindrical model to the point clouds of the dense vertical regions using a random sampling consensus algorithm, and obtaining the main pole structure parameters of the target power tower to be reconstructed based on the fitted cylindrical model. Horizontally distributed point cloud clusters are extracted from the final three-dimensional point cloud data of the power tower. Plane equations are used to fit the horizontally distributed point cloud clusters, and the crossarm structure parameters of the target power tower to be reconstructed are obtained based on the fitted plane equations. The conductor point cloud is extracted from the final three-dimensional point cloud data of the power tower, the conductor point cloud is fitted using the catenary equation, and the conductor structure parameters of the target power tower to be reconstructed are obtained based on the fitted catenary equation. The final three-dimensional point cloud data of the power tower is filtered based on a preset phase stability screening threshold to obtain stable scattering points; the stable scattering points are combined with the main pole structure parameters, the crossarm structure parameters and the conductor structure parameters to generate a three-dimensional reconstruction model of the target power tower to be reconstructed.

[0011] Another embodiment of the present invention provides a three-dimensional reconstruction device for power tower tomography, comprising: a data acquisition module, an initial reconstruction module, a model building module, and a joint optimization module; The system includes: a data acquisition module for acquiring raw SAR image data of the target power tower to be reconstructed; an initial reconstruction module for inputting the raw SAR image data into a pre-trained super-resolution reconstruction network for forward propagation to obtain initial temporal high-resolution SAR image data; a joint optimization module for repeatedly performing joint optimization operations based on the initial temporal high-resolution SAR image data to generate final 3D point cloud data of the power tower; and a model building module for extracting structural parameters of the target power tower to be reconstructed based on the final 3D point cloud data to obtain a 3D reconstruction model of the target power tower. The joint optimization operations include: Based on the current temporal high-resolution SAR image data, generate the current three-dimensional point cloud data of the power tower; wherein, the temporal high-resolution SAR image data when the joint optimization operation is first performed is the initial temporal high-resolution SAR image data. Based on the current 3D point cloud data of the power tower, determine whether the super-resolution reconstruction network meets the preset accuracy requirements; If so, the current 3D point cloud data of the power tower is used as the final 3D point cloud data of the power tower; otherwise, the current 3D point cloud data of the power tower is projected onto the 2D SAR image domain to generate simulation data, the residual between the simulation data and the current time-series high-resolution SAR image data is calculated, and backpropagation is performed based on the residual to update the parameters of the super-resolution reconstruction network. The raw SAR image data is then input again into the updated super-resolution reconstruction network to generate the temporal high-resolution SAR image data required for the next joint optimization operation.

[0012] Another embodiment of the present invention provides a terminal device, including: a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein when the processor executes the computer program, it implements the three-dimensional reconstruction method for power tower tomography of the present invention.

[0013] Another embodiment of the present invention provides a computer-readable storage medium item, including: a stored computer program, wherein, when the computer program is running, it controls the device where the computer-readable storage medium is located to execute the present invention's method for three-dimensional reconstruction of power tower tomography.

[0014] The embodiments of the present invention have the following beneficial effects: This invention proposes a method for 3D reconstruction of power towers using tomography. First, raw SAR image data of the target power tower to be reconstructed is acquired. Then, the raw SAR image data is input into a pre-trained super-resolution reconstruction network for forward propagation to obtain initial temporal high-resolution SAR image data. Subsequently, a joint optimization operation is repeatedly performed based on this initial data to generate the final 3D point cloud data of the power tower. Specifically, in the joint optimization operation, the current 3D point cloud data of the power tower is generated based on the current temporal high-resolution SAR image data. If the preset accuracy requirement is not met, the current 3D point cloud data of the power tower is projected onto a 2D SAR image domain to generate simulation data. The residual between the simulation data and the current temporal high-resolution SAR image data is calculated and backpropagated to update the parameters of the super-resolution reconstruction network. The raw SAR image data is then input again into the updated network to generate the data required for the next operation, until the accuracy requirement is met. Finally, structural parameters are extracted from the final 3D point cloud data of the power tower to obtain a 3D reconstruction model. This application obtains initial high-resolution data by inputting raw image data into a pre-trained network, and then extracts target features through subsequent iterative loops, which can provide a clearer and more reliable basic data source for power tower tomographic reconstruction. Specifically, this application constructs a joint optimization closed-loop mechanism based on 3D feedback to 2D: when performing repeated joint optimization operations, not only is a 3D point cloud generated based on the current high-resolution data, but more importantly, the simulation data is generated by forcibly projecting the current 3D point cloud back into the 2D SAR image domain, and the parameters of the super-resolution network are updated by backpropagation of the residual between the simulation data and the high-resolution image. This design based on 3D projection to verify the 2D network can effectively solve the problem that existing conventional reconstruction methods rely on a one-way open-loop process of front-end image enhancement and back-end tomographic inversion, and lack a dynamic feedback verification mechanism from 3D to 2D, resulting in the uncontrolled accumulation and divergence of front-end errors downstream. Furthermore, by organically integrating the super-resolution reconstruction network and 3D point cloud generation in the iterative closed loop, this application ensures that each update of the network parameters can be continuously verified and dynamically corrected for 3D physical authenticity, realizing the mutual promotion of 2D image enhancement and 3D structure inversion, and significantly improving the accuracy and reliability of high-precision 3D modeling of power towers in engineering scenarios. Attached Figure Description

[0015] To more clearly illustrate the technical solution of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0016] Figure 1 This is a flowchart illustrating a three-dimensional reconstruction method for power tower tomography provided in an embodiment of the present invention.

[0017] Figure 2 This is a schematic diagram of the apparatus for a three-dimensional reconstruction method for electric tower tomography provided in an embodiment of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0019] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains; the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the application; the terms “comprising” and “having”, and any variations thereof, in the specification, claims, and foregoing description of the drawings are intended to cover non-exclusive inclusion.

[0020] In the description of the embodiments of this application, technical terms such as "first" and "second" are used only to distinguish different objects and should not be construed as indicating or implying relative importance or implicitly specifying the number, specific order, or primary and secondary relationship of the indicated technical features. In the description of the embodiments of this application, "multiple" means two or more, unless otherwise explicitly defined.

[0021] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0022] In the description of the embodiments in this application, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this document generally indicates that the preceding and following related objects have an "or" relationship.

[0023] In the description of the embodiments of this application, the term "multiple" refers to two or more (including two), similarly, "multiple sets" refers to two or more (including two sets), and "multiple pieces" refers to two or more (including two pieces).

[0024] In the description of the embodiments of this application, unless otherwise expressly specified and limited, technical terms such as "installation," "connection," "joining," and "fixing" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. For those skilled in the art, the specific meaning of the above terms in the embodiments of this application can be understood according to the specific circumstances.

[0025] To address the problem that conventional power tower reconstruction methods in the prior art cannot meet the requirements of high-precision engineering modeling, an embodiment of the present invention provides a three-dimensional reconstruction method for power tower tomography, comprising: Step S1: Obtain the raw SAR image data of the target power tower to be reconstructed; In a preferred embodiment, acquiring the raw SAR image data of the target power tower to be reconstructed specifically includes: acquiring multiple time-series SAR complex image raw data covering the target power tower area by having a spaceborne synthetic aperture radar pass over the area at different times; based on this, to ensure the accuracy of subsequent time-series super-resolution and tomographic reconstruction, the acquired multiple time-series SAR image raw data needs to be precisely preprocessed to obtain aligned and clean time-series raw resolution SAR images of the target area, specifically including: performing high-precision registration of time-series images based on the strong scattering physical characteristics of the power tower in the SAR image; using a nonlocal mean algorithm combined with polarization features to suppress background clutter in the registered image; performing atmospheric phase correction on the denoised image based on a spatiotemporal filtering strategy, and outputting the processed time-series raw resolution SAR image.

[0026] Preferably, high-precision registration of time-series images is performed based on the strong scattering characteristics of power towers through the following steps: First, considering that power towers are typical man-made metal frame structures and exhibit high-brightness strong scattering characteristics in SAR images, a set of strong scattering points located in the target area of ​​the power tower is initially screened by setting a high-brightness pixel amplitude threshold for the SAR image. Then, the amplitude standard deviation of each scattering point in this set is calculated in the multi-temporal image sequence, points severely affected by the environment are removed, and points with amplitude changes below the set standard deviation threshold are retained as candidate control points.

[0027] Next, with each candidate control point as the center, the complex coherence coefficients of the main and auxiliary temporal images are calculated within a set neighborhood window (e.g., a 32×32 pixel window). The calculation formula is as follows: in, and These represent the complex pixel values ​​at corresponding positions in the main and auxiliary temporal images, respectively. For complex conjugate, i is the pixel index within the corresponding window. After calculation using the above formula, regions with severe temporal decoherence are removed, and high-stability points with complex coherence coefficients greater than a set threshold (preferably 0.7 in this embodiment) are strictly retained. Finally, an affine transformation model is constructed based on the obtained high-stability control points, and its basic form is as follows: in, For the spatial coordinates of the reference main image, The mapping coordinates of the auxiliary image to be registered are: and These represent the horizontal and vertical translations (i.e., offset parameters) of the spatial coordinate system. At least three pairs (preferably dozens of evenly distributed pairs in practical engineering) of control points are selected from the image to be registered and the reference image. The affine transformation matrix and offset matrix are calculated using the least squares method to determine the coefficients of the affine transformation model. Based on the obtained transformation matrix, the image to be registered is resampled, retaining the true phase information of the complex data, thus completing sub-pixel-level registration between time-series SAR images.

[0028] Preferably, background clutter suppression is achieved by combining nonlocal means with polarization features through the following steps: To preserve the delicate, perforated structure of the power tower to the greatest extent while smoothing noise, this embodiment employs a nonlocal means (NLM) filtering method that combines polarization information. The core idea of ​​the traditional NLM method is to utilize the nonlocal self-similarity present in the image, that is, the value of any pixel in the image can be obtained by weighted averaging of all pixels in the entire image that have a similar structure, and the initial basic weight between two neighboring pixels is used. The calculation is as follows: In the formula, represents the Euclidean distance between the image blocks containing pixels i and j, and h is an attenuation parameter controlling the smoothing effect of the filter. Specifically, background clutter such as vegetation and ground surfaces in the natural environment exhibits high randomness and strong anisotropy in microwave scattering; while artificial metal targets such as power towers typically show a strong secondary angular reflection effect in the cross-polarization channel (HV / VH) and possess the significant physical characteristic of extremely low polarization scattering entropy. Therefore, this embodiment extracts the polarization features of each pixel (including scattering entropy, anisotropy, and polarization ratio), calculates the Euclidean distance of polarization features between pixels, and superimposes it as a penalty term into the aforementioned basic formula. That is, when a significant difference is found between the polarization features of the target pixel (low-entropy metal point) and its neighboring pixels (high-entropy vegetation point), its smoothing weight is dynamically reduced, thereby assisting the NLM algorithm in more accurately allocating the weight matrix. This nonlocal reconstruction process of cascaded polarization features can effectively suppress clutter interference from the surrounding environment, thereby significantly enhancing the overall signal-to-noise ratio of the power tower area.

[0029] Preferably, atmospheric phase correction is performed through the following steps to obtain clean temporal phase data: When multi-phase SAR radar microwaves pass through the atmosphere, atmospheric phase delay (atmospheric phase screen) is generated due to changes in water vapor and the ionosphere. This embodiment separates and extracts the target phase based on the physical assumption that the geometric phase of the tower target exhibits high-frequency abrupt changes in space, while the atmospheric disturbance phase exhibits low-frequency smooth and gradual changes in space and high-frequency random characteristics in the temporal domain. Specifically, for the highly stable control points remaining after registration, their interferometric phase difference sequences are extracted. First, a high-pass filtering strategy is used in the time domain to remove the long-term trend phase caused by the slow deformation of the tower itself. Then, in the spatial dimension, a two-dimensional low-pass filtering or Kriging spatial interpolation algorithm is used to fit the residual phase after the high-pass filtering in the time domain, estimating a low-frequency atmospheric phase screen surface model covering the entire observation area. Finally, the corresponding atmospheric phase estimation value is subtracted pixel-by-pixel from the original phase of each registered complex image, completing the complete decoupling of the target phase and the atmospheric disturbance phase.

[0030] In this embodiment, by employing the data acquisition and preprocessing process tailored to the specific physical properties of the power tower target (strong scattering, low polarization entropy), precise spatial alignment of multi-phase complex SAR images at the sub-pixel level is achieved. Simultaneously, radar polarization parameters are cleverly integrated into the non-local mean filtering distance metric, eliminating interference from natural background noise such as vegetation and surface clutter at the source, highlighting the effective scattering of weak signals from the hollowed-out power tower. Furthermore, spatiotemporal separation correction addresses phase contamination caused by atmospheric convection, maximizing the restoration of the true radar echo geometric phase. This significantly improves the signal-to-noise ratio of the acquired time-series raw resolution SAR images, ensuring high stability and continuity of the phase over time. It smoothly eliminates data jumps caused by various external interferences, providing a high-confidence, physically pure, high-quality two-dimensional data benchmark for subsequent deep learning super-resolution networks (PG-SRNet). This effectively avoids the risk of deep learning getting trapped in local optima or producing artifacts during subsequent iterative optimization due to geometric deviations or phase contamination in the initial input data.

[0031] Step S2: Input the raw SAR image data into a pre-trained super-resolution reconstruction network for forward propagation to obtain initial temporal high-resolution SAR image data; In a preferred embodiment, the super-resolution reconstruction network includes a low-resolution phase information feature processing branch and a low-resolution amplitude information feature processing branch; the low-resolution phase information feature processing branch includes: a recurrent convolutional layer, a gated recurrent unit, and a phase unwrapping layer connected in sequence; the low-resolution amplitude information feature processing branch includes: a multi-level stride convolutional layer and a deconvolutional layer connected in sequence. The step of inputting the raw SAR image data into a pre-trained super-resolution reconstruction network for forward propagation to obtain initial temporal high-resolution SAR image data includes: processing the phase information in the raw SAR image data using a recurrent convolutional layer based on the low-resolution phase information feature processing branch to obtain spatial long-range dependent phase features; processing the spatial long-range dependent phase features using a gated recurrent unit to obtain temporal phase change features; unwrapping and reconstructing the temporal phase change features using a phase unwrapping layer to generate a high-resolution phase image; compressing the amplitude information in the raw SAR image data using a multi-level stride convolutional layer based on the low-resolution amplitude information feature processing branch to obtain spatial abstract features; upsampling the spatial abstract features using a deconvolutional layer to generate a high-resolution amplitude image; and combining the high-resolution phase image and the high-resolution amplitude image to obtain the initial temporal high-resolution SAR image data.

[0032] Preferably, the loss function used by the super-resolution reconstruction network during the pre-training phase is as follows: ; in, For the total loss function, These are the regularization coefficients for each term; The The basic amplitude loss is as follows: Where T is the total number of time-series nodes, and t is the current time-series node; For high-resolution amplitude information obtained from network prediction, For true high-resolution amplitude information; The The phase continuity loss term is as follows: ,in, To obtain the high-resolution phase for prediction, This is the high-resolution phase predicted at the previous moment; The The phase consistency loss term is as follows: Where LPF(·) is the low-pass filter operation, The low-resolution phase is the input; The For symmetric regularization, specifically: ;in, (·) indicates a vertical mirror flip operation.

[0033] Specifically, in order to achieve efficient feature extraction and physical constraints in the super-resolution reconstruction process described above, this embodiment provides in-depth constraints on the network's internal operating mechanism and the physical meaning of the loss function: For the low-resolution phase information feature processing branch, considering that the phase information in SAR images is directly related to surface deformation and the three-dimensional elevation geometry of the target, and that the phase of the original radar echo usually exhibits a periodic entanglement state in the range of [−π,π], which is very easy to generate artificial truncation edges, this embodiment first uses a recurrent convolutional layer (RCL) to process the input low-resolution phase image. By introducing a recurrent feedback connection, the network can not only focus on local features, but also capture the long-range dependence of phase features in the spatial dimension (such as the gradual gradient of the phase at the upper and lower ends of the main pole of the power tower). Subsequently, a gated recurrent unit (GRU) is used to perform temporal modeling on the multi-temporal SAR sequence to learn the small evolution law of the power tower phase over time (such as the small temporal deformation caused by thermal expansion and contraction or wind load). Finally, by embedding a phase unwrapping layer, the restricted entangled phase is mapped and unfolded into a real, continuous absolute phase value, eliminating the interference caused by phase jumps to the neural network convolution operation, thereby outputting a high-quality, entangle-free, high spatial resolution phase image through phase reconstruction.

[0034] For the low-resolution amplitude information feature processing branch, since the SAR amplitude image reflects the electromagnetic scattering intensity characteristics of the target, the hollow metal structure of the power tower (such as the tower material angle steel) will appear as a complex alternating texture of strong and weak in the amplitude image. This embodiment designs a downsampling module composed of multiple multi-level stride (e.g., the preferred stride=2) convolutional layers, which can quickly compress the spatial size of the feature map, while multiplying the receptive field of the neurons and extracting deeper and more abstract global topological features of the power tower. In the feature decoding stage, a series of deconvolutional layers are used for progressive upsampling operations. During this process, a cascaded batch normalization layer (BN) and a nonlinear activation function layer (ReLU) are preferred. The BN layer effectively suppresses the internal covariate shift of the deep layers of the network and accelerates the model convergence. The ReLU layer improves the network's nonlinear representation and recovery ability of local high-frequency scattering bright spots of the power tower (such as insulators and crossarm connections), and finally accurately restores the high-resolution amplitude image.

[0035] It should be noted that the core of the pre-training stage described in this embodiment is not to blindly pursue the visual smoothness of the image, but rather to achieve this through the combination of the above four loss functions ( The network output is forced to conform to the actual spatial and physical laws of power towers. The operating mechanism of each constraint is as follows: The base amplitude loss uses the L1 norm (absolute error) to preserve the sharpness of strong scattering points (high-frequency edges) in the amplitude image, avoiding the excessive penalty of extreme values ​​caused by using the L2 norm, which would lead to smoothing and blurring of the image. The phase continuity loss term applies a time consistency constraint by calculating the high-resolution phase difference between the current time t and the previous time t−1, ensuring that the power tower, as a rigid infrastructure, can conform to the objective physical law that deformation is continuous within adjacent short observation periods and that there will be no drastic phase jumps. The phase consistency loss term is introduced by low-pass filtering (LPF). Its physical meaning is that the essential task of the super-resolution network is to supplement high-frequency details and ensure that the macroscopic low-frequency base structure of the predicted high-resolution phase must be highly consistent with the low-resolution phase of the original input of the system after the high-frequency components are filtered out by low-pass filtering. This prevents the network from excessively creating illusions that change the basic elevation geometric benchmark of the original data. The symmetry regularization term is a key structural prior constraint. Considering that transmission towers generally have a vertically symmetrical geometry along their central axis in three-dimensional space, this embodiment uses... (·) The vertical mirror flip operation forces the verification of the left-right symmetry of the amplitude feature matrix during the training phase. Once the network generates false textures on one side, this regularization term will generate a huge penalty gradient, thereby guiding the entire super-resolution model to converge in the solution space in the direction that satisfies the symmetric physical topology of the electric tower.

[0036] In this embodiment, the dual-branch network architecture tailored to the characteristics of SAR complex data achieves decoupling and targeted enhancement of temporal phase variation and spatial amplitude texture. More importantly, this application abandons the traditional deep learning black-box fitting mode that relies solely on data-driven approaches and innovatively establishes a physical-guided prior mechanism composed of vertical symmetry, phase temporal continuity, and frequency band consistency. This mechanism, through a targeted multi-factor joint loss function, directly injects objective electromagnetic scattering and tower geometric topology during the network training phase. This enables the super-resolution reconstruction network to not only avoid generating false artifact textures when facing strong background clutter or weak texture regions, but also adaptively utilize symmetric priors and temporal correlations to complete the missing small crossarm and insulator structures. The resulting initial temporal high-resolution SAR image data achieves a substantial breakthrough in resolution and strictly preserves physical properties, thus providing a high-dimensional input foundation with rich details, continuous structure, and high physical reliability for subsequent sparse tensor decomposition tomography.

[0037] Step S3: Based on the initial temporal high-resolution SAR image data, repeatedly perform the joint optimization operation to generate the final 3D point cloud data of the power tower; wherein, the joint optimization operation includes: Based on the current temporal high-resolution SAR image data, generate the current three-dimensional point cloud data of the power tower; wherein, the temporal high-resolution SAR image data when the joint optimization operation is first performed is the initial temporal high-resolution SAR image data. Based on the current 3D point cloud data of the power tower, determine whether the super-resolution reconstruction network meets the preset accuracy requirements; If so, the current 3D point cloud data of the power tower is used as the final 3D point cloud data of the power tower; otherwise, the current 3D point cloud data of the power tower is projected onto the 2D SAR image domain to generate simulation data, the residual between the simulation data and the current time-series high-resolution SAR image data is calculated, and backpropagation is performed based on the residual to update the parameters of the super-resolution reconstruction network. The raw SAR image data is then input again into the updated super-resolution reconstruction network to generate the temporal high-resolution SAR image data required for the next joint optimization operation.

[0038] In a preferred embodiment, generating the current 3D point cloud data of the power tower based on the current temporal high-resolution SAR image data includes: extracting a binary mask of the target power tower based on the current temporal high-resolution SAR image data; forcing the 3D scattering tensor of the background region to zero based on the binary mask, and mapping the current temporal high-resolution SAR image data to the 3D scattering tensor of the target region; and solving the 3D scattering tensor of the target region under joint horizontal and vertical sparsity constraints to obtain the current 3D point cloud data of the power tower.

[0039] Preferably, the step of solving the three-dimensional scattering tensor of the target region under the combined horizontal and vertical sparsity constraints to obtain the current three-dimensional point cloud data of the power tower specifically includes: constructing an optimization objective function with the three-dimensional scattering tensor as the solution variable; obtaining the optimal three-dimensional complex scattering coefficient distribution by solving the optimization objective function, and using the three-dimensional complex scattering coefficient distribution as the current three-dimensional point cloud data of the power tower; The specific optimization objective function is as follows: Y is a SAR three-dimensional observation matrix containing the current temporal high-resolution SAR image data. For the observation imaging operator, * denotes the mode product of the tensor and the matrix. It is the F-norm. Let S be the regularization coefficient, S be the three-dimensional scattering tensor to be solved, and M(x,y) be the extracted binary mask. It is a mixed norm; The mixing norm The specific calculation formula is as follows: ; Where x and y are the pixel indices in the azimuth and range directions, respectively. , , representing the total number of pixels in the azimuth and range directions, respectively; z is the elevation resolution meta-index. Let S be the total resolution element in the elevation direction, and |S(x,y,z)| represent taking the modulus of the three-dimensional scattering tensor at the spatial location (x,y,z).

[0040] Preferably, the step of projecting the current 3D point cloud data of the power tower onto the 2D SAR image domain to generate simulation data, calculating the residual between the simulation data and the current temporal high-resolution SAR image data, and performing backpropagation based on the residual to update the parameters of the super-resolution reconstruction network specifically includes: multiplying the 3D complex scattering coefficients at each spatial location in the current 3D point cloud data of the power tower with the baseline phase factor of the corresponding elevation dimension, summing and integrating the multiplication result along the elevation dimension to obtain the pixel value at the corresponding pixel location in the 2D SAR image domain, and constructing the simulation data from all the pixel values ​​at the pixel locations; calculating the degree of difference between the pixel values ​​at each pixel location in the current temporal high-resolution SAR image data and the pixel values ​​at the corresponding pixel locations in the simulation data, and obtaining the residual in the 2D SAR image domain; using the residual in the 2D SAR image domain as a supervision signal for physical consistency constraints, calculating the gradient of the residual with respect to the parameters of the super-resolution reconstruction network through the backpropagation algorithm, and updating the parameters of the super-resolution reconstruction network based on the gradient.

[0041] Preferably, in the above joint optimization operation, the core purpose of extracting the binary mask M(x,y) of the target power tower and the background forced zeroing constraint mechanism is to reduce the solution space of the tomographic inversion and eliminate false scattering interference from complex terrain backgrounds. Since high-voltage transmission towers are usually erected in complex natural backgrounds (such as mountains, forests, or farmland), the volume scattering and surface scattering of the background region will severely overlap with the multiple scattering signals of the power tower under radar side-view imaging geometry. In this embodiment, the binary mask is dynamically generated by the amplitude information after super-resolution reconstruction. At the mathematical level, the boundary constraint condition (i.e., st S(x,y,z)=0 for M(x,y)=0) is explicitly added, which forcibly cuts off the possibility of the algorithm allocating electromagnetic energy in areas without power towers. This not only greatly reduces the computational complexity of three-dimensional tensor reconstruction, but also suppresses the false point cloud caused by the projection of background noise in the vertical elevation direction from the source.

[0042] Preferably, for the constructed sparse tensor decomposition optimization objective function, its first term To ensure data fidelity, considering that the thermal noise and environmental clutter of the radar system typically follow a complex Gaussian distribution, this embodiment uses the Frobenius norm (F-norm) to accurately measure the overall fitting error between the observation matrix and the synthetic observation matrix. This ensures that the three-dimensional complex scattering coefficients obtained by inversion, when projected back into the observation space, must infinitely approximate the multi-baseline echo signal actually received by the radar antenna, thus guaranteeing the physical fidelity of the reconstruction results.

[0043] Specifically, optimize the second term in the objective function. The mixed norm regularization term is the core technical feature of this embodiment that overcomes the bottleneck of structural fracture caused by the sparse vertical baseline in traditional TomoSAR imaging algorithms. Traditional compressed sensing (CS) tomography methods usually treat each pixel as an isolated one-dimensional elevation vector, relying solely on L1 norm optimization to obtain absolute sparsity at the point level. Even if the inverse problem of insufficient baseline can be solved, this operation disrupts the continuous topological relationship of the power tower as a rigid physical structure. This embodiment innovatively extends the scattering distribution into a three-dimensional tensor and applies L2,1 mixed norm constraints. The specific calculation formula is as follows: The internal L2 norm (i.e., the square root of the sum of squares) acts on the horizontal plane (azimuth x and range y), its physical meaning being to apply a group lasso effect to all horizontal pixels within the same elevation layer z, encouraging the metal frame of the power tower to maintain smooth continuity in the horizontal direction and avoiding isolated discontinuities. The external L1 norm (i.e., the sum of the absolute values ​​over the elevation z) acts on the vertical direction, its physical meaning being to force sparse selection characteristics between different elevation layers, ensuring that the scattered energy of the power tower is concentrated only at a few specific height layers (such as the main crossarm and ground wire support), while the energy of most suspended height layers is strictly compressed to zero. This dual physical constraint of horizontal continuity and vertical dispersion perfectly matches the actual three-dimensional grid structure properties of the power tower.

[0044] Furthermore, the joint iterative optimization process in this embodiment overturns the traditional open-loop reconstruction paradigm. In the step of projecting the current 3D point cloud data of the power tower onto the 2D SAR image domain to generate simulation data, its essence is a forward mathematical reproduction of the radar physical imaging mechanism. Since the 3D point cloud carries the accurate elevation layer z and the 3D complex scattering intensity S(x,y,z), when the 3D point cloud is multiplied by the phase factor of the corresponding baseline and integrated along the elevation, the generated simulation data actually represents a 2D ideal image that a 3D power tower should present from the radar sensor perspective, which absolutely conforms to the laws of physical space. When the system calculates the difference between the simulation data and the time-series high-resolution image currently output by the super-resolution network, the obtained 2D SAR image domain residual is no longer the fitting error generated by manual labeling in traditional deep learning, but a supervision signal of physical consistency constraint. This physical residual is transformed into a gradient through the backpropagation algorithm, and the weight parameters of the super-resolution reconstruction network (PG-SRNet) are updated accordingly. This is equivalent to using a rigorous radar 3D electromagnetic scattering model to perform a powerful dynamic correction of the 2D illusion of the neural network. During the next forward propagation, the network will be forced to generate more high-frequency image details that are more consistent with the physical reality of three-dimensional objects, until the two-dimensional image generated by the network can perfectly reflect a consistent three-dimensional point cloud (i.e., meet the preset accuracy stopping condition).

[0045] Furthermore, in order to ensure the efficient convergence of the above joint iterative optimization process and prevent the ineffective consumption of computing resources, this embodiment clearly defines the judgment mechanism for determining whether the super-resolution reconstruction network meets the preset accuracy requirements (or accuracy stopping conditions). Specifically, the preset accuracy requirement preferably includes one of the following judgment conditions or a combination of the following judgment conditions: Condition 1 (Two-dimensional residual convergence judgment): Calculate the sum of absolute values ​​of residuals or root mean square error (RMSE) in the two-dimensional SAR image domain under the current loop. When the rate of change of the residual error for K consecutive iterations (e.g., K=3) is less than the preset minimum residual threshold ϵ, it indicates that the image generated by the front-end two-dimensional super-resolution network and the physical projection generated by the back-end three-dimensional inversion have reached a high degree of consistency, and the network parameters have converged.

[0046] Condition 2 (3D Structural Stability Determination): Compare the current 3D point cloud data of the power tower generated in the current loop with the 3D point cloud data generated in the previous loop, and calculate the norm difference between the two 3D complex scattering coefficient distribution matrices. When this norm difference is less than the preset structural stability threshold δ, it indicates that the result of the 3D tensor tomography solution has no longer undergone substantial changes, and the spatial topology of the power tower has reached a stable state.

[0047] Condition 3 (Maximum Iteration Count Protection): To prevent the model from falling into an infinite loop due to oscillations in extremely noisy environments, the system presets a maximum number of iterations Nmax (e.g., Nmax = 50). When the number of times the joint optimization operation is executed reaches this limit, the system forcibly stops the iteration.

[0048] It is worth noting that in practical engineering applications, this embodiment preferably uses the AND logic of two-dimensional residual convergence determination and three-dimensional structural stability determination as the primary stopping condition. Through this dual-check accuracy determination mechanism, the iteration is considered successful only when the two-dimensional image no longer generates new false illusory textures and the geometric position and scattering energy of the point cloud in three-dimensional space no longer drift. Once the above accuracy requirements are met, the system immediately triggers a stop command to exit the loop and outputs the three-dimensional point cloud of the last iteration as the final three-dimensional point cloud data of the power tower downstream (step S4). By explicitly specifying this accuracy stopping condition, the risk of overfitting in the unsupervised test-time optimization (TTO) process of deep learning is eliminated, and the robustness and operational efficiency of the entire power tower three-dimensional reconstruction system are guaranteed from the perspective of algorithm engineering.

[0049] In this embodiment, the three-dimensional scattering tensor solution model constrained by the L2,1 mixing norm is constructed to accurately characterize the inherent geometric topology of the horizontal grid of the power tower, which is coherent and vertically layered and sparse. This overcomes the shortcomings of traditional isolated pixel inversion, which is prone to structural breaks and isolated noise in complex baseline environments, and ensures the structural self-consistency of the generated point cloud. At the same time, by introducing a test-time joint iterative optimization mechanism based on three-dimensional forward projection, the tensor inversion physical model is deeply coupled with a deep learning super-resolution network. The generated physical simulation residuals are used as self-supervised signals to guide the on-site update of network parameters. The resulting three-dimensional point cloud data of the power tower not only breaks through the physical diffraction limit of the original radar hardware in terms of spatial resolution, but also eliminates the risk of disordered accumulation of preprocessing errors in the one-way open-loop process. This provides a core three-dimensional data source with complete structure, physical reliability, and accuracy that meets engineering measurement requirements for the subsequent extraction of accurate structural parameters of key components of the power tower (main pole, crossarm, conductor) and the generation of high-fidelity three-dimensional reconstruction models.

[0050] Step S4: Construct a three-dimensional reconstruction model of the target power tower to be reconstructed based on the final three-dimensional point cloud data of the power tower; In a preferred embodiment, the step of constructing a three-dimensional reconstruction model of the target power tower based on the final three-dimensional point cloud data of the power tower includes: extracting point clouds of dense vertical regions from the final three-dimensional point cloud data of the power tower, fitting a cylindrical model to the point clouds of the dense vertical regions using a random sampling consensus algorithm, and obtaining the main pole structure parameters of the target power tower to be reconstructed based on the fitted cylindrical model. Horizontally distributed point cloud clusters are extracted from the final three-dimensional point cloud data of the power tower. Plane equations are used to fit the horizontally distributed point cloud clusters, and the crossarm structure parameters of the target power tower to be reconstructed are obtained based on the fitted plane equations. The conductor point cloud is extracted from the final three-dimensional point cloud data of the power tower, the conductor point cloud is fitted using the catenary equation, and the conductor structure parameters of the target power tower to be reconstructed are obtained based on the fitted catenary equation. The final three-dimensional point cloud data of the power tower is filtered based on a preset phase stability screening threshold to obtain stable scattering points; the stable scattering points are combined with the main pole structure parameters, the crossarm structure parameters and the conductor structure parameters to generate a three-dimensional reconstruction model of the target power tower to be reconstructed.

[0051] Preferably, after obtaining the final high-precision 3D point cloud data of the power tower, since the point cloud is essentially still a discrete set of spatial coordinates, in order to transform it into a physical engineering model that can be directly used by power grid operation and maintenance personnel for safety assessment, this embodiment designs a corresponding structural parameter extraction and fitting mechanism based on the geometric topological characteristics of different components of the power tower: For the main pole structure of the power tower, considering the typical vertical continuous extension and dense point cloud distribution of the main pole in space, this embodiment preferably uses the Random Sampling Consensus Algorithm (RANSAC) to fit the cylindrical model. Since a small number of outliers inevitably exist in the SAR 3D point cloud due to multipath effects, the RANSAC algorithm can effectively remove these outliers through iterative random sampling. The specific optimization objective is to minimize the sum of the vertical distances from all valid point clouds to the cylindrical surface. The radius r of the cylinder, the coordinates c of the cylinder's axis starting point, and the axis direction vector a are determined by solving the objective function. Using the extracted axis direction vector a and radius r, not only can the spatial skeleton of the main pole be accurately reconstructed, but the tilt rate and verticality deviation of the main pole can also be directly calculated, providing direct data support for tower tilt deformation monitoring.

[0052] For the crossarm structure of an electric tower, it typically appears in space as a cluster of horizontal or inclined planar points extending outward from the main pole. This embodiment separates the crossarm point cloud clusters using region growing or density clustering algorithms and then constructs a spatial planar equation. Perform fitting. Minimize the point cloud samples. The sum of squared orthogonal distances to the fitted plane is used to accurately solve for the coefficients a, b, c, d of the plane equation using principal component analysis (PCA) or least squares method. The resulting plane normal vector and spatial intercept can accurately reconstruct the absolute elevation, spatial orientation, and hanging point position of each layer of the crossarm of the power tower.

[0053] Regarding the conductor structure, since high-voltage transmission conductors sag naturally between two towers due to their own gravity, they strictly follow a catenary morphology in the real physical world. Therefore, in this embodiment, after segmenting the linear point cloud trajectory from the point cloud, the catenary equation is used. Nonlinear curve fitting is performed on the traverse point cloud. Specifically, a nonlinear least squares optimization algorithm (such as the Levenberg-Marquardt algorithm) is used to iteratively solve for parameter 'a' (which determines the degree of sag) and reference height offset parameter 'b' (which determines the absolute elevation position of the traverse), representing the traverse sag control. Successful extraction of these parameters allows the system to directly quantify and assess whether the traverse sag exceeds a safe threshold and determine its safe clearance distance from vegetation or buildings below.

[0054] Furthermore, to further ensure the high fidelity of the final 3D reconstruction model, this embodiment, in addition to parametric modeling, introduces a permanent scatterer (PS) point filtering mechanism based on temporal phase stability. Specifically, the system traverses the final 3D point cloud data, calculates the time-averaged amplitude of the complex scattering coefficient at each spatial location within a time period T, and retains only highly stable scattering points with a time-averaged amplitude greater than the threshold by setting a strict phase stability filtering threshold. The physical significance of this mechanism is to efficiently filter out occasional flocks of birds flying over the power tower, leaves around branches swaying violently in the wind, and random clutter points caused by transient atmospheric disturbances, while retaining rigid structural points with extremely stable electromagnetic scattering characteristics, such as the metal tower body and insulator strings.

[0055] Finally, the system rigidly aligns and combines the highly purified stable PS discrete point cloud with the main cylinder, crossarm plane, and catenary generated by the above parameterized fitting in spatial coordinate system, thereby generating a hybrid (parameter + discrete) three-dimensional reconstruction model of the target power tower to be reconstructed that contains both macroscopically accurate geometric parameters and retains microscopic real scattering details.

[0056] In this embodiment, by employing the geometric parameter fitting strategies set separately for each heterogeneous component of the power tower (main pole, crossarm, and conductor), a leap from irregular discrete 3D point clouds to a high-level physical entity model with clear engineering semantics is achieved. Furthermore, by combining the RANSAC algorithm with a time-series PS point threshold screening mechanism, the inherent local multipath noise interference of radar point clouds is fundamentally overcome, ensuring that even under extremely harsh outdoor electromagnetic and meteorological environments, geometric parameters representing the true attitude of the power tower can be robustly extracted. The resulting 3D reconstruction model of the power tower not only possesses extremely high spatial visual fidelity, but more importantly, its output quantitative indicators such as main pole tilt and conductor sag can be directly and seamlessly integrated into the power grid's safety assessment and disaster early warning system. This solves the practical application problem of moving from SAR remote sensing tomography technology to the engineering implementation of power line inspection, significantly improving the efficiency and accuracy of all-weather hazard investigation of transmission channels in complex scenarios such as typhoons and geological disasters.

[0057] like Figure 2 As shown, another embodiment of the present invention also provides a three-dimensional reconstruction device for power tower tomography, including: a data acquisition module, an initial reconstruction module, a model building module, and a joint optimization module; The system includes: a data acquisition module for acquiring raw SAR image data of the target power tower to be reconstructed; an initial reconstruction module for inputting the raw SAR image data into a pre-trained super-resolution reconstruction network for forward propagation to obtain initial temporal high-resolution SAR image data; a joint optimization module for repeatedly performing joint optimization operations based on the initial temporal high-resolution SAR image data to generate final 3D point cloud data of the power tower; and a model building module for extracting structural parameters of the target power tower to be reconstructed based on the final 3D point cloud data to obtain a 3D reconstruction model of the target power tower. The joint optimization operations include: Based on the current temporal high-resolution SAR image data, generate the current three-dimensional point cloud data of the power tower; wherein, the temporal high-resolution SAR image data when the joint optimization operation is first performed is the initial temporal high-resolution SAR image data. Based on the current 3D point cloud data of the power tower, determine whether the super-resolution reconstruction network meets the preset accuracy requirements; If so, the current 3D point cloud data of the power tower is used as the final 3D point cloud data of the power tower; otherwise, the current 3D point cloud data of the power tower is projected onto the 2D SAR image domain to generate simulation data, the residual between the simulation data and the current time-series high-resolution SAR image data is calculated, and backpropagation is performed based on the residual to update the parameters of the super-resolution reconstruction network. The raw SAR image data is then input again into the updated super-resolution reconstruction network to generate the temporal high-resolution SAR image data required for the next joint optimization operation.

[0058] It is understood that the above-described device embodiments correspond to the method embodiments of the present invention, and can realize the three-dimensional reconstruction method for electric tower tomography provided by any of the above-described method embodiments of the present invention.

[0059] It should be noted that the device embodiments described above are merely illustrative, and some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the device embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can specifically be implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0060] Based on the above-described embodiment of the three-dimensional reconstruction method for power tower tomography, another embodiment of the present invention provides a terminal device, which includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements any one of the three-dimensional reconstruction methods for power tower tomography of the present invention.

[0061] For example, in this embodiment, the computer program can be divided into one or more modules, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the computer program in the terminal device.

[0062] The terminal device may be a desktop computer, laptop, handheld computer, or cloud server, etc. The terminal device may include, but is not limited to, a processor and a memory.

[0063] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the terminal device, connecting all parts of the terminal device via various interfaces and lines.

[0064] Based on the above-described method embodiments, another embodiment of the present invention provides a computer-readable storage medium including a stored computer program, wherein, when the computer program is executed, it controls the device where the computer-readable storage medium is located to execute the three-dimensional reconstruction method for power tower tomography described in any of the above-described method embodiments of the present invention.

[0065] The modules / units integrated in the device / terminal equipment, if implemented as software functional units and sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.

[0066] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A method for three-dimensional reconstruction of power towers using tomography, characterized in that, include: Acquire raw SAR image data of the target power tower to be reconstructed; The raw SAR image data is input into a pre-trained super-resolution reconstruction network for forward propagation to obtain initial temporal high-resolution SAR image data. Based on the initial time-series high-resolution SAR image data, the joint optimization operation is repeatedly performed to generate the final three-dimensional point cloud data of the power tower. A three-dimensional reconstruction model of the target power tower to be reconstructed is constructed based on the final three-dimensional point cloud data of the power tower. The joint optimization operation includes: Based on the current temporal high-resolution SAR image data, generate the current three-dimensional point cloud data of the power tower; wherein, the temporal high-resolution SAR image data when the joint optimization operation is first performed is the initial temporal high-resolution SAR image data. Based on the current 3D point cloud data of the power tower, determine whether the super-resolution reconstruction network meets the preset accuracy requirements; If so, the current 3D point cloud data of the power tower is used as the final 3D point cloud data of the power tower; otherwise, the current 3D point cloud data of the power tower is projected onto the 2D SAR image domain to generate simulation data, the residual between the simulation data and the current time-series high-resolution SAR image data is calculated, and backpropagation is performed based on the residual to update the parameters of the super-resolution reconstruction network. The raw SAR image data is then input again into the updated super-resolution reconstruction network to generate the temporal high-resolution SAR image data required for the next joint optimization operation.

2. The method for three-dimensional reconstruction of power towers by tomography as described in claim 1, characterized in that, The super-resolution reconstruction network includes a low-resolution phase information feature processing branch and a low-resolution amplitude information feature processing branch; The low-resolution phase information feature processing branch includes: a recurrent convolutional layer, a gated recurrent unit, and a phase unwrapping layer connected in sequence; the low-resolution amplitude information feature processing branch includes: a multi-level stride convolutional layer and a deconvolutional layer connected in sequence. The step of inputting the raw SAR image data into a pre-trained super-resolution reconstruction network for forward propagation to obtain initial temporal high-resolution SAR image data includes: processing the phase information in the raw SAR image data using a recurrent convolutional layer based on the low-resolution phase information feature processing branch to obtain spatial long-range dependent phase features; processing the spatial long-range dependent phase features using the gated recurrent unit to obtain temporal phase change features; unwrapping and reconstructing the temporal phase change features using the phase unwrapping layer to generate a high-resolution phase image; compressing the amplitude information in the raw SAR image data using a multi-level stride convolutional layer based on the low-resolution amplitude information feature processing branch to obtain spatial abstract features; upsampling the spatial abstract features using the deconvolutional layer to generate a high-resolution amplitude image; and combining the high-resolution phase image and the high-resolution amplitude image to obtain the initial temporal high-resolution SAR image data.

3. The method for three-dimensional reconstruction of power towers by tomography as described in claim 2, characterized in that, The loss function used by the super-resolution reconstruction network during the pre-training phase is as follows: ； in, For the total loss function, These are the regularization coefficients for each term; The The basic amplitude loss is as follows: Where T is the total number of time-series nodes, and t is the current time-series node; For high-resolution amplitude information obtained from network prediction, For true high-resolution amplitude information; The The phase continuity loss term is as follows: ,in, To obtain the high-resolution phase for prediction, This is the high-resolution phase predicted at the previous moment; The The phase consistency loss term is as follows: Where LPF(·) is the low-pass filter operation, The low-resolution phase is the input; The For symmetric regularization, specifically: ;in, (·) indicates a vertical mirror flip operation.

4. The method for three-dimensional reconstruction of power towers by tomography as described in claim 3, characterized in that, The step of generating current 3D point cloud data of the power tower based on current time-series high-resolution SAR image data includes: extracting a binary mask of the target power tower based on the current time-series high-resolution SAR image data; forcing the 3D scattering tensor of the background region to zero based on the binary mask, and mapping the current time-series high-resolution SAR image data to the 3D scattering tensor of the target region; solving the 3D scattering tensor of the target region under joint horizontal and vertical sparsity constraints to obtain the current 3D point cloud data of the power tower.

5. The method for three-dimensional reconstruction of power towers by tomography as described in claim 4, characterized in that, The step of solving the three-dimensional scattering tensor of the target region under the combined horizontal and vertical sparsity constraints to obtain the current three-dimensional point cloud data of the power tower specifically includes: constructing an optimization objective function with the three-dimensional scattering tensor as the solution variable; obtaining the optimal three-dimensional complex scattering coefficient distribution by solving the optimization objective function, and using the three-dimensional complex scattering coefficient distribution as the current three-dimensional point cloud data of the power tower; The specific optimization objective function is as follows: Y is a SAR three-dimensional observation matrix containing the current temporal high-resolution SAR image data. For the observation imaging operator, * denotes the mode product of the tensor and the matrix. It is the F-norm. Let S be the regularization coefficient, S be the three-dimensional scattering tensor to be solved, and M(x,y) be the extracted binary mask. It is a mixed norm; The mixing norm The specific calculation formula is as follows: ； Where x and y are the pixel indices in the azimuth and range directions, respectively. , , representing the total number of pixels in the azimuth and range directions, respectively; z is the elevation resolution meta-index. Let S be the total resolution element in the elevation direction, and |S(x,y,z)| represent taking the modulus of the three-dimensional scattering tensor at the spatial location (x,y,z).

6. The method for three-dimensional reconstruction of power towers by tomography as described in claim 5, characterized in that, The process of projecting the current 3D point cloud data of the power tower onto the 2D SAR image domain to generate simulation data, calculating the residual between the simulation data and the current temporal high-resolution SAR image data, and performing backpropagation based on the residual to update the parameters of the super-resolution reconstruction network specifically includes: multiplying the 3D complex scattering coefficients at each spatial location in the current 3D point cloud data of the power tower with the baseline phase factor of the corresponding elevation dimension, summing and integrating the multiplication result along the elevation dimension to obtain the pixel value at the corresponding pixel location in the 2D SAR image domain, and constructing the simulation data from all the pixel values ​​at the pixel locations; calculating the degree of difference between the pixel values ​​at each pixel location in the current temporal high-resolution SAR image data and the pixel values ​​at the corresponding pixel locations in the simulation data to obtain the residual in the 2D SAR image domain; using the residual in the 2D SAR image domain as a supervision signal for physical consistency constraints, calculating the gradient of the residual with respect to the parameters of the super-resolution reconstruction network through the backpropagation algorithm, and updating the parameters of the super-resolution reconstruction network based on the gradient.

7. The method for three-dimensional reconstruction of power towers by tomography as described in claim 6, characterized in that, The step of constructing a three-dimensional reconstruction model of the target power tower based on the final three-dimensional point cloud data of the power tower includes: extracting point clouds of dense vertical regions from the final three-dimensional point cloud data of the power tower, fitting a cylindrical model to the point clouds of the dense vertical regions using a random sampling consensus algorithm, and obtaining the main pole structure parameters of the target power tower to be reconstructed based on the fitted cylindrical model. Horizontally distributed point cloud clusters are extracted from the final three-dimensional point cloud data of the power tower. Plane equations are used to fit the horizontally distributed point cloud clusters, and the crossarm structure parameters of the target power tower to be reconstructed are obtained based on the fitted plane equations. The conductor point cloud is extracted from the final three-dimensional point cloud data of the power tower, the conductor point cloud is fitted using the catenary equation, and the conductor structure parameters of the target power tower to be reconstructed are obtained based on the fitted catenary equation. The final three-dimensional point cloud data of the power tower is filtered based on a preset phase stability screening threshold to obtain stable scattering points; the stable scattering points are combined with the main pole structure parameters, the crossarm structure parameters and the conductor structure parameters to generate a three-dimensional reconstruction model of the target power tower to be reconstructed.

8. A three-dimensional reconstruction device for power tower tomography, characterized in that, include: The module includes a data acquisition module, an initial reconstruction module, a model building module, and a joint optimization module. The data acquisition module is used to acquire the original SAR image data of the target power tower to be reconstructed; the initial reconstruction module is used to input the original SAR image data into a pre-trained super-resolution reconstruction network for forward propagation to obtain initial temporal high-resolution SAR image data; the joint optimization module is used to repeatedly perform joint optimization operations based on the initial temporal high-resolution SAR image data to generate the final three-dimensional point cloud data of the power tower. The model building module is used to extract the structural parameters of the target power tower to be reconstructed based on the final 3D point cloud data of the power tower, and obtain a 3D reconstruction model of the target power tower to be reconstructed; wherein, the joint optimization operation includes: Based on the current temporal high-resolution SAR image data, generate the current three-dimensional point cloud data of the power tower; wherein, the temporal high-resolution SAR image data when the joint optimization operation is first performed is the initial temporal high-resolution SAR image data. Based on the current 3D point cloud data of the power tower, determine whether the super-resolution reconstruction network meets the preset accuracy requirements; If so, the current 3D point cloud data of the power tower is used as the final 3D point cloud data of the power tower; otherwise, the current 3D point cloud data of the power tower is projected onto the 2D SAR image domain to generate simulation data, the residual between the simulation data and the current time-series high-resolution SAR image data is calculated, and backpropagation is performed based on the residual to update the parameters of the super-resolution reconstruction network. The raw SAR image data is then input again into the updated super-resolution reconstruction network to generate the temporal high-resolution SAR image data required for the next joint optimization operation.

9. A terminal device, characterized in that, The method includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein when the processor executes the computer program, it implements the three-dimensional reconstruction method for power tower tomography as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, include: A stored computer program, wherein, when the computer program is executed, it controls the device containing the computer-readable storage medium to perform the three-dimensional reconstruction method of power tower tomography as described in any one of claims 1-7.

Images