A ground penetrating radar three-dimensional voxel semantic segmentation method based on dielectric constant field gradient constraint

Abstract

The application provides a ground penetrating radar three-dimensional voxel semantic segmentation method based on a dielectric constant field gradient constraint, which comprises the following steps: acquiring detection signals of each underground medium in an underground space, pre-processing the detection signals, and outputting a three-dimensional dielectric constant voxel field; calculating the gradient modulus value of each voxel in the three-dimensional dielectric constant voxel field, and determining the semantic segmentation boundary formed by the voxels in the three-dimensional dielectric constant voxel field based on the gradient modulus value; determining at least one target starting point meeting a preset condition in the three-dimensional dielectric constant voxel field according to the semantic segmentation boundary; for each target starting point, performing diffusion processing on the neighboring voxels adjacent to the target starting point with the target starting point as the center, and forming a target entity voxel cluster corresponding to the target starting point; extracting physical and geometric features of each target entity voxel cluster, performing topological semantic correction according to a preset material semantic mapping relationship, and converting the target entity voxel cluster into a structured semantic object corresponding to each underground medium in the underground space.

Description

Technical Field

[0001] This invention relates to the fields of geophysical exploration and underground engineering detection technology, specifically to a three-dimensional voxel semantic segmentation method for ground-penetrating radar based on dielectric constant field gradient constraints. Background Technology

[0002] With the acceleration of urbanization, the development and utilization of urban underground space has become increasingly frequent. Underground pipe networks are intricate and complex, and due to factors such as geological evolution and construction disturbances, hidden defects such as underground cavities and loosening are frequent, posing a serious potential threat to urban road safety. Ground Penetrating Radar (GPR) is a non-destructive testing tool that uses high-frequency electromagnetic waves to detect the distribution of underground media. GPR captures the signals corresponding to reflection, refraction, and scattering that occur when electromagnetic waves propagate through underground media and encounter interfaces with differences in dielectric constant or conductivity. Through signal processing and data analysis, it obtains spatial distribution information of underground structures.

[0003] However, raw ground-penetrating radar echo data is typically based on deep learning-based physical parameter inversion methods, which suffer from two problems: ambiguity of boundaries between various subsurface media and lack of semantic meaning. Specifically, while this method can recover the dielectric constant numerical field of the subsurface media, its output is essentially an unstructured three-dimensional floating-point matrix. This underlying numerical field data has two core defects: First, boundary ambiguity: limited by the smoothing effect of convolutional neural networks and the physical limitations of radar wavelengths, the inverted dielectric constant field often appears as a numerical gradient band several pixels wide at the boundaries between different materials (such as air and soil), rather than a true sharp physical interface, making it difficult to accurately quantify the geometric dimensions of the target. Second, lack of semantic meaning: the inversion results only provide physical parameters and lack clear semantic attributes.

[0004] Therefore, there is an urgent need for a segmentation method that can automatically transform the underlying fuzzy physical parameter field into a segmentation method with clear boundaries and explicit semantic labels. Summary of the Invention

[0005] This invention provides a three-dimensional voxel semantic segmentation method for ground-penetrating radar based on dielectric constant field gradient constraints, in order to solve the technical problem of inaccuracy when segmenting underground media in underground space.

[0006] The first aspect of the present invention provides a three-dimensional voxel semantic segmentation method for ground-penetrating radar based on dielectric constant field gradient constraints. The method includes: acquiring detection signals of various underground media in underground space; preprocessing the detection signals to obtain a three-dimensional echo data tensor; inputting the three-dimensional echo data tensor into a three-dimensional inversion model to output a three-dimensional dielectric constant voxel field; calculating the gradient magnitude of each voxel in the three-dimensional dielectric constant voxel field and determining the semantic segmentation boundary formed by the voxels in the three-dimensional dielectric constant voxel field based on the gradient magnitude; determining at least one target starting point in the three-dimensional dielectric constant voxel field that meets preset conditions according to the semantic segmentation boundary; for each target starting point, performing diffusion processing on neighboring voxels adjacent to the target starting point as the center to form a target entity voxel cluster corresponding to the target starting point; extracting physical and geometric features for each target entity voxel cluster, performing topological semantic correction according to a preset material semantic mapping relationship, and transforming the target entity voxel cluster into a structured semantic object describing various underground media in underground space.

[0007] In one embodiment of the present invention, the detection signal is preprocessed to obtain a three-dimensional echo data tensor, including: extracting multi-channel high-dimensional feature space from the detection signal to obtain multi-channel feature vectors, and performing weighted adjustment processing on the multi-channel feature vectors based on channel and spatial dimensions to obtain enhanced feature vectors; performing feature reconstruction processing on the enhanced feature vectors to output the reconstructed three-dimensional echo data tensor.

[0008] In one embodiment of the present invention, the gradient magnitude of each voxel in a three-dimensional dielectric constant voxel field is calculated, and a high gradient surface is extracted based on the gradient magnitude. The semantic segmentation boundary formed by each voxel in the three-dimensional dielectric constant voxel field is determined based on the high gradient surface. This includes: calculating the spatial gradient vector of each voxel in the three-dimensional dielectric constant voxel field using a three-dimensional gradient operator, and determining the semantic segmentation boundary based on the spatial gradient vector. Determine the gradient magnitude Among them, the gradient magnitude The calculation formula is: Where x, y, z represent the electric field intensity vector field of each voxel in the three-dimensional dielectric constant voxel field. The dielectric constant is represented by a voxel whose gradient magnitude exceeds an adaptive threshold in the three-dimensional dielectric constant voxel field, forming a high gradient surface, which is then used as the semantic segmentation boundary.

[0009] In one embodiment of the present invention, the method further includes: extracting features from the spatial gradient vector using filters of different angles and wavelengths to obtain gradient structure feature parameters; and using the gradient structure feature parameters, combined with the interval reflected by the dielectric constant gradient, to perform weighted correction on the spatial gradient vector.

[0010] In one embodiment of the present invention, determining at least one target starting point that satisfies preset conditions in a three-dimensional dielectric constant voxel field based on the semantic segmentation boundary includes: determining at least one neighborhood in a three-dimensional dielectric constant voxel field based on the semantic segmentation boundary; and selecting voxels with relative dielectric constants within a preset dielectric constant range and gradient magnitudes lower than a preset background noise level as target starting points by traversing the voxels in the neighborhood.

[0011] In one embodiment of the present invention, for each target starting point, a diffusion process is performed on the neighboring voxels adjacent to the target starting point, with the target starting point as the center, to form a target entity voxel cluster corresponding to the target starting point. This includes: calculating the difference in average dielectric constant between the neighboring voxels and the target starting point, and obtaining the gradient magnitude of the neighboring voxels; if the difference is less than a preset tolerance and the gradient magnitude is lower than an adaptive threshold, then the neighboring voxels are added to the target entity voxel cluster corresponding to the target starting point.

[0012] In one embodiment of the present invention, physical and geometric features are extracted for each target entity voxel cluster, and topological semantic correction is performed according to a preset material semantic mapping relationship to transform the target entity voxel cluster into a structured semantic object describing underground space. This includes: performing preliminary material labeling on the target entity voxel cluster based on the physical and geometric features corresponding to the target entity voxel cluster and according to a preset dielectric material semantic mapping relationship; and performing at least one of topological semantic correction by combining topological semantic correction rules, multi-dimensional joint semantic mapping, and multi-level verification mechanisms to transform the target entity voxel cluster with preliminary material labeling into a structured semantic object. The physical and geometric features include at least one of the median of the relative permittivity of each voxel in the target entity voxel cluster, aspect ratio, and compactness.

[0013] In one embodiment of the present invention, the method further includes: outputting a structured data file describing each underground medium in the underground space, wherein the structured data file includes: semantic category labels corresponding to the underground medium in the underground space, inferred material properties, geometric dimension parameters and spatial location coordinates.

[0014] A second aspect of the present invention provides a detection signal analysis and processing system, comprising: a data acquisition and preprocessing module for acquiring detection signals of various underground media in an underground space, preprocessing the detection signals to obtain a three-dimensional echo data tensor, inputting the three-dimensional echo data tensor into a three-dimensional inversion model, and outputting a three-dimensional dielectric constant voxel field; a physical field reconstruction module for calculating the gradient magnitude of each voxel in the three-dimensional dielectric constant voxel field; a semantic boundary locking module for determining the semantic segmentation boundary formed by voxels in the three-dimensional dielectric constant voxel field based on the gradient magnitude; a solidification segmentation module for determining at least one target starting point that meets preset conditions in the three-dimensional dielectric constant voxel field according to the semantic segmentation boundary; for each target starting point, performing diffusion processing on neighboring voxels adjacent to the target starting point as the center to form a target entity voxel cluster corresponding to the target starting point; and a physical semantic mapping and correction module for extracting physical and geometric features of each target entity voxel cluster, performing topological semantic correction according to a preset material semantic mapping relationship, and converting the target entity voxel cluster into a structured semantic object describing various underground media in the underground space.

[0015] A third aspect of the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the steps of the method for analyzing and processing detection signals as described in the third aspect of the present invention.

[0016] The beneficial effects of this invention are as follows: By calculating the spatial gradient vector and magnitude of the three-dimensional dielectric constant field, a high-gradient surface is locked as the semantic segmentation boundary. This gradient boundary serves as an external "potential barrier" constraint, combined with the internal "numerical consistency" driving force, to perform dual-criteria verification during the region growing process. Attached Figure Description The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application. It is obvious that the drawings described below are merely some embodiments of this application, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort.

[0017] In the attached diagram: Figure 1 This is a flowchart illustrating the method for analyzing and processing detection signals provided in one embodiment of the present invention; Figure 2 This is a schematic diagram comparing the technical effects of segmenting underground space based on dielectric constant field gradient constraints according to an embodiment of the present invention. Figure 3 This is a schematic diagram of the structure of a detection signal analysis and processing system provided in one embodiment of the present invention; Figure 4 This is a schematic diagram of the structure of an electronic device for analyzing and processing operational detection signals provided in one embodiment of the present invention. Detailed Implementation

[0018] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments. Various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. In the absence of conflict, the following embodiments and features in the embodiments can be combined with each other.

[0019] Before providing a detailed description of the embodiments of this application, we will first define and explain some of the nouns, terms, and custom concepts involved in the embodiments of this application to ensure the consistency and clarity of the terminology throughout the text.

[0020] 1) Standardized 3D echo data tensor D: refers to the 3D data matrix that is standardized and aligned in both spatial and temporal dimensions after the original 3D radar echo data volume (C-scan) has undergone zero-time correction, background removal, adaptive bandpass filtering, and clutter suppression model processing to eliminate surface reflection time differences, strong horizontal interference fringes, out-of-band noise, and incoherent clutter interference. It is represented as... ,in and Indicates the coordinates of the survey line. Indicates the time sampling point.

[0021] 2) Three-dimensional dielectric constant voxel field This refers to the three-dimensional voxel matrix representing the distribution of the relative permittivity of the subsurface medium, obtained by nonlinearly mapping and reconstructing the standardized three-dimensional echo data tensor D using a three-dimensional inversion model (such as 3DU-Net (three-dimensional U-shaped network)) and then performing super-resolution reconstruction using a conditional denoising diffusion model. It is represented as: The value of each voxel (also known as a three-dimensional voxel of ground-penetrating radar) corresponds to the relative permittivity. Here, voxels correspond to electromagnetic parameters, which can accurately describe the physical properties of the underground medium. 3) Spatial gradient vector : Refers to the three-dimensional dielectric constant voxel field calculated using a three-dimensional gradient operator (such as the three-dimensional Sobel operator), where each voxel is in the voxel field. , , The partial derivative vectors in three spatial directions are used to characterize the rate of change and direction of the dielectric constant in three-dimensional space.

[0022] 4) Gradient magnitude or : refers to the spatial gradient vector The modulus is used to quantitatively characterize the degree of drastic change of the dielectric constant in space, and its high value region corresponds to the interface between different dielectric materials.

[0023] 5) High gradient surface: refers to a continuous three-dimensional surface formed by connecting voxels with gradient magnitudes higher than the gradient magnitude, which is extracted by setting an adaptive threshold. This surface is the locked semantic segmentation boundary.

[0024] 6) Low-dielectric seed point: refers to the seed point with a relative permittivity selected by traversing the three-dimensional permittivity voxel field. Within a specific range (e.g.) to (between) and its gradient magnitude or A voxel with a background noise level below the preset level indicates that the voxel is located in a homogeneous internal region (such as inside an air cavity), and serves as the target starting point (high confidence starting point) for the region growing algorithm.

[0025] 7) Consistency criterion verification: During the region growth process, the difference between the relative permittivity of the candidate neighboring voxel and the average permittivity of the currently grown voxel cluster is calculated. If the difference is less than the preset tolerance, it is determined that the numerical consistency is met, which is one of the necessary conditions for allowing the voxel cluster to be incorporated.

[0026] 8) Barrier blocking criterion verification: During the region growth process, it is checked whether the gradient magnitude at the candidate neighboring voxel position is higher than a determined adaptive threshold. If it is higher, it is determined that the material interface barrier has been touched, and growth in that direction is forcibly stopped. This serves as an absolute constraint to prevent the algorithm from overflowing.

[0027] 9) Refined processing mechanism of gradient potential barrier: refers to the algorithm mechanism that combines nonmaximum suppression, multi-level barrier discrimination (hard barrier, soft barrier, non-barrier) and backtracking verification and secondary growth judgment to perform multi-scale and closed-loop control on the region growth boundary, which is used to solve the problem of undersegmentation or oversegmentation caused by the gradual transition zone of physical boundary in the inversion result.

[0028] 10) Topological semantic correction rules: refers to a set of rules that use spatial enclosing relationships and geometric constraints to perform advanced semantic reasoning on voxel clusters that have been initially labeled with material attributes. These rules are used to accurately identify structured entities with clear engineering significance, such as underground collapse cavities and underground pipelines.

[0029] 11) Multi-dimensional joint semantic mapping and multi-level verification mechanism: This refers to a comprehensive verification mechanism that introduces auxiliary physical features such as multi-polarization features and dispersion characteristics, and combines them with profile continuity verification, spatial regularity joint judgment and surface semantic association verification to eliminate the ambiguity caused by single physical feature judgment and improve the confidence of semantic recognition.

[0030] 12) Structured data files: These are standardized 3D model files that contain clear semantic category labels, inferred material properties, geometric dimension parameters, and spatial location coordinates, and can be directly imported into BIM (Building Information Modeling) systems or urban underground space digital twin platforms. Examples include CityGML (City Geography Markup Language) format, IFC (Industry Foundation Classes) format, or OBJ (Object) files with attribute tables.

[0031] The main technical problem that this invention aims to solve is that after processing radar signals corresponding to underground spaces through convolutional neural networks, the dielectric constant field of various underground media in the underground space is shown as a numerically gradual transition zone of several pixels in width at the interface of different underground media (such as air and soil, plastic and sand), rather than a real sharp physical interface. This results in blurred boundaries between various underground media and difficulty in accurately quantifying the geometric dimensions of the target. Furthermore, the dielectric constant field lacks clear semantic attribute encapsulation, creating a semantic gap between physical data and practical applications, making it impossible to directly import it into BIM systems or urban underground space digital twin platforms for observation, analysis, and use.

[0032] In existing technologies, traditional image segmentation algorithms (such as global thresholding and conventional region growing) are prone to overflowing into the background medium when dealing with images containing a certain width of gradual transition bands, resulting in severe oversegmentation; or they may terminate prematurely due to noise interference in weak signal areas, leading to undersegmentation. Traditional techniques present an irreconcilable contradiction between improving boundary sharpness and maintaining topological closure. The breakthrough of this technical solution lies in: introducing the spatial gradient vector and magnitude calculation of a three-dimensional dielectric constant field, locking high-gradient surfaces as semantic segmentation boundaries, and using these gradient boundaries as external "potential barrier" constraints, combined with the internal "numerical consistency" driver, to perform dual criterion verification during the region growing process.

[0033] By introducing a gradient barrier blocking mechanism, when the growth algorithm spreads to the boundary between different materials, even if the dielectric constant value at that point still meets the consistency requirement due to the smoothing effect of the inversion model, the high gradient barrier will force the growth in that direction to stop. This mechanism effectively prevents the growth algorithm from overflowing into the background medium by crossing the blurred gradient transition zone in the inversion result, thus obtaining a target entity voxel cluster with sharp edges and topological closure. Based on this, by extracting the physical and geometric features of the voxel cluster, and combining topological semantic correction rules, multi-dimensional joint semantic mapping, and multi-level verification mechanisms, the meaningless discrete physical parameter field is automatically transformed into a structured entity object with clear physical boundaries and explicit semantic labels, eliminating the semantic gap and realizing the automatic and accurate transformation from continuous physical fields to structured semantic objects.

[0034] To address the aforementioned problems, this application provides an analysis and processing method for the detection signal (a three-dimensional voxel semantic segmentation method for ground-penetrating radar based on dielectric constant field gradient constraints). The method includes: acquiring the detection signal (raw three-dimensional radar echo data) and performing preprocessing to obtain a standardized three-dimensional echo data tensor. Where X and Y represent the coordinates of the survey line, and T represents the time sampling point; the standardized three-dimensional echo data tensor D is input into the trained three-dimensional inversion model, and the output is a three-dimensional dielectric constant voxel field. In this context, the value of each voxel in the three-dimensional permittivity voxel field corresponds to the relative permittivity. ; Calculation of three-dimensional dielectric constant voxel field using three-dimensional gradient operator Spatial gradient vector of each voxel and gradient magnitude And based on gradient magnitude High-gradient surfaces are extracted as locked semantic segmentation boundaries. At least one high-confidence starting point (target starting point) is selected in the three-dimensional dielectric constant voxel field. A growth queue is established centered on the high-confidence starting point, and iterative diffusion (diffusion processing) is performed to neighboring voxels. When determining whether to incorporate neighboring voxels into the current voxel cluster, a dual criterion verification is performed: numerical consistency check and barrier blocking check based on the semantic segmentation boundary, to obtain the target entity voxel cluster. Physical and geometric features are extracted for each target entity voxel cluster. Preliminary material labeling is performed based on the preset dielectric material semantic mapping relationship. Topological semantic correction is then performed using topological semantic correction rules, multi-dimensional joint semantic mapping, and multi-level verification mechanisms to transform the target entity voxel cluster into a structured semantic object. A structured data file describing the underground space object is output, including: semantic category labels, inferred material properties, geometric dimension parameters, and spatial location coordinates.

[0035] Figure 1The diagram shows a flowchart of an analysis and processing method for the above-mentioned detection signal provided in an optional embodiment of this application. The analysis and processing method for the detection signal includes the following steps.

[0036] Step S101: Acquire the detection signals of various underground media in the underground space, preprocess the detection signals to obtain the three-dimensional echo data tensor, input the three-dimensional echo data tensor into the three-dimensional inversion model, and output the three-dimensional dielectric constant voxel field.

[0037] Step S102: Calculate the gradient magnitude of each voxel in the three-dimensional dielectric constant voxel field, and determine the semantic segmentation boundary formed by the voxels in the three-dimensional dielectric constant voxel field based on the gradient magnitude.

[0038] Step S103: Determine at least one target starting point that satisfies the preset conditions in the three-dimensional dielectric constant voxel field based on the semantic segmentation boundary.

[0039] Step S104: For each target starting point, diffuse the target starting point to the neighboring voxels adjacent to the target starting point to form a target entity voxel cluster corresponding to the target starting point.

[0040] Step S105: Extract physical and geometric features for each target entity voxel cluster, perform topological semantic correction based on the preset material semantic mapping relationship, and transform the target entity voxel cluster into a structured semantic object describing the various underground media in the underground space.

[0041] This application provides a flowchart illustrating a three-dimensional voxel semantic segmentation method for ground-penetrating radar based on dielectric constant field gradient constraints in an optional embodiment. The method is applied to a data processing terminal and specifically includes the following steps: Step S201: Acquire the raw 3D radar echo data and preprocess it to obtain a standardized 3D echo data tensor. The standardized 3D echo data tensor... , and Indicates the coordinates of the survey line. This indicates the time sampling point. In some embodiments, in this step, the area to be measured is first scanned in a grid pattern using a vehicle-mounted array ground-penetrating radar or a manually towed 3D ground-penetrating radar. To ensure a balance between detection depth and resolution, a center frequency of [value missing] is preferably used. to The radar antenna was used to acquire the raw three-dimensional radar echo data volume (C-scan). Subsequently, a series of data cleaning operations were performed on the raw data to improve the signal-to-noise ratio. Specifically, zero-time correction was first performed, adjusting the start time of each data track to eliminate differences in surface reflection time caused by uneven air layer thickness; next, background removal was performed, using moving average subtraction to remove strong horizontal interference fringes caused by direct ground waves and antenna coupling signals; finally, bandpass filtering was performed to filter out low-frequency drift interference and high-frequency random noise. After these processing steps, a standardized three-dimensional echo data tensor was finally obtained. .

[0042] Step S202: Input the standardized 3D echo data tensor into the trained 3D inversion model, and output a 3D dielectric constant voxel field, where the value of each voxel in the 3D dielectric constant voxel field corresponds to the relative dielectric constant. .

[0043] In some embodiments, this step aims to convert the radar signal into a dielectric constant field. This embodiment constructs a three-dimensional inversion model based on a fully convolutional network (FCN) architecture (e.g., a modified version of 3DU-Net). The normalized three-dimensional echo data tensor processed in step S101 is then used... The trained network is input, and the network output is a three-dimensional voxel field of dielectric constant, where the value of each voxel corresponds to the relative dielectric constant. At this point, we obtain... Although it reflects physical properties, due to the convolution and upsampling smoothing effects of neural networks, different materials (such as relative permittivity) may exhibit variations. The void and The boundary between soil surfaces is typically represented as a gradient transition band several pixels wide, rather than a sharp interface.

[0044] Step S203: Calculate the spatial gradient vector of each voxel in the three-dimensional dielectric constant voxel field using the three-dimensional gradient operator. and gradient magnitude And based on gradient magnitude High gradient surfaces are extracted and used as the locked semantic segmentation boundaries.

[0045] In some embodiments, to address the boundary ambiguity problem caused by the aforementioned "transition zone," this step introduces a three-dimensional gradient operator. For each voxel in the three-dimensional dielectric constant voxel field, its spatial gradient vector is calculated using the three-dimensional Sobel operator. In discrete data, partial derivatives can be approximated using the central difference method. The region of high gradient magnitude represents the spatial location where the dielectric constant changes drastically. This can be achieved by setting an adaptive threshold (e.g., taking the maximum value of the gradient field). The high gradient surface is extracted, which is the locked semantic boundary.

[0046] Step S204: Select a high-confidence seed point in the three-dimensional dielectric constant voxel field, establish a growth queue with the high-confidence seed point as the center, iteratively diffuse to the neighboring voxels, and when determining whether to merge the neighboring voxels into the current voxel cluster, perform a dual criterion check of numerical consistency check and barrier blocking check based on semantic segmentation boundary to obtain the target entity voxel cluster.

[0047] In some embodiments, this step aims to segment a continuously distributed permittivity field into independent geometric entities using physical property values ​​as an intrinsic driving force and gradient boundaries as extrinsic constraints. Specifically, a region growing algorithm constrained by the gradient field is employed. First, the system automatically selects high-confidence seed points in the three-dimensional voxel field. By traversing the voxel field, relative permittivity points are selected based on prior physical knowledge. Within a specific range (e.g.) to (between) and its gradient magnitude Voxels with gradient magnitude values ​​below a preset threshold are marked as low-dielectric seed points. A growth queue is then established centered on these seed points, iteratively expanding to neighboring voxels. When determining whether to incorporate neighboring voxels into the current voxel cluster, a dual-criteria check is performed: firstly, a consistency criterion check is performed, calculating the difference in average dielectric constant between the candidate neighboring voxels and the current voxel cluster. If this difference is less than a preset tolerance, numerical consistency is satisfied. Secondly, a barrier blocking criterion check is performed, checking the gradient magnitude at the location of the neighboring voxel. If the gradient magnitude at this location is higher than the boundary adaptive threshold determined in step S103, it is considered to have touched a "material interface barrier," and growth in that direction is forcibly stopped even if the consistency criterion is satisfied. This mechanism of introducing a gradient barrier effectively prevents the growth algorithm from "overflowing" into the background medium through the blurred transition zone in the inversion result, thus obtaining a target entity voxel cluster with sharp edges and a closed topology.

[0048] Step S205: Extract physical and geometric features for each target entity voxel cluster, perform preliminary material labeling based on the preset dielectric material semantic mapping relationship, and perform topological semantic correction by combining topological semantic correction rules, multi-dimensional joint semantic mapping and multi-level verification mechanism to transform the target entity voxel cluster into a structured semantic object.

[0049] This step transforms meaningless voxel clusters into semantic objects that computers and large models can understand. First, core feature extraction is performed: for each voxel cluster, its internal relative permittivity is calculated. the median of And geometric morphological features (such as aspect ratio, volume, compactness, etc.). Next, a lookup table mapping is performed: preliminary labeling is done based on a pre-defined dielectric-material semantic mapping library: if... The material is determined to be "air / gas"; if The material was determined to be "dry backfill soil"; if It may exhibit total reflection characteristics, classifying the material as "water / metal". Subsequent topological semantic correction is performed: using spatial enclosure relationships and geometric shape constraints, the initially labeled material is corrected: Rule 1 (void determination): If a voxel cluster is of material "air" and is completely enclosed in space by voxel clusters of material "soil" or "concrete", and its volume is greater than a threshold (e.g., ... If the material of a voxel cluster is "air" or "plastic", and its geometry is a long, thin cylinder (the length of the principal axis is much greater than the radius of the cross section), then its semantic label is corrected to "underground pipeline".

[0050] Step S206: Output a structured data file containing semantic category labels, inferred material properties, geometric dimension parameters, and spatial location coordinates.

[0051] After the above processing steps, the final generated 3D model is no longer the original ground-penetrating radar echo image or discrete physical parameter field, but a structured voxel model containing explicit semantic information. The system will output a structured data file with semantic category labels, inferred material properties, geometric dimensional parameters, and spatial location coordinates, such as CityGML format, IFC format, or OBJ file with attribute tables. This output can be directly imported into a BIM system or a digital twin platform for urban underground space to support automated geological modeling, underground pipeline collision detection, and cavity hazard risk assessment, thereby overcoming the technical shortcomings of traditional methods that rely on tedious image interpretation and manual drawing.

[0052] To enable those skilled in the art to more clearly understand the technical improvements and preferred implementation schemes of this application at each step, the detailed implementation of each step is described in detail below with reference to the accompanying drawings.

[0053] In the data acquisition and preprocessing step S201, zero-time correction is performed on the raw three-dimensional radar echo data to adjust the start time of each data track, thereby eliminating the differences in surface reflection time caused by uneven air layer thickness. Next, background removal is performed using a moving average subtraction method to remove strong horizontal interference fringes caused by direct ground waves and antenna coupling signals. Finally, bandpass filtering is applied to filter out low-frequency drift interference and high-frequency random noise.

[0054] To accommodate radar antennas with different center frequencies, this embodiment incorporates an adaptive frequency band adjustment mechanism. If the center frequency of the radar antenna is... Then the filtering range of the bandpass filter is set to to To accurately preserve the effective echo signal within this frequency band and avoid the loss of high-frequency details; if the center frequency of the radar antenna is to Then the filtering range of the bandpass filter is set to to This invention adaptively filters out out-of-band noise and retains effective band signals during wideband detection. In contrast, traditional bandpass filters typically use a fixed cutoff frequency, which cannot be adaptively adjusted according to changes in the antenna center frequency, easily leading to the incorrect filtering of high-frequency detail signals or the retention of low-frequency noise. This invention maximizes spatial resolution while ensuring detection depth by adaptively adjusting the filtering frequency band, providing a high-quality, standardized three-dimensional echo data tensor D for subsequent physical field reconstruction.

[0055] To further suppress strong reflection clutter generated by the steel mesh in reinforced concrete structures and random scattering interference caused by coarse aggregates, this embodiment introduces a dedicated clutter suppression model in the preprocessing stage. The preprocessed data is input into a pre-trained clutter suppression model. This model includes a multi-scale feature extraction module, a residual enhancement module, and a feature fusion module. First, the multi-scale feature extraction module extracts features from the data, using multiple two-dimensional convolutional and max-pooling layers to extract the initial spatial features of the input single-channel ground-penetrating radar data, and performs downsampling to map them to a multi-channel high-dimensional feature space, obtaining multi-channel feature vectors. Next, the residual enhancement module performs weighted adjustment of the multi-channel feature vectors based on channel and spatial dimensions. The residual enhancement module consists of multiple convolutional attention units connected by residuals. By combining channel attention and spatial attention mechanisms, it adaptively enhances important channels and key spatial locations in the feature map, obtaining enhanced feature vectors. Finally, the enhanced feature vectors are reconstructed using a feature fusion module. Multiple sets of two-dimensional deconvolution layers are used to gradually restore the feature map to its original spatial dimensions. A residual enhancement module is then applied again to emphasize key structural features, outputting clutter-suppressed three-dimensional echo data. In contrast, traditional background removal or principal component analysis methods can only eliminate horizontal direct waves, exhibiting extremely poor suppression capabilities for incoherent clutter caused by reinforcing bars and aggregates. This invention, through a cascaded network of multi-scale feature extraction, residual enhancement, and feature fusion, adaptively suppresses multi-source complex clutter, significantly improving the amplitude and contrast of weak echo signals from defects, laying a high signal-to-noise ratio data foundation for subsequent physical field reconstruction.

[0056] In the deep learning-based physics field reconstruction step S202, a 3DU-Net model based on a fully convolutional network (FCN) architecture was constructed as the 3D inversion model. This model consists of an encoder stage and a decoder stage, preserving low-dimensional spatial information through skip connections to achieve a nonlinear mapping from input (normalized 3D echo data tensor D) to output (3D dielectric constant voxel field E(x,y,z)). Since the true dielectric constant distribution (ground value) of underground real-world scenarios is extremely difficult to obtain, this embodiment uses the finite-difference time-domain (FDTD) method to generate a large-scale high-fidelity synthetic dataset. In the simulation modeling, dielectric constant distribution maps containing various underground models such as cavities, pipelines, and layered soil are constructed as labels. The corresponding radar echo image is simulated and used as input to supervise the training of the 3DU-Net network. After training, the measured standardized three-dimensional echo data tensor D is input into the model, which can quickly and automatically invert the high-precision three-dimensional dielectric constant voxel field E(x,y,z).

[0057] To overcome the shortcomings of purely data-driven deep learning models, such as lack of physical constraints and susceptibility to "false anomalies" that do not conform to physical laws, this embodiment introduces physical information constraints during the training process of the 3D inversion model. Specifically, the Maxwell's equations for the propagation of the electric and magnetic fields, which satisfy the electromagnetic wave propagation equation of ground-penetrating radar, are decoupled, and combined with the predicted electric field and dielectric constant, a physical residual formula satisfying the electromagnetic wave propagation equation is constructed: ,in, For electric field, Permeability, For conductivity, Where is the dielectric constant. For the electric field source term, the physical residual formula is used. This invention incorporates the physical loss function into the training loss function of the 3D inversion model, enabling jointly data-driven and physics-driven supervised training. In contrast, traditional deep learning inversion relies entirely on the statistical distribution of training samples; when there is a deviation between the measured data and the training set, the inversion results are easily distorted. This invention introduces Maxwell's equations as the physical loss function into the network training, forcibly constraining the network output to conform to the physical laws of electromagnetic wave propagation. This significantly improves the physical consistency and interpretability of the inversion results, accurately retrieving the true dielectric constant value even under low signal-to-noise ratio or complex geological conditions.

[0058] After outputting the three-dimensional dielectric constant voxel field E(x,y,z), and calculating the spatial gradient vector... Previously, to further improve the spatial resolution of the voxel field, this embodiment introduced a super-resolution reconstruction step based on a conditional denoising diffusion model. First, in the Gaussian noise diffusion stage, a... Add Gaussian noise of the same size as the input image, with noise intensity increasing sequentially. Compare the input image with... Adding the Gaussian noises separately, we get A single-channel image with noise. Next, [the image will be processed]. Each single-channel noisy image is channel-merged with the inference data corresponding to the input image to obtain... Two-channel noisy image data is input into a residual self-attention U-net network for noise estimation, yielding the estimated noise. During the inference phase, a pure noisy image of the same size as the input image is generated. This pure noisy image is then processed... The iteration process is performed at the... In the second iteration, the residual self-attention U-net network is used for the third iteration. Noise estimation is performed on the image obtained in the nth iteration, and from the nth iteration... By subtracting estimated noise from the subsequent iterations of the image, a high-resolution three-dimensional dielectric constant voxel field is finally reconstructed after super-resolution reconstruction. In contrast, traditional interpolation methods (such as bicubic interpolation) introduce severe blurring and jagged edges when the image is enlarged, and cannot recover lost high-frequency details. This invention uses a conditional denoising diffusion model for super-resolution reconstruction, which can adaptively recover high-frequency details such as tiny cracks and pipe edges in the dielectric constant field, improving the spatial resolution to the millimeter level and providing an excellent data source for subsequent high-precision boundary locking.

[0059] In the semantic boundary locking step S203, in order to transform the continuously distributed dielectric constant field into an entity with clear boundaries, this embodiment uses the three-dimensional Sobel operator to approximate the spatial gradient vector of each voxel in the three-dimensional dielectric constant voxel field E(x,y,z) using the central difference method. Gradient magnitude The calculation formula is: , In discrete voxel data, the partial derivatives , , The gradient magnitude is approximated using the central difference method. Corresponding gradient magnitude field Subsequently, in order to adaptively extract the boundary, this embodiment takes the gradient magnitude field. An adaptive threshold of 15% to 25% of the maximum value is used, for example, 20% of the maximum value. Regions with gradient magnitudes higher than this adaptive threshold are extracted as high-gradient surfaces, which serve as the locked semantic segmentation boundaries. In contrast, traditional fixed-threshold boundary extraction methods cannot adapt to differences in signal attenuation caused by different geological media and exploration depths, easily leading to overfitting of shallow boundaries and loss of deep boundaries. This invention extracts high-gradient surfaces using adaptive thresholds, dynamically locking boundaries based on the drastic changes in local dielectric constant, ensuring the integrity and accuracy of boundary extraction in variable underground media. It is understood that the above values ​​are exemplary, and this application embodiment does not limit the data.

[0060] As an alternative implementation, the spatial gradient vector is calculated. In addition to the three-dimensional Sobel operator, the three-dimensional gradient operator can also be the three-dimensional Prewitt operator or the three-dimensional Scharr operator. Taking the three-dimensional Scharr operator as an example, it has higher isotropic characteristics when calculating discrete partial derivatives and can more accurately capture weak dielectric constant abrupt changes at the interface.

[0061] The spatial gradient vector is calculated. Subsequently, the heterogeneity of the underground medium (such as gravel in the soil and uneven local water content) causes abnormal distortion of the gradient field, leading to deviations in boundary positioning. To eliminate this effect, this embodiment introduces a gradient correction mechanism based on a Gabor filter. Specifically, it utilizes different angles and wavelengths... Gabor filters for spatial gradient vectors Feature extraction is performed to obtain gradient structure feature parameters. The Gabor filter, as a linear local transform, has the following function:

[0062] in, , ,parameter Controlling the wavelength, Control the direction, Controlling phase shift, Controlling standard deviation, Control the aspect ratio. Utilize different angles and wavelengths. The Gabor filter is used to extract features from the spatial gradient vector to obtain gradient structure feature parameters. Subsequently, combined with the interval reflected by the dielectric constant gradient, the spatial gradient vector is... Weighted corrections are applied to adjust for anomalous distortions in the gradient field caused by the heterogeneity of the medium. In contrast, without gradient correction, local noise caused by the heterogeneous medium is misidentified as a physical interface, leading to premature termination of the region growing algorithm or the generation of incorrect segmentation fragments. This invention extracts multi-scale structural features using a Gabor filter and applies weighted corrections, eliminating local distortions caused by heterogeneity and significantly improving the positioning accuracy and structural consistency of the physical interface.

[0063] In the restricted solidification segmentation step S204 based on gradient potential barriers, to ensure the region growing algorithm has a high-confidence starting point, the system automatically selects seed points in the three-dimensional voxel field. This is achieved by traversing the three-dimensional dielectric constant voxel field E(x,y,z). It is between 1 and 1.5 (corresponding to the air medium region), and its gradient modulus is... Voxels with background noise levels below a preset threshold are marked as low-dielectric seed points. This ensures that seed points are located within a homogeneous internal region of the medium, avoiding the misselection of boundary transition zones or noise points as seed points. Subsequently, a growth queue is established centered on this seed point, iteratively spreading to adjacent neighboring voxels. When determining whether to incorporate a neighboring voxel into the current voxel cluster, a dual-criteria check is performed, including: 1) Consistency criterion check: Calculate the difference between the average dielectric constant of the candidate neighboring voxel and the current voxel cluster. If the difference is less than a preset tolerance, numerical consistency is satisfied, and incorporation is allowed; 2) Barrier blocking criterion check: Check the gradient magnitude at the location of the neighboring voxel. If the gradient magnitude at that location is higher than a determined adaptive threshold, it is determined that the material interface barrier has been reached, and growth in that direction is forcibly stopped. In contrast, traditional region growing algorithms rely solely on numerical consistency. When faced with a gradient transition zone of a certain width in the inversion result, it is very easy to "overflow" into the background medium through the fuzzy boundary, resulting in severe oversegmentation. This invention introduces a gradient boundary-based barrier blocking check, which organically combines the consistency of physical properties with the blocking properties of gradient boundaries, effectively preventing algorithm overflow and ensuring sharp edges of segmented entities.

[0064] Considering that the physical boundary in the inversion results usually presents as a gradual transition band with a certain width, rather than an ideal step interface, this embodiment further introduces a fine-grained gradient barrier processing mechanism in step S204. First, non-maximum suppression is performed: for the gradient magnitude field... Non-maximum suppression is performed, which means that only local maxima are retained along the principal gradient direction, compressing the gradient transition band of a certain thickness into a candidate boundary layer of a single voxel width, thereby accurately locating the most likely physical interface position. Secondly, multi-level barrier discrimination is performed: a multi-level barrier discrimination criterion is constructed, which selects gradient magnitudes higher than a first threshold. The region is defined as a hard barrier, where region growth absolutely stops; the gradient magnitude is set between the first threshold. With the second threshold The region between them is defined as a soft barrier, in which The growth algorithm here combines the consistency of surrounding voxels and morphological constraints for comprehensive judgment. If multiple consecutive directions touch the soft potential barrier, it is determined to be a true boundary; if the gradient magnitude is lower than the second threshold... The regions are defined as non-barrier areas, allowing normal growth. Finally, backtracking verification and secondary growth are performed: after region growth is completed, backtracking verification is performed on the boundary voxels of each voxel cluster. If a continuous low-gradient region is found around a boundary voxel but not included in the cluster, secondary growth is triggered to avoid undersegmentation caused by local anomalies in the gradient field. In contrast, if only a single hard threshold barrier is used, growth is easily terminated prematurely at noisy locations (undersegmentation) or overflows at weak boundaries (oversegmentation). This invention achieves fine-grained control of the growth boundary by suppressing the non-maximum value and compressing the boundary, combined with multi-level barriers and backtracking verification, greatly improving the topological closure and geometric accuracy of the segmented entities.

[0065] In the physical semantic mapping and correction step S205, to extract high-fidelity, multi-dimensional physical features, this embodiment introduces a multi-feature fusion mechanism based on Hilbert transform and wavelet transform. First, a Hilbert transform is performed on the single-channel signal of each target entity voxel cluster to extract the energy, instantaneous amplitude, and instantaneous phase features of each radar signal sampling point. The radar energy feature is the sum of squares of the amplitudes of the samples within a specified time window, and the instantaneous amplitude and instantaneous phase features are the amplitude and phase components of the complex signal obtained after performing the Hilbert transform on the radar signal, respectively. The calculation formula for the Hilbert transform is: Where t is time, and x(t) is the Hilbert transform of the real signal. This refers to the convolution of the signal x(t) with 1 / πt, where PV represents the Cauchy principal value and τ is the integration constant. These three eigenvalues ​​are stored in three three-dimensional matrices, synthesizing three feature data volumes. Next, the three feature data volumes are linearly normalized to eliminate dimensional differences. Then, wavelet transform is used for multi-feature fusion, specifically the db wavelet. The scale level and maximum value fusion rule are selected, and multi-feature fusion is performed on the longitudinal section of the linearly normalized three-dimensional feature data volume. Finally, inverse wavelet transform is performed to obtain the fused feature data volume. In contrast, traditional feature extraction relies only on single amplitude information, which cannot fully reflect the phase abrupt changes and energy attenuation characteristics of electromagnetic waves at the medium interface, easily leading to ambiguity in material identification. This invention extracts instantaneous multi-dimensional features through Hilbert transform and uses wavelet transform for multi-scale fusion, greatly enriching the feature dimensions and providing a high-fidelity feature data volume for subsequent high-precision material property mapping.

[0066] In the physical semantic mapping and correction step S105, the core feature extraction is first performed: for each target entity voxel cluster, the median of its internal relative permittivity is calculated. And geometric features such as aspect ratio and compactness. Next, a lookup table mapping is performed: preliminary material labeling is performed based on a preset dielectric material semantic mapping library: if the median of the relative permittivity... The material is determined to be air or gas; if The material was determined to be dry backfill soil; if It may exhibit total internal reflection characteristics, indicating the material is either water or metal. Subsequently, topological semantic correction is performed: using spatial enclosure relationships and geometric shape constraints, the initially labeled materials are corrected: Rule 1 (Void Detection): If a voxel cluster is made of air and is completely enclosed in space by voxel clusters of soil or concrete material, and its volume is greater than... If a voxel cluster is made of air or plastic and its geometry is a long, thin cylinder with a principal axis length much greater than its cross-sectional radius, then its semantic label is corrected to an underground pipeline. In contrast, traditional classification methods rely solely on a single dielectric constant value, failing to distinguish between targets with the same dielectric constant but completely different geometric shapes and spatial relationships (such as cavities in air and pipelines in air). This invention, by integrating the median dielectric constant with geometric features and introducing spatial topology correction rules, achieves automatic and accurate conversion from discrete physical fields to structured semantic objects, eliminating the semantic gap.

[0067] To further improve the accuracy of semantic mapping and reduce the ambiguity caused by single physical feature judgment, this embodiment introduces a multi-dimensional joint semantic mapping and multi-level verification mechanism in step S205. Specifically, the multi-dimensional joint semantic mapping includes: 1) Polarization feature extraction: If the ground penetrating radar system has multi-polarization acquisition capability, the reflection intensity ratio of voxel clusters in different polarization directions is extracted as a multi-polarization feature to distinguish isotropic targets (such as cavities) from anisotropic targets (such as directional pipelines); 2) Dispersion characteristic analysis: The dispersion characteristics in multi-band radar data or inversion results are used to help distinguish between water-bearing targets and dry targets. The multi-level verification mechanism includes: 3) Profile continuity verification: For cavity-type targets completely surrounded by the background medium, the semantic label "underground hidden cavity" is assigned only after the continuity verification of adjacent detection profiles confirms that it is not an isolated noise point; 4) Spatial regularity joint judgment: For multiple voxel clusters with similar physical characteristics, if they exhibit a regular spatial arrangement (such as equal spacing or parallel distribution), they are jointly judged as "pipeline group" or "pipe gallery structure"; 5) Surface semantic association verification: The spatial location of voxel clusters is associated with surface semantic targets for verification. If the voxel cluster is located within the projection range below the surface manhole cover target, its confidence as an "artificial facility" is enhanced. In contrast, traditional semantic mapping is prone to misjudging isolated noise points as defects or failing to distinguish between densely arranged pipelines and a single large cavity. This invention introduces auxiliary physical features such as multipolarization and dispersion, and combines multi-level verification with profile continuity, spatial regularity and surface semantic association, which greatly eliminates identification ambiguity and ensures high confidence and high accuracy of semantic mapping in complex underground spaces.

[0068] Please see Figure 2 , Figure 2 This diagram illustrates a comparison of the effects of techniques for segmenting underground space based on dielectric constant field gradient constraints. (Combined with...) Figure 2 This paper provides a direct explanation of the technical effects of the detection signal analysis and processing based on dielectric constant field gradient constraints involved in this invention.

[0069] Figure 2 This paper demonstrates the segmented target entity voxel clusters after processing using the detection signal analysis and processing method based on dielectric constant field gradient constraints of this invention. Semantic boundaries are locked through gradient field calculation, and non-maximum suppression is used to compress the transition zone into a candidate boundary layer of single voxel width. Combined with multi-level potential barriers (hard and soft barriers) constrained region growth, the final segmented target entity has sharp edges and topological closure, completely eliminating the smoothing effect of the transition zone. This achieves high-precision geometric quantization and significantly improves the boundary accuracy of underground target segmentation.

[0070] This application also provides an analysis and processing system 30 for detection signals based on dielectric constant field gradient constraints, the system comprising: The data acquisition and preprocessing module 301 is used to acquire detection signals of various underground media in the underground space, preprocess the detection signals, and obtain three-dimensional echo data tensors.

[0071] The physics field reconstruction module 302 is used to input the three-dimensional echo data tensor into the three-dimensional inversion model and output the three-dimensional dielectric constant voxel field.

[0072] The semantic boundary locking module 303 is used to calculate the gradient magnitude of each voxel in the three-dimensional dielectric constant voxel field, and determine the semantic segmentation boundary formed by the voxels in the three-dimensional dielectric constant voxel field based on the gradient magnitude.

[0073] The entity segmentation module 304 is used to determine at least one target starting point that meets preset conditions in the three-dimensional dielectric constant voxel field according to the semantic segmentation boundary; for each target starting point, diffusion processing is performed on the target starting point as the center to the neighboring voxels adjacent to the target starting point to form a target entity voxel cluster corresponding to the target starting point.

[0074] The physical semantic mapping and correction module 305 is used to extract physical and geometric features for each target entity voxel cluster, perform topological semantic correction according to the preset material semantic mapping relationship, and transform the target entity voxel cluster into a structured semantic object describing each underground medium in the underground space.

[0075] The result output module 306 is used to output a structured data file describing the various underground media in the underground space.

[0076] In some preferred embodiments, the data acquisition and preprocessing module further includes a clutter suppression submodule, which is used to input the preprocessed data into a pre-trained clutter suppression model and output clutter-suppressed three-dimensional echo data.

[0077] The physics field reconstruction module also includes a physics constraint training submodule, which is used to decouple the electric field and magnetic field propagation of the ground-penetrating radar electromagnetic wave propagation equations in the model training process, and combine the predicted electric field and dielectric constant to construct a physical residual formula that satisfies the electromagnetic wave propagation equations as a physical loss function and integrate it into the training loss function of the three-dimensional inversion model.

[0078] The physical field reconstruction module and the semantic boundary locking module are also connected by a super-resolution reconstruction sub-module, which is used to input the three-dimensional dielectric constant voxel field into the pre-trained conditional denoising diffusion model for super-resolution reconstruction and output the high-resolution three-dimensional dielectric constant voxel field after super-resolution reconstruction.

[0079] The semantic boundary locking module also includes a Gabor filter correction submodule, used to utilize different angles and wavelengths. The Gabor filter extracts features from the spatial gradient vector to obtain gradient structure feature parameters, and then performs weighted correction on the spatial gradient vector by combining the interval reflected by the dielectric constant gradient.

[0080] In some preferred embodiments, the restricted solidification segmentation module further includes a gradient barrier refinement processing submodule, which is used to perform non-maximum suppression processing on the gradient magnitude field, construct a multi-level barrier discrimination criterion, and perform backtracking verification and secondary growth judgment on the boundary voxels of each voxel cluster after the region growth is completed.

[0081] The physical semantic mapping and correction module also includes a multi-feature fusion submodule, which performs Hilbert transform on the single-channel signal of each voxel cluster, extracts the energy, instantaneous amplitude and instantaneous phase features of each radar signal sampling point, and uses wavelet transform to perform multi-feature fusion to obtain the fused feature data volume.

[0082] like Figure 4 As described above, the present invention provides a method for operation Figure 3 The electronic device 40 of the analysis and processing system shown may include a memory 401, a processor 402 and a bus, and may also include computer programs stored in the memory 401 and executable on the processor 402, such as the various functional modules of the analysis and processing system.

[0083] The memory 401 includes at least one type of readable storage medium, such as flash memory, portable hard drive, multimedia card, card-type memory (e.g., SD or DX memory), magnetic memory, magnetic disk, optical disk, etc. In some embodiments, the memory 401 can be an internal storage unit of the electronic device 40, such as the portable hard drive of the electronic device 40. In other embodiments, the memory 401 can also be an external storage device of the electronic device 40, such as a plug-in portable hard drive, smart media card (SMC), secure digital card (SD), flash card, etc., equipped on the electronic device 400. Furthermore, the memory 401 can include both internal and external storage units of the electronic device 40. The memory 401 can be used not only to store application software and various types of data installed on the electronic device 40, such as code for analysis and processing methods, but also to temporarily store data that has been output or will be output.

[0084] In some embodiments, the processor 402 may be composed of integrated circuits, such as a single packaged integrated circuit or multiple integrated circuits packaged with the same or different functions, including combinations of one or more central processing units (CPUs), microprocessors, digital processing chips, graphics processors, and various control chips. The processor 402 is the control unit of the electronic device 40, connecting various components of the electronic device 40 through various interfaces and lines. It executes programs or modules stored in the memory 401 and calls data stored in the memory 401 to perform various functions and process data of the electronic device 40.

[0085] The processor 402 executes the operating system and various installed applications of the electronic device 40. The processor 402 executes the applications to implement the steps in the above-described analysis and processing method.

[0086] For example, the computer program may be divided into one or more modules, which are stored in the memory 401 and executed by the processor 402 to complete this application. The one or more modules may be a series of computer program instruction segments capable of performing specific functions, which describe the execution process of the computer program in the electronic device 40.

[0087] The above embodiments are merely illustrative of the principles and effects of the present invention and are not intended to limit the invention. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of the present invention. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in the present invention should still be covered by the claims of the present invention.

Claims

1. A three-dimensional voxel semantic segmentation method for ground-penetrating radar based on dielectric constant field gradient constraints, characterized in that, The method includes: The detection signals of various underground media in the underground space are acquired, the detection signals are preprocessed to obtain a three-dimensional echo data tensor, the three-dimensional echo data tensor is input into a three-dimensional inversion model, and a three-dimensional dielectric constant voxel field is output. Calculate the gradient magnitude of each voxel in the three-dimensional dielectric constant voxel field, and determine the semantic segmentation boundary formed by the voxels in the three-dimensional dielectric constant voxel field based on the gradient magnitude. Based on the semantic segmentation boundary, at least one target starting point satisfying the preset conditions is determined in the three-dimensional dielectric constant voxel field; For each target starting point, a diffusion process is performed from the target starting point to the neighboring voxels adjacent to the target starting point to form a target entity voxel cluster corresponding to the target starting point; Physical and geometric features are extracted for each target entity voxel cluster, and topological semantic correction is performed based on a preset material semantic mapping relationship to transform the target entity voxel cluster into a structured semantic object describing each underground medium in the underground space.

2. The method according to claim 1, characterized in that, The preprocessing of the detected signal to obtain the three-dimensional echo data tensor includes: The detection signal is subjected to multi-channel high-dimensional feature space extraction to obtain multi-channel feature vectors, and the multi-channel feature vectors are subjected to weighted adjustment processing of channel and spatial dimensions to obtain enhanced feature vectors; The enhanced feature vector is subjected to feature reconstruction processing, and the reconstructed three-dimensional echo data tensor is output.

3. The method according to claim 1, characterized in that, The calculation of the gradient magnitude of each voxel in the three-dimensional dielectric constant voxel field, the extraction of high-gradient surfaces based on the gradient magnitudes, and the determination of the semantic segmentation boundary formed by each voxel in the three-dimensional dielectric constant voxel field based on the high-gradient surfaces include: The spatial gradient vector of each voxel in the three-dimensional dielectric constant voxel field is calculated using a three-dimensional gradient operator, and based on the spatial gradient vector... Determine the gradient magnitude , wherein the gradient magnitude The calculation formula is: Wherein, x, y, z represent the electric field intensity vector field of each voxel in the three-dimensional dielectric constant voxel field, and the... Indicates the dielectric constant; Voxels whose gradient magnitude exceeds an adaptive threshold in the three-dimensional dielectric constant voxel field are extracted to form a high gradient surface, and the high gradient surface constitutes the semantic segmentation boundary.

4. The method according to claim 3, characterized in that, Also includes: The spatial gradient vector is used to extract features using filters of different angles and wavelengths to obtain gradient structure feature parameters; The spatial gradient vector is weighted and corrected using the gradient structure characteristic parameters and the interval reflected by the dielectric constant gradient.

5. The method according to claim 1, characterized in that, The step of determining at least one target starting point satisfying preset conditions in the three-dimensional dielectric constant voxel field based on the semantic segmentation boundary includes: Based on the semantic segmentation boundary, at least one neighborhood is determined in the three-dimensional dielectric constant voxel field. By traversing the voxels in the neighborhood, the voxels whose relative dielectric constant is within a preset dielectric constant range and whose gradient magnitude is lower than a preset background noise level are selected as the target starting point.

6. The method according to claim 1, characterized in that, The step of performing a diffusion process, centered on each target starting point, outwards to neighboring voxels adjacent to the target starting point to form a target entity voxel cluster corresponding to the target starting point includes: Calculate the difference in average dielectric constant between the neighboring voxel and the target starting point, and obtain the gradient magnitude of the neighboring voxel; If the difference is less than the preset tolerance and the gradient magnitude is lower than the adaptive threshold, then the neighborhood voxel is added to the target entity voxel cluster corresponding to the target starting point.

7. The method according to claim 1, characterized in that, The step of extracting physical and geometric features for each target entity voxel cluster, performing topological semantic correction based on a preset material semantic mapping relationship, and transforming the target entity voxel cluster into a structured semantic object describing the underground space includes: Based on the physical and geometric features corresponding to the target entity voxel cluster, the target entity voxel cluster is initially material-marked according to the preset dielectric material semantic mapping relationship; By combining at least one of topological semantic correction rules, multi-dimensional joint semantic mapping and multi-level verification mechanisms to perform topological semantic correction, the target entity voxel cluster that has undergone preliminary material labeling is transformed into a structured semantic object. The physical and geometric features include at least one of the median of the relative permittivity, aspect ratio and compactness of each voxel in the target entity voxel cluster.

8. The method according to claim 1, characterized in that, Also includes: Output a structured data file describing each underground medium in the underground space. The structured data file includes: semantic category labels, inferred material properties, geometric dimension parameters, and spatial location coordinates corresponding to each underground medium in the underground space.

9. The method according to claim 1, characterized in that, The detection signal is acquired via ground-penetrating radar.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the three-dimensional voxel semantic segmentation method for ground-penetrating radar based on dielectric constant field gradient constraints as described in any one of claims 1 to 8.

Images