Methods, equipment, and media for constructing 3D models of tunnels and karst caves based on ground-penetrating radar images.

Abstract

This invention belongs to the field of image processing technology, specifically providing a method, device, and medium for constructing a 3D model of a tunnel cave based on ground-penetrating radar (GPR) images. The method includes: obtaining multiple wave train images based on GPR detection reports from advanced geological forecasting; constructing a wave train image dataset; obtaining the vertex coordinates of the 2D boundary contours of the cave's anomalous area; constructing a 3D model of the cave; and deploying the 3D model into mixed reality (MR) glasses to obtain a visual image of the cave's spatial morphology and location within the tunnel face. This invention utilizes an advanced Mask R-CNN instance segmentation algorithm optimized with Transformer and deformable convolutional networks to accurately extract anomalous areas of the cave and construct a high-precision 3D model. Deploying this model into MR glasses provides a more intuitive perception of the cave's 3D morphology, significantly improving the detection accuracy, modeling efficiency, and perception effect of tunnel caves.

Description

Technical Field

[0001] This invention belongs to the field of image processing technology and relates to a method, equipment and medium for constructing a three-dimensional model of a tunnel cave based on ground-penetrating radar images. Background Technology

[0002] In tunnel construction in karst-developed areas, karst caves are a major adverse geological feature affecting construction safety and structural stability. Ground-penetrating radar (GPR), as a non-destructive, high-efficiency, and high-resolution advanced geological prediction method, has been widely used for karst cave detection in front of the tunnel face, and can acquire two-dimensional wave train images and preliminary information on anomalous areas.

[0003] The core requirements for tunnel karst cave detection and modeling are to accurately identify the location and shape of karst caves, quickly construct three-dimensional models, and intuitively present their spatial distribution on the construction site, so as to provide a basis for construction decisions and safety measures.

[0004] Existing tunnel and karst cave detection and modeling techniques mainly fall into three categories: 1) Manual / semi-automatic identification of ground-penetrating radar images: By relying on the experience of technicians to interpret the reflection characteristics of wave train images, or by using basic algorithms such as threshold segmentation and edge detection to extract abnormal areas, preliminary identification of karst caves can be achieved.

[0005] 2) Deep learning-based two-dimensional anomaly region detection: Networks such as CNN and Mask R-CNN are used to perform target detection and instance segmentation on wave train images to improve the efficiency and accuracy of cave recognition, but this only involves two-dimensional plane extraction.

[0006] 3) Traditional 3D modeling of karst caves: It relies heavily on laser scanning, photogrammetry, and geological numerical simulation, requiring extensive manual intervention and parameter input. It is impossible to automatically generate a three-dimensional model directly from two-dimensional ground-penetrating radar data.

[0007] However, existing technologies have the following shortcomings: ① Data chain disruption: The extraction of abnormal areas from ground-penetrating radar wave train images and subsequent 3D modeling techniques were not effectively integrated, resulting in a disjointed process.

[0008] ② Insufficient 3D visualization capabilities: Existing methods are unable to directly construct 3D models from exploration data, and lack real-time and intuitive presentation of cave spatial morphology.

[0009] ③Low level of automation: It relies on manual intervention, has low modeling efficiency, and is difficult to meet the needs of rapid and accurate modeling in tunnel construction.

[0010] ④ Limited practicality: There is a lack of visualization tools that are adapted to the needs of the construction site, making it difficult to provide intuitive guidance for actual operations. Summary of the Invention

[0011] This invention aims to solve the technical problems in tunnel geological radar cave detection, such as strong subjectivity of manual interpretation, low accuracy of anomaly identification, inability to directly construct three-dimensional models from two-dimensional data, low degree of automation in modeling, and lack of immersive three-dimensional visualization methods at construction sites. It provides a method, equipment, and medium for constructing three-dimensional models of tunnel caves based on geological radar images, so as to achieve efficient and intelligent processing of the entire process from automatic extraction of geological radar data, accurate identification of caves, three-dimensional modeling to real-time MR visualization.

[0012] This invention provides a method for constructing a three-dimensional model of a tunnel or karst cave based on ground-penetrating radar images, comprising the following steps: Step 1: Based on the ground-penetrating radar detection report in advanced geological forecasting, obtain multiple wave train images; Step 2: Annotate the anomalous areas of the caves in multiple wave train images to obtain the annotation information of the anomalous areas of the caves; construct a wave train image dataset based on the wave train images and the annotation information of the anomalous areas of the caves. A cave instance segmentation model based on frequency-spatial domain dual-branch feature enhancement and dynamic adaptive receptive field network was constructed. The cave instance segmentation model based on frequency-spatial domain dual-branch feature enhancement and dynamic adaptive receptive field network was pre-trained using the wave train image dataset to obtain the trained cave instance segmentation model. Mask images of anomalous regions in karst caves are extracted from wave train images based on a trained karst cave instance segmentation model. The cave instance segmentation model based on frequency-spatial dual-branch feature enhancement and dynamic adaptive receptive field network includes a frequency-spatial dual-branch feature extraction network, a dynamic adaptive receptive field convolution module, a cave boundary perception map neural network module, and a cave morphology prior constraint loss function. Step 3: Perform data formatting on the mask image of the cave anomaly area to obtain the vertex coordinates of the two-dimensional boundary contour of the cave anomaly area; Three-dimensional contour point cloud data of the karst cave anomaly region is constructed based on the vertex coordinates of the two-dimensional boundary contour. Step 4: Introduce a cave modeling algorithm and construct a 3D model of the cave using 3D point cloud data of the cave's anomalous areas; Step 5: Deploy the 3D model of the cave into the mixed reality glasses to obtain a visual image of the spatial morphology and location of the cave within the tunnel face.

[0013] Furthermore, the specific process of step one is as follows: The entire ground-penetrating radar detection report is read and its format is recognized to obtain text and image information; The text information includes tunnel face mileage information, karst anomaly description information, geological condition description, karst cave risk level, and engineering treatment suggestions; and the extracted text information is structured, organized, and stored to form standardized data. The image information includes ground-penetrating radar wave train images; and the image information is preprocessed to obtain preprocessed wave train images.

[0014] Furthermore, the preprocessing of image information includes: image denoising, contrast enhancement, edge smoothing, and invalid region removal.

[0015] Furthermore, the specific process of step two is as follows: Step 2.1: Construct a wave train image dataset based on wave train image information and cave anomaly area annotation information; Step 2.2: Construct a cave instance segmentation model based on frequency-spatial domain dual-branch feature enhancement and dynamic adaptive receptive field network; The cave instance segmentation model based on frequency-spatial dual-branch feature enhancement and dynamic adaptive receptive field network includes a frequency-spatial dual-branch feature extraction network, a dynamic adaptive receptive field convolution module, a cave boundary perception map neural network module, and a cave morphology prior constraint loss function. Step 2.3: Use a cave instance segmentation model based on frequency-spatial dual-branch feature enhancement and dynamic adaptive receptive field network to extract the abnormal regions in the wave train image, and obtain the cave abnormal region mask in the ground-penetrating radar wave train image and the statistical data corresponding to the mask.

[0016] Furthermore, the specific process of constructing the wave train image dataset is as follows: Multiple wave train images containing information on cave size, shape, and distribution type were captured from the ground-penetrating radar report, thus obtaining wave train image information; The abnormal boundaries, abnormal ranges, and abnormal types of the karst cave abnormality areas are marked to obtain the karst cave abnormality area marking information; The wave train image information and the cave anomaly area annotation information are converted into a format to obtain the wave train image dataset.

[0017] Furthermore, the frequency-spatial dual-branch feature extraction network includes a spatial branch, a frequency branch, and a dual-branch adaptive fusion module; The spatial domain is divided into four residual blocks, each residual block consisting of two... The system consists of convolutional layers and incorporates batch normalization and ReLU activation functions. Furthermore, a boundary enhancement module is introduced between residual blocks. Boundary features are extracted using the Sobel edge detection operator and fused with the backbone features to obtain the local spatial features of the wave train image. The frequency domain branch is used to convert the wave train image to the frequency domain and extract the feature representation of the cave reflection signal in the frequency domain. The dual-branch adaptive fusion module obtains global statistical information of the spatial and frequency branches through global average pooling, and learns the fusion weights of the spatial and frequency branches through a fully connected layer to obtain the fusion features of the spatial and frequency branches. .

[0018] Furthermore, the specific process for obtaining the characteristic representation of the cave reflection signal is as follows: A two-dimensional fast Fourier transform is performed on the input wave train image to convert the wave train image from spatial domain data to frequency domain data; Logarithmic transformation and normalization are performed on the frequency domain data to obtain frequency domain data of different frequency components; Design a frequency domain attention module to weight frequency domain data of different frequency components using a learnable weight matrix to obtain frequency domain features; The frequency domain features are converted into spatial domain features by inverse Fourier transform, and the spatial domain features are fused with local spatial features to obtain the characteristic representation of the cave reflection signal.

[0019] Furthermore, fusion features The expression is as follows: ; ; ; in, The feature map of the spatial branch, Feature map of frequency domain branch, This is a global average pooling operation. It is a fully connected layer. It is the Sigmoid activation function. The fusion weights for the spatial branches, The fusion weights are for the frequency domain branches.

[0020] Furthermore, the dynamic adaptive receptive field convolution module includes a multi-scale parallel convolution structure, a scale-selective attention mechanism, and a hole convolution enhancement structure; The multi-scale parallel convolutional structure: Three convolutional kernels of different scales are designed to process the input features in parallel, resulting in three sets of feature maps of different scales; The scale-selection attention mechanism dynamically selects the corresponding convolution scale based on three sets of feature maps at different scales, obtains global statistical information of the input features through global average pooling, and learns the weights of the convolution kernels at each scale through two fully connected layers. ; The hole convolution enhancement structure: for The convolution kernel uses dilated convolution with a dilation rate of 2, thus expanding the actual receptive field to [value missing]. ;for The convolution kernel uses a standard convolution with a dilation rate of 1.

[0021] Furthermore, the cave boundary perception graph neural network module is used to convert the wave train image into a graph structure; the specific process is as follows: The Simple Linear Iterative Clustering algorithm is used to segment the wave train image into several superpixel regions, with each superpixel region serving as a graph node. For adjacent superpixel regions, establish edge connections between the corresponding graph nodes; The initial features of each graph node are obtained. The initial features of each graph node consist of the feature statistics of the corresponding superpixel region, and the initial features of each graph node include color mean, texture features and boundary intensity.

[0022] Furthermore, the specific process for obtaining the boundary strength is as follows: Boundary strength is obtained by calculating the average gradient value of the superpixel region boundary using the Sobel operator.

[0023] Furthermore, the prior constraint loss function for cave morphology The expression is as follows: ; ; ; ; For segmentation loss; This is the boundary smoothing loss, used to constrain the smoothness of the segmentation boundary and penalize drastic changes in the boundary. The morphological consistency loss is used to constrain the segmentation results to conform to the typical morphological characteristics of karst caves. This is the frequency domain consistency loss, used to constrain the consistency of frequency domain features between the predicted segmented region and the ground truth labeled region; Boundary smoothing loss The weighting coefficients, For morphological consistency loss The weighting coefficients, For frequency domain consistency loss Weighting coefficients; For segmentation mask; For the Laplace operator; This refers to the number of boundary pixels; This is a closure loss, used to penalize cases where the segmented region is not closed. Uniformity loss is used to constrain signal uniformity within the segmented region; Input image; For predicting the mask; True label mask; This is element-wise multiplication; For Fast Fourier Transform.

[0024] Furthermore, the specific process for obtaining the two-dimensional boundary contour data is as follows: Binarize the mask image, extract the abnormal areas of the karst caves, set the mask part to white, and set the background to black; Edge detection algorithms are used to identify the boundaries of abnormal areas in karst caves; Using a polygon fitting method, a smooth closed contour line is generated for the boundary of the karst cave anomaly area; the vertex coordinates of the two-dimensional boundary contour are obtained based on the closed contour line.

[0025] Furthermore, the specific process of constructing the three-dimensional contour point cloud data of the karst cave anomaly region based on the vertex coordinates of the two-dimensional boundary contour is as follows: Based on the pixel ratio of the wave train image, the vertex coordinates of the two-dimensional boundary contour are converted into a coordinate system of actual geological dimensions; Based on the distance between the vertex coordinates of the two-dimensional boundary contour of two adjacent wave train images of the same working face, the coordinate system of the actual geological dimensions is extended to a three-dimensional coordinate system to obtain the three-dimensional contour point cloud data of the cave anomalous area.

[0026] Furthermore, the specific process of step four is as follows: The three-dimensional contour point cloud data of the abnormal area of ​​the karst cave is filtered and denoised to obtain preprocessed point cloud data. The preprocessed 3D point cloud data is transformed into a continuous triangular mesh surface through a spatial subdivision algorithm, and a closed, hole-free, non-overlapping, and topologically reasonable karst cave 3D solid model is constructed. The real-world attribute information of the tunnel face is mapped onto the surface of the 3D solid model of the karst cave. Then, the 3D solid model of the karst cave is meshed to generate a geometrically accurate, structurally sound, and highly compatible 3D model of the karst cave.

[0027] As a further aspect of the present invention, the present invention also provides an electronic device, including a memory, one or more processors, and one or more programs stored in the memory, said one or more programs including instructions for executing the method for constructing a three-dimensional model of a tunnel cave based on ground-penetrating radar images as described above.

[0028] As a further aspect of the present invention, the present invention also provides a medium comprising one or more programs executable by one or more processors of a device, the one or more programs comprising instructions for performing the method for constructing a three-dimensional model of a tunnel cave based on ground-penetrating radar images as described above.

[0029] Compared with the prior art, the present invention has the following beneficial effects: This invention utilizes a 3D modeling method for karst caves based on ground-penetrating radar images. By employing an advanced Mask R-CNN instance segmentation algorithm optimized with Transformer and deformable convolutional networks, it accurately extracts abnormal regions of karst caves and constructs a high-precision 3D model. Furthermore, by deploying this model into MR glasses, it provides a more intuitive perception of the 3D morphology of karst caves, significantly improving the detection accuracy, modeling efficiency, and perception effect of tunnel karst caves, thereby solving the technical problems in the prior art.

[0030] In addition to the objectives, features, and advantages described above, the present invention has other objectives, features, and advantages. The invention will now be described in further detail with reference to the figures. Attached Figure Description

[0031] The accompanying drawings, which form part of this application, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is a flowchart illustrating a method for constructing a three-dimensional model of a tunnel cave based on ground-penetrating radar images, as described in Embodiment 1 of the present invention. Figure 2 This is a framework diagram of the FSD-DARF model in Embodiment 1 of the present invention; Figure 3 This is a schematic diagram of the palm surface observed by wearing MR glasses in the experimental example of this invention; Figure 4 This is a schematic diagram of the cavern at the face of the palm observed by wearing MR glasses in an experimental example of this invention. Detailed Implementation

[0032] To make the above-mentioned objectives, features, and advantages of the present invention clearer and easier to understand, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be noted that the accompanying drawings of the present invention are all in a simplified form and use non-precise proportions, and are only used to facilitate and clearly assist in illustrating the implementation of the present invention; the "several" mentioned in the present invention are not limited to the specific number shown in the examples in the accompanying drawings; the orientations or positional relationships indicated by terms such as "front," "middle," "rear," "left," "right," "up," "down," "top," "bottom," and "center" mentioned in the present invention are all based on the orientations or positional relationships shown in the accompanying drawings of the present invention, and do not indicate or imply that the device or component referred to must have a specific orientation, nor should they be construed as limitations on the present invention.

[0033] Example 1: See Figure 1 As shown, the present invention provides a method for constructing a three-dimensional model of a tunnel cave based on ground-penetrating radar images, comprising the following steps: Step 1: Obtain multi-wave train images by identifying the geological radar detection reports in advanced geological forecasting.

[0034] Preferably, the specific process of step one is as follows: Step 1.1: Read and recognize the overall format of the ground-penetrating radar detection report to obtain text and image information, so as to avoid interference between different types of data and ensure the accuracy and stability of subsequent information extraction; Step 1.2: Identify and match keywords in the text information to extract key engineering information; the key engineering information includes tunnel face mileage information, karst anomaly description information, geological condition description, karst cave risk level, and engineering treatment suggestions; The extracted text information is structured, organized, and stored to form standardized data that can be directly accessed, providing basic data support for subsequent information binding, annotation, display, and construction decision-making in 3D models; Step 1.3: Automatically identify, locate, and extract multiple wave train images from the image information, and preprocess the wave train images to obtain preprocessed wave train images.

[0035] As a further aspect of this embodiment, the specific methods for preprocessing the wave train image include image denoising, contrast enhancement, edge smoothing, and invalid region removal, in order to eliminate interference factors such as report background, text overlay, and image distortion, making the wave train image features clearer and ensuring that the subsequent abnormal region extraction algorithm can stably and accurately identify the cave reflection features.

[0036] Preferably, the specific process for preprocessing ground-penetrating radar wave train images refers to existing technologies.

[0037] Step 2: The cave instance segmentation algorithm based on frequency-spatial dual-branch feature enhancement and dynamic adaptive receptive field network (FSD-DARF algorithm) is used to accurately extract the cave anomalous areas from the ground-penetrating radar wave train image.

[0038] Preferably, the specific process of step two is as follows: Step 2.1: Annotate the cave anomaly areas in multiple wave train images to obtain cave anomaly area annotation information; Based on wave train images and annotation information of cave anomaly areas, a wave train image dataset is constructed. The wave train image dataset needs to cover a variety of geological environments and cave features, including the diversity of cave size, shape, and distribution type. Step 2.2: Construct a cave instance segmentation model based on frequency-spatial domain dual-branch feature enhancement and dynamic adaptive receptive field network (FSD-DARF model). This model adopts a three-stage architecture of "dual-branch feature extraction + dynamic receptive field adaptation + boundary awareness optimization". Its overall network structure includes a frequency-spatial domain dual-branch feature extraction network (FSD-Net), a dynamic adaptive receptive field convolutional module (DARF-Conv), a cave boundary awareness graph neural network module (Cavity-Boundary GNN), and a cave morphology prior constraint loss function. Step 2.3: Extract the anomalous regions from the wave train image based on the FSD-DARF model, and obtain the mask of the cave anomalous region in the ground-penetrating radar wave train image and the statistical data of the mask.

[0039] As a further embodiment, the specific process of step 2.1 is as follows: Step 2.1.1: Collection of wave train images; Multiple wave train images are extracted from the ground-penetrating radar detection report, and each wave train image is ensured to meet the requirements of model training in terms of resolution and quality. Step 2.1.2: Accurately label the anomalous areas of the karst caves in the wave train image, and clarify the anomalous boundaries, anomalous range and anomalous type; Step 2.1.3: Convert the data format of the wave train image information and the annotation information of the cave anomaly area into a standard dataset that can be used for model training to ensure that the model has stable and reliable recognition capabilities under different working conditions.

[0040] As a further solution in this embodiment, based on the unique frequency domain characteristics of wave train images (the reflection signals from caves exhibit specific frequency distribution patterns in the frequency domain, and caves of different types and sizes have distinguishable feature representations in the frequency domain), existing deep learning methods mainly focus on spatial domain feature extraction, neglecting the frequency domain information contained in wave train images. This invention introduces frequency domain analysis into the cave segmentation task of ground-penetrating radar wave train images for the first time, extracting both spatial and frequency domain features simultaneously through a dual-branch network structure to achieve a more comprehensive feature representation. Specifically, the frequency-spatial dual-branch feature extraction network (FSD-Net) includes a spatial branch, a frequency branch, and a dual-branch adaptive fusion module.

[0041] Preferably, the spatial branch employs an improved residual network as its backbone to extract local spatial features from the wave train image. The spatial branch is designed to capture the geometric features of the cave boundary, including its location, orientation, and curvature. Specifically, the spatial branch contains four residual blocks, each consisting of two 3×3 convolutional layers, with batch normalization and ReLU activation functions introduced. To enhance the perception of cave boundary details, a boundary enhancement module is introduced between the residual blocks, extracting boundary features using the Sobel edge detection operator and fusing them with the backbone features.

[0042] Preferably, the frequency branch is used to convert the wave train image to the frequency domain, and extract the feature representation of the cave reflection signal in the frequency domain. Specifically, the specific steps for obtaining the feature representation of the cave reflection signal based on the frequency domain distribution are as follows: A two-dimensional fast Fourier transform (2D-FFT) is performed on the input wave train image to convert the wave train image from spatial domain data to frequency domain data; Logarithmic transformation and normalization are performed on frequency domain data to enhance the visualization of low-frequency components; A frequency domain attention module is designed to weight different frequency components using a learnable weight matrix, highlighting the frequency range in which the cave reflection signal is located. The processed frequency domain features are converted into spatial domain features by inverse Fourier transform (IFFT), and the spatial domain features are fused with local spatial features to obtain the feature representation of the cave reflection signal.

[0043] Preferably, to effectively integrate the features of the spatial and frequency domain branches, this invention designs a dual-branch adaptive fusion module. This module obtains global statistical information of the features of each branch through global average pooling, and learns the fusion weights of the spatial and frequency domain branches through a fully connected layer to obtain the fused features of the spatial and frequency domain branches. .

[0044] Further optimized, fusion features The expression is as follows: ; ; ; in, Feature map of spatial branch; Feature maps for frequency domain branches; This is a global average pooling operation; It is a fully connected layer; Use the Sigmoid activation function; The fusion weights for the spatial domain branches are used to measure the feature maps of the spatial domain branches. The importance of fusion features The value range is [0,1], derived from the spatial branch feature map. It is obtained by global average pooling and a fully connected layer, followed by the Sigmoid activation function; The fusion weights for the frequency domain branches are used to measure the feature maps of the frequency domain branches. The importance of fusion features The value range is [0,1], derived from the frequency domain branch feature map. It is obtained by global average pooling and a fully connected layer, followed by the Sigmoid activation function; This is a fusion feature map of the spatial domain branch and the frequency domain branch, derived from the spatial domain branch feature map. Frequency domain branch feature map The result is obtained by weighting and summing the results according to their respective fusion weights.

[0045] Further optimized feature maps of spatial branches The specific acquisition process is as follows: The preprocessed wave train image is input into the spatial branch and sequentially undergoes convolution operations and feature extraction in four residual blocks. The number of channels in the output feature map of each residual block is 256, 512, 1024, and 2048, respectively. After passing through the boundary enhancement module, the feature map output by the fourth residual block is taken as the final output of the spatial branch. Its spatial resolution is 1 / 16 of the input image, and the number of channels is 2048.

[0046] A further preferred feature map of the frequency domain branch The specific acquisition process is as follows: After performing a 2D-FFT transform on the input wave train image, frequency domain data is obtained, with the same dimensions as the input image. Following logarithmic transformation and normalization, the frequency components are weighted using a frequency domain attention module. The weighted frequency domain data is then inversely transformed back to the spatial domain using an IFFT to obtain the spatial branch feature map. Frequency domain feature maps with the same spatial resolution and number of channels This is so that the subsequent dual-branch adaptive fusion module can perform feature fusion.

[0047] As a further aspect of this embodiment, since the scale of karst caves varies greatly in ground-penetrating radar wave train images, ranging from small cavities (diameter less than 0.5 meters) to large karst caves (diameter greater than 5 meters), and traditional fixed receptive field convolution kernels are difficult to adapt to the characteristics of karst caves of different scales simultaneously (small receptive field convolution kernels (such as 3×3) are insufficient in perceiving the overall morphology of large karst caves, while large receptive field convolution kernels (such as 7×7) have a weak ability to capture the boundary details of small karst caves), this invention designs a dynamic adaptive receptive field convolution module (DARF-Conv) to achieve dynamic adaptive adjustment of the receptive field size of the convolution kernel.

[0048] Preferably, the core design of the dynamically adaptive receptive field convolution module is to achieve dynamic adaptive adjustment of the receptive field size of the convolution kernel, which specifically includes a multi-scale parallel convolution structure, a scale-selective attention mechanism, and a hole convolution enhancement structure: Multi-scale parallel convolutional structure: Designing three different scale convolutional kernels to process fused features in parallel. : Convolution kernels are used to capture detailed features of small caverns. Convolution kernels are used to extract morphological features of medium-scale caves. Convolutional kernels are used to perceive the overall structure of large caverns. Three scales of convolutional kernels share input features, perform convolution operations independently, and generate three sets of feature maps at different scales. Scale-selective attention mechanism: based on fusion features To dynamically select an appropriate convolutional scale, a scale-selection attention mechanism was designed. This mechanism obtains global statistical information of the input features through global average pooling, and then learns the weights of the convolutional kernels at each scale through two fully connected layers. ; Further optimization, weight The expression is as follows: ; in, For input features, As the first fully connected layer, This is the second fully connected layer. for The weights of the convolution kernel, for The weights of the convolution kernel, for The weights of the convolution kernel.

[0049] Dilated Convolution Enhancement Structure: To further expand the receptive field without increasing the number of parameters, a dilated convolution enhancement structure is introduced based on the multi-scale parallel convolution structure; for The convolution kernel uses dilated convolution with a dilation rate of 2, thus expanding the actual receptive field to [value missing]. ;for The convolution kernel uses a standard convolution with a dilation rate of 1. The introduction of the dilated convolution enhancement structure enables the DARF-Conv module to perceive a wider range of contextual information, which has a significant advantage in identifying karst caves with greater burial depth and weaker signals.

[0050] As a further solution in this embodiment, since the boundary of a karst cave exhibits a gradual transition characteristic in the ground-penetrating radar wave train image, the boundary area often has problems such as blurring and breakage. Traditional convolutional neural networks, when processing the boundary area, are difficult to effectively utilize global boundary information due to the locality of convolution operations. Therefore, this invention applies graph neural networks to the karst cave segmentation task of ground-penetrating radar wave train images for the first time, and designs a karst cave boundary perception graph neural network module (Cavity-Boundary GNN).

[0051] Preferably, the process of converting the wave train image into a graph structure using the cave-boundary perception graph neural network module (Cavity-Boundary GNN) is as follows: ① Superpixel segmentation: The SLIC (Simple Linear Iterative Clustering) algorithm is used to segment the wave train image into several superpixel regions, with each superpixel region serving as a graph node. Superpixel segmentation can effectively reduce the number of graph nodes, lower computational complexity, and preserve the local structural information of the image.

[0052] ② Adjacency Establishment: For adjacent superpixel regions, establish an edge connection between the corresponding graph nodes. The adjacency relationship is determined based on the boundary sharing of the superpixel regions. If two superpixel regions have boundary contact, then there is an edge between their corresponding graph nodes.

[0053] ③ Node feature initialization: The initial features of each graph node consist of the feature statistics of the corresponding superpixel region, including color mean, texture features, boundary strength, etc.; the boundary strength is obtained by calculating the average gradient value of the superpixel region boundary using the Sobel operator.

[0054] Further preferably, to obtain boundary strength, this invention designs a boundary-aware message passing mechanism to enhance the feature propagation capability of boundary nodes. In the Cavity-Boundary GNN module, the graph structure constructed above is input into the graph neural network, and a boundary-aware message passing mechanism is designed. This mechanism is set in the node feature update layer of the graph neural network to perform feature propagation and aggregation on the adjacency relationships of graph nodes, thereby enhancing the feature propagation capability of boundary nodes. The message passing formula is: ; ; ; in, For nodes In the The feature vector of the layer, For nodes The set of adjacent nodes, The weight matrix is ​​a learnable matrix. For nodes To the node Attention weights and , For the boundary enhancement bias term, For nodes Boundary strength, For nodes Boundary strength, These are learnable weight parameters. Layer index for boundary-aware message passing mechanism. , This represents the total number of message passing layers. For nodes In the Feature vectors after layer message passing For nodes The initial feature vector (i.e., the 14-dimensional feature vector obtained in the node feature initialization step); For nodes In the The feature vector is updated after layer message passing; For nodes The set of adjacent nodes; This represents the dimension of the node feature vector. In this embodiment, it represents the number of message passing layers. Each layer corresponds to an independent weight matrix. , No. Weight matrix for layer message passing The values ​​are learned during model training via backpropagation, and the initial values ​​are uniformly initialized using Xavier. Specifically, Node In the eigenvectors of the layer from Linear mapping of dimensional feature space to A 3D feature space is used to transform and update node features.

[0055] The boundary enhancement bias term makes nodes with higher boundary strength have greater influence in message passing, thereby enhancing the propagation effect of boundary information.

[0056] As a further solution in this embodiment, based on the specific morphological characteristics of karst caves in ground-penetrating radar wave train images (the boundaries usually present closed curves, the internal signal is relatively uniform, and the morphology has a certain continuity and smoothness), traditional segmentation loss functions (such as cross-entropy loss and Dice loss) only focus on pixel-level classification accuracy and do not consider the morphological constraints of karst caves; therefore, this invention designs a karst cave morphology prior constraint loss function. By incorporating the morphological features of karst caves into the model training process, the consistency and accuracy of segmentation results can be improved.

[0057] Preferably, the prior constraint loss function for cave morphology The expression is as follows: ; ; ; ; in, The segmentation loss is a combination of Focal Loss and Dice Loss. Segmentation loss The expression is as follows: ; ; ; in, Focal loss function The balance coefficient, in this embodiment, Choose a value of 0.5; Dice loss function The balance coefficient, in this embodiment, Choose a value of 0.5; The true label for each pixel (1 represents the cave area, 0 represents the background area); To predict the probability that this pixel is a cave area for the model; In order to focus on the parameters, in this embodiment, Choose a value of 2; Number of boundary pixels; Focal loss function By reducing the loss weight of easily classified samples, the model can focus more on boundary pixels that are difficult to classify. As a smoothing factor, in this embodiment, A value of 1 is used to avoid the denominator being zero; Dice loss function The overall integrity of the segmented region is optimized by calculating the overlap between the predicted region and the actual region.

[0058] This is the boundary smoothing loss, used to constrain the smoothness of the segmentation boundary and penalize drastic changes in the boundary. The morphological consistency loss is used to constrain the segmentation results to conform to the typical morphological characteristics of karst caves. This is the frequency domain consistency loss, used to constrain the consistency of frequency domain features between the predicted segmented region and the ground truth labeled region; To divide the loss The weighting coefficients, Boundary smoothing loss The weighting coefficients, For frequency domain consistency loss Weighting coefficients; For segmentation mask; For the Laplace operator; This refers to the number of boundary pixels; This is a closure loss, used to penalize cases where the segmented region is not closed. Uniformity loss is used to constrain signal uniformity within the segmented region; Input image; For predicting the mask; True label mask; This is element-wise multiplication; For Fast Fourier Transform; Boundary smoothing loss Encourage the segmentation boundaries to have smooth curves and avoid jagged boundaries; Caves typically present closed regions with relatively uniform internal signals. This invention addresses this by incorporating a morphological consistency loss mechanism. The constraint segmentation results conform to the typical morphological characteristics of karst caves.

[0059] As a further aspect of this embodiment, the FSD-DARF model is pre-trained using the Boehringer Intensity Framework (BIF) image dataset to obtain a pre-trained cave instance segmentation model. Specifically, the steps for pre-training the FSD-DARF model using the BIF image dataset are as follows: The COCO format wave train image dataset was used as input, with wave train image information and cave anomaly area annotation information as input. Weights were pre-trained using the COCO format Porsche image dataset. Transfer learning is employed to improve training efficiency; the optimizer uses the AdamW optimizer with a multi-step decay learning rate adjustment strategy, and the loss function adopts a cave morphology prior constraint loss function. ; After pre-training, the segmentation accuracy (mAP) and recall of the trained cave instance segmentation model are evaluated using a test set.

[0060] This invention, through reasonable optimization strategies and loss function adjustments, enables the trained cave instance segmentation model to fully learn the reflection patterns, boundary features, and morphological features of caves in wave train images, ultimately obtaining a stable, efficient, and high-precision cave segmentation model.

[0061] As a further embodiment, the specific process of intelligent extraction of abnormal regions from wave train images based on the FSD-DARF model is as follows: Input: The wave train image automatically extracted from the previous stage is used as input, and after preprocessing, it is imported into the trained FSD-DARF model; Processing: The FSD-DARF model automatically generates segmentation results, including: Anomaly mask: Used to mark detected abnormal areas of caverns; Statistical data: including pixel area and coordinate range of abnormal regions; Output: The output image is the same size as the input image. Abnormal areas are covered with masks of different colors to visually show the location and extent of the cave area. Data storage: Save the test results as the following structured data: Masked image: Segmentation results of marked anomalous areas in the karst cave, PNG image format; Statistical data: Records the geometric information of the masked area to provide data support for subsequent 3D modeling. This completes the process of extracting abnormal regions from the wave train images, automatically acquiring the mask of the cave abnormal region in the ground-penetrating radar wave train images and the statistical data of the mask, providing data support for the subsequent 3D modeling stage.

[0062] Step 3: Perform data formatting on the mask image of the cave anomaly area to obtain two-dimensional boundary contour data of the cave anomaly area that can be used for 3D modeling; Two-dimensional boundary contour data is converted into three-dimensional contour point cloud data to achieve accurate conversion from two-dimensional image data to three-dimensional spatial data.

[0063] Preferably, the specific process for obtaining two-dimensional boundary contour data is as follows: Binarization: The mask image is binarized to extract the abnormal areas of the karst caves. The masked area is set to white and the background is set to black. Edge detection: Edge detection algorithms are used to identify the boundaries of abnormal areas in karst caves; Contour fitting: Using polygon fitting, a smooth closed contour line is generated for the boundary of the abnormal area of ​​the cave. The vertex coordinates of the two-dimensional boundary contour are obtained based on the closed contour line. This ensures that the contour truly reflects the actual shape of the cave and avoids problems such as edge breakage, discontinuity, and burrs from affecting the accuracy of subsequent modeling. Data output: The vertex coordinates of the two-dimensional boundary contour are stored in JSON format.

[0064] Preferably, the specific process for constructing the three-dimensional contour point cloud data of the karst cave anomaly region based on the vertex coordinates of the two-dimensional boundary contour is as follows: Contour coordinate normalization: Based on the pixel ratio of the wave train image, the vertex coordinates of the two-dimensional boundary contour are converted into a coordinate system of actual geological dimensions (in meters). Layered boundary generation: Based on the distance between the vertex coordinates of the two-dimensional boundary contour of two adjacent wave train images of the same working face, the coordinate system of the actual geological dimensions is extended to a three-dimensional coordinate system to obtain the three-dimensional contour point cloud data of the cave anomalous area. Output file format: 3D contour point cloud file, OBJ format.

[0065] This completes the processing of the contour data of the cave anomaly area, and a preliminary three-dimensional model is constructed based on the boundary contour data of the cave anomaly area from two wave train images.

[0066] Step 4: Introduce a cave modeling algorithm and use the 3D point cloud data of the cave's abnormal areas to construct a 3D model of the cave.

[0067] Preferably, the specific process of step four is as follows: Step 4.1: Filter and denoise the three-dimensional contour point cloud data of the karst cave anomaly area to eliminate boundary extraction errors or noise, and obtain the preprocessed point cloud data. Step 4.2: Using a combination of triangulation and voxel reconstruction, surface fitting and mesh generation are performed on the preprocessed 3D point cloud data. Specifically, the preprocessed 3D point cloud data is transformed into a continuous triangular mesh surface through a spatial subdivision algorithm to construct a closed, hole-free, non-overlapping, and topologically sound 3D solid model of the cave, so that the 3D solid model of the cave can realistically reflect the spatial morphology, volume, and extension direction of the cave. Mapping the real-world attribute information of the tunnel face onto the 3D solid model of the karst cave enhances the visualization effect of the 3D solid model of the karst cave. Step 4.3: Optimize the mesh of the 3D solid model of the cave. Mesh optimization includes mesh smoothing, simplification, hole repair, and normal correction. While maintaining the accuracy of the cave shape, reduce the complexity of the model, reduce the number of faces, and improve the running efficiency of the 3D solid model of the cave during rendering, loading, and interactive display, so as to adapt it to the real-time rendering requirements of mixed reality devices. Step 4.4: Generate a geometrically accurate, structurally sound, and highly compatible 3D model of the karst cave, which can be directly applied to subsequent mixed reality deployment, engineering demonstration, safety analysis, and construction guidance.

[0068] Step 5: Deploy the constructed 3D model of the cave into MR glasses to achieve real-time perception and interactive display of the cave's spatial morphology.

[0069] Preferably, step five aims to deploy the constructed 3D model of the cave into MR (Mixed Reality) glasses to achieve real-time perception and interactive display of the cave's spatial morphology, ultimately achieving an immersive on-site perception effect. Specifically, the process is as follows: Step 5.1: Convert the 3D model of the cave into a 3D file format compatible with MR glasses (GLTF, which supports efficient loading and real-time rendering) to adapt to the operating environment of MR glasses; Step 5.2: Using model lightweighting technology, the number of polygons in the converted 3D cave model is dynamically adjusted according to the computing power of the glasses to ensure smooth operation; The textures of the lightweighted 3D cave model are compressed in an appropriate format to reduce file size while maintaining visualization quality; Step 5.3: To provide an immersive experience, the 3D model of the cave needs to be integrated with the tunnel environment; Step 5.4: To enhance user experience, design intuitive interactive functions to allow users to view cave information in real time.

[0070] More preferably, the specific process of step 5.3 is as follows: 1. Tunnel Scene Reconstruction: (1) Based on the actual tunnel design drawings, construct a tunnel virtual environment model that is consistent with the real scene.

[0071] (2) The spatial alignment between the user's position and the tunnel model is calibrated by using the built-in sensor of the MR glasses.

[0072] 2. Cave model positioning: The 3D model of the cave is accurately positioned to the corresponding location of the virtual tunnel model based on the mileage information of its working face and the location data of the ground-penetrating radar wave train image.

[0073] More preferably, the specific process of step 5.4 is as follows: 1. Cave Information Display: (1) Bind the karst information and treatment suggestions extracted from the geological radar report to the relevant locations in the karst cave 3D model fused with the tunnel environment.

[0074] (2) When the user wears MR glasses to observe the tunnel face area, the 3D model of the karst cave fused with the tunnel environment and its related information will automatically appear in the field of view and be marked by semi-transparent labels.

[0075] 2. Cave Selection and Details: (1) Users select a cave model by using gesture control.

[0076] (2) The selected cave will be highlighted and an information panel will pop up, displaying karst description information and treatment suggestions for users to view in detail.

[0077] Through the above deployment, users can enter a virtual tunnel scene by wearing MR glasses. When looking at the tunnel face area, they can view the 3D model and related information of the karst cave in real time, intuitively feel the spatial shape and location of the cave, and obtain detailed karst descriptions and treatment suggestions through interactive functions. This method effectively improves the intuitiveness of geological information and decision-making efficiency in tunnel construction, providing strong technical support for safe construction.

[0078] Experimental example: Step 1: Obtaining the ground-penetrating radar detection report; The specific method is as follows: the tunnel advanced geological prediction unit uses the ground-penetrating radar method to detect the tunnel face, and requires the layout of two survey lines to form a complete advanced geological prediction report in PDF format.

[0079] Step 2: Automatic acquisition of ground-penetrating radar detection report data; The specific method is as follows: Receive advanced geological forecast reports from tunnel advanced geological forecasting units. Import the data from the ground-penetrating radar (GPR) detection report into the automatic data acquisition system. By identifying the GPR detection report in the advanced geological forecast, obtain the tunnel face mileage information, wave train images, karst information, and treatment recommendations. The tunnel face mileage information, karst information, and treatment recommendations are text data, while the wave train images are image data.

[0080] Specifically: (1) Use the Python toolkit markdown-pdf to recognize advanced geological forecast PDF files and obtain text information (text information includes tunnel face mileage information, karst information, and karst treatment suggestions); the specific steps are as follows: (1.1) Load the PDF file using the program compiler; (1.2) Use the markdown-pdf tool to automatically parse the text content of the PDF file, extract it and save it as a Markdown (.md) file.

[0081] (1.3) In the generated Markdown file, obtain and save the following information by filtering through keywords: Mileage information of the working face (using regular expressions to match keywords related to "mileage" or "working face");  Karst information (matching keywords related to "karst" or "abnormality");  Recommendations (matching keywords related to "recommendations" or "solutions"); (2) Use the Python toolkits PDFPlumber and Pillow to automatically identify advanced geological prediction PDF files and obtain wave train images; the specific steps are as follows: (2.1) Load the PDF file using the program compiler; (2.2) Traverse all pages of the PDF and extract the embedded images from each page; (2.3) Image filtering and recognition: Use image classification algorithm (convolutional neural network) to determine whether each extracted image belongs to the ground-penetrating radar wave train image, and filter out the two images that are classified as "wave train images".

[0082] (2.4) Save the filtered ground-penetrating radar wave train image.

[0083] (3) Output and storage; (3.1) Store the text information in JSON format. As follows: { Mileage: "YK16+456", "Karst Description": "A large-scale karst cave was discovered at YK16+460, initially identified as a medium-sized anomaly." Recommended treatment: Grouting should be used for this purpose. Specific parameters are detailed in the design documents. } (3.2) Save the wave train image as a PNG file, named YK16+456_1.png and YK16+456_2.png (named according to the mileage number) for easy subsequent processing.

[0084] This completes the data acquisition process for ground-penetrating radar (GPR) detection, automatically obtaining the text and image information from the GPR report.

[0085] Step 3: Extraction of abnormal regions from the wave train image; the specific process is as follows: The image information obtained in step two is further imported into the wave train image anomaly region extraction system; A cave instance segmentation algorithm based on frequency-spatial dual-branch feature enhancement and dynamic adaptive receptive field network (FSD-DARF algorithm) is employed to accurately extract anomalous areas of caves from ground-penetrating radar wave train images. Specifically: 3.1 After preprocessing, the two detected wave train images are sequentially input into the trained FSD-DARF-based model.

[0086] 3.2 The trained, improved FSD-DARF model begins automatic detection, generating anomaly region mask images and statistical data. The anomaly region mask image is used to mark detected anomalous areas within the caverns. The statistical data includes the pixel area and coordinate range of the anomalous areas. The anomaly region mask image data is saved (in PNG format).

[0087] Furthermore, it is worth mentioning that in step 3.1, the trained model based on the improved FSD-DARF was mentioned.

[0088] The FSD-DARF model consists of two parts: a frequency-spatial dual-branch feature extraction network (FSD-Net) and a dynamic adaptive receptive field convolutional module (DARF-Conv). The Frequency-Spatial Dual-Branch Feature Extraction Network (FSD-Net) transforms wave train images to the frequency domain using FFT, simultaneously extracting features in both the spatial and frequency domains. Existing Transformers only perform attention calculations in the spatial domain; this invention introduces a frequency domain branch for the first time. The Dynamic Adaptive Receptive Field Convolutional Module (DARF-Conv) dynamically selects 3×3, 5×5, and 7×7 convolutional kernels through a scale-selective attention mechanism. Deformable convolution adjusts the sampling position, and DARF-Conv adjusts the receptive field size.

[0089] The construction process of the optimized FSD-DARF model is not directly shown in the embodiments. This work is preparatory for implementing the present invention. In the formal application stage, steps 3.1 and 3.2 can be directly executed. Specifically, the training process of the optimized FSD-DARF network is as follows: (1) Dataset creation: Extract 200 wave train images from the ground-penetrating radar detection report, ensuring that the image resolution is above 150 dpi. The dataset needs to cover a variety of geological environments and cave features, including the diversity of cave size, shape, and distribution type. Using Labelme annotation software, manually annotate the cavernous anomaly areas in the wave train images, using polygon regions as the annotation type. Save the annotation results as JSON files, with one annotation file corresponding to each image; Use the labelme2coco tool to convert Labelme's annotation files into COCO dataset format. The converted data includes: training set annotation files, validation set annotation files, and test set annotation files.

[0090] (2) The training process includes: Data input: Using the COCO format dataset, input the wave train image and annotation information into the model; Model training: Using pre-trained weights from the COCO dataset for transfer learning improves training efficiency. The optimizer uses AdamW, and the learning rate adjustment strategy is multi-step decay. The loss functions include classification loss, bounding box regression loss, and mask loss. Model validation: After training, the model's segmentation accuracy (mAP) and recall are evaluated using a test set.

[0091] Step 4: Perform data formatting on the extracted abnormal areas to generate boundary contour information that can be used for 3D modeling.

[0092] Preferably, the specific process of step four is as follows: (1) Extract the contour data of the abnormal region mask image output in step three. The specific process is as follows: The specific method for boundary extraction is as follows: Binarization: The masked image is binarized to extract the abnormal areas of the cavern (the masked area is white and the background is black). Edge detection: The Canny edge detection algorithm is used to identify the boundaries of abnormal regions; Contour fitting: Using polygon fitting methods, smooth closed contour lines are generated; Data output: Save the vertex coordinates of the boundary contour (stored in JSON format).

[0093] (2) Construction of the three-dimensional format of boundary contour data; Contour coordinate normalization: Based on the pixel ratio of the wave train image, the extracted contour coordinates are converted into a coordinate system of actual geological dimensions (unit: m).

[0094] Layered boundary generation: Based on the distance between the boundary contour data of the cave anomaly area in two wave train images of the same working face, the boundary data is extended into a three-dimensional boundary.

[0095] Output file format: 3D contour point cloud file, OBJ format.

[0096] This completes the processing of the contour data of the cave anomaly area, and a preliminary three-dimensional model is constructed based on the boundary contour data of the cave anomaly area from two wave train images.

[0097] Step 5: 3D modeling of the cave; A cave modeling algorithm is introduced to generate a 3D model of the cave using contour data of anomaly areas. Specifically: (1) Data input and preprocessing; including: Input data: Input the 3D contour point cloud file generated in step four into the cave 3D modeling system.

[0098] Preprocessing: The point cloud data is filtered and denoised to eliminate boundaries and extract noise points.

[0099] (2) A three-dimensional model of the cave was constructed using a method combining Delaunay triangulation and Marching Cubes algorithm; the specific process is as follows: Delaunay triangulation; Divide the 3D point cloud data into non-overlapping triangular meshes.

[0100] Optimize the quality of the triangular mesh to make the surface of the cave smoother.

[0101] Marching Cubes algorithm processing:  Voxelization is performed on the karst cave area to convert the point cloud data into a mesh model.

[0102] Reconstruct the three-dimensional volume based on the elevation values ​​of the contour point cloud to generate the surface of a closed cave.

[0103] Texture mapping: Mapping the real-world attribute information of the tunnel face onto the surface of the 3D model enhances the visualization effect of the model.

[0104] (3) Optimization of the three-dimensional model; The Quadric Edge Collapse algorithm is used to reduce the number of triangles and optimize model complexity.

[0105] (4) Output of the 3D model; Outputs 3D models in OBJ format, compatible with mainstream 3D design and display platforms.

[0106] Step Six: 3D Model Visualization and Deployment; The constructed 3D model of the cave is deployed into MR glasses to achieve real-time perception and interactive display of the cave's spatial morphology. Specifically: (1) After completing the construction of the 3D model of the cave in step five, the model needs to be formatted and optimized to adapt to the operating environment of the MR glasses: Format conversion: Convert the cave model into a 3D file format (GLTF) compatible with MR glasses, which supports efficient loading and real-time rendering.

[0107] Model optimization: Simplified polygons: Using LOD (Level of Detail) technology, the number of polygons in the model is dynamically adjusted according to the computing power of the glasses to ensure smooth operation.

[0108] Texture compression: Compress the model's textures (Basis Universal format) to reduce file size while maintaining visualization quality.

[0109] (2) To provide an immersive experience, the 3D model of the cave needs to be integrated with the tunnel environment: Tunnel scene reconstruction: Based on the actual tunnel design drawings, construct a virtual tunnel environment model that is consistent with the real scene.

[0110] The spatial alignment between the user's position and the tunnel model is calibrated using the built-in sensors in the MR glasses.

[0111] Cave model positioning: The 3D model of the karst cave is precisely located to the corresponding position in the virtual tunnel model based on the mileage information of its working face and the location data of the ground-penetrating radar wave train image.

[0112] (3) To enhance user experience, intuitive interactive functions were designed to allow users to view cave information in real time; Cave Information Display: Bind the karst information and treatment recommendations extracted from the ground-penetrating radar report to the relevant locations in the cave model.

[0113] When a user wears MR glasses to observe the tunnel face area, the cave model and its related information will automatically appear in the field of view and be labeled with semi-transparent tags.

[0114] Cave selection and details: Users can select a specific cave model using gestures; The selected cave will be highlighted, and an information panel will pop up, displaying karst description information and treatment suggestions for users to view in detail.

[0115] With the above deployment, users can enter a virtual tunnel scene after wearing MR glasses. When looking at the tunnel face area, they can view the 3D model of the cave and related information in real time (see...). Figures 2 to 4As shown in the image, one can intuitively perceive the spatial morphology and location of the karst caves, and obtain detailed karst descriptions and treatment suggestions through interactive functions. This approach effectively improves the intuitiveness of geological information and decision-making efficiency during tunnel construction, providing strong technical support for safe construction.

[0116] Example 2: As a further embodiment of the present invention, the present invention also provides an electronic device, comprising: One or more processors; Storage device for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the aforementioned method.

[0117] In practical use, users can interact with servers, which also function as terminals, via a network to receive or send messages. Terminal devices are generally various devices equipped with a display and used through a human-computer interface, including but not limited to smartphones, tablets, laptops, and desktop computers. Various specific application software can be installed on these terminal devices as needed, including but not limited to web browsers, instant messaging software, social media platforms, and shopping apps.

[0118] Furthermore, the server is a network server used to provide various services, such as a backend server that provides corresponding calculation services for the ground-penetrating radar images transmitted from the terminal device, so as to realize the processing of the tunnel cave 3D model construction method based on ground-penetrating radar images, calculate the spatial morphology and location visual map of the cave in the tunnel face, and finally return it to the terminal device.

[0119] Example 3: As a further embodiment of the present invention, the present invention also provides a medium comprising one or more programs executable by one or more processors of a device, the one or more programs comprising instructions for performing the method for constructing a three-dimensional model of a tunnel cave based on ground-penetrating radar images as described above.

[0120] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for constructing a three-dimensional model of a tunnel / karst cave based on ground-penetrating radar images, characterized in that, Includes the following steps: Step 1: Based on the ground-penetrating radar detection report in advanced geological forecasting, obtain multiple wave train images; Step 2: Annotate the anomalous areas of the caves in multiple wave train images to obtain the annotation information of the anomalous areas of the caves; construct a wave train image dataset based on the wave train images and the annotation information of the anomalous areas of the caves. A cave instance segmentation model based on frequency-spatial domain dual-branch feature enhancement and dynamic adaptive receptive field network was constructed. The cave instance segmentation model based on frequency-spatial domain dual-branch feature enhancement and dynamic adaptive receptive field network was pre-trained using the wave train image dataset to obtain the trained cave instance segmentation model. Mask images of anomalous regions in karst caves are extracted from wave train images based on a trained karst cave instance segmentation model. The cave instance segmentation model based on frequency-spatial dual-branch feature enhancement and dynamic adaptive receptive field network includes a frequency-spatial dual-branch feature extraction network, a dynamic adaptive receptive field convolution module, a cave boundary perception map neural network module, and a cave morphology prior constraint loss function. Step 3: Perform data formatting on the mask image of the cave anomaly area to obtain the vertex coordinates of the two-dimensional boundary contour of the cave anomaly area; Three-dimensional contour point cloud data of the karst cave anomaly region is constructed based on the vertex coordinates of the two-dimensional boundary contour. Step 4: Introduce a cave modeling algorithm and construct a 3D model of the cave using the 3D contour point cloud data of the cave's abnormal areas; Step 5: Deploy the 3D model of the cave into the mixed reality glasses to obtain a visual image of the spatial morphology and location of the cave within the tunnel face.

2. The method for constructing a three-dimensional model of a tunnel / karst cave based on ground-penetrating radar images according to claim 1, characterized in that, The specific process of step one is as follows: The entire ground-penetrating radar detection report is read and its format is recognized to obtain text and image information; The text information includes tunnel face mileage information, karst anomaly description information, geological condition description, karst cave risk level, and engineering treatment suggestions; and the extracted text information is structured, organized, and stored to form standardized data. The image information includes ground-penetrating radar wave train images; and the image information is preprocessed to obtain preprocessed wave train images.

3. The method for constructing a three-dimensional model of a tunnel / karst cave based on ground-penetrating radar images according to claim 2, characterized in that, The preprocessing of image information includes: image denoising, contrast enhancement, edge smoothing, and invalid region removal.

4. The method for constructing a three-dimensional model of a tunnel / karst cave based on ground-penetrating radar images according to any one of claims 1-3, characterized in that, The specific process of step two is as follows: Step 2.1: Construct a wave train image dataset based on wave train image information and cave anomaly area annotation information; Step 2.2: Construct a cave instance segmentation model based on frequency-spatial domain dual-branch feature enhancement and dynamic adaptive receptive field network; The cave instance segmentation model based on frequency-spatial dual-branch feature enhancement and dynamic adaptive receptive field network includes a frequency-spatial dual-branch feature extraction network, a dynamic adaptive receptive field convolution module, a cave boundary perception map neural network module, and a cave morphology prior constraint loss function. Step 2.3: Use a cave instance segmentation model based on frequency-spatial dual-branch feature enhancement and dynamic adaptive receptive field network to extract the abnormal regions in the wave train image, and obtain the cave abnormal region mask in the ground-penetrating radar wave train image and the statistical data corresponding to the mask.

5. The method for constructing a three-dimensional model of a tunnel / karst cave based on ground-penetrating radar images according to claim 4, characterized in that, The specific process of constructing the wave train image dataset is as follows: Multiple wave train images containing information on cave size, shape, and distribution type were captured from the ground-penetrating radar report, thus obtaining wave train image information; The abnormal boundaries, abnormal ranges, and abnormal types of the karst cave abnormality areas are marked to obtain the karst cave abnormality area marking information; The wave train image information and the cave anomaly area annotation information are converted into a format to obtain the wave train image dataset.

6. The method for constructing a three-dimensional model of a tunnel / karst cave based on ground-penetrating radar images according to claim 5, characterized in that, The frequency-spatial dual-branch feature extraction network includes a spatial branch, a frequency branch, and a dual-branch adaptive fusion module. The spatial branch contains four residual blocks, each consisting of two... The system consists of convolutional layers and incorporates batch normalization and ReLU activation functions. Furthermore, a boundary enhancement module is introduced between residual blocks. Boundary features are extracted using the Sobel edge detection operator and fused with the backbone features to obtain the local spatial features of the wave train image. The frequency domain branch is used to convert the wave train image to the frequency domain and extract the feature representation of the cave reflection signal in the frequency domain. The dual-branch adaptive fusion module obtains global statistical information of the spatial and frequency branches through global average pooling, and learns the fusion weights of the spatial and frequency branches through a fully connected layer to obtain the fusion features of the spatial and frequency branches. .

7. The method for constructing a three-dimensional model of a tunnel / karst cave based on ground-penetrating radar images according to claim 6, characterized in that, The specific process for obtaining the characteristic representation of the cave reflection signal is as follows: A two-dimensional fast Fourier transform is performed on the input wave train image to convert the wave train image from spatial domain data to frequency domain data; Logarithmic transformation and normalization are performed on the frequency domain data to obtain frequency domain data of different frequency components; Design a frequency domain attention module to weight frequency domain data of different frequency components using a learnable weight matrix to obtain frequency domain features; The frequency domain features are converted into spatial domain features by inverse Fourier transform, and the spatial domain features are fused with local spatial features to obtain the characteristic representation of the cave reflection signal.

8. The method for constructing a three-dimensional model of a tunnel / karst cave based on ground-penetrating radar images according to claim 6, characterized in that, Fusion features The expression is as follows: ； ； ； in, The feature map of the spatial branch, Feature map of frequency domain branch, This is a global average pooling operation. It is a fully connected layer. It is the Sigmoid activation function. The fusion weights for the spatial branches, The fusion weights are for the frequency domain branches.

9. The method for constructing a three-dimensional model of a tunnel / karst cave based on ground-penetrating radar images according to any one of claims 6-8, characterized in that, The dynamic adaptive receptive field convolution module includes a multi-scale parallel convolution structure, a scale-selective attention mechanism, and a hole convolution enhancement structure. The multi-scale parallel convolutional structure: Three convolutional kernels of different scales are designed to process the input features in parallel, resulting in three sets of feature maps of different scales; The scale-selection attention mechanism dynamically selects the corresponding convolution scale based on three sets of feature maps at different scales, obtains global statistical information of the input features through global average pooling, and learns the weights of the convolution kernels at each scale through two fully connected layers. ; The hole convolution enhancement structure: for The convolution kernel uses dilated convolution with a dilation rate of 2, thus expanding the actual receptive field to [value missing]. ;for The convolution kernel uses a standard convolution with a dilation rate of 1.

10. The method for constructing a three-dimensional model of a tunnel / karst cave based on ground-penetrating radar images according to claim 9, characterized in that, The cave boundary sensing graph neural network module is used to convert wave train images into graph structures; the specific process is as follows: The Simple Linear Iterative Clustering algorithm is used to segment the wave train image into several superpixel regions, with each superpixel region serving as a graph node. For adjacent superpixel regions, establish edge connections between the corresponding graph nodes; The initial features of each graph node are obtained. The initial features of each graph node consist of the feature statistics of the corresponding superpixel region, and the initial features of each graph node include color mean, texture features and boundary intensity.

11. The method for constructing a three-dimensional model of a tunnel / karst cave based on ground-penetrating radar images according to claim 10, characterized in that, The specific process for obtaining the boundary strength is as follows: Boundary strength is obtained by calculating the average gradient value of the superpixel region boundary using the Sobel operator.

12. The method for constructing a three-dimensional model of a tunnel / karst cave based on ground-penetrating radar images according to claim 11, characterized in that, Prior constraint loss function for cave morphology The expression is as follows: ； ； ； ； in: For segmentation loss; This is the boundary smoothing loss, used to constrain the smoothness of the segmentation boundary and penalize drastic changes in the boundary. The morphological consistency loss is used to constrain the segmentation results to conform to the typical morphological characteristics of karst caves. This is the frequency domain consistency loss, used to constrain the consistency of frequency domain features between the predicted segmented region and the ground truth labeled region; Boundary smoothing loss The weighting coefficients, For morphological consistency loss The weighting coefficients, For frequency domain consistency loss Weighting coefficients; For segmentation mask; For the Laplace operator; This refers to the number of boundary pixels; This is a closure loss, used to penalize cases where the segmented region is not closed. Uniformity loss is used to constrain signal uniformity within the segmented region; Input image; For predicting the mask; True label mask; This is element-wise multiplication; For Fast Fourier Transform.

13. The method for constructing a three-dimensional model of a tunnel / karst cave based on ground-penetrating radar images according to claim 12, characterized in that, The specific process for obtaining two-dimensional boundary contour data is as follows: Binarize the mask image, extract the abnormal areas of the karst caves, set the mask part to white, and set the background to black; Edge detection algorithms are used to identify the boundaries of abnormal areas in karst caves; Using a polygon fitting method, a smooth closed contour line is generated for the boundary of the karst cave anomaly area; the vertex coordinates of the two-dimensional boundary contour are obtained based on the closed contour line.

14. The method for constructing a three-dimensional model of a tunnel / karst cave based on ground-penetrating radar images according to claim 13, characterized in that, The specific process of constructing the 3D contour point cloud data of the karst cave anomaly region based on the vertex coordinates of the 2D boundary contour is as follows: Based on the pixel ratio of the wave train image, the vertex coordinates of the two-dimensional boundary contour are converted into a coordinate system of actual geological dimensions; Based on the distance between the vertex coordinates of the two-dimensional boundary contour of two adjacent wave train images of the same working face, the coordinate system of the actual geological dimensions is extended to a three-dimensional coordinate system to obtain the three-dimensional contour point cloud data of the cave anomalous area.

15. The method for constructing a three-dimensional model of a tunnel / karst cave based on ground-penetrating radar images according to any one of claims 11-14, characterized in that, The specific process of step four is as follows: The three-dimensional contour point cloud data of the abnormal area of ​​the karst cave is filtered and denoised to obtain preprocessed point cloud data. The preprocessed 3D point cloud data is transformed into a continuous triangular mesh surface through a spatial subdivision algorithm, and a closed, hole-free, non-overlapping, and topologically reasonable karst cave 3D solid model is constructed. The real-world attribute information of the tunnel face is mapped onto the surface of the 3D solid model of the karst cave. Then, the 3D solid model of the karst cave is meshed to generate a geometrically accurate, structurally sound, and highly compatible 3D model of the karst cave.

16. An electronic device, characterized in that, It includes a memory, one or more processes, and one or more programs stored in the memory, said one or more programs including instructions for executing the method for constructing a three-dimensional model of a tunnel cave based on ground-penetrating radar images as described in any one of claims 1-15.

17. A medium, characterized in that, Includes one or more programs executed by one or more processors of the device, said one or more programs including instructions for executing the method for constructing a three-dimensional model of a tunnel cave based on ground-penetrating radar images as described in any one of claims 1-15.

Images