Road disease intelligent detection method and system based on three-dimensional ground penetrating radar

Abstract

The invention relates to the technical field of road detection, in particular to an intelligent road disease detection method and system based on a three-dimensional ground penetrating radar, and the method comprises the steps: obtaining and processing three-dimensional radar data, and obtaining a horizontal section image in the depth direction and vertical section images of a plurality of channels in the measuring line direction; determining the boundary position of the underground structure layer; selecting a horizontal slice image with a corresponding depth according to the boundary position, and inputting the horizontal slice image into a first disease detection model to obtain a preliminary disease candidate area; mapping the preliminary candidate region to a vertical profile image, intercepting a local vertical profile image of each channel, and inputting the local vertical profile image into a second disease detection model to obtain disease category information; fusion decision is carried out based on the transverse position of the candidate area and the disease category information of each channel, a final disease type is determined and is associated to the candidate area, consistency of disease positioning and type judgment is realized through cooperative utilization of the horizontal section image and the multi-channel vertical section image, and the accuracy and stability of detection are improved.

Description

Technical Field

[0001] This invention relates to the field of road inspection technology, and in particular to an intelligent method and system for detecting highway defects based on three-dimensional ground-penetrating radar. Background Technology

[0002] During long-term service, highways are prone to developing hidden defects such as interlayer voids, loosening, cracks, and water seepage under the influence of vehicle loads and environmental factors. These defects are difficult to detect from the surface, but they can significantly weaken the road's load-bearing capacity and even cause safety accidents such as collapses. Ground penetrating radar (GPR), as an efficient, non-destructive, and high-resolution underground detection technology, has become an important means of detecting highway defects. Traditional methods mainly rely on manual interpretation or simple analysis of single-channel vertical profile images acquired by two-dimensional GPR. However, the information in a single-dimensional image is limited, and defect features are easily confused with medium interface reflections and random noise, resulting in inherent limitations such as high misjudgment rate, many missed detections, and difficulty in three-dimensional positioning and quantitative evaluation.

[0003] With the development of 3D ground-penetrating radar technology, horizontal slice images reflecting horizontal distribution and multi-channel vertical profile images reflecting vertical structure can be acquired simultaneously, providing a data foundation for 3D identification of road defects. However, existing detection methods have failed to fully utilize the synergistic value of these multi-dimensional data: although horizontal slice images can clearly show the planar distribution of defects, they have weak textures and blurred edges, making it difficult for general target detection models to directly adapt and unable to determine the specific type of defect; while detection based on vertical profiles can provide category information, single-channel analysis results are greatly affected by noise, have poor spatial consistency, and lack effective correlation with horizontal positioning, resulting in key bottlenecks such as feature confusion, positioning deviation, and type misjudgment in the 3D defect identification process. These problems seriously restrict the automation level and reliability of highway defect detection.

[0004] The information disclosed in this background section is intended only to enhance the understanding of the general background of the invention and should not be construed as an admission or in any way implying that the information constitutes prior art known to those skilled in the art. Summary of the Invention

[0005] This invention provides a method and system for intelligent detection of highway defects based on three-dimensional ground-penetrating radar, thereby effectively solving the problems in the background technology.

[0006] To achieve the above objectives, the technical solution adopted by this invention is: an intelligent detection method for highway defects based on three-dimensional ground-penetrating radar, comprising the following steps: Acquire and process 3D radar data to obtain horizontal slice images distributed along the depth direction and vertical profile images of multiple channels distributed along the survey line direction; Based on multiple vertical cross-sectional images, the location of the boundary line of the underground structural layers is determined; Based on the location of the boundary line, select the horizontal slice image at the corresponding depth and input it into the first disease detection model to obtain the preliminary disease candidate region on the horizontal slice; The preliminary disease candidate region is mapped to the vertical profile image, and the local vertical profile image corresponding to the candidate region in each channel is extracted; The local vertical profile images of each channel are input into the second disease detection model to obtain disease category information for each channel; Based on the lateral position of the preliminary disease candidate area and the disease category information of each channel, a fusion decision is made to determine the final disease type and associate it with the preliminary disease candidate area.

[0007] Furthermore, making integrated decisions includes: The disease category information detected in each channel and its coverage width within the horizontal range of the preliminary disease candidate area are statistically analyzed. Calculate the comprehensive score for each disease type based on the preset severity weights and the coverage width; Based on the comprehensive score, the primary disease type and / or secondary disease type are determined as the final disease type.

[0008] Furthermore, the overall score is calculated using the following formula: ; In the formula, For the overall score; Different weights are assigned based on the severity of the disease type; To calculate the number of channels with the same disease type.

[0009] Furthermore, the first disease detection model includes a C3WC module, which extracts high-frequency edge features and low-frequency texture features from horizontal slice images simultaneously by embedding wavelet transform into convolution operations, thereby enhancing the ability to identify diseases with weak edges and large scale.

[0010] Furthermore, the first disease detection model or the second disease detection model includes an OEM module; the OEM module generates an attention weight map through the following steps: Perform edge detection on the input image to generate an initial edge mask; The initial edge mask is normalized and mapped with Gaussian weights to generate an attention weight map with high center weights and gradually decreasing edge weights. The attention weight map is multiplied with the feature map of the first disease detection model or the second disease detection model to enhance the feature response that matches the prior morphology of interlayer diseases.

[0011] Furthermore, obtaining preliminary disease candidate regions on the horizontal slice includes: Sliding window detection is performed on the horizontal slice image to generate multiple detection windows; Input the detection window into the first disease detection model to obtain multiple initial disease candidate boxes; The multiple initial disease candidate boxes are merged and filtered to form the preliminary disease candidate region set.

[0012] Further, the step of mapping the preliminary disease candidate region to the vertical profile image and extracting the local vertical profile image corresponding to the candidate region in each channel includes: Based on the lateral coordinate range and channel number of the preliminary disease candidate region in the horizontal slice image, determine its corresponding lateral position range in the vertical cross-sectional image of each channel. Based on the preset vertical analysis depth, image blocks are extracted from the vertical profile images of each channel along the horizontal position interval to form the local vertical profile image.

[0013] Furthermore, obtaining the disease category information for each channel includes: The local vertical profile images of each channel are input into the second disease detection model; Obtain the disease category and related attribute information corresponding to each local area, output by the second disease detection model; Based on the horizontal coordinates of the preliminary disease candidate regions, the disease categories and related attribute information are stored by channel.

[0014] Furthermore, determining the location of the boundary lines of the underground structural layers includes: Multiple vertical cross-sectional images are superimposed pixel-by-pixel to obtain a superimposed image; The superimposed image is smoothed. Calculate the vertical gradient of the smoothed image to obtain gradient information; Based on the gradient information, local peak values ​​are detected, and effective peak values ​​are selected as the boundary line positions of the underground structural layer according to preset conditions.

[0015] This invention also includes a highway defect intelligent detection system based on three-dimensional ground-penetrating radar, the system comprising: The data acquisition and preprocessing module is used to acquire and process three-dimensional radar data to obtain horizontal slice images distributed along the depth direction and vertical profile images of multiple channels distributed along the survey line direction. The underground stratum determination module is used to determine the location of the boundary line of the underground structural layer based on multiple vertical profile images; The horizontal slice disease initial localization module is used to select a horizontal slice image of corresponding depth according to the position of the boundary line, and input it into the first disease detection model to obtain the preliminary disease candidate area on the horizontal slice. The vertical profile image cropping module is used to map the preliminary disease candidate region to the vertical profile image and crop the local vertical profile image corresponding to the candidate region in each channel; The vertical profile disease classification module is used to input the local vertical profile images of each channel into the second disease detection model to obtain disease category information for each channel. The multi-dimensional information fusion decision module is used to make fusion decisions based on the lateral position of the preliminary disease candidate area and the disease category information of each channel, determine the final disease type, and associate it with the preliminary disease candidate area.

[0016] The beneficial effects of this invention are as follows: By acquiring and processing 3D radar data, the location of the boundary line of the underground structural layer is determined based on the vertical profile image. The horizontal slice image at the corresponding depth is selected and input into the first disease detection model to obtain the preliminary disease candidate area. The image is then mapped to the vertical profile image to extract a local image and input into the second disease detection model to obtain disease category information. Finally, a fusion decision is made to determine the final disease type. This solves the problems of high misjudgment rate, frequent missed detection, and inability to locate in 3D by traditional highway disease detection methods. It has efficient and accurate highway disease detection capabilities, can significantly reduce the misjudgment rate, and can achieve accurate 3D spatial positioning of diseases.

[0017] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 The flowchart is shown below for a method for intelligent detection of highway defects based on three-dimensional ground-penetrating radar. Figure 2 This is a flowchart illustrating the overall process of a smart detection method for highway defects based on three-dimensional ground-penetrating radar. Figure 3 This is a schematic diagram illustrating the detection principle of ground-penetrating radar. Figure 4 This is a schematic diagram of the OEM-HWNet network structure; Figure 5 This is a schematic diagram of the C3WC structure; Figure 6 This is a schematic diagram of the OEM structure; Figure 7 A visual comparison of the OEM-HWNet highway interlayer defect detection network with other methods; Figure 8 A schematic diagram illustrating the characteristics and spatial location of highway defects; Figure 9 This is a schematic diagram of a highway defect intelligent detection system based on three-dimensional ground-penetrating radar. Detailed Implementation

[0020] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0021] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0022] Example 1: like Figures 1 to 8 As shown, this application provides an intelligent detection method for highway defects based on three-dimensional ground-penetrating radar. The method includes: S10: Acquire and process 3D radar data to obtain horizontal slice images distributed along the depth direction and vertical profile images of multiple channels distributed along the survey line direction; S20: Determine the location of the boundary line of the underground structural layers based on multiple vertical profile images; S30: Based on the location of the boundary line, select the horizontal slice image at the corresponding depth and input it into the first disease detection model to obtain the preliminary disease candidate area on the horizontal slice; S40: Map the preliminary disease candidate areas to the vertical profile image, and extract the local vertical profile images corresponding to the candidate areas in each channel; S50: Input the local vertical profile images of each channel into the second disease detection model to obtain disease category information for each channel; S60: Based on the lateral position of the preliminary disease candidate area and the disease category information of each channel, a fusion decision is made to determine the final disease type and associate it with the preliminary disease candidate area.

[0023] First, a 3D ground-penetrating radar (GPR) scan is used to acquire 3D radar data containing multi-channel radar echo information. This data is then preprocessed and reconstructed to generate multi-layered horizontal slice (C-slice) images distributed along the depth direction and vertical profile (B-scan) images distributed along the survey line direction. Next, based on multiple vertical profile images of the same road segment, gradient features reflecting changes in different underground media layers are extracted through image overlay, smoothing, and vertical gradient calculation. Local peaks are detected in these gradient features to determine the boundary lines of the underground structural layers. Based on this, horizontal slice images within the corresponding depth range are selected according to the boundary line locations, and these images are input into a first disease detection model for analysis to obtain preliminary disease candidate areas in the horizontal slice dimension. Finally, based on the lateral coordinate range of the preliminary disease candidate areas in the horizontal slice images and the radar channel number information, the preliminary disease candidate areas are mapped to corresponding... In the vertical profile image, along the horizontal position interval and the preset vertical analysis depth, local vertical profile images corresponding to the candidate areas are extracted from the vertical profile images of each channel. Then, the local vertical profile images of each channel are input into the second disease detection model to obtain disease category information and related attribute information of the corresponding local area of ​​each channel, and the multi-channel detection results are organized according to the horizontal position of the candidate area. Finally, based on the horizontal position of the preliminary disease candidate area and the disease category information of each channel, the spatial distribution of diseases in the candidate area of ​​different channels is fused and decided to comprehensively determine the final disease type. The final disease type is then associated with the corresponding preliminary disease candidate area, thereby outputting the fused highway disease detection result. Through the above technical solution, the multi-dimensional information collaborative utilization based on three-dimensional ground penetrating radar data is realized, giving full play to the advantages of horizontal slicing in the initial screening of disease location and the advantages of multi-channel vertical profile in disease type determination, effectively improving the accuracy and reliability of highway disease detection.

[0024] By collaboratively utilizing horizontal slice images and multi-channel vertical profile images from 3D radar data, a unified detection process consisting of initial screening of disease location, spatial mapping, multi-channel verification, and fusion decision-making is constructed. This effectively overcomes the instability problem caused by relying on single-dimensional or single-channel information for disease determination in existing methods. It fully leverages the advantages of horizontal slice images in disease planar localization and the advantages of multi-channel vertical profile images in disease structure representation and type identification. By associating and fusing multi-dimensional and multi-channel detection results at the same spatial location, the disease localization results and type determination are consistent, thereby significantly reducing the impact of noise interference and medium interface reflection on the detection results and improving the accuracy and stability of disease identification.

[0025] In this embodiment, in step S10, a three-dimensional ground-penetrating radar is used to detect underground information of the highway to be detected, such as... Figure 3 As shown, (a) illustrates the working principle of ground-penetrating radar, where ε represents the dielectric constant of the medium; (b) represents the A-scan signal recorded by the receiving antenna; and (c) represents the B-scan image of the ground-penetrating radar formed by acquiring the A-scan signal along the survey line direction, thus obtaining three-dimensional radar data. The three-dimensional radar data is then cut along the z-axis and x-axis directions to obtain horizontal slices (C-slice) of different depths and vertical profiles (B-scan) images of multiple channels. Among them, the superposition is pixel-level superposition; the smoothing process uses Gaussian smoothing; the vertical gradient is calculated using the Sobel vertical gradient operator; the preset conditions include the minimum interval threshold between peaks and the gradient intensity noise threshold; after selecting the effective peaks as the boundary line positions, it also includes: fusing the boundary line positions obtained from multiple survey lines to generate the average layer line.

[0026] Because diseases in horizontal slices exhibit weak texture edges and large size variations, they differ significantly from ordinary target detection tasks. Furthermore, inter-layer disease detection based on B-scan images often suffers from poor localization accuracy and difficulty adapting to large-scale diseases. For specific domain problems, existing models are difficult to apply directly and often exhibit suboptimal results. Therefore, this invention designs the OEM-HWNet model, such as... Figure 4 As shown, it is mainly divided into three parts: Backbone, Neck, and Head.

[0027] The backbone consists of the CBS module, the C3WC module, and the SPPF module. The CBS module includes a standard convolutional layer, a batch normalization layer (BN), and the SiLU activation function. The original C3 module, which consisted of three convolutional layers and several bottleneck modules, has been replaced by the C3WC module. This module integrates WTConv into the bottleneck to obtain a larger receptive field and improve the ability to capture low-frequency features. The SPPF module is an optimized version of the Spatial Pyramid Pooling (SPP) module. The Neck part achieves semantic information fusion from low to high level by combining the Feature Pyramid Network (FPN) and the Path Aggregation Network (PANet). It is worth noting that we inserted a prior knowledge-based module called OEM between the Neck and Backbone parts. This module enhances the model's learning of interlayer diseases, suppresses background interference, and significantly improves the model's learning ability and ability to locate interlayer diseases. The head section includes four detection heads with dimensions of 10×10, 20×20, 40×40, and 80×80, enabling the network to perform high-precision detection of interlayer defects with large-scale variations, reducing missed detections. Key improvements include: The C3WC module was designed by introducing wavelet convolution, such as... Figure 5 As shown, this module obtains WCBottleneck by introducing wavelet convolution into the Bottleneck module. It consists of three convolutions and multiple WCBottleneck modules. This module can significantly increase the network's receptive field and its ability to capture low-frequency information, thereby improving its adaptability to targets with large-scale changes. The OEM module was designed based on prior knowledge from human input, such as... Figure 6 As shown, the shape and texture features of interlayer diseases are used as prior knowledge. First, the edge detection Canny operator is used to extract the target edge. Then, the morphological dilation algorithm is introduced as a supplement to the Canny operator to connect adjacent line segments to form the target region and separate it from the background. Finally, the target region is weighted to enhance the target features. It is then embedded into the target detection network to locate the interlayer disease region and enhance its features. A fourth detection head has been added to the existing one to improve the ability to identify large-scale targets; The OEM-HWNet model designed in this invention is compared with other advanced models, such as... Figure 7As shown in Case 2, at the location indicated by the yellow arrow, OEM-HWNet was able to detect large-scale interlayer defects with accurate localization and high confidence, while YOLOv5s and YOLOv7 could not. Although RT-DETR, YOLOv3, and YOLOv8s could detect large-scale interlayer defects, their detection results were not accurate. The results indicate that the model needs a larger receptive field to accurately detect large-scale defects. Figure 7 As shown in Case 1, some diseases were not detected in the YOLOv5s detection results, while OEM-HWNet was able to detect all interlayer diseases. These results indicate that OEM, based on prior knowledge, can help the model to accurately locate and learn in a focused manner. In summary, OEM-HWNet performs best in the detection of interlayer diseases.

[0028] Horizontal slice and vertical profile datasets were constructed and divided into training, validation and test sets in an 8:1:1 ratio. Two object detection networks, called OEM-HWNet_C-slice and OEM-HWNet_B-scan, were trained based on the improved OEM-HWNet network and used for disease detection in horizontal slices and vertical profiles, respectively. As a preferred embodiment of the above, the first disease detection model includes a C3WC module. The C3WC module extracts high-frequency edge features and low-frequency texture features from the horizontal slice image by embedding wavelet transform into convolution operation, thereby enhancing the ability to identify diseases with weak edges and large scale.

[0029] Specifically, the first disease detection model is used to identify disease regions in horizontal slice images. Given that diseases in horizontal slice images typically exhibit weak edges, large scale variations, and low texture contrast, reflecting abnormalities, feature extraction methods based solely on spatial domain convolution are insufficient to simultaneously consider both disease boundary information and overall structural information. Therefore, in this embodiment, the first disease detection model includes a C3WC module to enhance the joint modeling capability of different frequency features in the horizontal slice image. The C3WC module is located in the feature extraction stage of the first disease detection model and processes the input horizontal slice image or its intermediate feature maps. Its core idea is to embed wavelet transform into the convolution operation process, enabling the model to simultaneously acquire high-frequency edge features and low-frequency texture features during a single forward propagation, thereby improving the ability to identify weak-edge, large-scale diseases. In a specific implementation, the C3WC module first receives the feature map from the previous network layer as input and performs discrete wavelet transform on the feature map, decomposing the input features into multiple sub-band components, including at least sub-band components representing high-frequency information and sub-band components representing low-frequency information. The model employs a C3WC module to extract discriminative features from different frequency bands. High-frequency subbands highlight edge variations and abrupt changes in reflection within the diseased area of ​​the horizontal slice image, while low-frequency subbands preserve the overall shape, scale, and background texture of the diseased area on the plane. Convolutional operations are then applied to each subband component to extract discriminative features at the corresponding frequency bands. After convolution, the feature results from different frequency bands are fused to form a joint feature representation containing multi-frequency information. This joint feature is then passed as the output of the C3WC module to subsequent network layers. This approach allows the model to simultaneously focus on local edge information and global structural information within a single module, avoiding recognition biases caused by relying solely on high-frequency or low-frequency features. In practical applications, when a horizontal slice image contains diseased areas with unclear edges, low reflection contrast, but large spatial scale, traditional convolutional features often struggle to form stable responses in shallow networks. However, by introducing the C3WC module, the model can capture the overall contour of the diseased area in the low-frequency components and enhance subtle reflection changes at the diseased boundary in the high-frequency components, thus creating a more significant and stable representation of the diseased area in the feature space.

[0030] In this embodiment, the first disease detection model or the second disease detection model includes an OEM module; the OEM module generates an attention weight map through the following steps: Perform edge detection on the input image to generate an initial edge mask; The initial edge mask is normalized and mapped with Gaussian weights to generate an attention weight map with high weights at the center and gradually decreasing weights at the edges. The attention weight map is multiplied with the feature map of the first disease detection model or the second disease detection model to enhance the feature response that matches the prior morphology of interlayer diseases.

[0031] Specifically, the first and / or second disease detection models include an OEM module to introduce attention constraints that conform to the morphological characteristics of underground diseases during the feature extraction and expression stages. Given that underground diseases in highway images typically appear as abnormal reflective areas distributed along the interlayer interface with certain continuity and edge features, relying solely on the self-learning attention mechanism of convolutional networks is easily affected by noise reflections or non-disease structure edges, making it difficult to stably highlight the true disease areas. Therefore, in this embodiment, edge prior information is explicitly introduced through the OEM module to guide and strengthen the model's feature response. The OEM module is set in the feature extraction or feature fusion stage of the disease detection model to generate an attention weight map based on the structural edge information of the input image and apply this attention weight map to the model feature map, thereby improving the feature response intensity that conforms to the prior morphology of the interlayer diseases. In one specific implementation, the OEM module first performs edge detection on the input image to obtain an initial edge mask reflecting changes in the image structure. The input image can be a horizontal slice image, a vertical cross-sectional image, or a feature map output from an intermediate layer of the model. Edge detection can employ gradient operators, direction operators, or other methods that reflect pixel intensity abrupt changes, highlighting the interlayer interface and reflection abrupt change locations in the edge mask. Subsequently, the initial edge mask is normalized to map edge responses at different locations to a uniform numerical range, eliminating the influence of differences in image brightness or contrast. After normalization, a Gaussian weight mapping is applied to the edge mask, causing the attention weight to reach a higher value in the edge center region and gradually decrease along the edge normal direction, thereby generating an attention weight map with high center weight and gradually decreasing edge weight. This attention weight map reflects the model's attention to features at different locations in space. After the attention weight map is generated, the OEM module performs element-wise multiplication between the attention weight map and the feature map in the disease detection model, enhancing the feature responses in high attention weight regions and relatively weakening the feature responses in low attention weight regions. This approach allows the model to focus more on regions in the feature space that are consistent with the morphology of the interlayer interfaces and possess continuous edge features. This suppresses the interference of random noise reflections or isolated strong reflection points unrelated to the disease on the detection results. In actual detection, when interlayer disease reflections and other structural interface reflections coexist in the radar image, the attention weight map generated by the OEM module enables the model to prioritize edge structures that match the prior morphology of the interlayer disease, thus forming a more concentrated response region during the feature extraction stage. Compared to detection models without the OEM module, this approach helps reduce the response intensity of false detection regions and improves the separability of disease regions in the feature space.

[0032] In step S20, determining the location of the boundary line of the underground structural layers includes: S21: Superimpose multiple vertical cross-sectional images at the pixel level to obtain a superimposed image; S22: Smooth the superimposed images; S23: Calculate the vertical gradient of the smoothed image to obtain gradient information; S24: Detect local peaks based on gradient information, and select effective peaks as the boundary lines of underground structural layers according to preset conditions.

[0033] First, multiple B-scan images of the same road segment are overlaid pixel by pixel: ; In the formula, in the formula, This represents the gray value of the i-th B-scan image at pixel coordinates (x, y); This represents the grayscale value of the superimposed radar image at pixel coordinates (x, y), where N represents the number of B-scan images involved in the superposition; x is the horizontal pixel coordinate; and y is the vertical pixel coordinate.

[0034] The vertical gradient is calculated using the Sobel vertical operator: ; In the formula, To superimpose the vertical gradient intensity of the image at pixel (x,y), This is the Sobel vertical convolution kernel used to calculate the gradient value in the vertical direction; "*" indicates the convolution operation; O(x,y) is the superimposed radar image; Local peaks are detected in the gradient intensity sequence, and effective peaks are selected as underground layer boundaries based on preset parameters such as minimum interval and noise filtering. The mean fusion of the layer lines of each road segment is performed to generate a stable layer boundary line to determine the horizontal slice depth; Specifically, firstly, vertical profile images of multiple adjacent channels are selected from 3D ground-penetrating radar data collected from the same road segment, and spatial registration is performed on the vertical profile images to ensure consistency in the lateral and depth directions of each channel. Then, the registered vertical profile images are superimposed pixel-level to enhance the stable reflection characteristics caused by the road structure layer interface and reduce the influence of random noise and local anomalous reflections, resulting in a superimposed image. Based on this, the superimposed image is smoothed so that the structural layer reflection appears as a continuous and smooth brightness variation region, thereby reducing noise interference to subsequent analysis. Next, gradient information is calculated along the vertical direction of the smoothed superimposed image to characterize the variation in reflection intensity at different depths. Due to the significant change in dielectric constant at the structural layer interface, it manifests as a distinct local extremum in the gradient information. Finally, based on the gradient information, local peak detection is performed on the gradient sequence at each lateral position, and effective peaks are selected by combining preset gradient amplitude thresholds, peak spacing, or continuity constraints. The positions corresponding to the effective peaks are determined as the boundary lines of the underground structural layers.

[0035] As a preferred embodiment of the above embodiment, in step S30, obtaining preliminary disease candidate regions on the horizontal slice includes: S31: Perform sliding window detection on the horizontal slice image to generate multiple detection windows; S32: Input the detection window into the first disease detection model to obtain multiple initial disease candidate boxes; S33: Merge and filter multiple initial disease candidate boxes to form a preliminary set of disease candidate regions.

[0036] Specifically, firstly, based on the spatial resolution of the horizontal slice image and the possible scale range of the disease, the horizontal slice image is divided into sliding windows. The window size and the overlap ratio between adjacent windows are set, and the windows are slid sequentially along the horizontal and vertical directions of the image to generate multiple detection windows covering the entire horizontal slice image, so as to avoid missed disease detection due to window boundary truncation. Then, each detection window is used as input and fed into the first disease detection model in batches to obtain multiple initial disease candidate boxes containing location and confidence information. Since the same disease area may be repeatedly identified in different detection windows, the obtained initial disease candidate boxes have a certain degree of spatial overlap. On this basis, the initial disease candidate boxes are fused and filtered. By analyzing the spatial overlap relationship between candidate boxes, candidate boxes belonging to the same disease area are merged. At the same time, redundant candidate boxes that obviously do not conform to the disease characteristics are removed according to preset size thresholds, confidence thresholds, or morphological constraints, thereby forming a stable and reliable preliminary disease candidate region set in the horizontal slice dimension.

[0037] Based on the range of the layer boundary line, horizontal slice images at multiple depths are extracted and input into the trained OEM-HWNet_C-slice two-dimensional slice disease detection model for sliding window detection. Then, the candidate boxes detected by horizontal slices at different depths are fused and the detection results are output. The result is the bounding box of the location of the disease in the two-dimensional horizontal slice, which serves as the preliminary disease candidate region. The sliding window detection performs a horizontal sliding window along the width direction of the horizontal slice, with a detection window interval of: ; In the formula, The overlap rate, The horizontal step size between adjacent sliding windows; The width of a single sliding window area; The sliding window overlap ratio is 0 to 1, such as 0.3 representing a 30% overlap area; Candidate box fusion employs strategies such as IOU, inclusion rate, and edge contact merging to achieve unified fusion and obtain preliminary disease candidate regions.

[0038] In step S30, the fusion and filtering process includes: performing box fusion based on at least one of the cross-union ratio, inclusion rate and edge contact relationship between candidate boxes, and removing redundant candidate boxes based on preset size threshold and confidence threshold.

[0039] In this embodiment, in step S30, the first disease detection model is a neural network model based on the improved OEM-HWNet network structure, specifically designed for horizontal slice image detection. Each candidate region in the preliminary disease candidate region set is represented by its absolute horizontal coordinate on the horizontal slice.

[0040] In this embodiment, in step S40, the preliminary disease candidate region is mapped to a vertical profile image, and the local vertical profile image corresponding to the candidate region in each channel is extracted, including: S41: Based on the lateral coordinate range and channel number of the preliminary disease candidate region in the horizontal slice image, determine its corresponding lateral position range in the vertical profile image of each channel. S42: Based on the preset longitudinal analysis depth, extract image blocks from the vertical profile images of each channel along the lateral position interval to form a local vertical profile image.

[0041] First, the lateral coordinate range of the preliminary disease candidate areas in the horizontal slice image is obtained to characterize the spatial location of the disease in the lateral direction of the road. Combined with the channel layout relationship of the 3D ground-penetrating radar system, the coordinate correspondence between the horizontal slice image and the vertical profile image of each channel is established. Based on the lateral coordinate range and channel number information, the position of the disease candidate area in the horizontal slice image is mapped to the vertical profile image of each channel to determine its corresponding lateral position interval in each channel vertical profile image. Subsequently, according to the preset longitudinal analysis depth, the corresponding image block is extracted from the vertical profile image of each channel along the depth direction within the lateral position interval to form a multi-channel local vertical profile image corresponding to the space of the disease candidate area.

[0042] The candidate bounding boxes are located in the corresponding multiple channel regions on the B-scan image, and multi-channel local B-scan images are captured. The extracted B-scan local image is input into the trained OEM-HWNet_B-scan disease detection model to obtain disease category information for different channels.

[0043] In step S50, the disease category information for each channel is obtained, including: S51: Input the local vertical profile images of each channel into the second disease detection model; S52: Obtain the disease category and related attribute information corresponding to each local area, output by the second disease detection model; S53: Based on the horizontal coordinates of the preliminary disease candidate area, store disease category and related attribute information by channel.

[0044] First, the local vertical profile images of each channel obtained in step S40 are used as input and processed by the second disease detection model either by channel or by batch. The second disease detection model analyzes the reflection structure features in the local vertical profile images and outputs disease category information and related attribute information for the corresponding local areas. The disease category information is used to characterize the type of disease, and the related attribute information is used to characterize the confidence level, spatial size, or structural features of the disease. Subsequently, based on the horizontal coordinates of the preliminary disease candidate areas in the horizontal slice image, the disease category information and related attribute information obtained from different channels are uniformly organized and stored, ensuring a clear one-to-one correspondence between the identification results of the same disease candidate area in different channels. In this way, the disease identification results obtained from multi-channel vertical profiles can be managed in a structured form, providing a clear and reliable data foundation for the consistency analysis and comprehensive judgment of multi-channel disease information in the subsequent fusion decision stage, thereby improving the stability and reliability of disease type identification results.

[0045] Based on the global horizontal coordinates and channel number of the candidate bounding boxes, the width of the lesions detected in the local images of different channels using B-scan is statistically analyzed. Then, the final lesion type at that location is determined by weighting the severity and number of lesions, and it is then associated with the candidate bounding boxes of the C-slice. The specific operation is as follows: First, calculate the coverage length of the disease frame in each channel within the horizontal range of the candidate frame, and select the disease with the largest width as the disease type for each channel. As a preferred embodiment of the above, step S60 includes making a fusion decision: S61: Statistically analyze the disease category information detected in each channel and its coverage width within the horizontal range of the preliminary disease candidate area; S62: Calculate the comprehensive score for each disease type based on the preset severity weight and coverage width; S63: Based on the overall score, determine the primary and / or secondary disease types as the final disease types.

[0046] Specifically, since the same underground disease may exhibit different reflection intensities, structural continuity, and identifiable ranges in different radar channels, relying solely on a single channel or a simple majority voting method is easily affected by local noise or occasional abnormal reflections. Therefore, in this embodiment, a weighted fusion mechanism is introduced to determine the disease type by comprehensively considering the spatial coverage of the disease in multiple channels and the severity of different disease types. For the same preliminary disease candidate area, the coverage width of each disease category within the lateral range of the candidate area is statistically analyzed. The coverage width is used to characterize the continuous distribution range of the disease type in the lateral direction of the road. By introducing the coverage width index, the identification results of diseases with a larger spatial range and more continuous structure have a higher weight in the fusion decision, thereby reducing the interference of local false detections on the final result. For each type of defect, a severity weight value is pre-set according to its impact on road structural safety. For example, defect types with a greater impact on bearing capacity are given a higher weight, while defect types with a relatively smaller impact on structure are given a lower weight. Then, the severity weight is combined with the coverage width of the corresponding defect type to obtain the comprehensive score of each defect type. The defect type with the highest comprehensive score is determined as the primary defect type. When the comprehensive scores of multiple defect types exceed the preset threshold, the defect type with the second highest comprehensive score can be further determined as the secondary defect type, thus forming a combined judgment result of "primary defect type and / or secondary defect type". Finally, the primary defect type and / or secondary defect type are used as the final defect type corresponding to the preliminary defect candidate area and are associated with the candidate area for output.

[0047] In this embodiment, in step S62, the overall score is calculated using the following formula: ; In the formula, This is the overall score for disease category k; Weights are assigned based on the severity of the disease; This refers to the number of adjacent channels with the same disease type.

[0048] Specifically, to quantitatively fuse multi-channel detection results of different disease types within the same preliminary disease candidate area, a comprehensive score is calculated for each disease type. This comprehensive score is obtained as follows: First, based on the disease category identification results stored by channel, the number of channels detecting the same disease type is counted to characterize the consistency and spatial stability of the disease in the multi-channel vertical profile. Simultaneously, corresponding severity weights are pre-set according to the degree of impact of different disease types on road structural safety. Then, the severity weight is multiplied by the number of channels corresponding to the disease type to obtain the comprehensive score for that disease type. This formulaic calculation method ensures that disease types that repeatedly appear in multiple channels and have a high degree of engineering hazard receive higher comprehensive scores, thus being prioritized as primary disease types in subsequent fusion decisions. Disease types that appear less frequently or have lower severity are treated as secondary disease types or not output. Therefore, by organically combining multi-channel consistency information with disease engineering attributes, the quantification and interpretability of the disease type fusion judgment process are achieved, significantly improving the stability and engineering applicability of the final disease type judgment results.

[0049] By comparing comprehensive scores, the results from multiple channels are fused at the primary and secondary levels in the form of a combination of "primary disease + secondary disease" to form the final disease type description. The final disease type is associated with the horizontal slice candidate box, that is, the multi-dimensional information is integrated and output as a unified disease detection result file.

[0050] Figure 8 This document presents a visualization example of detection results based on 3D ground-penetrating radar data. First, underground strata are determined through vertical profiles, and appropriate horizontal slices are selected. Then, the OEM-HWNet_C-slice model is used to detect and merge horizontal slices at different depths to obtain candidate disease regions. For example... Figure 8 The detection results for the upper and middle parts are shown. Subsequently, using the corresponding positions of the candidate regions in the vertical profile, local regions are extracted from the multi-channel image, and then the OEM-HWNet_B-scan model is used to detect information such as the type of disease. Figure 8 The results of the detection in the lower half are shown; finally, statistical analysis of these results is performed to obtain the final interlayer defect detection results based on three-dimensional ground penetrating radar.

[0051] Example 2: This invention also includes an intelligent highway defect detection system based on three-dimensional ground-penetrating radar, such as... Figure 9 As shown, the system includes: The data acquisition and preprocessing module is used to acquire and process three-dimensional radar data to obtain horizontal slice images distributed along the depth direction and vertical profile images of multiple channels distributed along the survey line direction. The underground stratum determination module is used to determine the location of the boundary line of the underground structural layers based on multiple vertical profile images; The horizontal slice disease initial localization module is used to select the horizontal slice image at the corresponding depth according to the boundary line position and input it into the first disease detection model to obtain the preliminary disease candidate area on the horizontal slice. The vertical profile image cropping module is used to map the preliminary disease candidate area to the vertical profile image and crop the local vertical profile image corresponding to the candidate area in each channel. The vertical profile disease classification module is used to input the local vertical profile images of each channel into the second disease detection model to obtain disease category information for each channel. The multi-dimensional information fusion decision module is used to make fusion decisions based on the horizontal position of the preliminary disease candidate area and the disease category information of each channel, to determine the final disease type and associate it with the preliminary disease candidate area.

[0052] The adjustment system described above in this invention can effectively realize the intelligent detection method for highway defects based on three-dimensional ground penetrating radar, and the technical effects it can achieve are as described in the above embodiments, and will not be repeated here.

[0053] Similarly, the above-mentioned optimization schemes for the system can also achieve the optimization effects corresponding to the methods in Embodiment 1, which will not be repeated here.

[0054] Although this application has been described in conjunction with specific features and embodiments, it is obvious that various modifications and combinations can be made thereto without departing from the spirit and scope of this application. Accordingly, this specification and accompanying drawings are merely exemplary illustrations of the application as defined herein, and are to be considered as covering any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from its scope. Thus, if such modifications and modifications fall within the scope of this application and its equivalents, this application intends to include such modifications and modifications.

Claims

1. A method for intelligent detection of highway defects based on three-dimensional ground-penetrating radar, characterized in that, The method includes: Acquire and process 3D radar data to obtain horizontal slice images distributed along the depth direction and vertical profile images of multiple channels distributed along the survey line direction; Based on multiple vertical cross-sectional images, the location of the boundary line of the underground structural layers is determined; Based on the location of the boundary line, select the horizontal slice image at the corresponding depth and input it into the first disease detection model to obtain the preliminary disease candidate region on the horizontal slice; The preliminary disease candidate region is mapped to the vertical profile image, and the local vertical profile image corresponding to the candidate region in each channel is extracted; The local vertical profile images of each channel are input into the second disease detection model to obtain disease category information for each channel; Based on the lateral position of the preliminary disease candidate area and the disease category information of each channel, a fusion decision is made to determine the final disease type and associate it with the preliminary disease candidate area.

2. The intelligent detection method for highway defects based on three-dimensional ground-penetrating radar according to claim 1, characterized in that, Making integration decisions includes: The disease category information detected in each channel and its coverage width within the horizontal range of the preliminary disease candidate area are statistically analyzed. Calculate the comprehensive score for each disease type based on the preset severity weights and the coverage width; Based on the comprehensive score, the primary disease type and / or secondary disease type are determined as the final disease type.

3. The intelligent detection method for highway defects based on three-dimensional ground-penetrating radar according to claim 2, characterized in that, The overall score is calculated using the following formula: ； In the formula, This is the overall score for disease category k; Weights are assigned based on the severity of the disease; This refers to the number of adjacent channels with the same disease type.

4. The intelligent detection method for highway defects based on three-dimensional ground-penetrating radar according to claim 1, characterized in that, The first disease detection model includes a C3WC module, which extracts high-frequency edge features and low-frequency texture features from horizontal slice images simultaneously by embedding wavelet transform into convolution operations, thereby enhancing the ability to identify diseases with weak edges and large scale.

5. The intelligent detection method for highway defects based on three-dimensional ground-penetrating radar according to claim 1, characterized in that, The first disease detection model or the second disease detection model includes an OEM module; the OEM module generates an attention weight map through the following steps: Perform edge detection on the input image to generate an initial edge mask; The initial edge mask is normalized and mapped with Gaussian weights to generate an attention weight map with high center weights and gradually decreasing edge weights. The attention weight map is multiplied with the feature map of the first disease detection model or the second disease detection model to enhance the feature response that matches the prior morphology of interlayer diseases.

6. The intelligent detection method for highway defects based on three-dimensional ground-penetrating radar according to claim 1, characterized in that, The process of obtaining preliminary disease candidate regions on horizontal slices includes: Sliding window detection is performed on the horizontal slice image to generate multiple detection windows; Input the detection window into the first disease detection model to obtain multiple initial disease candidate boxes; The multiple initial disease candidate boxes are merged and filtered to form the preliminary disease candidate region set.

7. The intelligent detection method for highway defects based on three-dimensional ground-penetrating radar according to claim 1, characterized in that, The step of mapping the preliminary disease candidate region to the vertical profile image and extracting the local vertical profile image corresponding to the candidate region in each channel includes: Based on the lateral coordinate range and channel number of the preliminary disease candidate region in the horizontal slice image, determine its corresponding lateral position range in the vertical cross-sectional image of each channel. Based on the preset vertical analysis depth, image blocks are extracted from the vertical profile images of each channel along the horizontal position interval to form the local vertical profile image.

8. The intelligent detection method for highway defects based on three-dimensional ground-penetrating radar according to claim 1, characterized in that, The acquisition of disease category information for each channel includes: The local vertical profile images of each channel are input into the second disease detection model; Obtain the disease category and related attribute information corresponding to each local area, output by the second disease detection model; Based on the horizontal coordinates of the preliminary disease candidate regions, the disease categories and related attribute information are stored by channel.

9. The intelligent detection method for highway defects based on three-dimensional ground-penetrating radar according to claim 1, characterized in that, Determining the location of the boundary lines of the underground structural layers includes: Multiple vertical cross-sectional images are superimposed pixel-by-pixel to obtain a superimposed image; The superimposed image is smoothed. Calculate the vertical gradient of the smoothed image to obtain gradient information; Based on the gradient information, local peak values ​​are detected, and effective peak values ​​are selected as the boundary line positions of the underground structural layer according to preset conditions.

10. A highway defect intelligent detection system based on three-dimensional ground-penetrating radar, characterized in that, The system includes: The data acquisition and preprocessing module is used to acquire and process three-dimensional radar data to obtain horizontal slice images distributed along the depth direction and vertical profile images of multiple channels distributed along the survey line direction. The underground stratum determination module is used to determine the location of the boundary line of the underground structural layer based on multiple vertical profile images; The horizontal slice disease initial localization module is used to select a horizontal slice image of corresponding depth according to the position of the boundary line, and input it into the first disease detection model to obtain the preliminary disease candidate area on the horizontal slice. The vertical profile image cropping module is used to map the preliminary disease candidate region to the vertical profile image and crop the local vertical profile image corresponding to the candidate region in each channel; The vertical profile disease classification module is used to input the local vertical profile images of each channel into the second disease detection model to obtain disease category information for each channel. The multi-dimensional information fusion decision module is used to make fusion decisions based on the lateral position of the preliminary disease candidate area and the disease category information of each channel, determine the final disease type, and associate it with the preliminary disease candidate area.

Images