Underground disease detection method and system based on three-dimensional multi-channel ground penetrating radar image

Abstract

The invention discloses an underground disease detection method and system based on a three-dimensional multi-channel ground penetrating radar image, and belongs to the technical field of underground disease detection. When the underground disease detection model is trained, not only is a disease target trained and recognized, but also a conventional target is included, the conventional target is prevented from being recognized as the disease target, so that the accuracy of disease recognition is improved, and by setting different bounding boxes for the two types of targets, it is guaranteed that the efficiency and accuracy requirements of model training are balanced. By setting a cache mechanism, the model is prevented from operating the same image for multiple times, and the operation overhead of the model is reduced; a channel difference threshold value and a mileage difference threshold value are set to identify a single disease, so that the single disease is prevented from being recorded for multiple times, and meanwhile, the disease identification accuracy can also be improved; and through a minimum continuous channel number threshold value, narrowband noise artifacts triggered by few channels are automatically eliminated.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of underground disease detection, and particularly relates to an underground disease detection method and system based on a three-dimensional multi-channel ground penetrating radar image. BACKGROUND

[0002] In recent years, a large number of existing highways have induced collapse disasters due to pavement structure aging, heavy traffic, underground cavities, disengagement and pipeline instability and other diseases, affecting driving safety. Therefore, periodic exploration of highways is needed to provide early warning of disasters and avoid affecting driving safety.

[0003] The exploration methods for the environment around the highway include core drilling exploration and radar exploration. The core drilling exploration has high precision but is highly destructive, has limited coverage and affects traffic. The radar exploration scans a large range of roadbed in a non-destructive manner at one time, has high exploration efficiency and does not affect traffic. Among them, the single-channel B-scan ground penetrating radar collects radar profiles along the direction of travel through a single antenna, and when the electromagnetic wave encounters a cavity, pipeline or well wall with significant dielectric difference, it shows hyperbolic reflection characteristics on the time-distance curve, which can be used for preliminary judgment of diseases. However, only a single profile is obtained at one time, and there is a lack of continuous data in the horizontal direction, and the real size and volume of the defect in the width and mileage directions of the road cannot be quantified. The multi-channel ground penetrating radar synchronously collects dozens of profiles through a parallel array antenna, significantly improves the horizontal sampling density, can generate a 3D radar wave field in real time on site, and improves the detection efficiency. However, the current multi-channel radar exploration still has problems of unclear defect category and inaccurate defect identification after exploration.

[0004] In summary, the existing multi-channel radar exploration method has the problems of unclear defect category and inaccurate defect identification. SUMMARY

[0005] The technical problem to be solved by the present application is to provide an underground disease detection method and system based on a three-dimensional multi-channel ground penetrating radar image in view of the deficiencies in the prior art.

[0006] To solve the above technical problems, the technical scheme adopted by the present application is as follows: An underground disease detection method based on a three-dimensional multi-channel ground penetrating radar image, comprising the following steps: S1, acquiring a B-scan image of the road surface and performing image preprocessing; S2, performing abnormal category labeling on the electromagnetic wave image, and constructing an abnormal feature data set according to the electromagnetic wave image and the abnormal category label; the abnormal category includes disease targets and conventional targets; S3, constructing an underground disease detection model according to a target detection algorithm, and training the underground disease detection model using the abnormal feature data set; Input the electromagnetic wave image to be detected into the trained underground disease detection model and output the corresponding label file; cache the electromagnetic wave image and the corresponding label file; the label file includes the label type and the label bounding box pixels; the label bounding box includes the diseased bounding box and the regular target bounding box; Before the underground disease detection model processes the input image, it first compares the input electromagnetic wave image with the cached electromagnetic wave image. If they are consistent, the pre-stored electromagnetic wave image label file is used as the input electromagnetic wave image label file and output; otherwise, it enters the underground disease detection model for processing; S4, based on the cached electromagnetic wave image and the corresponding label file, converting the label bounding box pixel position into the real position according to the conversion relationship between pixel coordinates and real position, and using it as the real position of the target; S5, merging two disease bounding boxes in the same image below the first distance difference threshold and below the depth threshold into a single disease area; Merge two disease bounding boxes on different images below the second distance difference threshold and the channel difference threshold into a single disease area; When the channel span in a single diseased area is less than the minimum channel threshold, the single diseased area is deleted; check whether the single diseased area overlaps with the conventional target bounding box. If so, delete the overlapping part; Output the final single disease area as the final detection result.

[0007] Furthermore, in step S2, during marking, fine frames are used for the damaged targets, and rectangular frames are used for the regular targets; the damaged targets include voids, cavities, and looseness, and the regular targets include pipelines and inspection wells.

[0008] Furthermore, the image preprocessing process includes sequentially performing direct wave removal, background removal, vertical in-band filtering and gain operation processing. Furthermore, the target retrieval algorithm adopts the YOLOv10 algorithm. The YOLOv10 algorithm includes three parts: Backbone, Neck, and Head. The Backbon part is used for feature extraction, the Neck part is used for feature fusion of the extracted features, and the Head part is used to generate the final output; The CBAM module is inserted after the output node of Neck's feature fusion. The CBAM module processes the feature map using channel attention and spatial attention respectively, and adds the processed feature maps to obtain the final output feature map.

[0009] Furthermore, the conversion relationship of step S4 is:

[0010]

[0011] in,( , ) is the real position coordinate, ( , ) pixel position coordinates, is the conversion coefficient of the mileage direction pixel, is the conversion coefficient of pixels in the depth direction.

[0012] Furthermore, the channel difference threshold is 3 channels to 4 channels, the first mileage difference threshold and the second mileage difference threshold are both 0.4m to 0.5m, and the depth difference threshold is less than the smaller height value of the two inspection boxes.

[0013] Furthermore, the minimum channel threshold is 1 or 2.

[0014] Furthermore, a spatial three-dimensional distribution image of the detection section is established based on the road width, actual road section and depth data; and a corresponding disease image is generated on the spatial three-dimensional distribution image based on the final detection result data.

[0015] The present invention also discloses an underground disease detection system based on three-dimensional multi-channel ground penetrating radar images, comprising: An image acquisition unit, configured to acquire a B-scan image of the road surface using a vehicle-mounted three-dimensional ground penetrating radar, and use the acquired image as input to a data processing unit; The data acquisition unit is used to process the acquired images according to the above-mentioned underground disease detection method based on three-dimensional multi-channel ground penetrating radar images and output the final detection results.

[0016] Furthermore, it also includes a visualization unit for establishing a spatial three-dimensional distribution image of the detection section based on the road width, actual road section and depth data; and generating a corresponding disease image on the spatial three-dimensional distribution image based on the final detection result data.

[0017] Compared with the prior art, the present invention has the following advantages: When training the underground disease detection model, the present invention not only trains the recognition of disease targets but also includes regular targets, avoiding the misidentification of regular targets as disease targets, thereby improving the accuracy of disease identification. By setting different bounding boxes for the two types of targets, the efficiency and accuracy requirements of model training are balanced. A caching mechanism is established to prevent the model from performing multiple calculations on the same image, reducing the model's operating overhead. Channel difference thresholds and mileage difference thresholds are set to identify a single disease, avoiding multiple recordings of a single disease and also improving disease identification accuracy. A minimum continuous channel number threshold is used to automatically eliminate narrowband noise artifacts triggered by very few channels.

[0018] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 Schematic diagram of the process of an embodiment of the underground disease detection method based on three-dimensional multi-channel ground penetrating radar images of the present invention; Figure 2 A ground penetrating radar image of an embodiment of the method for detecting underground diseases based on three-dimensional multi-channel ground penetrating radar images of the present invention; Figure 3 This is a B-scan image annotation diagram of an embodiment of the underground disease detection method based on three-dimensional multi-channel ground penetrating radar images of the present invention; Figure 4 A disease detection XML file for an embodiment of the underground disease detection method based on three-dimensional multi-channel ground penetrating radar images of the present invention; Figure 5 This is a YOLOv10x network structure diagram of an embodiment of the underground disease detection method based on three-dimensional multi-channel ground penetrating radar images of the present invention; Figure 6 This is a diagram of the damage detection results of an embodiment of the underground damage detection method based on three-dimensional multi-channel ground penetrating radar images of the present invention. DETAILED DESCRIPTION

[0020] Example of underground disease detection method based on three-dimensional multi-channel ground penetrating radar image: like Figures 1-6 As shown, the underground disease detection method based on three-dimensional multi-channel ground penetrating radar images includes the following steps: S1. Acquire a B-scan image of the road surface and perform image preprocessing. B-scan is a mode in electromagnetic wave detection technology that converts the amplitude of reflected waves at different depths into brightness points to generate a two-dimensional cross-sectional image, usually perpendicular to the scanning direction.

[0021] like Figure 2As shown, the inspection range is determined based on the road material type, the road structure layer combination and thickness, and the desired defect type. A 3D GPR system with an antenna of appropriate frequency is selected and mounted on the rear mounting bracket of the inspection vehicle. The GPR acquisition parameters, including the time window, sampling mode, number of sampling points, and sampling interval, are adjusted. The inspection vehicle travels at a normal speed on the road being inspected. The onboard 3D GPR collects B-scan images of the road surface, which are then used as input for the data processing unit. The vehicle travels along the length of the road, collecting images using multiple channels; preferably, 15 channels are provided along the width of the road. The onboard GPR system uses antennas to transmit electromagnetic waves to detect underground structures. Fifteen pairs of antennas operate simultaneously, corresponding to 15 equally spaced sampling lines in the horizontal direction. This allows the collection of reflected echo signals from the underground medium at multiple locations in the horizontal direction, providing richer data for underground structure analysis. Images can also be collected in 10-channel increments. B-scan GPR images are derived from all channels at 5-meter intervals.

[0022] The image preprocessing process includes direct wave removal, background removal, vertical bandpass filtering and gain operation processing in sequence. It can make the waveform abnormal area visible, and then improve the contrast of the image to make the abnormal features more obvious. Specifically, vertical bandpass filtering: suppresses the radar low-frequency drift and high-frequency noise, retains the main frequency part of the radar signal to improve the signal-to-noise ratio, highlights the energy of the layer reflection signal, and improves the resolution of the radar profile signal. Zero line correction: removes the first arrival wave received by the receiving antenna that does not enter the ground and is transmitted through the air. Background removal: select the obvious horizontal interference signal area in the radar profile, calculate the average of all channel data in this section as the background noise, and calculate the difference between all channels of the radar profile and the background noise to achieve the purpose of removing the background noise.

[0023] S2. Mark abnormal categories on the electromagnetic wave image and construct an abnormal feature dataset based on the electromagnetic wave image and abnormal category labels; like Figure 3 As shown, to prevent non-defective equipment from affecting the accuracy of defective target detection, abnormal labeling includes both defective and regular targets. Defective targets are labeled with fine boxes, while regular targets are labeled with rectangular boxes. Defects include voids, holes, and looseness, while regular targets include pipelines and manholes. Specifically, fine box labeling, also known as polygonal labeling, uses polygonal boxes to precisely delineate target objects. Rectangular box labeling involves drawing a rectangular box within the image to enclose the target object. It is convenient and efficient, and the labeling is output as a txt file.

[0024] S3. Build an underground disease detection model based on the target detection algorithm, and use the abnormal feature data set to divide it into a training set, a validation set, and a test set according to an 8:1:1 ratio to train the underground disease detection model.

[0025] like Figure 5 As shown, the target retrieval algorithm adopts the YOLOv10 algorithm. Preferably, the YOLOv10 algorithm adopts the YOLOv10x algorithm. The YOLOv10x algorithm includes three parts: Backbone, Neck, and Head. The Backbon part is used for feature extraction, the Neck part is used for feature fusion of the extracted features, and the Head part is used to generate the final output.

[0026] A CBAM (convolutional attention) module is inserted after the backbone feature fusion output node SPPF. The CBAM module processes the feature map using channel attention and spatial attention, respectively, and then adds the processed feature maps to produce the final output feature map. By introducing the CBAM attention module at each stage of the backbone, the channel attention branch automatically highlights the most valuable feature channels for disease detection based on global statistics, suppressing irrelevant background interference. Furthermore, the spatial attention branch generates a hotspot location map on the feature map, effectively enhancing the response to small, low-contrast targets. This model not only more accurately captures weak signals such as fine cracks, corrosion, and voids in underground pipelines, but also significantly reduces the false alarm rate caused by background clutter, thereby improving overall detection accuracy and robustness.

[0027] The backbone uses a C2f module to split the input feature map into two parallel paths along the channel direction. One path uses a 1×1 convolutional layer to halve the number of channels. The other path uses multiple layers of C2fCIB to extract rich features, and finally splices them together to restore the original channels. Subsequently, the model uses SCDown, first adjusting the channels with a 1×1 convolutional layer, and then performing spatial downsampling with a 3×3 convolution, effectively reducing computational redundancy and accelerating inference. SPPF is introduced on the deepest feature map, using 5×5, 9×9, and 13×13 pooling kernels to fuse multi-scale context. This is followed by a CBAM module that sequentially recalibrates channel and spatial attention to enhance salient features. Next, the PSA layer finely models global dependencies and dynamically generates channel weighted mappings through multi-head attention, further enhancing long-range semantic fusion.

[0028] Neck receives feature maps of different scales from P3, P4, and P5, and efficiently fuses them through Upsample, Concat, and C2f / C2fCIB modules. The NMS-free v10Detect detection head then outputs accurate object bounding boxes and category predictions end-to-end.

[0029] Specifically, shallow, high-resolution P3 features are first upsampled by a factor of two and concatenated with the next-level P4 features in the channel dimension. The concatenated features are then refined and their residuals reused via the C2f / C2fCIB modules. Similarly, the fused mid-level features are downsampled back to P4 resolution and concatenated with the deeper P5 features. A second round of C2fCIB modules further enhances semantic expression. Finally, the deep P5 features, while maintaining their original resolution, are finally integrated with the mid- and shallow-level features through upsampling and concat. This entire process not only balances shallow details with deep semantics, but also leverages the cross-stage connections and compact bottleneck design of the C2f series modules to achieve efficient feature reuse and dimensionality unification, providing the detection head with both rich and compact multi-scale feature input.

[0030] The Detect detection head generates rich candidate boxes through one-to-many branches in the training phase, and only retains the one-to-one branch in the inference phase to achieve end-to-end NMS-free detection, thereby significantly reducing post-processing delay and improving accuracy. In terms of loss function, the model uses binary cross entropy for classification, combines DFL and DIoU for bounding box regression, and aligns the supervision signals of the two branches through consistent matching metrics to enhance training robustness. During the training process, each round of iteration automatically updates the network weights through backpropagation according to the above comprehensive loss until the loss converges; the final model only enables the one-to-one branch during inference to achieve fast and accurate end-to-end target detection.

[0031] Input the electromagnetic wave image to be detected into the trained underground disease detection model and output the corresponding label file; cache the electromagnetic wave image and the corresponding label file to a specified directory for subsequent use; Figure 4 As shown in the figure, the label file includes the label type and the label bounding box pixels. The label bounding box includes the disease bounding box and the regular target bounding box.

[0032] In order to meet the input image size requirements of the underground disease detection model, mosaic data enhancement, adaptive anchor box calculation, and adaptive image scaling operations are performed on the image before input.

[0033] Before the underground disease detection model processes the input image, it first compares the input electromagnetic wave image with the cached electromagnetic wave image. If they are consistent, the label file of the pre-stored electromagnetic wave image is used as the label file of the input electromagnetic wave image and output; otherwise, it enters the underground disease detection model for processing.

[0034] S4. Based on the cached electromagnetic wave image and the corresponding label file, the pixel position of the label bounding box is converted to the real position according to the conversion relationship between pixel coordinates and real position, which is used as the real position of the disease. Specifically, the label file generated by batch inference is read, the pixel coordinates of each bounding box are extracted, the corresponding image area is binarized and the contour is extracted using OpenCV, and the Tesseract-OCR open source plug-in is called to automatically identify the mileage scale text and depth scale text of the B-scan image. Combining the mileage and pixel ratio of the image with the depth and pixel ratio of the side, the pixel span is converted into the actual mileage distance and depth value. The mileage unit is meters, and the depth unit is centimeters.

[0035] The conversion relationship of step S4 is:

[0036]

[0037] in,( , ) is the real position coordinate, ( , ) pixel position coordinates, is the conversion coefficient of the mileage direction pixel, is the conversion coefficient of pixels in the depth direction.

[0038] Specifically, during caching, the data are summarized into a new Excel workbook by category and saved in five worksheets named voids, cavities, looseness, pipelines, and inspection wells. Each worksheet has the following five columns in order: maximum mileage distribution, minimum mileage distribution, channel number span range, maximum depth distribution, and minimum depth distribution.

[0039] S5, such as Figure 6 As shown, two disease bounding boxes in the same image below the first distance difference threshold and the depth threshold are merged into a single disease area.

[0040] In order to avoid mis-segmentation caused by slight fluctuations in the mileage scale, two defect bounding boxes on different images below the second mileage difference threshold and the channel difference threshold are merged into a single defect area.

[0041] To filter out excessively narrow noise, if the channel span within a single defect region is less than the minimum channel threshold, the single defect region is deleted. The minimum channel threshold is 1 or 2, preferably 1. Because the defects discussed in this invention are relatively large and unlikely to be located within a single channel, defects located within a single channel are considered narrowband noise or artifacts triggered by only a few channels.

[0042] In order to avoid the influence of conventional targets on disease targets, the single disease area is detected to see if it overlaps with the conventional target bounding box. If so, the overlapping part is deleted.

[0043] Output the final single disease area as the final detection result.

[0044] Mileage refers to the distance from the detection starting point to the detection point along the length of the road. Because a single defect may be segmented by two or more channels during scanning, the same defect may appear in two images. Therefore, a channel threshold and a second mileage difference threshold are used. For example, voids, holes, and loose compaction have a relatively large distribution range and are prone to being detected by multiple channel images.

[0045] The channel difference threshold is 3 channels to 4 channels, and the first mileage difference threshold and the second mileage difference threshold are both 0.4m to 0.5m. Preferably, the channel difference threshold is 3 channels, and the first mileage difference threshold and the second mileage difference threshold are both 0.5m.

[0046] A spatial three-dimensional distribution image of the inspection section is established based on the road width, actual road section and depth data; and a corresponding disease image is generated on the spatial three-dimensional distribution image based on the final inspection result data.

[0047] Example of underground disease detection system based on three-dimensional multi-channel ground penetrating radar images: The underground disease detection system based on three-dimensional multi-channel ground penetrating radar images includes: an image acquisition unit for collecting B-scan images of the road surface using a vehicle-mounted three-dimensional ground penetrating radar, and using the collected images as input to a data processing unit.

[0048] The data acquisition unit is used to process the acquired images according to the above-mentioned embodiment of the underground disease detection method based on three-dimensional multi-channel ground penetrating radar images and output the final detection results.

[0049] It also includes a visualization unit for establishing a spatial three-dimensional distribution image of the detection section based on the road width, actual road section and depth data; and generating a corresponding disease image on the spatial three-dimensional distribution image based on the final detection result data.

[0050] This embodiment is implemented with reference to the above-mentioned embodiment of the underground disease detection method based on three-dimensional multi-channel ground penetrating radar image, which will not be described in detail here.

[0051] The above description is only a preferred embodiment of the present invention and does not limit the present invention in any way. Any simple modification, change and equivalent structural change made to the above embodiment based on the technical essence of the present invention shall still fall within the scope of protection of the technical solution of the present invention.

Claims

1. A method for detecting underground diseases based on three-dimensional multi-channel ground penetrating radar images, characterized by: The following steps are included: S1, obtaining a B-scan image of the road surface and performing image preprocessing; S2. Mark abnormal categories on the electromagnetic wave image and construct an abnormal feature dataset based on the electromagnetic wave image and abnormal category labels; The abnormal categories include diseased targets and regular targets; S3. Construct an underground disease detection model based on the target detection algorithm and train the underground disease detection model using the abnormal feature dataset; Input the electromagnetic wave image to be detected into the trained underground disease detection model and output the corresponding label file; cache the electromagnetic wave image and the corresponding label file; the label file includes the label type and the label bounding box pixels; the label bounding box includes the diseased bounding box and the regular target bounding box; Before the underground disease detection model processes the input image, it first compares the input electromagnetic wave image with the cached electromagnetic wave image. If they are consistent, the label file of the pre-stored electromagnetic wave image is used as the label file of the input electromagnetic wave image and outputs it; Otherwise, enter the underground disease detection model for processing; S4, based on the cached electromagnetic wave image and the corresponding label file, converting the label bounding box pixel position into the real position according to the conversion relationship between pixel coordinates and real position, as the real position of the target; S5, merging two disease bounding boxes in the same image below the first distance difference threshold and below the depth threshold into a single disease area; Merge two disease bounding boxes on different images below the second distance difference threshold and the channel difference threshold into a single disease area; When the channel span in a single diseased area is less than the minimum channel threshold, the single diseased area is deleted; Detect whether a single diseased area overlaps with the regular target bounding box. If so, delete the overlapping part. Output the final single disease area as the final detection result.

2. The underground disease detection method based on three-dimensional multi-channel ground penetrating radar images according to claim 1, characterized in that: In step S2, when marking, fine frames are used to mark the defective targets, and rectangular frames are used to mark the regular targets; the defects include voids, cavities, and looseness, and the regular targets include pipelines and inspection wells.

3. The underground disease detection method based on three-dimensional multi-channel ground penetrating radar images according to claim 1, characterized in that: The image preprocessing process includes direct wave removal, background removal, vertical band filtering and gain operation processing in sequence.

4. The underground disease detection method based on three-dimensional multi-channel ground penetrating radar images according to claim 1, characterized in that: The target retrieval algorithm adopts the YOLOv10 algorithm. The YOLOv10 algorithm includes three parts: Backbone, Neck, and Head. The Backbon part is used for feature extraction, the Neck part is used to fuse the extracted features, and the Head part is used to generate the final output; The CBAM module is inserted after the output node of Neck's feature fusion. The CBAM module processes the feature map using channel attention and spatial attention respectively, and adds the processed feature maps to obtain the final output feature map.

5. The underground disease detection method based on three-dimensional multi-channel ground penetrating radar images according to claim 1, characterized in that: The conversion relationship of step S4 is: in,( , ) is the real position coordinate, ( , ) pixel position coordinates, is the conversion coefficient of the mileage direction pixel, is the conversion coefficient of pixels in the depth direction.

6. The method for detecting underground diseases based on three-dimensional multi-channel ground penetrating radar images according to claim 1, characterized in that: The channel difference threshold is 3 channels to 4 channels, the first mileage difference threshold and the second mileage difference threshold are both 0.4m to 0.5m, and the depth difference threshold is less than the smaller height value in the two inspection boxes.

7. The method for detecting underground diseases based on three-dimensional multi-channel ground penetrating radar images according to claim 1, characterized in that: The minimum channel threshold is 1 or 2.

8. The underground disease detection method based on three-dimensional multi-channel ground penetrating radar images according to claim 1, characterized in that: Establish a spatial three-dimensional distribution image of the detection section based on road width, actual road section and depth data; The corresponding disease image is generated on the spatial three-dimensional distribution image according to the final detection result data.

9. A system for detecting underground diseases based on three-dimensional multi-channel ground-penetrating radar images, characterized by: include: An image acquisition unit, configured to acquire a B-scan image of the road surface using a vehicle-mounted three-dimensional ground penetrating radar, and use the acquired image as input to a data processing unit; The data acquisition unit is used to process the acquired images according to the underground disease detection method based on three-dimensional multi-channel ground penetrating radar images as described in any one of claims 1 to 8, and output the final detection results.

10. The underground disease detection system based on three-dimensional multi-channel ground penetrating radar images according to claim 9, characterized in that: It also includes a visualization unit for establishing a spatial three-dimensional distribution image of the detection section based on the road width, actual road section and depth data; and generating a corresponding disease image on the spatial three-dimensional distribution image based on the final detection result data.

Images