Road defect identification method and electronic equipment

Abstract

The invention provides a road defect identification method and electronic equipment, and the method comprises the steps: obtaining three-dimensional ground penetrating radar data corresponding to the interior of a target road, carrying out the voxel segmentation of the three-dimensional ground penetrating radar data, and obtaining a plurality of three-dimensional voxels corresponding to the three-dimensional ground penetrating radar data; the multiple three-dimensional voxels are input into a feature extraction model obtained through pre-training, multiple sets of initial feature maps are output through processing of the feature extraction model, each set of initial feature maps comprises multiple initial feature sub-maps, and each initial feature sub-map corresponds to a feature containing a target detection direction; analyzing and processing the plurality of groups of initial feature maps to obtain an initial defect direction; determining a target weight corresponding to each detection direction according to the initial defect direction; according to the target weight and the multiple sets of initial feature maps, target defect data corresponding to the interior of the target road are determined through a second defect detection model obtained through pre-training. According to the invention, accurate identification of the internal defects of the road is realized.

Description

Technical Field

[0001] This disclosure relates to the field of data processing, and more particularly to a method and electronic device for identifying road defects. Background Technology

[0002] With increasing road traffic volume, regular, efficient, and accurate non-destructive testing of road structures is crucial. 3D ground-penetrating radar (GPR) has become a cutting-edge technology in this field due to its ability to quickly acquire three-dimensional spatial data beneath the pavement structure. The key to current technological development lies in how to automatically and accurately identify defects such as voids, looseness, and cracks from massive amounts of complex 3D radar data.

[0003] Currently, defect detection is usually performed by combining dimensionality reduction slicing with two-dimensional convolutional network recognition. However, dimensionality reduction slicing forces the 3D radar data to be sliced, which disrupts the spatial continuity of defects and easily misjudges real defects as local anomalies, thus leading to inaccurate defect identification and limited recognition accuracy. Summary of the Invention

[0004] In view of this, the purpose of this disclosure is to propose a road defect identification method and electronic device to solve or partially solve the above problems.

[0005] To achieve the above objectives, a first aspect of this disclosure provides a road defect identification method, comprising: Acquire three-dimensional ground-penetrating radar data corresponding to the interior of the target road, and perform voxel segmentation on the three-dimensional ground-penetrating radar data to obtain multiple three-dimensional voxels corresponding to the three-dimensional ground-penetrating radar data. The multiple three-dimensional voxels are input into a pre-trained feature extraction model. After processing by the feature extraction model, multiple sets of initial feature maps are output. Each set of initial feature maps includes multiple initial feature sub-maps, and each initial feature sub-map corresponds to a feature containing the target detection direction. The initial defect direction is obtained by analyzing and processing the multiple sets of initial feature maps; Based on the initial defect direction, determine the target weight corresponding to each detection direction; Based on the target weights and the multiple sets of initial feature maps, the target defect data corresponding to the interior of the target road is determined using the pre-trained second defect detection model.

[0006] Based on the same inventive concept, a second aspect of this disclosure proposes a road defect identification device, comprising: The data acquisition module is configured to acquire three-dimensional ground-penetrating radar data corresponding to the interior of the target road, perform voxel segmentation on the three-dimensional ground-penetrating radar data, and obtain multiple three-dimensional voxels corresponding to the three-dimensional ground-penetrating radar data. The initial feature map determination module is configured to input the plurality of three-dimensional voxels into a pre-trained feature extraction model, and through the feature extraction model, output multiple sets of initial feature maps, wherein each set of initial feature maps includes multiple initial feature sub-maps, and each initial feature sub-map corresponds to a feature containing the target detection direction; The defect direction determination module is configured to analyze and process the multiple sets of initial feature maps to obtain the initial defect direction; The target weight determination module is configured to determine the target weight corresponding to each detection direction based on the initial defect direction; The defect data determination module is configured to determine the target defect data corresponding to the interior of the target road based on the target weights and the multiple sets of initial feature maps, using a pre-trained second defect detection model.

[0007] Based on the same inventive concept, a third aspect of this disclosure proposes an electronic device including a memory, a processor, and a computer program stored in the memory and executable by the processor, wherein the processor implements the method described above when executing the computer program.

[0008] Based on the same inventive concept, a fourth aspect of this disclosure provides a non-transitory computer-readable storage medium that stores computer instructions for causing a computer to perform the methods described above.

[0009] As can be seen from the above, this disclosure proposes a road defect identification method and electronic device. It acquires three-dimensional ground-penetrating radar (GPR) data corresponding to the interior of a target road, performs voxel segmentation on the GPR data to obtain multiple three-dimensional voxels. These multiple voxels are input into a pre-trained feature extraction model, which processes the data to output multiple sets of initial feature maps. Each set of initial feature maps includes multiple initial feature sub-maps, each corresponding to a feature containing the target detection direction. The multiple sets of initial feature maps are analyzed to obtain initial defect directions. Based on these initial defect directions, target weights corresponding to each detection direction are determined. Based on the target weights and the multiple sets of initial feature maps, a pre-trained second defect detection model is used to determine the target defect data corresponding to the interior of the target road. By segmenting the three-dimensional GPR data into voxels and then directly analyzing these voxels to obtain the final target defect data, the method avoids slicing the three-dimensional GPR data to reduce it to two-dimensional data, preserving complete three-dimensional information and effectively distinguishing between real defects and noise. Meanwhile, after determining the initial defect direction, the target weight corresponding to each detection direction is determined based on the initial defect direction. Then, the target defect data is determined based on the target weight and multiple sets of initial feature maps. This allows the second defect detection model to utilize the preliminary judgment of the defect morphology and integrate prior knowledge into the final defect identification, thereby improving the accuracy of the target defect data determination. Attached Figure Description

[0010] To more clearly illustrate the technical solutions in this disclosure or related technologies, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only embodiments of this disclosure. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 This is a flowchart of a road defect identification method according to an embodiment of the present disclosure; Figure 2 This is a schematic diagram of the structure of a defect identification model according to another embodiment of the present disclosure; Figure 3 This is a flowchart of a road defect identification method according to another embodiment of the present disclosure; Figure 4 This is a schematic diagram of the model network structure according to another embodiment of the present disclosure; Figure 5 This is a schematic diagram of the road defect identification device according to an embodiment of the present disclosure; Figure 6 This is a schematic diagram of the structure of an electronic device according to an embodiment of the present disclosure. Detailed Implementation

[0012] To make the objectives, technical solutions, and advantages of this disclosure clearer, the following detailed description is provided in conjunction with specific embodiments and the accompanying drawings.

[0013] It should be noted that, unless otherwise defined, the technical or scientific terms used in the embodiments of this disclosure should have the ordinary meaning understood by one of ordinary skill in the art to which this disclosure pertains. The terms "first," "second," and similar terms used in the embodiments of this disclosure do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0014] The following are definitions of terms used in this disclosure: GPR: Ground Penetrating Radar (GPR) is a non-destructive testing method that uses high-frequency electromagnetic waves to detect the distribution of underground media.

[0015] 3D GPR: Three-dimensional ground penetrating radar, a ground penetrating radar system capable of acquiring three-dimensional spatial data volumes of underground targets, whose data is usually represented as a three-dimensional data volume (Data Cube).

[0016] B-Scan: A two-dimensional depth profile obtained by GPR continuously acquiring data along a survey line, reflecting electromagnetic wave reflection information in the vertical plane below the survey line.

[0017] C-Scan: A two-dimensional planar image of GPR at a certain depth, synthesized from multiple parallel B-Scan data, reflecting electromagnetic wave reflection information on the horizontal plane at that depth.

[0018] Transformer: A deep learning model based on self-attention mechanism, originally used for natural language processing, and now widely used in fields such as computer vision.

[0019] Self-Attention: A mechanism for calculating the correlation and importance between different positions in a sequence or dataset.

[0020] Voxel: short for volume element, is a point in three-dimensional space, corresponding to a pixel in a two-dimensional image. A 3D GPR data volume can be regarded as being composed of a large number of voxels.

[0021] Anisotropy refers to the property of a substance that its physical properties differ in different directions.

[0022] With increasing road traffic volume, regular, efficient, and accurate non-destructive testing of road structures is crucial. 3D ground-penetrating radar (GPR) has become a cutting-edge technology in this field due to its ability to quickly acquire three-dimensional spatial data beneath the pavement structure. The key to current technological development lies in how to automatically and accurately identify defects such as voids, looseness, and cracks from massive amounts of complex 3D radar data.

[0023] Currently, defect detection typically employs a combination of dimensionality reduction slicing and 2D convolutional network recognition. The specific process involves acquiring a 3D GPR data volume. This 3D data volume is then sliced ​​into hundreds or even thousands of independent 2D B-Scan images along a specific direction (usually the direction of travel along a road). Each B-Scan image is input into a pre-trained 2D convolutional neural network, such as U-Net, YOLO, or Faster R-CNN. The network extracts features from the 2D images and identifies regions where defects may exist, outputting the defect location and category for each 2D image. By summarizing the recognition results of all 2D slices and combining them with their spatial coordinate information, the approximate shape and location of the defect in 3D space are reconstructed manually or through algorithms.

[0024] First, because 3D GPR data is dimensionality reduced and sliced ​​into a large number of independent 2D images (such as B-Scan) for processing, this method breaks the spatial continuity of the data and loses the key correlation information across different slices, making it difficult to accurately distinguish between real, three-dimensional defects and random noise, resulting in limited recognition accuracy.

[0025] Meanwhile, underground defects (such as horizontal cracks and vertical settlement) exhibit significant directionality and irregularity, with their signal characteristics (reliability) varying greatly across radar scan profiles in different directions. Existing models typically treat all dimensions of data equally, failing to adjust analysis strategies based on the defect's morphological characteristics (anisotropy). The lack of a mechanism for dynamically and selectively analyzing data leads to the easy omission or misjudgment of defects with specific morphologies.

[0026] Finally, the computational complexity of the standard global self-attention mechanism increases quadratically with the number of data points. When directly applied to massive amounts of 3D GPR voxel data, it generates unbearable computational overhead and memory usage, making it difficult to effectively train and deploy advanced model architectures in practice.

[0027] To address the aforementioned problems, this embodiment proposes a road defect identification method, such as... Figure 1 As shown, the method includes: Step 101: Obtain the three-dimensional ground-penetrating radar data corresponding to the interior of the target road, and perform voxel segmentation on the three-dimensional ground-penetrating radar data to obtain multiple three-dimensional voxels corresponding to the three-dimensional ground-penetrating radar data.

[0028] In practice, three-dimensional ground-penetrating radar (GPR) data corresponding to the interior of the target road is acquired, and the GPR data is then segmented into voxels. The main principle of three-dimensional GPR is based on the different propagation characteristics of high-frequency electromagnetic waves in different media. When the high-frequency electromagnetic waves emitted by the transmitting antenna encounter an interface with a different dielectric constant, a reflected echo is generated. Based on the time and form of the reflected echo received by the receiving antenna, the distance to the reflecting interface can be determined, and the potential nature of the reflector can be assessed.

[0029] In this embodiment, because the 3D ground-penetrating radar data is represented in the form of a 3D data volume, it can obtain specific data information of the target road's internal profile in different directions and positions. Therefore, voxel segmentation is performed on the 3D ground-penetrating radar data, which means dividing the 3D data volume into multiple continuous and non-overlapping small 3D data blocks in space, resulting in a voxel sequence. The voxel sequence contains multiple 3D voxels, each corresponding to each of the aforementioned small 3D data blocks, and each 3D voxel contains its spatial location encoding information.

[0030] In this embodiment, before voxel segmentation of the 3D ground-penetrating radar data, the acquired 3D ground-penetrating radar data can be preprocessed to eliminate noise in the data. The preprocessing methods include denoising and gain compensation.

[0031] Step 102: Input the multiple three-dimensional voxels into the pre-trained feature extraction model. After processing by the feature extraction model, output multiple sets of initial feature maps. Each set of initial feature maps includes multiple initial feature sub-maps, and each initial feature sub-map corresponds to a feature containing the target detection direction.

[0032] In practice, a pre-trained feature extraction model is obtained, and the multiple three-dimensional voxels are input into the feature extraction model. After processing by the feature extraction model, multiple sets of initial feature maps are output. Each set of initial feature maps includes multiple initial feature sub-maps, and each initial feature sub-map corresponds to a feature containing the target detection direction.

[0033] In this embodiment, the target detection direction refers to the detection direction when the three-dimensional ground penetrating radar performs detection, that is, the situation inside the target road detected from different spatial dimensions and spatial angles.

[0034] Step 103: Analyze and process the multiple sets of initial feature maps to obtain the initial defect direction.

[0035] In practice, the multiple sets of initial feature maps are analyzed and processed to obtain initial defect directions, whereby the initial defect directions represent the defect distribution directions corresponding to defects within the target road surface. For example, defect directions include horizontal distribution, vertical distribution, and clumping. It is understood that the aforementioned horizontal and vertical directions are directions corresponding to the road surface of the target road.

[0036] Step 104: Determine the target weight corresponding to each detection direction based on the initial defect direction.

[0037] In practice, because the initial defect direction characterizes the distribution direction of the defect, the corresponding radar data obtained when observing the defect from different spatial angles will have different characterization features. That is, the same defect will present different signal characteristics on radar scan profiles in different directions, which is the anisotropy of the defect. In this embodiment, anisotropy specifically refers to the significant differences in echo signal characteristics of underground defects when scanning in different directions due to their different shapes and orientations.

[0038] For example, if the defects are distributed horizontally, then when observing in a direction parallel to the road surface of the target road, the radar data should show a relatively long segment consisting entirely of defects. Conversely, when observing in a direction perpendicular to the road surface of the target road, the radar data should show a shorter segment consisting entirely of defects. Therefore, based on the initial defect direction, the target weight corresponding to each detection direction is determined.

[0039] Specifically, to improve the recognition accuracy of the subsequent second defect detection model, the weight of the initial feature sub-image in the same direction as the initial defect direction should be set to a slightly larger weight value, so that the subsequent second defect detection model can pay more attention to the features in that direction.

[0040] Step 105: Based on the target weights and the multiple sets of initial feature maps, the target defect data corresponding to the interior of the target road is determined using the pre-trained second defect detection model.

[0041] In specific implementation, a pre-trained second defect detection model is obtained, and target defect data corresponding to the inside of the target road is determined using the pre-trained second defect detection model based on the target weights and the multiple sets of initial feature maps.

[0042] In this embodiment, the target defect data is descriptive information about defects inside the target road. The target defect data includes the three-dimensional spatial coordinates of the defect, its geometric dimensions, volume, type, and direction. The defect types include cavities, looseness, cracks, and voids.

[0043] The above scheme acquires 3D ground-penetrating radar (GPR) data corresponding to the interior of the target road. The GPR data is then segmented into voxels to obtain multiple 3D voxels. These voxels are input into a pre-trained feature extraction model, which processes the data to output multiple sets of initial feature maps. Each initial feature map includes multiple initial feature sub-maps, each corresponding to a feature containing the target detection direction. The initial defect directions are then analyzed to obtain the initial defect directions. Based on these initial defect directions, the target weights corresponding to each detection direction are determined. Based on the target weights and the multiple initial feature maps, a pre-trained second defect detection model is used to determine the target defect data corresponding to the interior of the target road. By segmenting the 3D GPR data into voxels and directly analyzing these voxels to obtain the final target defect data, the scheme avoids slicing the 3D GPR data to reduce it to 2D data, preserving complete 3D information and effectively distinguishing between real defects and noise. Meanwhile, after determining the initial defect direction, the target weight corresponding to each detection direction is determined based on the initial defect direction. Then, the target defect data is determined based on the target weight and multiple sets of initial feature maps. This allows the second defect detection model to utilize the preliminary judgment of the defect morphology and integrate prior knowledge into the final defect identification, thereby improving the accuracy of the target defect data determination.

[0044] In some embodiments, the feature extraction model includes a first feature extraction module, a second feature extraction module, a third feature extraction module, and a fourth feature extraction module. In step 102, the plurality of three-dimensional voxels are input into the pre-trained feature extraction model, and after processing by the feature extraction model, multiple sets of initial feature maps are output, specifically including: Step 1021: Encode each three-dimensional voxel to obtain the three-dimensional voxel code corresponding to each three-dimensional voxel.

[0045] Step 1022: Input multiple three-dimensional voxel codes into the first feature extraction module, and process them through the first feature extraction module to obtain the first initial feature map.

[0046] Step 1023: Input the first initial feature map into the second feature extraction module, and process it through the second feature extraction module to obtain the second initial feature map.

[0047] Step 1024: Input the second initial feature map into the third feature extraction module, and process it through the third feature extraction module to obtain the third initial feature map.

[0048] Step 1025: Input the third initial feature map into the fourth feature extraction module, and process it through the fourth feature extraction module to obtain the fourth initial feature map.

[0049] Step 1026: The first initial feature map, the second initial feature map, the third initial feature map, and the fourth initial feature map are used as multiple sets of initial feature maps.

[0050] In practice, each three-dimensional voxel is encoded, that is, spatial feature encoding is performed on each three-dimensional voxel to obtain the three-dimensional voxel code corresponding to each three-dimensional voxel.

[0051] Multiple 3D voxel codes are input into a first feature extraction module, and processed by the first feature extraction module to obtain a first initial feature map. The first feature extraction module includes a window multi-head self-attention unit and a shifted window multi-head self-attention unit.

[0052] The first initial feature map is input to the second feature extraction module, and processed by the second feature extraction module to obtain a second initial feature map. The second feature extraction module includes a first feature merging unit and a first feature extraction unit.

[0053] The second initial feature map is input to the third feature extraction module, and processed by the third feature extraction module to obtain the third initial feature map. The third feature extraction module includes a second feature merging unit and a second feature extraction unit.

[0054] The third initial feature map is input to the fourth feature extraction module, and processed by the fourth feature extraction module to obtain the fourth initial feature map. The fourth feature extraction module includes a third feature merging unit and a third feature extraction unit.

[0055] The first initial feature map, the second initial feature map, the third initial feature map, and the fourth initial feature map are used as multiple sets of initial feature maps.

[0056] In this embodiment, the network structures corresponding to the first feature extraction unit, the second feature extraction unit, and the third feature extraction unit can be the same or different. Preferably, the network structures corresponding to the first feature extraction unit, the second feature extraction unit, and the third feature extraction unit are different to further improve the accuracy of the final multiple sets of initial feature maps.

[0057] In some embodiments, the first feature extraction module includes a window multi-head self-attention unit and a shifted window multi-head self-attention unit. In step 1022, the plurality of three-dimensional voxel codes are input to the first feature extraction module, and processed by the first feature extraction module to obtain a first initial feature map, specifically including: Step 10221: Input the multiple three-dimensional voxel codes into the window multi-head self-attention unit, and process them through the window multi-head self-attention unit to obtain the target extraction feature map; Step 10222: The target extracted feature map is shifted, and a first initial feature map is determined based on the shifted window multi-head self-attention unit.

[0058] In practice, the multiple 3D voxel codes are input into a window multi-head self-attention unit (WMI). The WMI processes these codes to obtain the target extraction feature map. Specifically, for each 3D voxel code, the WMI divides it into non-overlapping fixed windows. Self-attention calculations are performed within each window, and different windows do not interact directly; the focus is primarily on extracting local information.

[0059] Specifically, the process of obtaining the target extraction feature map by the multi-head self-attention unit of the window includes: Step A: Input each 3D voxel code into the window multi-head self-attention unit, and process it through the window multi-head self-attention unit to obtain the target extraction features corresponding to each 3D voxel code; Step B involves concatenating all extracted target features to obtain the target extracted feature map.

[0060] In practice, for each three-dimensional voxel code, the three-dimensional voxel code is input into a window multi-head self-attention unit, and processed by the window multi-head self-attention unit to obtain the target extraction features corresponding to each three-dimensional voxel code.

[0061] In this embodiment, different 3D voxel codes belong to different windows. The multi-head self-attention unit of the window includes multiple heads, with each 3D voxel code corresponding to one head. This enables the multi-head self-attention unit of the window to independently calculate the attention weight for each 3D voxel code using the head corresponding to the 3D voxel code, thereby obtaining the target extraction features.

[0062] All target extracted features are concatenated and linearly transformed to obtain a target extracted feature map, wherein the target extracted feature map contains local information corresponding to each three-dimensional voxel encoding.

[0063] The target extracted feature map is shifted, and a first initial feature map is determined based on the shifted window multi-head self-attention unit. The shifted window multi-head self-attention unit shifts windows, allowing data that were originally adjacent but belonged to different windows to participate in the self-attention calculation of the same window. This method increases information interaction between windows, thereby enhancing the model's ability to perceive global information.

[0064] Specifically, the process by which the shifted window multi-head self-attention unit determines the first initial feature map includes: Step a: Shift the extracted feature map of the target to obtain a shifted feature map of the target; Step b: Divide the target displacement feature map into multiple displacement feature sub-maps; Step c: Input each shift feature sub-image into the shift window multi-head self-attention unit, and process it through the shift window multi-head self-attention unit to obtain the target shift feature corresponding to each shift feature sub-image; Step d: Concatenate all target shift features to obtain the first initial feature map.

[0065] In specific implementation, the target extraction feature map is shifted to obtain a target shifted feature map. Specifically, each target extraction feature contained in the target extraction feature map is shifted. The purpose of the shifting is to enable target extraction features that are originally adjacent but belong to different windows to participate in the self-attention calculation of the same window. The shifting step size is half the window size corresponding to the 3D voxel encoding.

[0066] The target shift feature map is divided into multiple shift feature sub-maps. That is, windows are re-divided on the target shift feature map; these re-divided windows are the same size as the windows corresponding to the 3D voxel encoding, but in different positions.

[0067] Each shift feature sub-image is input into a shift window multi-head self-attention unit, which includes multiple heads, with each shift feature sub-image corresponding to one head. This allows the shift window multi-head self-attention unit to independently calculate attention weights for each shift feature sub-image using the head corresponding to that shift feature sub-image, thereby obtaining the target shift feature.

[0068] All target shift features are concatenated and linearly transformed to obtain a first initial feature map. The first initial feature map not only contains local information corresponding to each three-dimensional voxel encoding, but also increases the information interaction between shift feature sub-maps through shift operations, thereby enhancing the perception of global information.

[0069] In some embodiments, the second feature extraction module includes a first feature merging unit and a first feature extraction unit. Step 1023, where the first initial feature map is input to the second feature extraction module and processed by the second feature extraction module to obtain a second initial feature map, specifically includes: Step 10231: Use the first feature merging unit to downsample the first initial feature map to obtain the first merged feature map; Step 10232: Input the first merged feature map into the first feature extraction unit, process it through the first feature extraction unit, and output the second initial feature map.

[0070] In specific implementation, the first feature merging unit is used to downsample the first initial feature map to reduce the spatial resolution and expand the receptive field, thereby obtaining the first merged feature map.

[0071] The first merged feature map is input to the first feature extraction unit, and processed by the first feature extraction unit. That is, the first feature extraction unit performs efficient local attention calculation on the downsampled first merged feature map to obtain the second initial feature map.

[0072] In this embodiment, the network structures corresponding to the second feature extraction module, the third feature extraction module, and the fourth feature extraction module can be the same or different. When the network structures are the same, the processing procedures are also the same. That is, the way the third feature extraction module processes the second initial feature map and the way the fourth feature extraction module processes the third initial feature map are the same as the way the second feature extraction module processes the first initial feature map, and will not be described again here.

[0073] In this embodiment, to further improve the accuracy of the final obtained multiple sets of initial feature maps, the network structures corresponding to the second feature extraction module, the third feature extraction module, and the fourth feature extraction module can be different, specifically including: The second feature extraction module includes a first feature merging unit and a first feature extraction unit; the third feature extraction module includes six second feature merging units and six second feature extraction units; and the fourth feature extraction module includes six third feature merging units and six third feature extraction units. The processing method of each second feature merging unit and each third feature merging unit is the same as that of the first feature merging unit. The processing method of each second feature extraction unit and each third feature extraction unit is the same as that of the first feature extraction unit, and will not be described again here.

[0074] The above scheme utilizes window-based multi-head self-attention units and shift-window-based multi-head self-attention units to solve the problem of huge computational overhead caused by global self-attention by calculating correlations within local windows and realizing information interaction using a shift-window strategy. This reduces computational complexity to linear, making it possible to process high-resolution, large-scale 3D GPR data. Furthermore, the combined use of a second, third, and fourth feature extraction module enables the model to learn features at different scales simultaneously, achieving excellent detection results for various defects with significant size differences.

[0075] In some embodiments, step 103 involves analyzing and processing the multiple sets of initial feature maps to obtain the initial defect direction, specifically including: Step 1031: The multiple sets of initial feature maps are fused to obtain an initial target feature map; Step 1032: Input the initial target feature map into the first defect detection model, process it through the first defect detection model, and output the initial defect direction.

[0076] In specific implementation, the multiple sets of initial feature maps are fused together, that is, the first initial feature map obtained by the first feature extraction module, the second initial feature map obtained by the second feature extraction module, the third initial feature map obtained by the third feature extraction module, and the fourth initial feature map obtained by the fourth feature extraction module are fused together to obtain the initial target feature map.

[0077] A pre-trained first defect detection model is obtained, the initial target feature map is input into the first defect detection model, and the initial defect direction is output after processing by the first defect detection model.

[0078] In some embodiments, step 104, which determines the target weight corresponding to each detection direction based on the initial defect direction, specifically includes: Step 1041: Obtain the pre-trained weight generation model; Step 1042: Input the initial defect direction into the weight generation model, process it through the weight generation model, and output the target weight corresponding to each detection direction.

[0079] In practice, a pre-trained weight generation model is obtained, the initial defect direction is input into the weight generation model, and after processing by the weight generation model, the target weight corresponding to each detection direction is output, wherein the target weight corresponds to the importance of the detection direction.

[0080] In some embodiments, step 105, based on the target weights and the multiple sets of initial feature maps, uses a pre-trained second defect detection model to determine the target defect data corresponding to the interior of the target road, specifically including: Step 1051: Obtain the target detection direction contained in each initial feature sub-map of the multiple sets of initial feature maps; for each initial feature sub-map, determine the target feature sub-map according to the target detection direction and the target weight, wherein the target feature sub-map is a feature map containing the target weight; Step 1052: Input all target feature sub-images into the pre-trained second defect detection model, process them through the second defect detection model, and output the target defect data corresponding to the inside of the target road.

[0081] In specific implementation, the target detection direction contained in each initial feature sub-map of the multiple initial feature maps is obtained. For each initial feature sub-map, a target feature sub-map is determined according to the target detection direction and the target weight, that is, the target weight is assigned to each target feature sub-map, and then the target feature sub-map is a feature map containing the target weight.

[0082] A pre-trained second defect detection model is obtained, and all target feature sub-images are input into the pre-trained second defect detection model. After processing by the second defect detection model, the target defect data corresponding to the interior of the target road is output.

[0083] In some embodiments, for the feature extraction model, the first defect detection model, the weight generation model, and the second defect detection model in the above embodiments, each of the above models can be trained separately, that is, the initial feature extraction model, the initial first defect detection model, the initial weight generation model, and the initial second defect detection model can be trained separately to obtain the final feature extraction model, the first defect detection model, the weight generation model, and the second defect detection model.

[0084] In some embodiments, the same historical dataset can be used to train the initial feature extraction model, the initial first defect detection model, the initial weight generation model, and the initial second defect detection model simultaneously. In this case, the initial feature extraction model, the initial first defect detection model, the initial weight generation model, and the initial second defect detection model can be regarded as a whole, i.e., the initial defect recognition model. The initial defect recognition model is trained using the historical dataset until the preset training conditions are met, thus obtaining the defect recognition model, i.e., the final feature extraction model, the first defect detection model, the weight generation model, and the second defect detection model.

[0085] In this embodiment, the preset training conditions include at least one of the following: determining that all training data in the historical dataset is input into the defect identification model for training, determining that the loss function of the defect identification model converges to a preset convergence threshold, or determining that the defect identification model is iteratively trained to a preset number of iterations.

[0086] For example, the preset training termination condition is to determine that all training data has been input into the defect recognition model for training: The training data consists of ten sets of data, each set including historical 3D ground-penetrating radar data and historical defect data. The preset training termination condition is that all data in the training data has been input into the defect identification model for training. That is, when all ten sets of data have been input into the defect identification model, there is no longer any data in the training data that has not yet been input into the defect identification model. At this point, the defect identification model training is considered complete, and the defect identification model is obtained.

[0087] Another example is that the preset training termination condition is to determine that the loss function of the defect recognition model converges to a preset convergence threshold: Training data is input into the defect identification model for training, and the training results are output. A loss function is determined based on the training results and the historical defect data. The loss function may include at least one of the following: mean squared error loss function, cross-entropy loss function, logarithmic loss function, exponential loss function, squared loss function, or absolute value loss function, etc. When the loss function converges to a preset convergence threshold, a preset training termination condition is satisfied, and the defect identification model is obtained.

[0088] Another example is that the preset training termination condition is to determine the defect identification model to be trained iteratively to a preset number of iterations.

[0089] The training data is input into the defect identification model for iterative training. The number of iterations is recorded. When the number of iterations equals the preset number of iterations, the preset training termination condition is met, and the defect identification model is obtained.

[0090] Based on the same inventive concept, another embodiment of this disclosure provides a road defect identification method, applied to a defect identification model, the structure of which is as follows: Figure 2 As shown, the input of the defect recognition model is the three-dimensional ground-penetrating radar data corresponding to the interior of the target road, and the output is the target defect data corresponding to the interior of the target road. The target defect data includes precise three-dimensional defect information (location, shape, size, etc.). The defect recognition model includes a hierarchical three-dimensional spatial perception module and an anisotropic dynamic weighting module.

[0091] In this embodiment, the 3D GPR data volume is the three-dimensional ground-penetrating radar data, such as... Figure 3 As shown, the method specifically includes: The preprocessed 3D GPR data volume (e.g., denoising, gain compensation) is used as the direct input to the model. This 3D data volume is spatially divided into multiple continuous and non-overlapping small 3D data blocks, known as voxel sequences. Each voxel sequence contains multiple 3D voxels, and each voxel contains its spatial location encoding information.

[0092] The hierarchical 3D spatial perception module performs local spatial feature encoding, i.e., multi-stage pyramid aggregation. The multiple 3D voxels are input into a pre-trained feature extraction model, which processes the data and outputs multiple sets of initial feature maps.

[0093] Self-attention computation is strictly confined to independent computation within each 3D window, rather than being performed globally across the entire data volume, thus significantly reducing computational complexity from quadratic to linear. To enable information flow between windows, a shifted window strategy is employed between consecutive layers of the network. By periodically translating the window grid, attention computation in adjacent layers can cross the boundaries of the previous layer's window, achieving effective fusion of global information.

[0094] To efficiently achieve global information fusion, a multi-stage hierarchical aggregation approach is adopted. For example... Figure 4 As shown, Figure 4 The model network structure diagram of this embodiment is shown. The network contains multiple subsequent stages, each consisting of a feature merging layer and a three-dimensional window self-attention module. The feature merging layer is the first feature merging unit, the second feature merging unit, and the third feature merging unit in the above embodiment. The three-dimensional window self-attention module is the first feature extraction unit, the second feature extraction unit, and the third feature extraction unit in the above embodiment.

[0095] The feature merging layer is responsible for downsampling the feature map from the previous stage (e.g., merging adjacent 2x2x2 features into one) to reduce spatial resolution and expand the receptive field. The subsequent attention module then performs efficient local attention calculations on the downsampled feature map. Through 2-3 iterations, the network obtains a small feature map containing global semantic information at the top layer, and outputs a preliminary prediction of the defect based on this map, i.e., the initial defect direction in the above embodiment.

[0096] The initial prediction is input into a lightweight weight generation subnetwork, namely the weight generation model in the above embodiment. This subnetwork generates a set of dynamic weight coefficients based on the predicted defect direction (e.g., whether it is horizontally distributed, vertically distributed, or clumped). This set of weights corresponds to the importance of different spatial dimensions or feature channels. For example, if the initial judgment is a horizontal crack, higher weights will be assigned to the vertical profile dimension, which better reflects this feature.

[0097] The generated dynamic weights are applied to the previously obtained initial eigenmap for weighted adjustment. This process enhances the representation of the feature dimensions most relevant to the defect morphology, while suppressing features in secondary or interfering dimensions.

[0098] The feature map, refined by dynamic weights, is then input into the final prediction head. The prediction head outputs precise and final descriptive information about the defect, including its three-dimensional spatial coordinates, geometric dimensions, volume, and category (such as voids or porosity).

[0099] The above approach significantly improves recognition accuracy, preserves complete 3D information, and effectively distinguishes between real defects and noise. It demonstrates stronger robustness to complex defects, with a dynamic weighting strategy enabling the model to "intelligently" handle defects of different shapes and orientations. The innovative 3D window local self-attention mechanism and pyramid structure reduce computational complexity to linear levels, making it possible to process high-resolution, large-scale 3D GPR data. It possesses multi-scale analysis capabilities; the pyramid-like hierarchical structure allows the model to simultaneously learn features at different scales, resulting in excellent detection performance for various defects with significant size differences. The processing flow is highly efficient and simplified, with the end-to-end model integrating the complex "slicing-recognition-reconstruction" process. The output information is richer, directly outputting the 3D morphological parameters of the defects, which is more valuable than fragmented 2D annotation information, providing more intuitive and comprehensive data support for subsequent road maintenance decisions.

[0100] It should be noted that the method of this disclosure embodiment can be executed by a single device, such as a computer or server. The method of this embodiment can also be applied to a distributed scenario, where multiple devices cooperate to complete the task. In such a distributed scenario, one of these devices may execute only one or more steps of the method of this disclosure embodiment, and the multiple devices will interact with each other to complete the method described.

[0101] It should be noted that the above description describes some embodiments of this disclosure. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recorded in the claims can be performed in a different order than that shown in the above embodiments and still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0102] Based on the same inventive concept, another embodiment of this disclosure proposes a road defect identification device, such as... Figure 5 As shown, it specifically includes: The data acquisition module 201 is configured to acquire three-dimensional ground-penetrating radar data corresponding to the interior of the target road, perform voxel segmentation on the three-dimensional ground-penetrating radar data, and obtain multiple three-dimensional voxels corresponding to the three-dimensional ground-penetrating radar data. The feature map determination module 202 is configured to input the plurality of three-dimensional voxels into a pre-trained feature extraction model, and output a plurality of initial feature maps through the feature extraction model. Each initial feature map includes a plurality of initial feature sub-maps, and each initial feature sub-map corresponds to a feature containing the target detection direction. The defect direction determination module 203 is configured to analyze and process the multiple sets of initial feature maps to obtain the initial defect direction. The weight determination module 204 is configured to determine the target weight corresponding to each detection direction based on the initial defect direction; The defect data determination module 205 is configured to determine the target defect data inside the target road based on the target weight and the multiple sets of initial feature maps, using a pre-trained second defect detection model.

[0103] In some embodiments, the feature extraction model includes a first feature extraction module, a second feature extraction module, a third feature extraction module, and a fourth feature extraction module, and the feature map determination module 202 is configured to: Each three-dimensional voxel is encoded to obtain the three-dimensional voxel code corresponding to each three-dimensional voxel; Multiple three-dimensional voxel codes are input into the first feature extraction module, and processed by the first feature extraction module to obtain the first initial feature map; The first initial feature map is input into the second feature extraction module, and after processing by the second feature extraction module, a second initial feature map is obtained. The second initial feature map is input into the third feature extraction module, and processed by the third feature extraction module to obtain the third initial feature map; The third initial feature map is input into the fourth feature extraction module, and processed by the fourth feature extraction module to obtain the fourth initial feature map; The first initial feature map, the second initial feature map, the third initial feature map, and the fourth initial feature map are used as multiple sets of initial feature maps.

[0104] In some embodiments, the first feature extraction module includes a window multi-head self-attention unit and a shifted window multi-head self-attention unit, and the feature map determination module 202 is configured to: The multiple 3D voxel codes are input into a window multi-head self-attention unit, and processed by the window multi-head self-attention unit to obtain the target extraction feature map; The target extracted feature map is shifted, and a first initial feature map is determined based on the shifted window multi-head self-attention unit.

[0105] In some embodiments, the feature map determination module 202 is configured to: Each 3D voxel code is input into a window multi-head self-attention unit, and processed by the window multi-head self-attention unit to obtain the target extraction features corresponding to each 3D voxel code; All extracted target features are concatenated to obtain the target extracted feature map.

[0106] In some embodiments, the feature map determination module 202 is configured to: The target extracted feature map is shifted to obtain a target shifted feature map; The target displacement feature map is divided into multiple displacement feature sub-maps; Each shift feature sub-image is input into the shift window multi-head self-attention unit, and processed by the shift window multi-head self-attention unit to obtain the target shift feature corresponding to each shift feature sub-image; All target shift features are concatenated to obtain the first initial feature map.

[0107] In some embodiments, the second feature extraction module includes a first feature merging unit and a first feature extraction unit, and the feature map determination module 202 is configured to: The first feature map is downsampled using the first feature merging unit to obtain the first merged feature map; The first merged feature map is input to the first feature extraction unit, and after processing by the first feature extraction unit, a second initial feature map is output.

[0108] In some embodiments, the defect direction determination module 203 is specifically configured as follows: The multiple sets of initial feature maps are fused to obtain the initial target feature map; The initial target feature map is input into the first defect detection model, and after processing by the first defect detection model, the initial defect direction is output.

[0109] In some embodiments, the weight determination module 204 is specifically configured as follows: Obtain the pre-trained weight generation model; The initial defect direction is input into the weight generation model, and after processing by the weight generation model, the target weight corresponding to each detection direction is output.

[0110] In some embodiments, the defect data determination module 205 is specifically configured to: Obtain the target detection direction contained in each initial feature sub-map of the multiple sets of initial feature maps. For each initial feature sub-map, determine the target feature sub-map based on the target detection direction and the target weight, wherein the target feature sub-map is a feature map containing the target weight. All target feature sub-images are input into a pre-trained second defect detection model. After processing by the second defect detection model, the target defect data corresponding to the interior of the target road is output.

[0111] For ease of description, the above apparatus is described in terms of its functions, divided into various modules. Of course, in implementing this disclosure, the functions of each module can be implemented in one or more software and / or hardware.

[0112] The apparatus described above is used to implement the corresponding road defect identification method in any of the following embodiments, and has the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0113] Based on the same inventive concept, corresponding to any of the above embodiments, this disclosure also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the road defect identification method described in any of the above embodiments.

[0114] Figure 6 This embodiment illustrates a more specific hardware structure of an electronic device. The device may include a processor 1010, a memory 1020, an input / output interface 1030, a communication interface 1040, and a bus 1050. The processor 1010, memory 1020, input / output interface 1030, and communication interface 1040 are interconnected internally via the bus 1050.

[0115] The processor 1010 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this specification.

[0116] The memory 1020 can be implemented in the form of ROM (Read Only Memory), RAM (Random Access Memory), static storage device, dynamic storage device, etc. The memory 1020 can store the operating system and other applications. When the technical solutions provided in the embodiments of this specification are implemented by software or firmware, the relevant program code is stored in the memory 1020 and is called and executed by the processor 1010.

[0117] The input / output interface 1030 is used to connect input / output modules to realize information input and output. Input / output modules can be configured as components within the device (not shown in the figure) or externally connected to the device to provide corresponding functions. Input devices may include keyboards, mice, touchscreens, microphones, various sensors, etc., while output devices may include displays, speakers, vibrators, indicator lights, etc.

[0118] The communication interface 1040 is used to connect a communication module (not shown in the figure) to enable communication between this device and other devices. The communication module can communicate via wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.).

[0119] Bus 1050 includes a pathway for transmitting information between various components of the device, such as processor 1010, memory 1020, input / output interface 1030, and communication interface 1040.

[0120] It should be noted that although the above-described device only shows the processor 1010, memory 1020, input / output interface 1030, communication interface 1040, and bus 1050, in specific implementations, the device may also include other components necessary for normal operation. Furthermore, those skilled in the art will understand that the above-described device may only include the components necessary for implementing the embodiments of this specification, and not necessarily all the components shown in the figures.

[0121] The electronic devices described above are used to implement the corresponding road defect identification methods in any of the foregoing embodiments, and have the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0122] Based on the same inventive concept, corresponding to the methods of any of the above embodiments, this disclosure also provides a non-transitory computer-readable storage medium storing computer instructions for causing the computer to execute the road defect identification method as described in any of the above embodiments.

[0123] The computer-readable medium of this embodiment includes permanent and non-permanent, removable and non-removable media, and information storage can be implemented by any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transfer medium that can be used to store information accessible by a computing device.

[0124] The computer instructions stored in the storage medium of the above embodiments are used to cause the computer to execute the road defect identification method as described in any of the above embodiments, and have the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0125] It is understood that before using the technical solutions of the various embodiments in this disclosure, users will be informed of the type, scope of use, and usage scenarios of the personal information involved in an appropriate manner, and user authorization will be obtained.

[0126] For example, upon receiving a user's active request, a prompt message is sent to the user to explicitly inform them that the requested operation will require the acquisition and use of the user's personal information. This allows the user to independently choose, based on the prompt message, whether to provide personal information to the software or hardware such as electronic devices, applications, servers, or storage media performing the operations of this disclosed technical solution.

[0127] As an optional but not limited implementation, in response to a user's active request, sending a prompt message to the user can be done via a pop-up window, where the prompt message can be presented in text format. Furthermore, the pop-up window can also include a selection control allowing the user to choose "agree" or "disagree" to provide personal information to the electronic device.

[0128] It is understood that the above notification and user authorization process are merely illustrative and do not constitute a limitation on the implementation of this disclosure. Other methods that comply with relevant laws and regulations may also be applied to the implementation of this disclosure.

[0129] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of this disclosure (including the claims) is limited to these examples; within the framework of this disclosure, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of different aspects of the embodiments of this disclosure as described above, which are not provided in detail for the sake of brevity.

[0130] Additionally, to simplify the description and discussion, and to avoid obscuring the embodiments of this disclosure, the provided drawings may or may not show well-known power / ground connections to integrated circuit (IC) chips and other components. Furthermore, the apparatus may be shown in block diagram form to avoid obscuring the embodiments of this disclosure, and this also takes into account the fact that the details of implementation of these block diagram apparatuses are highly dependent on the platform on which the embodiments of this disclosure will be implemented (i.e., these details should be fully understood by those skilled in the art). While specific details (e.g., circuits) have been set forth to describe exemplary embodiments of this disclosure, it will be apparent to those skilled in the art that the embodiments of this disclosure can be implemented without these specific details or with variations thereof. Therefore, these descriptions should be considered illustrative rather than restrictive.

[0131] Although this disclosure has been described in conjunction with specific embodiments thereof, many substitutions, modifications, and variations of these embodiments will be apparent to those skilled in the art from the foregoing description. For example, other memory architectures (e.g., dynamic RAM (DRAM)) may be used with the embodiments discussed.

[0132] This disclosure is intended to cover all such substitutions, modifications, and variations that fall within the broad scope of the appended claims. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A method for identifying road defects, characterized in that, include: Acquire three-dimensional ground-penetrating radar data corresponding to the interior of the target road, and perform voxel segmentation on the three-dimensional ground-penetrating radar data to obtain multiple three-dimensional voxels corresponding to the three-dimensional ground-penetrating radar data. The multiple three-dimensional voxels are input into a pre-trained feature extraction model. After processing by the feature extraction model, multiple sets of initial feature maps are output. Each set of initial feature maps includes multiple initial feature sub-maps, and each initial feature sub-map corresponds to a feature containing the target detection direction. The initial defect direction is obtained by analyzing and processing the multiple sets of initial feature maps; Based on the initial defect direction, determine the target weight corresponding to each detection direction; Based on the target weights and the multiple sets of initial feature maps, the target defect data corresponding to the interior of the target road is determined using the pre-trained second defect detection model.

2. The method according to claim 1, characterized in that, The feature extraction model includes a first feature extraction module, a second feature extraction module, a third feature extraction module, and a fourth feature extraction module. The process involves inputting the multiple three-dimensional voxels into a pre-trained feature extraction model, which then processes the data to output multiple sets of initial feature maps, including: Each three-dimensional voxel is encoded to obtain the three-dimensional voxel code corresponding to each three-dimensional voxel; Multiple three-dimensional voxel codes are input into the first feature extraction module, and processed by the first feature extraction module to obtain the first initial feature map; The first initial feature map is input into the second feature extraction module, and after processing by the second feature extraction module, a second initial feature map is obtained. The second initial feature map is input into the third feature extraction module, and processed by the third feature extraction module to obtain the third initial feature map; The third initial feature map is input into the fourth feature extraction module, and processed by the fourth feature extraction module to obtain the fourth initial feature map; The first initial feature map, the second initial feature map, the third initial feature map, and the fourth initial feature map are used as multiple sets of initial feature maps.

3. The method according to claim 2, characterized in that, The first feature extraction module includes a window multi-head self-attention unit and a shifted window multi-head self-attention unit. The step of inputting the multiple three-dimensional voxel codes into the first feature extraction module, and processing them through the first feature extraction module to obtain a first initial feature map, includes: The multiple 3D voxel codes are input into a window multi-head self-attention unit, and processed by the window multi-head self-attention unit to obtain the target extraction feature map; The target extracted feature map is shifted, and a first initial feature map is determined based on the shifted window multi-head self-attention unit.

4. The method according to claim 3, characterized in that, The step of inputting the multiple 3D voxel codes into a window multi-head self-attention unit, and processing them through the window multi-head self-attention unit to obtain the target extraction feature map includes: Each 3D voxel code is input into a window multi-head self-attention unit, and processed by the window multi-head self-attention unit to obtain the target extraction features corresponding to each 3D voxel code; All extracted target features are concatenated to obtain the target extracted feature map.

5. The method according to claim 3, characterized in that, The step of shifting the extracted feature map of the target and determining a first initial feature map based on the multi-head self-attention unit of the shift window includes: The target extracted feature map is shifted to obtain a target shifted feature map; The target displacement feature map is divided into multiple displacement feature sub-maps; Each shift feature sub-image is input into the shift window multi-head self-attention unit, and processed by the shift window multi-head self-attention unit to obtain the target shift feature corresponding to each shift feature sub-image; All target shift features are concatenated to obtain the first initial feature map.

6. The method according to claim 2, characterized in that, The second feature extraction module includes a first feature merging unit and a first feature extraction unit. The step of inputting the first initial feature map into the second feature extraction module and processing it to obtain the second initial feature map includes: The first feature map is downsampled using the first feature merging unit to obtain the first merged feature map; The first merged feature map is input to the first feature extraction unit, and after processing by the first feature extraction unit, a second initial feature map is output.

7. The method according to claim 1, characterized in that, The process of analyzing and processing the multiple sets of initial feature maps to obtain the initial defect direction includes: The multiple sets of initial feature maps are fused to obtain the initial target feature map; The initial target feature map is input into the first defect detection model, and after processing by the first defect detection model, the initial defect direction is output.

8. The method according to claim 1, characterized in that, The step of determining the target weight corresponding to each detection direction based on the initial defect direction includes: Obtain the pre-trained weight generation model; The initial defect direction is input into the weight generation model, and after processing by the weight generation model, the target weight corresponding to each detection direction is output.

9. The method according to claim 1, characterized in that, The step of determining the target defect data corresponding to the interior of the target road using a pre-trained second defect detection model based on the target weights and the multiple sets of initial feature maps includes: Obtain the target detection direction contained in each initial feature sub-map of the multiple sets of initial feature maps. For each initial feature sub-map, determine the target feature sub-map based on the target detection direction and the target weight, wherein the target feature sub-map is a feature map containing the target weight. All target feature sub-images are input into a pre-trained second defect detection model. After processing by the second defect detection model, the target defect data corresponding to the interior of the target road is output.

10. An electronic device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the method as described in any one of claims 1 to 9.

Images