Road hidden disease detection method based on artificial intelligence and double-view guidance

Abstract

This invention discloses a method for detecting hidden road defects based on artificial intelligence and dual-view guidance, comprising the following steps: Step 1: acquiring radar data; Step 2: preprocessing dual-view images; Step 3: constructing jointly labeled samples; Step 4: building a dual-stream detection network; Step 5: performing cross-view feature fusion; Step 6: completing defect target detection and output. The method of this invention can fully utilize the complementary structural information of the longitudinal and transverse views in ground-penetrating radar images, improving the completeness of defect target representation and recognition accuracy. It effectively solves the problems of incomplete features, cross-view inconsistencies, and high false negative rates in traditional single-view detection, and can be widely applied to intelligent defect recognition tasks in urban roads, airport runways, and other underground structure scenarios.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence and underground structural defect detection, and in particular to a method for detecting hidden road defects based on artificial intelligence and dual-view guidance. Background Technology

[0002] With the increasing intensity of urban underground space development, the loads on underground structures such as roads, airport runways, and rail transit are continuously increasing. Under the influence of long-term loads, groundwater changes, freeze-thaw cycles, and construction disturbances, underground structural layers are prone to various hidden defects such as cracks, cavities, water damage, and loosening. Figure 1 As shown in the image, these types of defects have complex causes, are hidden in their distribution, and evolve rapidly. If not detected in time, they will seriously threaten the safety of road structures and the stability of public transportation operations. However, accurately identifying these hidden defects can effectively provide early warning of road structural risks, reduce the accident rate, and provide crucial data support for infrastructure maintenance and management.

[0003] First, accurate damage detection results help management to understand the true health status of the structure, thereby scientifically formulating maintenance plans, slowing down the aging process of the road, avoiding unnecessary excavation, and reducing maintenance costs.

[0004] Secondly, by accurately identifying the location and type of underground defects, the operational safety level of urban roads can be improved, especially in transportation hubs, old areas and areas prone to subsidence, thereby enhancing the foresight and systematic nature of accident prevention.

[0005] Furthermore, with the development of smart cities and digital infrastructure, the high-precision detection results of underground defects also provide a high-value data source for emerging applications such as automated inspection and facility digital twins, helping to upgrade facility management to a smarter level.

[0006] Therefore, accurate and efficient detection technology for hidden underground diseases has become an important part of the urban infrastructure safety assurance system, with broad application prospects and practical needs.

[0007] Current methods for detecting hidden underground diseases mostly rely on deep learning-based image target detection frameworks, and mainstream research can be divided into the following categories:

[0008] The first category is based on single-view detection models. These methods typically use two-dimensional radar images, such as B-Scan or C-Scan, as input and rely on models such as YOLO and Faster R-CNN for disease identification.

[0009] The main idea is to extract features from a vertical or horizontal single-view image and then use a convolutional neural network to locate and classify the target. This type of method is simple to implement, has high inference efficiency, and is suitable for single-view scenarios with sufficient labeled data.

[0010] The second category is image stitching-based multi-view detection methods;

[0011] The main idea is to stitch together the horizontal and vertical images and use the result as input to the existing object detection model in order to improve the multi-object detection capability. This type of method reuses the existing detection architecture with minimal changes, and the idea is intuitive, but it has limitations in terms of view structure alignment and semantic consistency modeling.

[0012] The third category is shallow feature fusion methods. These methods use a two-stream neural network structure, inputting two types of views into different encoders, and fusing them in the middle or at the end of the network through splicing, weighting, or attention. For example, in radar runway defect detection, structures such as MV-GPRNet use C-Scan and B-Scan feature maps stacked as inputs, and complete the fusion of multi-view information through a cross-channel fusion module.

[0013] In their paper "MV-GPRNet: multi-view subsurface defect detection network for airport runway inspection based on GPR" (Remote Sensing, 2022), Li et al. proposed that by fusing three-dimensional radar C-Scan images and two-dimensional B-Scan images to construct a multi-view detection network, cross-view feature fusion was achieved in airport runway defect detection.

[0014] In addition, the underground target recognition method proposed by Kim et al. (IEEE JSTARS, 2019) uses a three-way slicing strategy to process 3D GPR images and trains single-stream detection models for the three types of views respectively, which improves the coverage of disease recognition to a certain extent.

[0015] However, the above methods have the following drawbacks in detecting underground diseases:

[0016] (1) The single-view method has limited information and incomplete features, and cannot fully reflect the spatial morphology of the disease. In particular, it is easy to miss or misjudge multi-target superposition and weak texture areas.

[0017] (2) Direct splicing or feature stacking of multi-view images failed to uncover the deep structural correspondence between horizontal and vertical views, resulting in a lot of redundant information, inconsistent targets across views, and unsatisfactory fusion effect;

[0018] (3) Existing detection models lack cross-view semantic guidance mechanisms, which cannot achieve complementary guidance of horizontal and vertical structural information in target discrimination, making it difficult to improve the model's ability to distinguish complex diseases, especially in the case of interference background and adjacent disease target scenarios.

[0019] (4) Existing methods cannot balance accuracy and speed. The fusion models often have problems such as large number of parameters and slow inference speed, which are not conducive to their promotion and application in actual road inspection and engineering deployment.

[0020] Therefore, in order to solve the above-mentioned technical problems, this invention constructs a method for detecting underground hidden diseases based on dual-view image-guided fusion, which improves the accuracy of disease identification and the completeness of multi-target detection. Summary of the Invention

[0021] In view of this, the present invention provides a method for detecting hidden road defects based on artificial intelligence and dual-view guidance.

[0022] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0023] The method for detecting hidden road defects based on artificial intelligence and dual-view guidance includes the following steps:

[0024] Step 1: Conduct radar data acquisition

[0025] (1a) Conduct on-site scanning of the road section to obtain continuous radar image data of the road structure layer;

[0026] (1b) After data cleaning, data correction and adaptive filtering to filter out the outliers, the data is visualized and the longitudinal section image B-Scan and the cross section image C-Scan are extracted respectively.

[0027] Step 2: Perform dual-view image preprocessing

[0028] (2a) Extract the image frames corresponding to the key disease depth layers from the B-Scan and C-Scan images after spatial and temporal registration;

[0029] (2b) The original C-Scan image is resized to the same resolution as the B-Scan image to form a radar image sample pair with cross-view alignment;

[0030] (2c) The upper part of the processed image sample is a cross-sectional view and the lower part is a longitudinal section view, which constitutes a complementary information source;

[0031] Step 3: Construct jointly labeled samples

[0032] (3a) By analyzing a continuous B-Scan image sequence, the key depth segment where the disease occurs is located, and the corresponding C-Scan image is extracted based on the depth to complete the spatial matching of the horizontal and vertical views;

[0033] (3b) The above horizontal and vertical images are stitched together or paired to form joint labeled samples; auxiliary algorithms are used to complete the pre-labeling of high-confidence regions;

[0034] (3c) For areas with weak differences between disease types or many interfering factors, three rounds of manual verification and boundary calibration shall be carried out;

[0035] (3d) For image samples with coexisting diseases and overlapping boundaries, overlapping label boxes are allowed, and a corresponding discrimination mechanism is proposed during the network training phase;

[0036] Step 4: Build a dual-stream detection network

[0037] Construct a mainstream and secondary network;

[0038] Step 5: Perform cross-view feature fusion

[0039] We propose the Dual-view Cross-Information Fusion Transformer module to perform a bidirectional cross-view attention mechanism between the main and auxiliary flows;

[0040] Step 6: Complete the detection and output of disease targets.

[0041] (6a) The features after the main and auxiliary streams are fused are weighted and integrated to generate the final joint feature map used for detection;

[0042] (6b) Output structured detection results including location bounding box, disease type and confidence level;

[0043] (6c) The final test results can be used for structured visualization of road defects, alarm prompts and subsequent maintenance decisions.

[0044] Preferably, in step 3(b), an auxiliary algorithm is used to pre-label high-confidence areas, such as clear targets like transverse and longitudinal cracks and water damage.

[0045] Preferably, in step 4, the mainstream network uses the Swin Transformer as the feature extraction backbone network, and takes the original size B-Scan image as input to model the longitudinal depth variation and global structural features of the disease.

[0046] The auxiliary flow network uses ResNet50 as the feature extraction network and takes the resampled C-Scan image as input to model fine-grained information such as the lateral distribution of the disease and boundary texture.

[0047] The characteristics of the main stream and auxiliary stream outputs are represented as follows: and .

[0048] Preferably, step 5 specifically includes the following steps:

[0049] (5a) Query-key-value vector construction of Q / K / V

[0050] Mainstream features F Li and auxiliary flow characteristics F Hi Flatten it into a sequence, and generate a query vector Q, a key vector K, and a value vector V for attention mechanism computation through linear mapping;

[0051] (5b) Cross-view attention interaction guided by mainstream principles

[0052] During cross-view interaction, the mainstream and auxiliary features not only achieve information fusion in a single attention calculation, but also repeatedly exchange deep feature information through a bidirectional guidance mechanism. The vertical global structural features extracted by the mainstream network Swin Transformer provide the auxiliary flow with a reference for the depth of the disease and the global spatial relationship during the interaction. The horizontal high-frequency detail features extracted by the auxiliary network ResNet50 in turn enhance the mainstream's ability to perceive local information such as boundaries and textures. Through multi-head cross-attention calculation, the two feature flows form a synergistic effect at the spatial and semantic levels, thereby improving the detection accuracy of hidden diseases at multiple types and scales.

[0053] Guided by the mainstream feature B-Scan path, the generated query vector The bond generated with the auxiliary flow Value vector Perform cross-attention computation; achieve feature fusion by calculating similarity weights;

[0054] (5c) Cross-view attention interaction guided by auxiliary flow

[0055] The auxiliary features also generate query vectors themselves and perform attention calculations with the key and value vectors of the mainstream features to complete the response to the mainstream information.

[0056] Preferably, the expression for step (5a) is:

[0057]

[0058]

[0059]

[0060] Where I represents the input feature sequence, W Q W K W V These are the projection matrices for the query, key, and value, respectively;

[0061] The expression for step (5b) is:

[0062]

[0063]

[0064] Here, Softmax represents the normalization of the dot product between the query and the key. is the scaling factor for the key vector dimension, and FEN is the feedforward neural network module. This indicates the mainstream features after cross-view fusion;

[0065] The expression for step (5c) is:

[0066]

[0067]

[0068] Preferably, step 6 specifically includes:

[0069] The features obtained after the fusion of the main and auxiliary streams are weighted and integrated to generate a joint feature map for final detection.

[0070]

[0071] in, For learnable gating coefficients, Features after fusion;

[0072] The output includes structured detection results containing location bounding boxes, disease types, and confidence levels, in the following format:

[0073]

[0074] in, This indicates the position and size of the i-th detection box. Indicates the type of disease. Rate its confidence level;

[0075] The final test results can be used for structured visualization of road defects, alarm prompts, and subsequent maintenance decisions.

[0076] The present invention achieves the following technical effects compared to the prior art:

[0077] (1) The present invention uses a main and auxiliary flow structure to process the longitudinal view B-Scan and the transverse view C-Scan respectively, which effectively solves the problem of large differences in the appearance of diseases in different views and incomplete features in a single view;

[0078] (2) This invention proposes a DCT module to realize cross-view bidirectional attention modeling, which enables the model to establish dynamic connections between global structure and local details, thereby enhancing the feature recognition capability in multi-target scenarios;

[0079] (3) This invention proposes a fusion strategy based on a gating mechanism, which improves the network’s response to key features while controlling the contribution of cross-view information, and avoids the information interference and redundancy problems caused by the traditional element-by-element fusion method. The overall architecture maintains better inference speed and model deployment flexibility while significantly improving detection accuracy, and is feasible for practical engineering implementation.

[0080] (4) This invention effectively overcomes the technical bottlenecks of traditional underground disease detection methods in terms of feature extraction, view fusion, detection accuracy and system efficiency, and provides a more stable and universal technical path for road structure layer safety detection, with broad application prospects and industrialization value. Attached Figure Description

[0081] Figure 1 This is a typical challenge scenario diagram for detecting hidden underground diseases;

[0082] Figure 2 Schematic diagram of radar image visualization and dual-view joint annotation;

[0083] Among them, 2(a) is the radar image data of the road structure layer; 2(b) is the longitudinal depth variation and lateral spatial distribution of the disease; 2(c)-2(e) are the processed image samples;

[0084] Figure 3 Schematic diagram of the dual-view joint annotation method;

[0085] Among them, 3(a) shows the difference in the appearance of different diseases in the longitudinal view and the transverse view; 3(b) shows the situation where the loose layer and the water-rich layer are difficult to distinguish in the longitudinal view; 3(c) shows the non-disease interference information in the transverse view; and 3(d) is a schematic diagram of multi-view joint annotation that integrates longitudinal and transverse image features.

[0086] Figure 4 A schematic diagram of the combined view-guided dual-flow detector structure of the present invention. Detailed Implementation

[0087] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0088] This invention discloses a method for detecting hidden road defects based on artificial intelligence and dual-view guidance, comprising the following steps:

[0089] Step 1: Conduct radar data acquisition

[0090] (1a) such as Figure 2 As shown in (a), the present invention utilizes a road defect detection vehicle equipped with a three-dimensional ground-penetrating radar (3D-GPR) device to conduct on-site scanning of a total of 30 kilometers of typical road sections of Pingxing Expressway and Zhongjiang Expressway in Guangdong Province, and obtains continuous road structure layer radar image data.

[0091] (1b) The radar equipment can output echo data from multiple directions, and after synchronous processing, the electromagnetic response spectrum of the structural layer in different dimensions can be obtained. After data cleaning, data correction, and adaptive filtering to filter outliers, visualization is performed, and the longitudinal section image B-Scan and the cross-sectional image C-Scan are extracted respectively, such as... Figure 2 As shown in (b), the two types of images reflect the longitudinal depth changes and lateral spatial distribution of the disease, respectively.

[0092] Step 2: Perform dual-view image preprocessing

[0093] (2a) Extract the image frames corresponding to the key disease depth layers from the B-Scan and C-Scan images after spatial and temporal registration;

[0094] (2b) To ensure the comparability of the horizontal and vertical views in terms of spatial scale, the original C-Scan image is resized to the same resolution (height × width) as the B-Scan image to form a radar image sample pair with cross-view alignment;

[0095] (2c) Processed image samples as follows Figure 2 As shown in (c, d, e), the upper part is a cross-sectional view and the lower part is a longitudinal section view, which constitutes a complementary information source;

[0096] Step 3: Construct jointly labeled samples

[0097] To improve the accuracy and consistency of disease detection annotation, this invention designs a dual-view joint annotation method guided by depth position consistency, as follows: Figure 3 As shown, the main steps include:

[0098] (3a) By analyzing a continuous B-Scan image sequence, the key depth segment where the disease occurs is located, and the corresponding C-Scan image is extracted based on the depth to complete the spatial matching of the horizontal and vertical views;

[0099] (3b) The above horizontal and vertical images are stitched or paired to form a joint annotation sample. In the preliminary annotation stage, auxiliary algorithms are used to complete the pre-annotation of high-confidence areas, such as clear targets like horizontal and vertical cracks and water damage;

[0100] (3c) For areas with weak differences between disease types or many interfering factors, such as loose and water-rich areas, three rounds of manual verification and boundary calibration are carried out to improve label stability and labeling consistency.

[0101] (3d) For image samples with coexisting diseases and overlapping boundaries, overlapping label boxes are allowed, and a corresponding discrimination mechanism is proposed during the network training stage to improve the generalization ability of the model;

[0102] Step 4: Build a dual-stream detection network

[0103] A dual-stream detection network with a joint view guidance mechanism is constructed, consisting of a main stream network and a secondary stream network, for jointly modeling structural information and texture details in B-Scan and C-Scan images;

[0104] The mainstream network uses the Swing Transformer as the backbone network for feature extraction, and takes the original size B-Scan image as input to model the vertical depth variation and global structural features of the disease.

[0105] The auxiliary flow path uses ResNet50 as the feature extraction network, taking resampled C-Scan images as input to model fine-grained information such as the lateral distribution of disease and boundary texture. The feature representations output by the main flow and auxiliary flow are as follows: and .

[0106] The overall network structure is as follows Figure 4 As shown.

[0107] Step 5: Perform cross-view feature fusion

[0108] To enhance meaning guidance and cross-interaction between horizontal and vertical views, this invention proposes the Dual-view Cross-Information Fusion Transformer (DCT) module, which performs a bidirectional cross-view attention mechanism between the main and auxiliary streams.

[0109] (5a) Query-key-value vector construction (Q / K / V)

[0110] First, the mainstream feature F Li and auxiliary flow characteristics F Hi Flattened into a sequence, and generated using linear mapping, a query vector (Q), a key vector (K), and a value vector (V) are used for attention mechanism computation:

[0111]

[0112]

[0113]

[0114] Where I represents the input feature sequence, W Q W K W V These are the projection matrices for the query, key, and value, respectively, which can be learned through training.

[0115] The purpose of this step is to map image features to a unified attention space, which facilitates subsequent similarity matching between views;

[0116] (5b) Cross-view attention interaction guided by mainstream principles

[0117] During cross-view interaction, the mainstream and auxiliary features not only achieve information fusion in a single attention calculation but also repeatedly exchange deep feature information through a bidirectional guidance mechanism. The vertical global structural features extracted by the mainstream network (Swin Transformer) provide the auxiliary flow with a reference for disease depth and global spatial relationships during the interaction; the horizontal high-frequency detail features extracted by the auxiliary network (ResNet50) in turn enhance the mainstream's ability to perceive local information such as boundaries and textures. Through multi-head cross-attention calculation, the two feature flows synergize at the spatial and semantic levels, thereby improving the detection accuracy of hidden diseases across multiple types and scales.

[0118] Guided by the mainstream features (B-Scan path), the generated query vector is... With the key and value vectors generated by the auxiliary flow ( , Perform cross-attention computation. This operation achieves feature fusion by calculating similarity weights:

[0119]

[0120]

[0121] Here, Softmax represents the normalization of the dot product between the query and the key. is a scaling factor for the key vector dimension, used to avoid gradient vanishing; FEN (Feed-Forward Network) is a feedforward neural network module used to enhance feature representation capabilities.

[0122] Final output This indicates the mainstream features after cross-view fusion;

[0123] (5c) Cross-view attention interaction guided by auxiliary flow (symmetric interaction)

[0124] Similarly, auxiliary features also generate query vectors themselves and perform attention calculations with the key-value vectors of the mainstream features to complete the response to mainstream information:

[0125]

[0126]

[0127] This step enables the model to achieve true two-way information interaction, allowing the features of the horizontal and vertical views to guide and complement each other in the semantic space, thereby improving the feature fusion effect and the disease discrimination ability.

[0128] Step 6: Complete the detection and output of disease targets.

[0129] The features from the fusion of the main and auxiliary streams are weighted and integrated to generate a joint feature map for final detection:

[0130]

[0131] in, These are learnable gating coefficients, representing the importance weights of the dominant features. The fused features The data is input into a target detection head (e.g., a YOLO-based prediction module), and the output is a structured detection result containing the location box, disease type, and confidence level, in the following format:

[0132]

[0133] in, This indicates the position and size of the i-th detection box. Indicates the type of disease. Rate its confidence level;

[0134] The final test results can be used for structured visualization of road defects, alarm prompts, and subsequent maintenance decisions.

[0135] Example 1:

[0136] 1. Simulation conditions

[0137] This invention was conducted using the deep learning framework PyTorch 1.12 on a CPU consisting of an Intel(R) Core(TM) i9-14900K CPU, an NVIDIA GeForce RTX 3090 GPU, and a Windows 11 operating system.

[0138] This experiment uses the 3D-GPR3K real road defect monitoring dataset, which contains 3000 pairs of paired C-Scan and B-Scan image samples, covering five typical underground hidden defects such as transverse cracks, longitudinal cracks, cavities, water damage, and loosening. Each pair of samples has a high-quality joint annotation box.

[0139] The training configuration is as follows:

[0140] Input image size: 1280×512;

[0141] Training / validation / test set split ratio: 7:2:1

[0142] Optimizer: AdamW, initial learning rate 0.001;

[0143] Batch size: 2, Number of training rounds: 150

[0144] Learning rate scheduling: warm-up for the first 2000 steps, and cosine decay for the last 5 rounds.

[0145] 2. Comparative Experiment Setup

[0146] To verify the detection performance of the method of the present invention, the following comparative methods were designed:

[0147] Single-view detectors include models such as Faster R-CNN, YOLOv8, and Deformable DETR, which use either a vertical view (B-Scan) or a horizontal view (C-Scan) as input.

[0148] Joint view detectors, such as YOLOv8 (joint-view) and Deformable DETR (joint-view), feed two types of views directly into a single-stream network by stacking them on the input channel.

[0149] Dual-view input models: Methods such as GAFF, FT, and XMCMT use parallel structures to process two views;

[0150] 3. Evaluation Indicators Explanation

[0151] The following metrics were used to evaluate the detection performance in the experiment:

[0152] mAP@[.5:.95]: AP is calculated every 0.05 within the IoU range of 0.5 to 0.95, and the average value is taken to measure the overall detection accuracy;

[0153] AP@50 / AP@75: Average accuracy at IoU=0.50 and IoU=0.75 respectively, often used for evaluating actual detection results;

[0154] FPS: Frames per second, reflecting the efficiency of model operation;

[0155] Params: Total number of model parameters, reflecting deployment cost.

[0156] 4. Experimental Comparison Results

[0157] Table 1 Comparative Experimental Study of Disease Detection Tasks .

[0158] As shown in Table 1, the input view and structural design have a significant impact on detection performance, exhibiting the following overall trend:

[0159] First, stitched views are superior to single views. Compared to using B-Scan or C-Scan images alone, stitching horizontal and vertical images together can fuse multi-angle information to a certain extent, improving the model's ability to detect multi-target diseases. YOLOv8's AP@75 improved from 29.0 (B-Scan) to 33.8 (joint-view).

[0160] Secondly, the dual-stream structure outperforms the stitched view. Methods such as GAFF and FT extract horizontal and vertical features independently through parallel branches, avoiding information aliasing and further improving detection accuracy. FT improves mAP to 40.2, higher than most stitching strategies.

[0161] Finally, the method of this invention outperforms other two-stream models. By implementing semantic guidance and cross-attention fusion between views through the proposed DCT module, and optimizing feature weight allocation using a gating mechanism, the ability to identify complex diseases is effectively improved. Ultimately, it achieves 88.5% AP@50 and 54.3% mAP on 3D-GPR3K, outperforming existing methods in all evaluation metrics.

[0162] 5. Ablation Experiment Analysis

[0163] To verify the independent contribution of each key structural component in this invention to performance improvement, a module ablation experiment was designed.

[0164] The results are shown in Table 2.

[0165] Table 2 Ablation Experiment Study of Detection Module

[0166] .

[0167] The results show that:

[0168] After adding the DJV module, mAP increased from 48.0% to 52.7%, an increase of 4.7 percentage points, indicating that the dual-view input structure effectively enhanced feature integrity and information coverage, especially with the most significant improvement at AP@50, indicating that multi-target diseases are more easily perceived;

[0169] After further integrating the DCT module proposed in this invention, mAP was further improved to 62.9%, and AP@75 was improved by 6.5 percentage points, verifying the advantages of the cross-view semantic guidance mechanism in fine boundary recognition and complex structure parsing;

[0170] The overall accuracy is highest when combined, indicating that each module of the present invention has good independent gain and cooperative gain.

[0171] 6. Horizontal Method Comparison Experiment

[0172] To further verify the leading performance of the method of this invention in real engineering scenarios, we selected a variety of representative optimized detection structures in the current community and conducted a systematic comparison. The results are shown in Table 3.

[0173] Table 3 Comparison of detection performance with existing methods

[0174] Compared to methods such as YOLOv3-FDL and Improve YOLOv8, this invention improves mAP by nearly 16 percentage points, demonstrating stronger characterization capabilities for disease structures in multi-objective and complex contexts. Compared to the top-performing Swin-GPR model, this invention shows improvements of approximately 10% in AP@50, AP@75, and mAP, particularly in high IoU (AP@75) scenarios, indicating its advantage in fine-grained disease boundary identification. Despite having a larger number of parameters, this method maintains a runtime efficiency of 15.6 FPS, meeting the real-time requirements of practical deployment scenarios.

[0175] Compared with existing detection methods based on single views or simple stitching and fusion, this invention makes full use of the structural complementarity and semantic consistency between the longitudinal view (B-Scan) and the transverse view (C-Scan) in three-dimensional ground penetrating radar images, thereby achieving higher accuracy in disease identification;

[0176] Furthermore, this invention employs a dual-stream detection structure, inputting the original-size B-Scan image into the Transformer main stream to model global structural features, and inputting the resampled C-Scan image into the CNN auxiliary stream to extract boundary texture features. Combined with the designed cross-view cross-guided module (DCT module), information interaction and guidance enhancement between horizontal and vertical features are realized, significantly improving the detection effect of multiple types of diseases.

[0177] Meanwhile, this invention optimizes the fusion efficiency of the network structure and the model deployment overhead, while taking into account both detection accuracy and inference speed, making it suitable for efficient disease identification tasks in real-world scenarios such as underground road structures.

[0178] The method of this invention effectively overcomes the limitations of traditional single-view models, such as incomplete features and cross-view limitations. Figure 1 Technical bottlenecks such as weak consistency modeling capabilities and high fusion redundancy have been overcome, improving the monitoring system's ability to perceive complex structural scenarios and fine-grained diseases and its robustness.

[0179] The above description is merely a preferred embodiment of the present invention and does not constitute any limitation on the technical scope of the present invention. Therefore, any minor modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention shall still fall within the scope of the technical solution of the present invention.

Claims

1. A method for detecting hidden road defects based on artificial intelligence and dual-view guidance, characterized in that, Includes the following steps: Step 1: Conduct radar data acquisition (1a) Conduct on-site scanning of the road section to obtain continuous radar image data of the road structure layer; (1b) After data cleaning, data correction and adaptive filtering to filter out the outliers, the data is visualized and the longitudinal section image B-Scan and the cross section image C-Scan are extracted respectively. Step 2: Perform dual-view image preprocessing (2a) Extract the image frames corresponding to the key disease depth layers from the B-Scan and C-Scan images after spatial and temporal registration; (2b) The original C-Scan image is resized to the same resolution as the B-Scan image to form a radar image sample pair with cross-view alignment; (2c) The upper part of the processed image sample is a cross-sectional view and the lower part is a longitudinal section view, which constitutes a complementary information source; Step 3: Construct jointly labeled samples (3a) By analyzing a continuous B-Scan image sequence, the key depth segment where the disease occurs is located, and the corresponding C-Scan image is extracted based on the depth to complete the spatial matching of the horizontal and vertical views; (3b) The above horizontal and vertical views are spliced ​​or paired to form a joint annotation sample; auxiliary algorithms are used to complete the pre-annotation of high-confidence regions; (3c) For areas with weak differences between disease types or many interfering factors, three rounds of manual verification and boundary calibration shall be carried out; (3d) For image samples with coexisting diseases and overlapping boundaries, overlapping label boxes are allowed, and a corresponding discrimination mechanism is proposed during the network training phase; Step 4: Build a dual-stream detection network Construct a mainstream and secondary network; Step 5: Perform cross-view feature fusion We propose the Dual-view Cross-Information Fusion Transformer module to perform a bidirectional cross-view attention mechanism between the main and auxiliary flows; Step 5 specifically includes the following steps: (5a) Query-key-value vector construction of Q / K / V The main stream feature F Li and the auxiliary stream feature F Hi are flattened into sequences, and the query vector Q, the key vector K and the value vector V for attention mechanism calculation are generated through a linear mapping. (5b) Cross-view attention interaction guided by mainstream principles During cross-view interaction, the mainstream and auxiliary features not only achieve information fusion in a single attention calculation, but also repeatedly exchange deep feature information through a bidirectional guidance mechanism. The vertical global structural features extracted by the mainstream network Swin Transformer provide the auxiliary flow with a reference for the depth of the disease and the global spatial relationship during the interaction. The horizontal high-frequency detail features extracted by the auxiliary network ResNet50 in turn enhance the mainstream's ability to perceive boundary and texture local information. Through multi-head cross-attention calculation, the two feature flows form a synergistic effect at the spatial and semantic levels, thereby improving the detection accuracy of hidden diseases at multiple types and scales. Guided by the mainstream feature B-Scan path, the generated query vector The bond generated with the auxiliary flow Value vector Perform cross-attention computation; achieve feature fusion by calculating similarity weights; (5c) Cross-view attention interaction guided by auxiliary flow The auxiliary features also generate query vectors themselves and perform attention calculations with the key and value vectors of the mainstream features to complete the response to the mainstream information; Step 6: Complete the detection and output of disease targets. (6a) The features after the main and auxiliary streams are fused are weighted and integrated to generate the final joint feature map used for detection; (6b) Output structured detection results including location bounding box, disease type and confidence level; (6c) The final test results can be used for structured visualization of road defects, alarm prompts and subsequent maintenance decisions.

2. The method for detecting hidden road defects based on artificial intelligence and dual-view guidance according to claim 1, characterized in that, In step 3(b), an auxiliary algorithm is used to complete the pre-labeling of high-confidence regions.

3. The method for detecting hidden road defects based on artificial intelligence and dual-view guidance according to claim 1, characterized in that, In step 4, the mainstream network uses the Swing Transformer as the backbone network for feature extraction, and takes the original size B-Scan image as input to model the longitudinal depth variation and global structural features of the disease. The auxiliary flow network uses ResNet50 as the feature extraction network and takes the resampled C-Scan image as input to model the lateral distribution of the disease and fine-grained information of the boundary texture. The characteristics of the main stream and auxiliary stream outputs are represented as follows: and .

4. The method for detecting hidden road defects based on artificial intelligence and dual-view guidance according to claim 1, characterized in that, The expression for step (5a) is: Where I represents the input feature sequence, W Q W K W V These are the projection matrices for the query, key, and value, respectively; The expression for step (5b) is: Here, Softmax represents the normalization of the dot product between the query and the key. is the scaling factor for the key vector dimension, and FEN is the feedforward neural network module. This indicates the mainstream features after cross-view fusion; The expression for step (5c) is: 。 5. The method for detecting hidden road defects based on artificial intelligence and dual-view guidance according to claim 1, characterized in that, Step 6 specifically includes: The features obtained after the fusion of the main and auxiliary streams are weighted and integrated to generate a joint feature map for final detection. in, For learnable gating coefficients, Features after fusion; The output includes structured detection results containing location bounding boxes, disease types, and confidence levels, in the following format: in, This indicates the position and size of the i-th detection box. Indicates the type of disease. Rate its confidence level; The final test results can be used for structured visualization of road defects, alarm prompts, and subsequent maintenance decisions.

Images