Urban road underground disease detection method and system based on image recognition

Abstract

The invention provides an urban road underground disease detection method and system based on image recognition, and belongs to the technical field of urban road intelligent detection. The method comprises the following steps: firstly, acquiring a road disease image data set in a field acquisition and forward modeling mode; then, through combined preprocessing, disease features are enhanced, and data quality is improved; secondly, through a Cascade R-CNN deep learning intelligent detection model, intelligently identifying a disease area and carrying out disease category marking; thirdly, for missed disease areas, obtaining a complete disease area identification data set through manual leak repairing; and finally, through disease attribute editing and risk level calculation, a final inspection report is generated and stored. Reliable technical support is provided for digital and standardized management of municipal road underground diseases, and the problems that traditional manual detection is low in efficiency, high in omission ratio and the like can be effectively solved.

Description

Technical Field

[0001] This invention relates to the field of intelligent detection technology for urban roads, specifically to a method and system for detecting underground defects in urban roads based on image recognition. Background Technology

[0002] With the large-scale development of urban underground space, the potential risks to road infrastructure are becoming increasingly prominent. Frequent incidents of ground subsidence, pipeline damage, and leakage have posed a serious threat to the safety of citizens. Currently, my country's municipal road hazard monitoring system still relies primarily on manual inspections and public feedback. This passive management model has significant shortcomings in early hazard identification, precise location, and preventative management. Therefore, promoting the application of intelligent, high-precision non-destructive testing methods in the periodic inspection of municipal roads and building a technology-enabled preventative maintenance mechanism has become an inevitable choice for improving the safety of urban infrastructure. However, the amount of image data collected during urban road defect detection is extremely large, approximately 12GB / 100km. Furthermore, the processing and interpretation of ground-penetrating radar images still heavily relies on manual operation, leading to challenges such as insufficient timeliness, low processing efficiency, and lack of precision. At the same time, the lack of unified operating standards makes it difficult to adapt to the continuously increasing demand for urban road maintenance projects. Therefore, automatic identification of urban road defects using ground-penetrating radar images has become an urgent problem to be solved in my country's road construction and maintenance management.

[0003] Currently, many scholars have attempted to use deep learning-based target detection algorithms to detect and identify ground-penetrating radar images. However, the following problems still exist when applying them to the identification of urban road defects: (1) The above results have not deeply explored the attribute features of different road defect radar images, and the sample size is small, which leads to model overfitting and the recognition accuracy cannot be guaranteed. (2) Urban road defect ground-penetrating radar images are easily affected by overpasses, streetlights, manhole covers, traffic signs, etc. The road defect features and interference information features are very similar, so some deep learning recognition algorithms cannot be directly applied to this research topic. (3) At present, the automatic identification system for urban road defects based on ground-penetrating radar images still needs further development. Therefore, there is an urgent need for a method that can improve the accuracy of ground-penetrating radar image recognition of urban road defects for small sample data and to develop an automatic identification system that can automatically and quickly detect urban road defects. Summary of the Invention

[0004] To address the problems in the existing technology, this invention provides a method and system for detecting underground road defects based on image recognition. The invention first acquires a road defect image dataset through field data collection and forward simulation; then, it enhances defect features and improves data quality through joint preprocessing; secondly, it intelligently identifies defect areas and labels defect categories using a Cascade R-CNN deep learning intelligent detection model; thirdly, it manually fills in any missed defect areas to obtain a complete defect area identification dataset; finally, it generates and stores a final inspection report through defect attribute editing and risk level calculation. The specific technical solution is as follows:

[0005] On one hand, the present invention provides a method for detecting underground defects in urban roads based on image recognition, the method comprising:

[0006] S1. Use a ground-penetrating radar vehicle to collect real radar image datasets in road scenarios.

[0007] S2. Based on engineering examples, obtain a dataset of simulated radar images of a single disease using forward modeling.

[0008] S3. Based on the real radar image dataset and the single disease simulated radar image dataset, a preprocessed ground radar image dataset is obtained through a preprocessing joint algorithm.

[0009] S4. Based on the preprocessed ground-penetrating radar image dataset, the disease category and regional location are obtained through the Cascade R-CNN deep learning intelligent detection model.

[0010] S5. Based on the disease category and regional location, complete disease area identification dataset is obtained through manual supplementation.

[0011] S6. Based on the complete disease area identification dataset, the final inspection report dataset is obtained through disease attribute editing and risk level calculation.

[0012] Optionally, in S3, based on the real radar image dataset and the single-disease simulated radar image dataset, a preprocessed ground-penetrating radar image dataset is obtained through a joint preprocessing algorithm, including:

[0013] S31. Based on the real radar image dataset and the simulated radar image dataset of a single disease, the DC component is removed using the R_DC algorithm to obtain the first-stage preprocessed ground radar image dataset.

[0014] S32. Based on the first-stage preprocessed ground-penetrating radar image dataset, the second-stage preprocessed ground-penetrating radar image dataset is obtained through a background noise removal algorithm.

[0015] S33. Based on the second-stage preprocessed ground-penetrating radar image dataset, the third-stage preprocessed ground-penetrating radar image dataset is obtained through the exponential gain amplification algorithm.

[0016] S34. Based on the third-stage preprocessed ground-penetrating radar image dataset, the fourth-stage preprocessed ground-penetrating radar image dataset is obtained through the artificial second-gain algorithm.

[0017] S35. Based on the fourth-stage preprocessed ground-penetrating radar image dataset, specific noise is suppressed through two-dimensional frequency domain bandpass filtering to obtain a target-enhanced preprocessed ground-penetrating radar image dataset.

[0018] Optionally, in S4, based on the preprocessed ground-penetrating radar image dataset, the Cascade R-CNN deep learning intelligent detection model is used to obtain the disease category and area location, including:

[0019] S41. Based on the preprocessed ground-penetrating radar image dataset, the intelligent detection preprocessed image dataset is obtained through preprocessing model.

[0020] S42. Based on the intelligent detection preprocessed image dataset, the first-stage intelligent detection dataset is obtained through the backbone neural network model.

[0021] S43. Based on the first-stage intelligent detection dataset, the second-stage intelligent detection dataset is obtained through a multi-level cascaded model.

[0022] S44. Based on the second-stage intelligent detection dataset, the disease category and regional location are obtained through the output layer model.

[0023] Optionally, in step S32, based on the first-stage preprocessed ground-penetrating radar image dataset, a second-stage preprocessed ground-penetrating radar image dataset is obtained through a background noise removal algorithm, including:

[0024] S321. Based on the first-stage preprocessed ground-penetrating radar image dataset, the first-stage noise-removed radar image dataset is obtained through the SVD background noise and target reflection signal separation algorithm.

[0025] S322. Based on the radar image dataset with noise removed in the first stage, the second stage preprocessed ground-penetrating radar image dataset is obtained by using the sliding window DWB adaptive local background noise removal algorithm.

[0026] Optionally, in S43, based on the first-stage intelligent detection dataset, a second-stage intelligent detection dataset is obtained through a multi-level cascaded model, including:

[0027] S431. Based on the first-stage intelligent detection dataset, a primary screening dataset is obtained through the primary detector model;

[0028] S432. Based on the primary screening dataset, the second-stage screening dataset is obtained through a dynamic negative sample filtering algorithm in a cascaded stage.

[0029] S433. Based on the second-stage screening dataset, the screening dataset for fine-tuning bounding boxes is obtained through the bounding box regression optimization model of the cascade stage.

[0030] S434. Based on the filtered dataset of the fine-tuned bounding boxes, a higher quality detection box dataset is obtained through the inter-level dynamic threshold adjustment algorithm.

[0031] S435. Based on this higher quality detection box dataset, a higher-level screening dataset is obtained by using a detector model that focuses on difficult samples in a hierarchical manner.

[0032] S436. Based on the higher-level filtered dataset, repeat steps S432-S435 in a multi-level series to obtain the second-stage intelligent detection dataset.

[0033] On the other hand, the present invention provides an image recognition-based urban road underground defect detection system, which is applied to an image recognition-based urban road underground defect detection method. The system includes:

[0034] The real radar image data acquisition module is used to collect real radar image datasets in road scenes using a ground-penetrating radar vehicle.

[0035] The forward modeling module is used to obtain a dataset of simulated radar images of a single disease based on engineering examples using the forward modeling method.

[0036] The joint preprocessing module is used to obtain a preprocessed ground-penetrating radar image dataset by using a joint preprocessing algorithm based on the real radar image dataset and the single disease simulated radar image dataset.

[0037] The deep learning intelligent detection module is used to obtain the disease category and area location based on the preprocessed ground-penetrating radar image dataset through the CascadeR-CNN deep learning intelligent detection model;

[0038] The manual patching module is used to obtain a complete disease area identification dataset by manually patching up the gaps based on the disease category and area location.

[0039] The disease attribute editing and risk level calculation module is used to obtain the final inspection report dataset based on the complete disease area identification dataset through disease attribute editing and risk level calculation.

[0040] Compared with the prior art, the technical solution of the present invention has at least the following beneficial effects:

[0041] (1) The recognition accuracy was guaranteed even with a small sample input.

[0042] (2) The recognition process can eliminate the influence of interference factors such as overpasses, streetlights, manhole covers, and traffic signs, and the recognition compatibility is higher;

[0043] (3) The identification process involves human assistance, which improves the accuracy of road defect identification.

[0044] (4) The identification process realizes an intelligent process from collection, identification to saving the diagnostic report, which can effectively solve the pain points of low efficiency and high false negative rate of traditional manual detection. Attached Figure Description

[0045] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0046] Figure 1 This is a flowchart of an embodiment of the urban road underground disease detection method based on image recognition of the present invention;

[0047] Figure 2 This is a flowchart of the joint preprocessing process in an embodiment of the image recognition-based method for detecting underground road defects of the present invention.

[0048] Figure 3 This is a flowchart illustrating the background noise removal process in an embodiment of the image recognition-based method for detecting underground road defects of the present invention.

[0049] Figure 4 This is a flowchart of the Cascade R-CNN deep learning intelligent detection method for detecting underground road defects based on image recognition, as described in this invention.

[0050] Figure 5 This is a flowchart of the multi-level cascaded module processing of an embodiment of the urban road underground disease detection method based on image recognition of the present invention;

[0051] Figure 6 This is a system functional module design diagram of the automatic detection system software for underground road defects based on image recognition, according to an embodiment of the present invention.

[0052] Figure 7This is an overall software architecture diagram of the automatic detection system software for underground road defects based on image recognition, according to an embodiment of the present invention.

[0053] Figure 8 This is the core function display interface of the automatic detection system software for underground road defects based on image recognition, as an embodiment of the urban road underground defect detection method of the present invention.

[0054] Figure 9 This is a schematic diagram showing the urban road underground disease detection report, based on an embodiment of the image recognition-based urban road underground disease detection method of the present invention.

[0055] Figure 10 This is a system block diagram of an embodiment of the urban road underground defect detection system based on image recognition of the present invention. Detailed Implementation

[0056] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0057] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0058] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0059] like Figure 1 The flowchart shown is an embodiment of the image recognition-based urban road underground defect detection method of the present invention. The present invention provides an image recognition-based urban road underground defect detection method, which is implemented by an image recognition-based urban road underground defect detection system. The method includes:

[0060] S1. Use a ground-penetrating radar vehicle to collect real radar image datasets in road scenarios.

[0061] Specifically, during the field survey, technicians primarily collected data from various districts and counties. The work area covered multiple road types, mainly semi-rigid asphalt pavements, with a small portion consisting of pedestrian walkways paved with bricks. All field data strictly adhered to standardized acquisition procedures: the equipment mainly used a 500MHz center frequency antenna for detection, with a 125mm lateral sampling step size, a recording window of 0.02ns for each sampling point, and a single-channel acquisition mode to ensure signal fidelity. The image dataset included data on four types of targets: loose, cavitary, degraded, and normal soil.

[0062] S2. Based on engineering examples, obtain a dataset of simulated radar images of a single disease using forward modeling.

[0063] Specifically, the GprMax forward modeling method was used to simulate the radar image attribute characteristics of different urban road defects (including loosening, voids and cavities). The influence of layer structure, geometry and filling medium on the three defects in the simulation process was studied, revealing the formation mechanism and radar image attribute characteristics of the three defects.

[0064] S3. Based on the real radar image dataset and the single disease simulated radar image dataset, a preprocessed ground radar image dataset is obtained through a preprocessing joint algorithm.

[0065] Specifically, such as Figure 2 The flowchart shown is a joint preprocessing process for an embodiment of the image recognition-based urban road underground disease detection method of the present invention. In step S3, based on the real radar image dataset and the single disease simulated radar image dataset, a preprocessed ground-penetrating radar image dataset is obtained through a joint preprocessing algorithm, including:

[0066] S31. Based on the real radar image dataset and the simulated radar image dataset of a single disease, the DC component is removed using the R_DC algorithm to obtain the first-stage preprocessed ground radar image dataset.

[0067] S32. Based on the first-stage preprocessed ground-penetrating radar image dataset, the second-stage preprocessed ground-penetrating radar image dataset is obtained through a background noise removal algorithm.

[0068] Furthermore, such as Figure 3 This is a flowchart illustrating the background noise removal process in an embodiment of the image recognition-based urban road underground disease detection method of the present invention. Step S32, based on the first-stage preprocessed ground-penetrating radar image dataset, uses a background noise removal algorithm to obtain a second-stage preprocessed ground-penetrating radar image dataset, including:

[0069] S321. Based on the first-stage preprocessed ground-penetrating radar image dataset, the first-stage noise-removed radar image dataset is obtained through the SVD background noise and target reflection signal separation algorithm.

[0070] The formula for separating SVD background noise from target reflection signal is shown in Equation 1:

[0071] (1)

[0072] In the formula: The signal after DC removal, The signal before DC transmission. This represents the total number of sampling points for the signal.

[0073] S322. Based on the radar image dataset with noise removed in the first stage, the second stage preprocessed ground radar image dataset is obtained by using the sliding window DWB adaptive local background noise removal algorithm.

[0074] Furthermore, for the non-uniform interface waveform remaining after SVD processing, a sliding window DWB algorithm is used for spatially adaptive local background estimation and removal. DWB is an adaptive background removal method based on local statistical characteristics. It uses sliding window background elimination technology to extract the mean value of adjacent channels as the environmental reference value, which is suitable for signal enhancement in non-stationary noise environments. This data processing strategy is particularly suitable for high-frequency antenna systems or scenarios with significant lateral interference targets. In specific implementation, the system constructs a movable reference window along the data channel and calculates the signal mean value of the window coverage area in real time as the background reference, thereby effectively suppressing random interference signals. The advantage of this method is that it can dynamically adjust the background reference value according to the characteristics of the target area, which is particularly suitable for handling electromagnetic interference problems with spatial distribution characteristics. After parameter comparison, the sliding window size was finally determined to be 3 to process waveforms in the range of 1 to 1200 sampling points, i.e., the number of rows, and finally achieve smoothing of non-uniform areas.

[0075] S33. Based on the second-stage preprocessed ground-penetrating radar image dataset, the third-stage preprocessed ground-penetrating radar image dataset is obtained through the exponential gain amplification algorithm.

[0076] Furthermore, the formula for exponential gain is Equation 2. Exponential gain can handle signals of different intensities more flexibly, which is better than the single amplification of linear gain. While preserving the details of high-amplitude signals, it enhances the discernibility of low-amplitude signals and reduces the risk of saturation distortion in hardware or algorithms by suppressing excessively strong signals.

[0077] (2)

[0078] Where s is the gain factor before the exponential gain, which cannot be 0, and is set to 1; e is the exponent, which cannot be 0, and is set to 2; the gain factor of the starting sampling point is set to 1, and the gain factor of the ending sampling point is set to 1000; x is the horizontal axis, and y is the vertical axis.

[0079] S34. Based on the third-stage preprocessed ground-penetrating radar image dataset, the fourth-stage preprocessed ground-penetrating radar image dataset is obtained through the artificial second-gain algorithm.

[0080] Furthermore, the artificial quadratic gain algorithm is a piecewise linear function. Its principle is to realize the gain using two parameters: the breakpoint position and the gain factor corresponding to the breakpoint position. The calculation method between points is linear interpolation. In one embodiment of the present invention, in addition to the first and last sampling points, three breakpoint positions are set as 2000, 3200, and 5000 respectively, and the gain factors of the five positions are 1, 1800, 2000, 4000, and 8000.

[0081] S35. Based on the fourth-stage preprocessed ground-penetrating radar image dataset, specific noise is suppressed through two-dimensional frequency domain bandpass filtering to obtain a target-enhanced preprocessed ground-penetrating radar image dataset.

[0082] Furthermore, the parameters of the two-dimensional frequency domain bandpass filter include Lg(m) and Ls(m), where Lg(m) represents the length of the lateral low-frequency signal, and Ls(m) is the maximum displacement of the system's micro-vibrations. After parameter testing, we selected Lg(m) equal to 0.5 and Ls(m) equal to 0.05. Similar to horizontal filtering, the two-dimensional frequency domain bandpass filter suppresses low-frequency information in the horizontal direction. This means that the algorithm weakens multiple waves while suppressing low-frequency interference in the horizontal direction, such as layer clutter. Through frequency domain wavenumber bandpass design, the two-dimensional frequency domain bandpass filter accurately suppresses low-frequency interference such as horizontal multiple waves and layer clutter in GPR data, while retaining effective target signals such as tilted or curved holes or gaps. This achieves controllable energy loss from diffraction targets and is suitable for subsequent high-precision target detection scenarios.

[0083] S4. Based on the preprocessed ground-penetrating radar image dataset, the disease category and regional location are obtained through the Cascade R-CNN deep learning intelligent detection model.

[0084] Specifically, such as Figure 4 This is a flowchart of the Cascade R-CNN deep learning intelligent detection method for detecting underground road defects based on image recognition, according to an embodiment of the present invention. Step S4, based on the preprocessed ground-penetrating radar image dataset, uses the Cascade R-CNN deep learning intelligent detection model to obtain the defect category and region location, including:

[0085] S41. Based on the preprocessed ground-penetrating radar image dataset, the intelligent detection preprocessed image dataset is obtained through preprocessing model.

[0086] S42. Based on the intelligent detection preprocessed image dataset, the first-stage intelligent detection dataset is obtained through the backbone neural network model.

[0087] S43. Based on the first-stage intelligent detection dataset, the second-stage intelligent detection dataset is obtained through a multi-level cascaded model.

[0088] Furthermore, such as Figure 5 This is a flowchart of the multi-level cascaded model processing of an embodiment of the urban road underground defect detection method based on image recognition of the present invention. Step S43, based on the first-stage intelligent detection dataset, obtains the second-stage intelligent detection dataset through the multi-level cascaded model, including:

[0089] S431. Based on the first-stage intelligent detection dataset, a primary screening dataset is obtained through the primary detector model;

[0090] S432. Based on the primary screening dataset, the second-stage screening dataset is obtained through a dynamic negative sample filtering algorithm in a cascaded stage.

[0091] Furthermore, dynamic negative sample filtering filters background signals (negative samples) at each cascade stage, effectively alleviating the sample imbalance problem during training and enabling the model to focus on high-value samples.

[0092] S433. Based on the second-stage screening dataset, the screening dataset for fine-tuning bounding boxes is obtained through the bounding box regression optimization model of the cascade stage.

[0093] S434. Based on the filtered dataset of the fine-tuned bounding boxes, a higher quality detection box dataset is obtained through the inter-level dynamic threshold adjustment algorithm.

[0094] S435. Based on the higher quality detection box dataset, a higher-level screening dataset is obtained through a hierarchical detector model that focuses on difficult samples.

[0095] S436. Based on the higher-level filtered dataset, repeat steps S432-S435 in a multi-level series to obtain the second-stage intelligent detection dataset.

[0096] S44. Based on the second-stage intelligent detection dataset, the disease category and regional location are obtained through the output layer model.

[0097] S5. Based on the disease category and regional location, complete disease area identification dataset is obtained through manual supplementation.

[0098] S6. Based on the complete disease area identification dataset, the final inspection report dataset is obtained through disease attribute editing and risk level calculation.

[0099] Furthermore, the software for the automatic detection system for underground defects in urban roads according to embodiments of the present invention is as follows:

[0100] 1) Functional modules

[0101] The automatic detection system for underground road defects is primarily intended for use by technicians performing image annotation and data analysis and processing. The required functions include intelligent identification and detection of road defects, modification and annotation of detected information, and completion and export of defect card information. The system functional modules of the automatic detection system software for underground road defects are designed as follows: Figure 6 It mainly consists of four modules.

[0102] The intelligent detection module, as the core processing unit of the system, employs an image analysis engine based on a deep learning-based target detection algorithm to achieve batch intelligent detection of radar images. Technicians can import images in multiple formats such as PNG / JPEG into folders, triggering the system to automatically execute the following processing flow: First, the image size is adaptively scaled and coordinates are calibrated for preprocessing. Then, the Cascade R-CNN algorithm is used to automatically locate and select diseased areas, simultaneously completing intelligent classification and labeling of diseases such as cavities, looseness, and delamination. To ensure the integrity of the detection, a manual intervention mechanism is specifically designed, supporting manual supplementary labeling of missed areas using rectangle / polygon tools. The labeled data is synchronized in real time to the following process modules and the MySQL database.

[0103] The disease attribute editing and risk calculation module aims to build a dynamic interactive platform for detection data, using a dual-view display mode of tables and disease details.

[0104] The inspection report generation and data management module strictly adheres to the "Technical Specification for Urban Road Maintenance" (CJJ36-2016) standard, and is designed with a structured output template that meets project acceptance requirements. The soil disease information card uses a tabbed layout, including: a basic information area (defect type, length, width, depth, area, pavement condition, defect location, coordinates, etc.); a disease characteristic area (defect radar map, defect geographical location diagram, on-site photos, diagram of pipelines surrounding the defect, information on pipelines surrounding the defect, etc.); a risk assessment area (probability of disease risk occurrence, risk consequence classification, underground disease risk level); and a treatment suggestion area (cause analysis, treatment suggestions, etc.). Technical personnel can upload on-site photos, pipeline distribution maps, and other attachments, and supplement cause analysis and treatment notes using a rich text editor. After data submission, the system performs dual verification (format review and logical verification). Upon successful verification, the data is synchronously stored in the MySQL database, and a log file containing a timestamp and operator ID is generated for later review.

[0105] The user management module is primarily used by user administrators to assign software operation permissions and model management functions to users. User administrators can also assign login and operation functions to other accounts. Model management mainly involves upgrading and replacing existing models.

[0106] 2) System Architecture

[0107] The automatic detection system for underground road defects in this embodiment utilizes two GUI development frameworks, PySimpleGUI and PyDearGUI, for complementary development. It also uses a target detection algorithm based on the PyTorch framework to identify defects in the underground space of urban roads. The overall software architecture of the system can be divided into three parts: image acquisition layer, image processing layer, and image application layer. The overall software architecture is as follows: Figure 7 As shown.

[0108] The image acquisition layer utilizes Swedish MALÅ SIR series ground-penetrating radar vehicles (500MHz antenna configuration) to conduct underground space exploration operations. This equipment, with an effective detection depth of 1-5 meters, accurately covers areas prone to urban road defects. The system's basic data is collected from on-site data collection on municipal roads in a district of Beijing, achieving non-destructive scanning and raw data capture of underground structures through the vehicle-mounted radar system.

[0109] The data processing layer undertakes the core preprocessing functions, including three-way collaborative processing: the radar signal conversion module parses the collected electromagnetic wave data into grayscale images using professional software; the user management system realizes dynamic configuration of user operation permissions through role-based hierarchical classification; and the model selection unit provides multiple algorithm switching interfaces to build a standardized input environment for subsequent intelligent detection.

[0110] The image application layer integrates full-process business function modules: the algorithm detection module uses the Cascade R-CNN deep learning framework to achieve intelligent disease identification, display the detection status in real time and output structured results; the disease information card system automatically generates standardized detection reports, supports visual preview and Word format export; the data management center uses a MySQL database to permanently store detection records, and is equipped with operation log recording and management functions to ensure the traceability of system operation.

[0111] 3) Core Function Demonstration

[0112] The core functional display interface of the automatic detection system for underground road defects in this embodiment is as follows: Figure 8 The intelligent detection module comprises three core functions: intelligent detection task management, detection result management, and missed detection area annotation. Task management is primarily responsible for generating and updating intelligent detection tasks; result management encompasses the visualization and structured storage of detection results; and the missed detection area annotation function mainly involves secondary annotation of defects missed by the target detection algorithm.

[0113] 4) Final report output

[0114] The urban road underground defect detection system software ultimately outputs a detection report. Taking the soil defect information card detection report as an example, the schematic diagram of the urban road underground defect detection report is as follows: Figure 9 The report generation function shown in the image strictly adheres to engineering acceptance specifications, automatically generating standardized soil defect information card reports. The report content includes: all basic information about the defect, on-site photos and diagrams, causal mechanisms, and suggested repair and construction plans.

[0115] like Figure 10 The diagram shown is a system block diagram of an embodiment of the urban road underground defect detection system based on image recognition of the present invention. The present invention provides an urban road underground defect detection system based on image recognition, which is applied to an urban road underground defect detection method based on image recognition. The system includes: a real radar image data acquisition module, a forward simulation module, a joint preprocessing module, a Cascade R-CNN deep learning intelligent detection module, a manual defect filling module, and a defect attribute editing and risk level calculation module. Specifically:

[0116] The real radar image data acquisition module is used to collect real radar image datasets in road scenes using a ground-penetrating radar vehicle.

[0117] The forward modeling module is used to obtain a dataset of simulated radar images of a single disease based on engineering examples using the forward modeling method.

[0118] The joint preprocessing module is used to obtain a preprocessed ground-penetrating radar image dataset by using a joint preprocessing algorithm based on the real radar image dataset and the single disease simulated radar image dataset.

[0119] The deep learning intelligent detection module is used to obtain the disease category and area location based on the preprocessed ground-penetrating radar image dataset through the CascadeR-CNN deep learning intelligent detection model;

[0120] The manual patching module is used to obtain a complete disease area identification dataset by manually patching up the gaps based on the disease category and area location.

[0121] The disease attribute editing and risk level calculation module is used to obtain the final inspection report dataset based on the complete disease area identification dataset through disease attribute editing and risk level calculation.

[0122] This invention provides a method and system for detecting underground road defects based on image recognition. First, it acquires a dataset of road defect images through field data collection and forward simulation. Then, it enhances defect features and improves data quality through joint preprocessing. Next, it uses a Cascade R-CNN deep learning intelligent detection model to intelligently identify defect areas and label defect categories. Third, for any missed defect areas, manual correction is performed to obtain a complete defect area identification dataset. Finally, through defect attribute editing and risk level calculation, a final inspection report is generated and stored. This invention provides reliable technical support for the digital and standardized management of underground road defects, effectively solving the pain points of low efficiency and high missed detection rate in traditional manual detection.

[0123] It is understood that the present invention has been described through the above embodiments and should not be construed as limiting the implementation and scope of the present invention. Those skilled in the art will recognize that various changes or equivalent substitutions can be made to these features and embodiments without departing from the spirit and scope of the present invention. Furthermore, under the teachings of the present invention, these features and embodiments can be modified to adapt to specific situations and materials without departing from the spirit and scope of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the protection scope of the present invention.

Claims

1. An image recognition-based urban road underground disease detection method, characterized in that, The method comprises: S1, using a ground penetrating radar vehicle to collect in a road scene to obtain a real radar image dataset; S2, according to an engineering example, a single disease simulation radar image dataset is obtained by forward simulation method; S3, according to the real radar image dataset and the single disease simulation radar image dataset, a pretreated ground penetrating radar image dataset is obtained by a pretreatment joint algorithm; S4, according to the pretreated ground penetrating radar image dataset, a disease category and area positioning are obtained by a Cascade R-CNN deep learning intelligent detection model; S5, according to the disease category and area positioning, a complete disease area recognition dataset is obtained by artificial leak repair; S6, according to the complete disease area recognition dataset, a final inspection report dataset is obtained by disease attribute editing and risk level calculation. 2.The image recognition-based urban road underground disease detection method according to claim 1, characterized in that, In S3, according to the real radar image dataset and the single disease simulation radar image dataset, a pretreated ground penetrating radar image dataset is obtained by a pretreatment joint algorithm, which comprises: S31, according to the real radar image dataset and the single disease simulation radar image dataset, a first-stage pretreated ground penetrating radar image dataset is obtained by removing direct current components through an R_DC algorithm; S32, according to the first-stage pretreated ground penetrating radar image dataset, a second-stage pretreated ground penetrating radar image dataset is obtained by removing background noise through an algorithm; S33, according to the second-stage pretreated ground penetrating radar image dataset, a third-stage pretreated ground penetrating radar image dataset is obtained by an exponential gain amplification algorithm; S34, according to the third-stage pretreated ground penetrating radar image dataset, a fourth-stage pretreated ground penetrating radar image dataset is obtained by an artificial secondary gain algorithm; S35, according to the fourth-stage pretreated ground penetrating radar image dataset, a target-enhanced pretreated ground penetrating radar image dataset is obtained by two-dimensional frequency domain band-pass filtering to suppress specific noise. 3.The image recognition-based urban road underground disease detection method according to claim 1, characterized in that, In S4, according to the pretreated ground penetrating radar image dataset, a disease category and area positioning are obtained by a Cascade R-CNN deep learning intelligent detection model, which comprises: S41, according to the pretreated ground penetrating radar image dataset, an intelligent detection pretreated image dataset is obtained by a pretreatment model processing; S42, according to the intelligent detection pretreated image dataset, a first-stage intelligent detection dataset is obtained by a backbone neural network model; S43, according to the first-stage intelligent detection dataset, a second-stage intelligent detection dataset is obtained by a multi-level cascade model; S44, according to the second-stage intelligent detection dataset, a disease category and area positioning are obtained by an output layer model. 4.The image recognition-based urban road underground disease detection method according to claim 2, characterized in that, In S32, according to the first-stage pretreated ground penetrating radar image dataset, a second-stage pretreated ground penetrating radar image dataset is obtained by a background noise removal algorithm, which comprises: S321, according to the first stage of pre-processing geological radar image data set, by SVD background noise and target reflection signal separation algorithm, get the first stage of removing noise radar image data set; S322, according to the first stage of removing noise radar image data set, by sliding window DWB adaptive local background noise removal algorithm, get the second stage of pre-processing geological radar image data set. 5.The image recognition-based urban road underground disease detection method according to claim 3, characterized in that, In the S43, according to the first stage of intelligent detection data set, through the multi-stage cascade model, the second stage of intelligent detection data set is obtained, including: S431, according to the first stage of intelligent detection data set, through the primary detector model, the primary screening data set is obtained; S432, according to the primary screening data set, through the dynamic negative sample filtering algorithm of hierarchical cascade stage, the second stage screening data set is obtained; S433, according to the second stage screening data set, through the bounding box regression optimization model of hierarchical cascade stage, the screening data set of fine-tuning bounding box is obtained; S434, according to the screening data set of fine-tuning bounding box, through the inter-stage dynamic threshold adjustment algorithm, the higher quality detection frame data set is obtained; S435, according to the higher quality detection frame data set, through the detector model of focusing on difficult samples in hierarchical progression, the higher level screening data set is obtained; S436, according to the higher level screening data set, multi-stage series repeat steps S432-S435, get the second stage of intelligent detection data set.

6. An image recognition-based urban road underground disease detection system for implementing the image recognition-based urban road underground disease detection method according to any one of claims 1 to 5, characterized by, The system comprises: Real radar image data acquisition module, for using geological radar car in road scene in the field acquisition, get real radar image data set; Forward simulation module, for according to engineering example, through forward simulation method, get single disease simulation radar image data set; Joint preprocessing module, for according to the real radar image data set and the single disease simulation radar image data set, through the preprocessing joint algorithm, get the pre-processing geological radar image data set; Deep learning intelligent detection module, for according to the pre-processing geological radar image data set, through the Cascade R-CNN deep learning intelligent detection model, get the disease category and area positioning; Artificial leak repair module, for according to the disease category and area positioning, through artificial leak repair, get complete disease area recognition data set; Disease attribute editing and risk level calculation module, for according to the complete disease area recognition data set, through disease attribute editing and risk level calculation, get the final inspection report data set.

Images