A highway defect visual recognition system and method based on artificial intelligence

Abstract

The application discloses a highway defect visual identification system and method based on artificial intelligence, and relates to the technical field of highway defect detection.The method comprises the following steps: inputting a road surface image and metadata, outputting an initial defect candidate set through a benchmark artificial intelligence model, calculating a context decay factor based on the metadata to obtain a preliminary calibration confidence, constructing a spatial correlation graph according to the initial defect candidate set, calculating a dynamic risk preliminary evaluation value, identifying a deep verification defect subset based on the dynamic risk preliminary evaluation value, the preliminary calibration confidence and a frame risk baseline value, analyzing the physical characteristics of a difference map by using a normal road surface generation model to obtain a physical rationality score, generating an optimized confidence, extracting the feature fingerprint of a defect satisfying the confidence condition of the optimized confidence, and performing similar case matching in a historical defect library; if a target historical case with a similarity exceeding a similarity threshold is matched, a risk final value is generated, and a historical attribution conclusion is generated; and a structured diagnosis report is output.

Description

Technical Field

[0001] This invention relates to the field of highway defect detection technology, specifically to a highway defect visual recognition system and method based on artificial intelligence. Background Technology

[0002] Visual identification of highway defects is a core component of refined highway maintenance. Current technologies largely rely on single AI models for initial defect screening, which has numerous limitations and fails to fully leverage the value of metadata. Existing technologies do not match lane types based on the geographical location of road surface images with pre-defined road information; they can only locate the general road surface area of ​​defects, failing to precisely correspond to specific lanes. This results in a lack of targeted maintenance operations and significantly reduces maintenance efficiency. Furthermore, current technologies use a single confidence calibration method, failing to incorporate metadata to quantify influencing factors and calculate attenuation factors, making them prone to false defect identification. Risk assessment does not consider spatial relationships between defects, relying solely on single features, leading to insufficient accuracy. Moreover, they do not utilize historical defect cases for similarity matching and risk correction, failing to establish non-spatial relationships between defects, resulting in low utilization of historical data. This hinders continuous improvement in identification accuracy and risk assessment levels, making it difficult to meet the actual needs of refined and efficient highway maintenance. Summary of the Invention

[0003] The purpose of this invention is to provide a visual recognition system and method for highway defects based on artificial intelligence, so as to solve the problems raised in the prior art.

[0004] To achieve the above objectives, the present invention provides the following technical solution: a visual recognition method for highway defects based on artificial intelligence, the method comprising:

[0005] S100: Input road surface image and metadata, perform initial defect screening through benchmark artificial intelligence model, and output an initial defect candidate set including the type, boundary and original confidence of each candidate defect; calculate the context decay factor based on the metadata, and calibrate the original confidence to obtain the preliminary calibrated confidence;

[0006] S200. Construct a spatial association graph based on the initial defect candidate set, and calculate the dynamic risk initial assessment value by comprehensively considering defect type, size, original confidence level, and topological importance; generate a verification trigger threshold and identify a deep verification defect subset based on the dynamic risk initial assessment value, the preliminary calibration confidence level, and the frame risk baseline value obtained based on all defect dynamic risk initial assessment values.

[0007] S300. Using the normal road surface generation model, the defect areas in the depth verification defect subset are reconstructed, the physical characteristics of the difference map are analyzed, the physical rationality score is obtained, and the initial calibration confidence is recalibrated to generate the optimized confidence.

[0008] S400. Extract the feature fingerprint of the defect that meets the confidence condition for the optimized confidence level, set weights based on type and dynamic risk initial assessment value, and perform similar case matching in the historical defect database; if a target historical case with similarity exceeding the similarity threshold is matched, adjust the dynamic risk initial assessment value according to the outcome of the target historical case, generate the risk final value, and generate historical attribution conclusions.

[0009] S500: Update the feature fingerprint, spatial and attribute information, risk final value and optimized confidence of the current defect to the historical defect database, and establish non-spatial association;

[0010] S600 outputs a structured diagnostic report containing defect type, precise boundary, optimized confidence level, risk final value, and the historical attribution conclusions.

[0011] According to the above scheme, step S100 includes:

[0012] S110. Input the road surface image collected by the road inspection and the metadata associated with the road inspection. The metadata includes the acquisition timestamp of the road surface image, the geographical location of the acquisition device, and the attitude data. The acquisition timestamp is used to characterize the acquisition time of the road surface image, the geographical location of the acquisition device is used to characterize the acquisition spatial location of the road surface image, and the attitude data is used to characterize the shooting attitude of the acquisition device.

[0013] S120. Input the road surface image into the benchmark artificial intelligence model for defect detection and segmentation. Defect detection and segmentation involves performing full-domain pixel traversal recognition on the road surface image to distinguish between normal road surface areas and suspected defect areas, identifying suspected defect areas in the road surface, and outputting an initial defect candidate set. Each candidate defect in the initial defect candidate set corresponds to a unique defect type, defect boundary, and original confidence level. The defect type is the morphological category of road surface damage, the defect boundary is the contour range of the suspected defect area, and the original confidence level is the quantified value of the benchmark artificial intelligence model's determination that the area is a defect.

[0014] S130. Parse the metadata, determine the light intensity level based on the acquisition timestamp, the light intensity level is used to characterize the ambient lighting conditions at the acquisition time, calculate the relative tilt angle between the image acquisition plane and the road surface based on the posture data, the relative tilt angle between the image acquisition plane and the road surface is used to characterize the spatial relationship between the acquisition viewpoint and the road surface, and perform lane type matching based on the geographical location and road information; quantify the influence of each parameter in the metadata on the defect recognition accuracy, calculate the context decay factor, the context decay factor is a quantization adjustment coefficient used to correct the original confidence level; the context decay factor is used to characterize the degree of interference of the acquisition environment on the defect recognition result.

[0015] S140. Based on the context decay factor and the original confidence level corresponding to each candidate defect, perform quantitative calibration of the original confidence level. Quantitative calibration is to normalize and correct the original confidence level through the context decay factor, eliminate the identification bias caused by the acquisition environment, and obtain the preliminary calibration confidence level of each candidate defect.

[0016] According to the above scheme, step S200 includes:

[0017] S210. Based on the boundaries of each candidate defect in the initial defect candidate set, determine the spatial location of each candidate defect in the road image, treat each candidate defect as an independent node, and use each independent node to uniquely correspond to a single candidate defect. Establish the association relationship between nodes according to the spatial distance between each node. The association relationship between nodes is used to characterize the spatial proximity between candidate defects. Construct a spatial association graph. The topological importance of a node in the spatial association graph is the quantified value of the association relationship between the node and other nodes.

[0018] S220. Extract the defect type, size, and original confidence level corresponding to each candidate defect. Combine the topological importance of the candidate defect in the spatial correlation graph and obtain the initial dynamic risk assessment value corresponding to each candidate defect through weighted calculation. The weighted calculation is to perform a fusion operation after assigning corresponding weights to the defect type, size, original confidence level, and topological importance.

[0019] S230. Perform statistical analysis on the initial dynamic risk assessment values ​​of all candidate defects. The statistical analysis involves performing an overall quantitative calculation on all initial dynamic risk assessment values ​​to characterize the overall risk level of a single frame image and obtain a frame risk baseline value. The frame risk baseline value is a quantitative representation of the overall risk level of defects in a single frame road image.

[0020] S240. Input the dynamic risk initial assessment value, preliminary calibration confidence level and frame risk baseline value of each candidate defect into the trigger function to calculate the corresponding verification trigger threshold; and identify the candidate defects that meet the requirement that the combined evaluation value of their dynamic risk initial assessment value and preliminary calibration confidence level exceeds their corresponding verification trigger threshold. The combined evaluation value is a comprehensive evaluation parameter obtained by fusing and quantifying the dynamic risk initial assessment value and preliminary calibration confidence level, and is identified as a member of the deep verification defect subset.

[0021] According to the above scheme, step S300 includes:

[0022] S310. Using the normal road surface generation model, with the boundary of each defect in the depth verification defect subset as the benchmark, the defect area and the surrounding road surface background area are selected as the generation input. The surrounding road surface background area is used to provide contextual reference for the reconstruction of the normal road surface morphology. The normal road surface morphology is reconstructed for the area corresponding to each defect, and the normal road surface reconstruction image corresponding to each defect area is output.

[0023] S320. Perform pixel-level comparison between the original road surface image region corresponding to each defect in the deep verification defect subset and the corresponding normal road surface reconstruction image. The pixel-level comparison is to perform difference calculation on the corresponding pixels of the original road surface image region and the normal road surface reconstruction image to generate a difference map corresponding to each defect region.

[0024] S330. Physical features are extracted and analyzed for the difference map corresponding to each defect area. The physical features are used to characterize the real damage features of the defect area to distinguish between real defects and interference factors, and to obtain the physical rationality score corresponding to each defect. The physical features are pixel grayscale distribution, edge gradient consistency and texture continuity in the difference map.

[0025] S340. Use the physical rationality score corresponding to each defect as a calibration coefficient, and recalibrate the initial calibration confidence level corresponding to the defect. The recalibration is to combine the physical rationality score to make a second correction to the initial calibration confidence level, further improve the accuracy of the confidence level, and obtain the optimized confidence level corresponding to each defect.

[0026] According to the above scheme, step S400 includes:

[0027] S410. Determine whether the optimized confidence level of each defect in the deep verification defect subset meets the confidence level condition. The confidence level condition is a preset confidence level qualification criterion, which is used to screen defects with matching value and select defects that meet the confidence level condition as defects to be matched.

[0028] S420. Extract the feature fingerprint of each defect to be matched. The feature fingerprint is a quantized set of the geometric morphology descriptor, the difference map texture statistics and the defect boundary contour features of the defect to be matched. The geometric morphology descriptor is used to characterize the external structural features of the defect, and the difference map texture statistics are used to characterize the texture distribution features of the defect.

[0029] S430. Based on the defect type and initial dynamic risk assessment value of each defect to be matched, set corresponding weights for each feature dimension in the feature fingerprint to form a feature matching weight matrix; the feature matching weight matrix is ​​used to differentiate and highlight the key feature dimensions corresponding to different defect types.

[0030] S440. Perform a weighted similarity calculation between the feature fingerprint of each defect to be matched and the feature fingerprint of all historical defect cases in the historical defect database. The weighted similarity calculation is a weighted fusion operation on the similarity of each feature dimension based on the feature matching weight matrix. If the calculated similarity exceeds the similarity threshold, then the historical defect case is determined as the target historical case.

[0031] S450. Extract the defect outcome information corresponding to the target historical case. The defect outcome information is used to characterize the subsequent development status of the historical defect. Determine the risk adjustment coefficient based on the defect outcome information. The risk adjustment coefficient is used to quantify the risk level of the defect. Based on the risk adjustment coefficient and the dynamic risk preliminary assessment value of the defect to be matched, generate the risk final value corresponding to the defect to be matched. At the same time, combine the defect causes of the target historical case to generate the historical attribution conclusion of the defect to be matched.

[0032] According to the above scheme, step S500 includes:

[0033] S510. For defects whose optimized confidence meets the confidence condition, their feature fingerprint, spatial and attribute information, risk final value and optimized confidence are merged into a new historical defect record; the new historical defect record is the carrier of full-dimensional information integration of the current defect.

[0034] Spatial information refers to the actual spatial coordinates corresponding to the current defect. These coordinates are obtained by converting the geographical location of the acquisition device during road image acquisition. Based on the geographical location, lane type matching is performed with preset road information, and the matched lane type is incorporated into the spatial information. Attribute information includes the defect type, size, physical rationality score, and historical attribution conclusions of the current defect. The update information of each current defect is organized according to a preset data format.

[0035] S520. Store the new historical defect record in the historical defect database;

[0036] S530. Based on the feature fingerprint, calculate the feature similarity between the new historical defect record and the existing historical defect records in the historical defect database; the feature similarity is used to characterize the degree of feature matching between the new record and the existing record.

[0037] S540. If the feature similarity exceeds the association threshold, a non-spatial association based on feature similarity and defect cause association is established between the new historical defect record and the corresponding existing historical defect record.

[0038] An artificial intelligence-based visual recognition system for highway defects includes: an input screening module, a risk verification module, a matching and attribution module, a database update module, and an output module.

[0039] The input screening module is used to input road surface images and metadata, perform initial defect screening through a benchmark artificial intelligence model to obtain an initial defect candidate set, calculate the context decay factor and calibrate the original confidence to obtain a preliminary calibration confidence, and at the same time perform lane type matching based on the geographical location and preset road information to determine the lane information corresponding to the defect.

[0040] The risk verification module is used to construct a spatial correlation map based on the initial defect candidate set, calculate the dynamic risk initial assessment value and frame risk baseline value, generate a verification trigger threshold and identify a deep verification defect subset, reconstruct the defect area through a normal road surface generation model, analyze the physical characteristics of the difference map to obtain a physical rationality score, and recalibrate the initial calibration confidence to generate an optimized confidence.

[0041] The matching attribution module is used to screen and optimize defects that meet the confidence conditions, extract their feature fingerprints and set weights based on defect type and dynamic risk initial assessment value, perform weighted similarity matching in the historical defect database, and adjust the dynamic risk initial assessment value to generate the risk final value and generate historical attribution conclusions after the matching meets the criteria.

[0042] The library update module is used to merge defect-related information into historical defect records and store them in the historical defect library, calculate the feature similarity between the new record and the existing record, and establish a non-spatial association based on feature similarity and defect cause association when the similarity reaches the standard.

[0043] The output module is used to output a structured diagnostic report that includes defect type, precise boundary, optimization confidence level, risk final value, and historical attribution conclusions.

[0044] According to the above scheme, the input screening module includes a metadata unit, a spatial association unit, and a dynamic triggering unit. The metadata unit is used to calculate the context decay factor based on metadata and calibrate the original confidence level to obtain a preliminary calibration confidence level. The spatial association unit is used to construct a spatial association graph of the initial defect candidate set and calculate the dynamic risk initial assessment value by comprehensively considering the defect type, size, original confidence level, and topological importance in the association graph. The dynamic triggering unit is used to dynamically generate a verification triggering threshold based on the dynamic risk initial assessment value, the preliminary calibration confidence level, and the frame risk baseline value determined according to the dynamic risk initial assessment values ​​of all defects, so as to identify a subset of deep verification defects.

[0045] According to the above scheme, the matching attribution module includes a fingerprint retrieval unit and a dynamic correction unit; the fingerprint retrieval unit is used to extract the feature fingerprint of the matching defect, and set weights based on the defect type and the initial dynamic risk assessment value, and perform weighted similarity matching in the historical defect database; the dynamic correction unit is used to increase or decrease the initial dynamic risk assessment value of the current defect according to the outcome of the matched target historical cases, generate the final risk value, and generate historical attribution conclusions.

[0046] According to the above scheme, the library update module includes an encapsulation and storage unit and an association construction unit. The encapsulation and storage unit is used to encapsulate the feature fingerprint, spatial and attribute information, risk final value and optimized confidence of the current defect into a new historical defect record and store it in the historical defect library. The association construction unit is used to establish a non-spatial association between the new historical defect record and the existing records in the historical library based on the feature fingerprint similarity and defect cause association.

[0047] Compared with the prior art, the beneficial effects of the present invention are:

[0048] 1. This application uses the geographical location of the data collection device to match lane types with preset road information, which can accurately determine the lane where the defect is located, thereby improving the efficiency and precision of highway maintenance.

[0049] 2. This application adopts a multi-level confidence calibration method that combines metadata calculation context decay factor with physical rationality score to reduce false defect misjudgment rate and improve the accuracy and reliability of defect identification;

[0050] 3. This application continuously enriches the historical defect knowledge base by establishing non-spatial associations between defects based on feature similarity and causes, thereby achieving continuous improvement in defect identification accuracy and risk assessment accuracy. Attached Figure Description

[0051] Figure 1 This is a flowchart illustrating the steps of a visual recognition method for highway defects based on artificial intelligence, as described in this invention.

[0052] Figure 2 This is a schematic diagram of the structure of a road defect visual recognition system based on artificial intelligence according to the present invention. Detailed Implementation

[0053] 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.

[0054] Example: Figures 1-2 As shown, the present invention provides a technical solution, a visual recognition method for highway defects based on artificial intelligence, the method comprising the following steps:

[0055] S100: Input road surface image and metadata, perform initial defect screening through benchmark artificial intelligence model, and output an initial defect candidate set including the type, boundary and original confidence of each candidate defect; calculate the context decay factor based on the metadata, and calibrate the original confidence to obtain the preliminary calibrated confidence;

[0056] S200. Construct a spatial association graph based on the initial defect candidate set, and calculate the dynamic risk initial assessment value by comprehensively considering defect type, size, original confidence level, and topological importance; generate a verification trigger threshold and identify a deep verification defect subset based on the dynamic risk initial assessment value, the preliminary calibration confidence level, and the frame risk baseline value obtained based on all defect dynamic risk initial assessment values.

[0057] S300. Using the normal road surface generation model, the defect areas in the depth verification defect subset are reconstructed, the physical characteristics of the difference map are analyzed, the physical rationality score is obtained, and the initial calibration confidence is recalibrated to generate the optimized confidence.

[0058] S400. Extract the feature fingerprint of the defect that meets the confidence condition for the optimized confidence level, set weights based on type and dynamic risk initial assessment value, and perform similar case matching in the historical defect database; if a target historical case with similarity exceeding the similarity threshold is matched, adjust the dynamic risk initial assessment value according to the outcome of the target historical case, generate the risk final value, and generate historical attribution conclusions.

[0059] S500: Update the feature fingerprint, spatial and attribute information, risk final value and optimized confidence of the current defect to the historical defect database, and establish non-spatial association;

[0060] S600 outputs a structured diagnostic report containing defect type, precise boundary, optimized confidence level, risk final value, and the historical attribution conclusions.

[0061] Specifically, in this embodiment, a vehicle-mounted line array camera is used to conduct road surface inspection on a national highway at a collection frequency of 20 frames per second, and road surface images and associated metadata are acquired simultaneously. The acquisition device is equipped with a high-precision inertial measurement unit and a differential GPS positioning module. The timestamp corresponding to each frame of the image is accurate to the millisecond level. The geographical location information includes latitude and longitude coordinates, and the attitude data includes the camera's pitch angle θ, roll angle φ, and yaw angle ψ.

[0062] Furthermore, step S100 includes:

[0063] S110. Input the road surface image collected by the road inspection and the metadata associated with the road inspection. The metadata includes the acquisition timestamp of the road surface image, the geographical location of the acquisition device, and the attitude data. The acquisition timestamp is used to characterize the acquisition time of the road surface image, the geographical location of the acquisition device is used to characterize the acquisition spatial location of the road surface image, and the attitude data is used to characterize the shooting attitude of the acquisition device.

[0064] Specifically, input the collected road surface images and metadata, where the collection timestamp is 2026-01-01 10:00:00.000, the geographical location of the collection device is 118° East longitude and 32° North latitude, and the attitude data are pitch angle -2.3°, roll angle 0.8°, and yaw angle 45.6°;

[0065] S120. Input the road surface image into the benchmark artificial intelligence model for defect detection and segmentation. Defect detection and segmentation involves performing full-domain pixel traversal recognition on the road surface image to distinguish between normal road surface areas and suspected defect areas, identifying suspected defect areas in the road surface, and outputting an initial defect candidate set. Each candidate defect in the initial defect candidate set corresponds to a unique defect type, defect boundary, and original confidence level. The defect type is the morphological category of road surface damage, the defect boundary is the contour range of the suspected defect area, and the original confidence level is the quantified value of the benchmark artificial intelligence model's determination that the area is a defect.

[0066] Specifically, the road surface image is input into the benchmark artificial intelligence model. In this embodiment, the benchmark artificial intelligence model adopts the Mask R-CNN architecture, with ResNet-50 as its backbone network. It is pre-trained and fine-tuned on the public road defect dataset Cityscapes and a self-built road surface disease dataset. The benchmark artificial intelligence model performs instance segmentation on the road surface image and outputs an initial defect candidate set. For example, the output result of a certain candidate defect is: defect type: transverse crack, defect boundary is a polygon coordinate sequence: [(120,345),(122,350),...,(118,348)], and the original confidence is 0.87. This is only an example and is not a limitation.

[0067] S130. Parse the metadata, determine the light intensity level based on the acquisition timestamp, the light intensity level is used to characterize the ambient lighting conditions at the acquisition time, calculate the relative tilt angle between the image acquisition plane and the road surface based on the posture data, the relative tilt angle between the image acquisition plane and the road surface is used to characterize the spatial relationship between the acquisition viewpoint and the road surface, and perform lane type matching based on the geographical location and road information; quantify the influence of each parameter in the metadata on the defect recognition accuracy, calculate the context decay factor, the context decay factor is a quantization adjustment coefficient used to correct the original confidence level; the context decay factor is used to characterize the degree of interference of the acquisition environment on the defect recognition result.

[0068] Specifically, based on the data collection timestamp 10:00:00.000, the meteorological data and solar altitude angle model for that day were queried to determine the light intensity level as Level 2. Light intensity levels are divided into 0-4 levels, with lower values ​​indicating worse lighting conditions. The classification criteria for light intensity levels are determined based on the statistical distribution of the signal-to-noise ratio (SNR) of images from different time periods in historical data collection. Time periods with an SNR below 10dB are classified as Level 0, 10-15dB as Level 1, 15-20dB as Level 2, 20-25dB as Level 3, and above 25dB as Level 4. Based on the attitude data, the relative tilt angle between the image acquisition plane and the road surface was calculated using the spatial geometric transformation formula: θ_rel=arccos(cosφ×cosθ); where θ _rel is the relative tilt angle, φ is the roll angle, and θ is the pitch angle; substituting φ=0.8° and θ=-2.3°, we calculate θ_rel=2.44°; based on the geographical location of 118°E and 32°N, spatial matching is performed with the preset high-precision road electronic map to determine that the current road segment is a two-way four-lane highway, the current data acquisition device is located in the second lane, and the lane type matching result is the driving lane; the value of the lane type influence coefficient Lane_n is obtained by statistical analysis of the defect detection rate of different lanes in historical inspection data. In this embodiment, based on the three-year average defect detection rate ratio of fast lane, driving lane, and slow lane of 0.85:1.00:1.15, after normalization, it is taken as 0.1, 0.2, and 0.3 respectively;

[0069] The influence of quantitative parameters on defect identification accuracy is assessed. This embodiment uses a predefined weighted attenuation model to calculate the context attenuation factor, with the formula: α = 1 - (w_1 × L_n + w_2 × θ_n + w_3 × Lane_n); where α is the context attenuation factor, L_n is the normalized illumination intensity level, θ_n is the normalized relative tilt angle, Lane_n is the lane type influence coefficient, and w_1, w_2, and w_3 are preset weight coefficients. The normalized illumination intensity level L_n is obtained by normalizing the illumination intensity level to its maximum and minimum values. The normalization benchmark is determined based on the complete distribution range of illumination intensity levels 0-4 in historical data. The normalized relative tilt angle... The angle θ_n is obtained by normalizing the relative tilt angle to its maximum and minimum values. The normalization benchmark is determined based on the measured range of the relative tilt angle from 0° to 15° in the historical data. The values ​​of the weighting coefficients w_1, w_2, and w_3 are obtained by performing multiple linear regression analysis on the fluctuation range of the original confidence level of the same road segment and the same defect under different acquisition conditions in the historical inspection data. In this embodiment, they are taken as 0.5, 0.3, and 0.2, respectively. In the current frame, L_n=0.5, θ_n=0.12, Lane_n=0.2, and α=1-(0.5×0.5+0.3×0.12+0.2×0.2)=0.674. This is only an example and is not a limitation.

[0070] S140. Based on the context decay factor and the original confidence level corresponding to each candidate defect, perform quantitative calibration of the original confidence level. Quantitative calibration is to normalize and correct the original confidence level through the context decay factor, eliminate the identification bias caused by the acquisition environment, and obtain the preliminary calibration confidence level of each candidate defect.

[0071] Specifically, the context decay factor α=0.674 is multiplied by the original confidence level 0.87 to obtain the preliminary calibration confidence level C_cal=0.87×0.674=0.586. This is only an example and is not a limitation.

[0072] Furthermore, step S200 includes:

[0073] In this embodiment, taking the identification of 12 candidate defects in the current frame image as an example;

[0074] S210. Based on the boundaries of each candidate defect in the initial defect candidate set, determine the spatial location of each candidate defect in the road image, treat each candidate defect as an independent node, and use each independent node to uniquely correspond to a single candidate defect. Establish the association relationship between nodes according to the spatial distance between each node. The association relationship between nodes is used to characterize the spatial proximity between candidate defects. Construct a spatial association graph. The topological importance of a node in the spatial association graph is the quantified value of the association relationship between the node and other nodes.

[0075] Specifically, the geometric center coordinates of each candidate defect boundary are extracted, with each defect treated as an independent node. A spatial proximity threshold of 50 pixels is set, determined through statistical analysis of the distance distribution between adjacent defects in historical inspection data, with the 80th percentile of the distance distribution used as the spatial proximity threshold. The Euclidean distance between any two nodes is calculated, and connecting edges are established between node pairs with a distance less than 50 pixels. A spatial association graph G=(V,E) is constructed, where V represents 12 nodes and E is the set of connecting edges between nodes. Node degree is used as a topological importance index; node degree is the number of edges connected to that node. For example, a crack defect located in a dense rut area has a spatial association with four surrounding defects, and its topological importance T=4. This is only an example and is not a limitation.

[0076] S220. Extract the defect type, size, and original confidence level corresponding to each candidate defect. Combine the topological importance of the candidate defect in the spatial correlation graph and obtain the initial dynamic risk assessment value corresponding to each candidate defect through weighted calculation. The weighted calculation is to perform a fusion operation after assigning corresponding weights to the defect type, size, original confidence level, and topological importance.

[0077] Specifically, the defect type, size, and original confidence level of each candidate defect are extracted. In this embodiment, a weighted summation model is used to calculate the initial dynamic risk assessment value R_init: R_init = a × C_type + b × S_norm + c × C_conf + d × T_norm; where C_type is the defect type risk coefficient, which is obtained by regression analysis of the final repair level and expansion rate of different types of defects in the historical defect database. In this embodiment, pits are taken as 0.9, transverse cracks as 0.6, longitudinal cracks as 0.5, and crazing as 0.8. S_norm is the normalized defect size, and the normalization benchmark is determined based on the distribution range of defect pixel area in historical data, linearly mapping the 0-5000 pixel range to the 0-1 range; C_conf is the original confidence level, and T_norm is the normalized topology. Importance is determined by the normalization benchmark based on the theoretical maximum value of the node degree in the spatial correlation graph and the actual observed distribution, linearly mapping the 0-10 interval to the 0-1 interval; a, b, c, and d are preset weights, which are determined by principal component analysis or analytic hierarchy process (AHP) on the correlation between defect risk level and various parameters in historical inspection data. In this embodiment, they are taken as 0.3, 0.2, 0.3, and 0.2, respectively; For the aforementioned transverse crack defect, its C_type=0.6, S_norm=0.15, C_conf=0.87, T_norm=0.33, and R_init=0.3×0.6+0.2×0.15+0.3×0.87+0.2×0.33=0.18+0.03+0.261+0.066=0.537; This is only an example and is not a limitation.

[0078] S230. Perform statistical analysis on the initial dynamic risk assessment values ​​of all candidate defects. The statistical analysis involves performing an overall quantitative calculation on all initial dynamic risk assessment values ​​to characterize the overall risk level of a single frame image and obtain a frame risk baseline value. The frame risk baseline value is a quantitative representation of the overall risk level of defects in a single frame road image.

[0079] Specifically, the average of the initial dynamic risk assessment values ​​of all 12 defects is calculated to obtain the frame risk baseline value R_base=(ΣR_init_i) / 12=0.412; where R_init_i is the initial dynamic risk assessment value of the i-th candidate defect;

[0080] S240. Input the dynamic risk initial assessment value, preliminary calibration confidence level and frame risk baseline value of each candidate defect into the trigger function to calculate the corresponding verification trigger threshold; and identify the candidate defects that meet the requirement that the combined evaluation value of their dynamic risk initial assessment value and preliminary calibration confidence level exceeds their corresponding verification trigger threshold. The combined evaluation value is a comprehensive evaluation parameter obtained by fusing and quantifying the dynamic risk initial assessment value and preliminary calibration confidence level, and is identified as a member of the deep verification defect subset.

[0081] Specifically, the dynamic risk initial assessment value R_init, the preliminary calibration confidence level C_cal, and the frame risk baseline value R_base of each candidate defect are input into the trigger function. In this embodiment, a ratio-based trigger function is used to calculate the verification trigger threshold, with the formula: θ_trigger=(R_init×C_cal) / R_base; where θ_trigger is the verification trigger threshold. The form and parameters of this trigger function are obtained by fitting the decision boundary of falsely detected defects and correctly detected defects in the two-dimensional distribution of R_init and C_cal in historical inspection data. For a defect, R... With _init=0.537, C_cal=0.586, and R_base=0.412, we calculate θ_trigger=(0.537×0.586) / 0.412=0.764. In this embodiment, the combined evaluation value is the product of R_init and C_cal, which is 0.315. This value does not exceed the verification trigger threshold of 0.764, so this defect does not enter the deep verification defect subset. In this frame, there are a total of 4 defects that meet the condition that the combined evaluation value exceeds the corresponding verification trigger threshold and are identified as deep verification defect subsets. This is only an example and is not a limitation.

[0082] Furthermore, step S300 includes:

[0083] In this embodiment, a generative adversarial network architecture is used to train a normal road surface generation model. The generator adopts a U-Net structure, and the discriminator adopts PatchGAN. 200 rounds of iterative training are completed on a training set containing 100,000 defect-free road surface images.

[0084] S310. Using the normal road surface generation model, with the boundary of each defect in the depth verification defect subset as the benchmark, the defect area and the surrounding road surface background area are selected as the generation input. The surrounding road surface background area is used to provide contextual reference for the reconstruction of the normal road surface morphology. The normal road surface morphology is reconstructed for the area corresponding to each defect, and the normal road surface reconstruction image corresponding to each defect area is output.

[0085] Specifically, taking a pothole defect in the deep verification defect subset as an example, its defect boundary rectangle is [200, 150, 80, 60], which represents the x-coordinate of the top left corner, the y-coordinate of the top left corner, the width, and the height. This rectangle is expanded outward by 20 pixels. The number of pixels to be expanded outward is determined based on the sensitivity analysis of the context region width during the training of the normal road surface generation model. In this embodiment, 20 pixels are used to obtain the generated input region [180, 130, 120, 100]. This region is input into the trained normal road surface generation model, and the corresponding normal road surface reconstruction image is output with a size of 120×100 pixels.

[0086] S320. Perform pixel-level comparison between the original road surface image region corresponding to each defect in the deep verification defect subset and the corresponding normal road surface reconstruction image. The pixel-level comparison is to perform difference calculation on the corresponding pixels of the original road surface image region and the normal road surface reconstruction image to generate a difference map corresponding to each defect region.

[0087] Specifically, the [180,130,120,100] region in the original road surface image is compared with the reconstructed normal road surface image at the pixel level to generate a difference map D(x,y)=|I_org(x,y)-I_rec(x,y)|, where I_org is the pixel gray value of the original image, I_rec is the pixel gray value of the reconstructed image, D(x,y) is the pixel value of the difference map, and x and y are the horizontal and vertical coordinates of the pixel in the image;

[0088] S330. Physical features are extracted and analyzed for the difference map corresponding to each defect area. The physical features are used to characterize the real damage features of the defect area to distinguish between real defects and interference factors, and to obtain the physical rationality score corresponding to each defect. The physical features are pixel grayscale distribution, edge gradient consistency and texture continuity in the difference map.

[0089] Specifically, physical features are extracted from the difference map. This embodiment extracts three physical features: the mean pixel grayscale distribution μ_D, edge gradient consistency C_edge, and texture continuity T_tex. A support vector machine regression model is used to calculate the physical plausibility score, with the formula: P=sigmoid(β_0+β_1×μ_D+β_2×C_edge+β_3×T_tex); where P is the physical plausibility score, and the sigmoid function maps the output to the 0-1 interval. The regression model is trained on a dataset of difference maps labeled with real and fake defects. The input features are the three physical features mentioned above, and the output is a physical rationality label of 0-1. β_0, β_1, β_2, and β_3 are the regression coefficients after the model training converges. In this embodiment, they are taken as -1.2, 0.8, 1.5, and 1.1, respectively. After calculation, the μ_D of the pit defect is 45.3, C_edge is 0.76, and T_tex is 0.62. Substituting these values, we get P=0.83. This is only an example and is not a limitation.

[0090] S340. Use the physical rationality score corresponding to each defect as the calibration coefficient, and recalibrate the initial calibration confidence level corresponding to the defect. The recalibration is to combine the physical rationality score to make a second correction to the initial calibration confidence level, further improve the accuracy of the confidence level, and obtain the optimized confidence level corresponding to each defect.

[0091] Specifically, the physical plausibility score P=0.83 of the defect is used as the calibration coefficient, and its initial calibration confidence C_cal=0.62 is used for recalibration. This embodiment uses a multiplicative calibration model, with the formula: C_opt=C_cal×(1+γ×(P-P_th)); where C_opt is the optimized confidence, γ is the calibration strength coefficient, which is determined by fitting the correlation between the physical plausibility score and the recognition accuracy in the validation set data. In this embodiment, it is set to 0.5. P_th is the physical plausibility benchmark threshold, which is determined by statistically analyzing the equal error rate points in the physical plausibility score distribution of cases marked as real defects and cases marked as false defects in the historical defect database. In this embodiment, it is set to 0.5. Substituting these values ​​into the calculation, we get C_opt=0.62×(1+0.5×(0.83-0.5))=0.62×1.165=0.722. This is only an example and is not a limitation.

[0092] Furthermore, step S400 includes:

[0093] In this embodiment, the historical defect database has accumulated more than 5,000 historical defect case records. Each record contains a feature fingerprint vector, defect outcome information, and defect cause field.

[0094] S410. Determine whether the optimized confidence level of each defect in the deep verification defect subset meets the confidence level condition. The confidence level condition is a preset confidence level qualification criterion, which is used to screen defects with matching value and select defects that meet the confidence level condition as defects to be matched.

[0095] Specifically, the confidence condition is an optimized confidence level C_opt ≥ 0.7. This confidence threshold is determined by backtracking and verifying cases with complete outcome records in the historical defect database, using the balance between recall and precision as the basis for threshold selection. In this embodiment, it is set to 0.7. In the deep verification defect subset, 3 out of 4 defects meet this condition and are selected as defects to be matched.

[0096] S420. Extract the feature fingerprint of each defect to be matched. The feature fingerprint is a quantized set of the geometric morphology descriptor, the difference map texture statistics and the defect boundary contour features of the defect to be matched. The geometric morphology descriptor is used to characterize the external structural features of the defect, and the difference map texture statistics are used to characterize the texture distribution features of the defect.

[0097] Specifically, the feature fingerprint of each defect to be matched is extracted. This embodiment uses a 68-dimensional feature vector, including: an 8-dimensional geometric descriptor, namely aspect ratio, rectangularity, circularity, and the first 5 orders of the Fourier descriptor; 30-dimensional difference map texture statistics, namely six statistics such as energy, entropy, contrast, and inverse difference moment of the gray-level co-occurrence matrix, extracted in four directions (0°, 45°, 90°, and 135°) and two window sizes (5×5 and 9×9); and 30-dimensional defect boundary contour features, namely the mean, variance, skewness, kurtosis, and 26 curvature values ​​sampled at equal intervals along the boundary. This is only an example and is not a limitation.

[0098] S430. Based on the defect type and initial dynamic risk assessment value of each defect to be matched, set corresponding weights for each feature dimension in the feature fingerprint to form a feature matching weight matrix; the feature matching weight matrix is ​​used to differentiate and highlight the key feature dimensions corresponding to different defect types.

[0099] Specifically, feature weights are set based on defect type and initial dynamic risk assessment. In this embodiment, a decision tree model is used to determine the initial weights of each feature dimension. The decision tree model is trained on a dataset labeled with defect type classification tags in a historical defect database. The decrease in Gini impurity of each feature when splitting at a tree node is used as the basis for the value of w_j_base. Then, dynamic adjustment is performed based on the initial dynamic risk assessment, using the formula: w_j = w_j_base × (1 + λ × R_init); where w_j is the final weight of the j-th feature dimension, w_j_base is the Gini importance of this feature in the decision tree model, λ is the risk adjustment coefficient, which is determined by grid search optimization of the distinguishability between high-risk and low-risk defects in the feature space in historical data. In this embodiment, it is set to 0.3, and R_init is the initial dynamic risk assessment value of the defect to be matched. After normalizing all feature dimension weights, a feature matching weight matrix W is formed. This is only an example and is not a limitation.

[0100] S440. Perform a weighted similarity calculation between the feature fingerprint of each defect to be matched and the feature fingerprint of all historical defect cases in the historical defect database. The weighted similarity calculation is a weighted fusion operation on the similarity of each feature dimension based on the feature matching weight matrix. If the calculated similarity exceeds the similarity threshold, then the historical defect case is determined as the target historical case.

[0101] Specifically, weighted cosine similarity is used to calculate the feature similarity between the defect to be matched and historical defect cases: Sim=(∑w_j×f_j×g_j) / [(∑w_j×f_j 2 ) (1 / 2) ×(∑w_j×g_j 2 ) (1 / 2)In this context, Sim represents the feature similarity, f_j is the j-th dimension of the feature fingerprint of the defect to be matched, g_j is the j-th dimension of the feature fingerprint of historical defect cases, and w_j is the corresponding weight. A similarity threshold θ_sim = 0.75 is set. This threshold is determined by receiver operating characteristic (ROC) curve analysis of the similarity distribution between known similar and dissimilar defects in the historical defect database, with the maximum Youden index used as the threshold value. Calculations show that a defect to be matched has a similarity of 0.81 with case number H-2023-0001 in the historical defect database, exceeding the similarity threshold. Therefore, this case is identified as the target historical case. This is only an example and is not intended to impose any limitations.

[0102] S450. Extract defect outcome information corresponding to the target historical case. The defect outcome information is used to characterize the subsequent development status of the historical defect. Determine the risk adjustment coefficient based on the defect outcome information. The risk adjustment coefficient is used to quantify the risk level of the defect. Based on the risk adjustment coefficient and the dynamic risk initial assessment value of the defect to be matched, generate the risk final value corresponding to the defect to be matched. At the same time, combine the defect causes of the target historical case to generate the historical attribution conclusion of the defect to be matched.

[0103] Specifically, the defect outcome information of historical case H-2023-0001 is extracted. This case involves a similar pothole defect discovered three years ago. Records show that it expanded from a diameter of 15cm to 45cm within 6 months, eventually developing into severe structural damage, with an emergency repair level. The risk adjustment coefficient γ_risk is determined based on the defect outcome information: γ_risk = 1 + δ × (S_severity + E_rate); where δ is the adjustment step size (0.2 in this embodiment), S_severity is the severity level of the outcome, determined according to the priority of pavement repair in highway maintenance specifications, combined with the actual urgency level of repair work orders in the historical case database (0 for minor, 1 for moderate, 2 for severe, and 3 for emergency in this embodiment), and E_rate is the rate of expansion level, determined based on the change in defect area over time in the historical case database. The percentile of the slope distribution is determined as follows: 0 for slow, 1 for moderate, 2 for fast, and 3 for rapid. In this case, S_severity=3 and E_rate=2, so γ_risk=1+0.2×(3+2)=2.0 is calculated. This coefficient is multiplied by the initial dynamic risk assessment value R_init=0.65 of the defect to be matched (another pothole defect) to generate the final risk value R_final=0.65×2.0=1.30. At the same time, the defect cause fields of base layer water seepage and heavy traffic are extracted from the historical cases of this target. Combined with the similarity of 0.81 between the current defect and the historical case, the historical attribution conclusion is generated: the defect morphology is highly similar to the historical case number H-2023-0001. This case is confirmed as a heavy-load fatigue crack caused by base layer water seepage, which eventually developed into a serious pothole. It is recommended to focus on inspecting drainage facilities and heavy-load lanes. This is only an example and is not a limitation.

[0104] Furthermore, step S500 includes:

[0105] S510. For defects whose optimized confidence meets the confidence condition, their feature fingerprint, spatial and attribute information, risk final value and optimized confidence are merged into a new historical defect record; the new historical defect record is the carrier of full-dimensional information integration of the current defect.

[0106] Furthermore, the spatial information is the actual spatial coordinates corresponding to the current defect. These coordinates are obtained by converting the geographical location of the acquisition device when the road image is acquired. Based on the geographical location, lane type matching is performed with preset road information, and the matched lane type is incorporated into the spatial information. The attribute information includes the defect type, size, physical rationality score, and historical attribution conclusions of the current defect. The update information of each current defect is organized according to a preset data format.

[0107] Specifically, for defects whose optimized confidence level meets the confidence level conditions, their feature fingerprint, spatial and attribute information, risk final value of 1.30, and optimized confidence level of 0.722 are merged into a new historical defect record. The spatial information includes: WGS-84 coordinates (118°E, 32°N) obtained by converting the geographic location of the acquisition device, and the driving lane determined by lane type matching. The attribute information includes: defect type is pothole, size is 1860 pixels, physical plausibility score is 0.83, and historical attribution conclusion text.

[0108] S520. Store the new historical defect record in the historical defect database;

[0109] Specifically, new historical defect records are stored in the historical defect database, and the system automatically assigns a unique case number H-2026-0001 to them;

[0110] S530. Based on the feature fingerprint, calculate the feature similarity between the new historical defect record and the existing historical defect records in the historical defect database; the feature similarity is used to characterize the degree of feature matching between the new record and the existing record.

[0111] Specifically, using the feature fingerprint of the new record as the query vector, and employing Euclidean distance as the metric, the feature similarity between the new record and all 5001 existing records in the historical defect database is calculated: d_E=[∑(f_j_new-f_j_old) 2 ] (1 / 2) Where d_E represents the Euclidean distance, f_j_new represents the j-th dimension feature value in the feature fingerprint of a new historical defect record, and f_j_old represents the j-th dimension feature value in the feature fingerprint of an existing record in the historical defect database. The Euclidean distance is converted into a similarity score in the range of 0-1: Sim'=1 / (1+d_E); where Sim' is the feature similarity obtained based on the Euclidean distance conversion.

[0112] S540. If the feature similarity exceeds the association threshold, a non-spatial association based on feature similarity and defect cause association is established between the new historical defect record and the corresponding existing historical defect record.

[0113] Specifically, in this embodiment, the association threshold θ_assoc=0.70 is set. This threshold is determined by statistical learning of the discriminative power of defect pairs with clear evolutionary associations and random defect pairs in the feature similarity distribution in the historical defect database, with the optimization objective of maximizing the F1 score. After calculation, the feature similarity Sim'=0.78 between the new record and the historical case H-2023-0001 exceeds the association threshold. Therefore, a non-spatial association edge is established between the new record and the historical case H-2023-0001. This association edge is stored in a graph data structure, representing that although the two defects are spatially separated by tens of kilometers, they have a high degree of commonality in morphological features and causal mechanisms.

[0114] This invention provides another technical solution: a visual recognition system for highway defects based on artificial intelligence. The system includes: an input screening module, a risk verification module, a matching and attribution module, a database update module, and an output module.

[0115] The input screening module is used to input road surface images and metadata, perform initial defect screening through a benchmark artificial intelligence model to obtain an initial defect candidate set, calculate the context decay factor and calibrate the original confidence to obtain a preliminary calibration confidence, and at the same time perform lane type matching based on the geographical location and preset road information to determine the lane information corresponding to the defect.

[0116] The risk verification module is used to construct a spatial correlation map based on the initial defect candidate set, calculate the dynamic risk initial assessment value and frame risk baseline value, generate a verification trigger threshold and identify a deep verification defect subset, reconstruct the defect area through a normal road surface generation model, analyze the physical characteristics of the difference map to obtain a physical rationality score, and recalibrate the initial calibration confidence to generate an optimized confidence.

[0117] The matching attribution module is used to screen and optimize defects that meet the confidence conditions, extract their feature fingerprints and set weights based on defect type and dynamic risk initial assessment value, perform weighted similarity matching in the historical defect database, and adjust the dynamic risk initial assessment value to generate the risk final value and generate historical attribution conclusions after the matching meets the criteria.

[0118] The library update module is used to merge defect-related information into historical defect records and store them in the historical defect library, calculate the feature similarity between the new record and the existing record, and establish a non-spatial association based on feature similarity and defect cause association when the similarity reaches the standard.

[0119] The output module is used to output a structured diagnostic report that includes defect type, precise boundary, optimization confidence level, risk final value, and historical attribution conclusions.

[0120] Furthermore, the input screening module includes a metadata unit, a spatial association unit, and a dynamic triggering unit. The metadata unit is used to calculate the context decay factor based on metadata and calibrate the original confidence level to obtain a preliminary calibration confidence level. The spatial association unit is used to construct a spatial association graph of the initial defect candidate set and calculate the dynamic risk initial assessment value by comprehensively considering the defect type, size, original confidence level, and topological importance in the association graph. The dynamic triggering unit is used to dynamically generate a verification triggering threshold based on the dynamic risk initial assessment value, the preliminary calibration confidence level, and the frame risk baseline value determined according to the dynamic risk initial assessment values ​​of all defects, so as to identify a subset of deep verification defects.

[0121] Furthermore, the matching attribution module includes a fingerprint retrieval unit and a dynamic correction unit; the fingerprint retrieval unit is used to extract the feature fingerprints of the matching defects, and set weights based on the defect type and the initial dynamic risk assessment value, and perform weighted similarity matching in the historical defect database; the dynamic correction unit is used to increase or decrease the initial dynamic risk assessment value of the current defect according to the outcome of the matched target historical cases, generate the final risk value, and generate historical attribution conclusions.

[0122] Furthermore, the library update module includes an encapsulation and storage unit and an association construction unit; the encapsulation and storage unit is used to encapsulate the feature fingerprint, spatial and attribute information, risk final value and optimized confidence of the current defect into a new historical defect record and store it in the historical defect library; the association construction unit is used to establish a non-spatial association between the new historical defect record and the existing records in the historical library based on the feature fingerprint similarity and defect cause association.

[0123] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A visual recognition method for highway defects based on artificial intelligence, characterized in that: The method includes: S100. Input road surface image and metadata, the metadata including the acquisition timestamp of the road surface image, the geographical location of the acquisition device, and attitude data; perform lane type matching based on the geographical location and preset road information to determine the lane information corresponding to the defect; perform initial defect screening through a benchmark artificial intelligence model, and output an initial defect candidate set including the type, boundary, and original confidence of each candidate defect; quantify the influence of each parameter in the metadata on the defect identification accuracy, calculate the context decay factor, the context decay factor is a quantization adjustment coefficient used to correct the original confidence, and calibrate the original confidence to obtain a preliminary calibrated confidence; S200. Construct a spatial association graph based on the initial defect candidate set, and calculate the dynamic risk initial assessment value by comprehensively considering defect type, size, original confidence level, and topological importance; generate a verification trigger threshold and identify a deep verification defect subset based on the dynamic risk initial assessment value, the preliminary calibration confidence level, and the frame risk baseline value obtained based on all defect dynamic risk initial assessment values. S300. Using the normal road surface generation model, the defect areas in the depth verification defect subset are reconstructed, the physical characteristics of the difference map are analyzed, the physical rationality score is obtained, and the initial calibration confidence is recalibrated to generate the optimized confidence. S400. Extract the feature fingerprint of the defect that meets the confidence condition for the optimized confidence level, set weights based on type and dynamic risk initial assessment value, and perform similar case matching in the historical defect database; if a target historical case with similarity exceeding the similarity threshold is matched, adjust the dynamic risk initial assessment value according to the outcome of the target historical case, generate the risk final value, and generate historical attribution conclusions. S500: Update the feature fingerprint, spatial and attribute information, risk final value and optimized confidence of the current defect to the historical defect database, calculate the feature similarity between the new record and the existing record, and establish a non-spatial association based on feature similarity and defect cause association when the similarity meets the standard. S600 outputs a structured diagnostic report containing defect type, precise boundary, optimized confidence level, risk final value, and the historical attribution conclusions.

2. The method for visual recognition of highway defects based on artificial intelligence according to claim 1, characterized in that: Step S100 includes: S110, Input the road surface images collected during road inspection and the metadata associated with the road inspection; S120. Input the road surface image into the benchmark artificial intelligence model for defect detection and segmentation, identify the suspected defect area in the road surface, and output an initial defect candidate set. Each candidate defect in the initial defect candidate set corresponds to a unique defect type, defect boundary and original confidence level. The defect type is the morphological category of road surface damage, the defect boundary is the outline range of the suspected defect area, and the original confidence level is the quantified value of the confidence that the benchmark artificial intelligence model determines that the area is a defect. S130. Parse the metadata, determine the light intensity level based on the acquisition timestamp, calculate the relative tilt angle between the image acquisition plane and the road surface based on the attitude data, and perform lane type matching based on the geographical location and road information; quantify the influence of each parameter in the metadata on the defect identification accuracy, and calculate the context decay factor, which is a quantization adjustment coefficient used to correct the original confidence level. S140. Based on the context decay factor and the original confidence level corresponding to each candidate defect, perform quantitative calibration of the original confidence level to obtain the preliminary calibration confidence level of each candidate defect.

3. The method for visual recognition of highway defects based on artificial intelligence according to claim 1, characterized in that: Step S200 includes: S210. Based on the boundaries of each candidate defect in the initial defect candidate set, determine the spatial location of each candidate defect in the road image, treat each candidate defect as an independent node, establish the association relationship between nodes according to the spatial distance between each node, and construct a spatial association graph. The topological importance of a node in the spatial association graph is the quantified value of the association relationship between the node and other nodes. S220. Extract the defect type, size and original confidence level of each candidate defect, and combine the topological importance of the candidate defect in the spatial association graph to obtain the initial dynamic risk assessment value of each candidate defect through weighted calculation. S230. Statistical analysis is performed on the initial dynamic risk assessment values ​​of all candidate defects to obtain the frame risk baseline value. The frame risk baseline value is a quantitative representation of the overall risk level of defects in a single frame road image. S240. Input the dynamic risk initial assessment value, preliminary calibration confidence level and frame risk baseline value of each candidate defect into the trigger function to calculate the corresponding verification trigger threshold; and identify the candidate defect whose combined assessment value of its dynamic risk initial assessment value and preliminary calibration confidence level exceeds its corresponding verification trigger threshold as a member of the deep verification defect subset.

4. The method for visual recognition of highway defects based on artificial intelligence according to claim 1, characterized in that: Step S300 includes: S310. Using the normal road surface generation model, with the boundary of each defect in the depth verification defect subset as the benchmark, select the defect area and the surrounding road background area as the generation input, reconstruct the normal road surface morphology of the area corresponding to each defect, and output the normal road surface reconstruction image corresponding to each defect area. S320. Compare the original road surface image region corresponding to each defect in the deep verification defect subset with the corresponding normal road surface reconstruction image at the pixel level to generate a difference map corresponding to each defect region. S330. Physical features are extracted and analyzed from the difference map corresponding to each defect area to obtain the physical rationality score corresponding to each defect. The physical features are pixel grayscale distribution, edge gradient consistency and texture continuity in the difference map. S340. Use the physical rationality score corresponding to each defect as a calibration coefficient, and recalibrate the initial calibration confidence level corresponding to the defect to obtain the optimized confidence level corresponding to each defect.

5. The method for visual recognition of highway defects based on artificial intelligence according to claim 1, characterized in that: Step S400 includes: S410. Determine whether the optimized confidence of each defect in the deep verification defect subset meets the confidence condition, and select the defects that meet the confidence condition as the defects to be matched. S420. Extract the feature fingerprint of each defect to be matched. The feature fingerprint is a quantized set of the geometric morphology descriptor, difference map texture statistics and defect boundary contour features of the defect to be matched. S430. Based on the defect type and dynamic risk initial assessment value of each defect to be matched, set corresponding weights for each feature dimension in the feature fingerprint to form a feature matching weight matrix. S440. Perform a weighted similarity calculation between the feature fingerprint of each defect to be matched and the feature fingerprint of all historical defect cases in the historical defect database. If the calculated similarity exceeds the similarity threshold, then the historical defect case is determined as the target historical case. S450. Extract the defect outcome information corresponding to the target historical case, and determine the risk adjustment coefficient based on the defect outcome information; based on the risk adjustment coefficient and the dynamic risk preliminary assessment value of the defect to be matched, generate the risk final value corresponding to the defect to be matched; at the same time, combine the defect causes of the target historical case to generate the historical attribution conclusion of the defect to be matched.

6. The method for visual recognition of highway defects based on artificial intelligence according to claim 5, characterized in that: Step S500 includes: S510. For defects whose optimized confidence meets the confidence condition, their feature fingerprint, spatial and attribute information, risk final value and optimized confidence are merged into a new historical defect record. S520. Store the new historical defect record in the historical defect database; S530. Based on the feature fingerprint, calculate the feature similarity between the new historical defect record and the existing historical defect records in the historical defect database; S540. If the feature similarity exceeds the association threshold, a non-spatial association based on feature similarity and defect cause association is established between the new historical defect record and the corresponding existing historical defect record.

7. A visual recognition system for highway defects based on artificial intelligence, characterized in that: The system includes: an input screening module, a risk verification module, a matching and attribution module, a database update module, and an output module; The input screening module is used to input road surface images and metadata. The metadata includes the acquisition timestamp of the road surface image, the geographical location of the acquisition device, and attitude data. An initial defect candidate set is obtained by performing initial defect screening through a benchmark artificial intelligence model. The influence of each parameter in the metadata on the defect recognition accuracy is quantified. A context decay factor is calculated and the original confidence is calibrated to obtain a preliminary calibration confidence. The context decay factor is a quantization adjustment coefficient used to correct the original confidence. At the same time, lane type matching is performed based on the geographical location and preset road information to determine the lane information corresponding to the defect. The risk verification module is used to construct a spatial correlation map based on the initial defect candidate set, calculate the dynamic risk initial assessment value and frame risk baseline value, generate a verification trigger threshold and identify a deep verification defect subset, reconstruct the defect area through a normal road surface generation model, analyze the physical characteristics of the difference map to obtain a physical rationality score, and recalibrate the initial calibration confidence to generate an optimized confidence. The matching attribution module is used to screen and optimize defects that meet the confidence conditions, extract their feature fingerprints and set weights based on defect type and dynamic risk initial assessment value, perform weighted similarity matching in the historical defect database, and after the matching is successful, adjust the dynamic risk initial assessment value according to the outcome of the target's historical cases, generate the risk final value and generate historical attribution conclusions. The library update module is used to merge defect-related information into historical defect records and store them in the historical defect library, calculate the feature similarity between the new record and the existing record, and establish a non-spatial association based on feature similarity and defect cause association when the similarity reaches the standard. The output module is used to output a structured diagnostic report that includes defect type, precise boundary, optimization confidence level, risk final value, and historical attribution conclusions.

8. The artificial intelligence-based visual recognition system for highway defects according to claim 7, characterized in that: The input screening module includes a metadata unit, a spatial association unit, and a dynamic triggering unit; The metadata unit is used to calculate the context decay factor based on the metadata and to calibrate the original confidence level to obtain a preliminary calibration confidence level. The spatial association unit is used to construct a spatial association graph of the initial defect candidate set, and to calculate the initial dynamic risk assessment value by comprehensively considering the defect type, size, original confidence level, and topological importance in the association graph. The dynamic triggering unit is used to dynamically generate a verification triggering threshold based on the initial dynamic risk assessment value, the initial calibration confidence level, and the frame risk baseline value determined according to the initial dynamic risk assessment values ​​of all defects, so as to identify a subset of deep verification defects.

9. The artificial intelligence-based visual recognition system for highway defects according to claim 7, characterized in that: The matching attribution module includes a fingerprint retrieval unit and a dynamic correction unit; The fingerprint retrieval unit is used to extract the feature fingerprints of matching defects, and set weights based on the defect type and the initial dynamic risk assessment value, and perform weighted similarity matching in the historical defect database. The dynamic correction unit is used to increase or decrease the initial dynamic risk assessment value of the current defect based on the outcome of the matched target historical cases, generate the final risk value, and generate historical attribution conclusions.

10. The artificial intelligence-based visual recognition system for highway defects according to claim 7, characterized in that: The library update module includes a library encapsulation unit and an association construction unit; The encapsulation and storage unit is used to encapsulate the feature fingerprint, spatial and attribute information, risk final value and optimized confidence of the current defect into a new historical defect record and store it in the historical defect database. The association construction unit is used to establish a non-spatial association between the new historical defect record and existing records in the historical database based on the similarity of feature fingerprints and the association of defect causes.

Images