Road and bridge surface defect rapid evaluation system based on deep learning

By constructing a component reference table, image overlay mapping, and blind spot recognition, a deep learning-based rapid assessment system for surface defects in roads and bridges solves the problem of automatic identification and supplementary inspection of components that were not effectively photographed during bridge inspections, thereby improving the completeness and efficiency of inspection results.

CN123071276APending Publication Date: 2026-11-10NANJING XIAOZHUANG UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611251176.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-18
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Existing technologies cannot automatically identify which structural components on the surface of a bridge are not effectively photographed during inspections, resulting in the inability to automatically verify the integrity of the inspection results. Furthermore, defect detection algorithms cannot consider the integrity of the input images, making it easy to miss critical areas.

Method used

A rapid assessment system for surface defects of roads and bridges based on deep learning is adopted, including a component reference table construction module, an image coverage mapping module, a blind spot identification module, and a supplementary inspection guidance generation module. The system determines the coverage effectiveness through component semantic segmentation and image localization, identifies blind spots, and generates supplementary inspection guidance.

Benefits of technology

It enables the completeness verification of inspection results, extending from pre-inspection route planning to post-inspection coverage acceptance, reducing the workload of manual verification, improving the accuracy of coverage statistics, and providing automatic supplementary inspection guidance. It is suitable for old bridges that have not established digital models.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123071276A_ABST
    Figure CN123071276A_ABST
Patent Text Reader

Abstract

This invention discloses a rapid assessment system for road and bridge surface defects based on deep learning, relating to the field of road maintenance technology. The system includes: a component reference table construction module, an image coverage mapping module, a blind spot identification module, and a supplementary inspection guidance generation module. The rapid assessment system for road and bridge surface defects proposed in this invention extends the verification of inspection integrity from pre-inspection path planning to the coverage acceptance stage after inspection. This fills the gap in existing technologies where it is impossible to automatically identify which component surfaces were not effectively photographed after inspection. It provides a quantifiable verification method for the integrity of inspection results. By introducing an image quality assessment to distinguish between "captured" and "obvious" images, it can be used for defect identification. By incorporating shooting distance, observation angle, and exposure sufficiency into the coverage effectiveness judgment, the coverage statistics are made closer to actual detection capabilities, avoiding overestimation of inspection integrity due to simply counting whether an image covers a certain area.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of road maintenance technology, specifically to a rapid assessment system for surface defects of roads and bridges based on deep learning. Background Technology

[0002] Regular visual inspections of roads and bridges are a fundamental part of maintenance management. Currently, the common practice is to use drones or manual cameras to capture images along the bridge surface, and then use detection algorithms or manual methods to identify surface defects from the images. To ensure complete inspection coverage, existing technologies mainly focus on two aspects: first, planning drone flight paths or pre-setting shooting points based on a 3D model of the bridge before the inspection to reduce omissions by optimizing the flight path; second, improving the recognition accuracy of defect detection algorithms to extract more complete information about defects from the collected images. In actual inspections, the execution of preset routes often deviates due to factors such as temporary equipment failures, obstructed views by pipelines under bridges, or operational errors, resulting in some structural components not being effectively photographed. Existing path planning methods are all pre-emptive strategies and cannot perform structured coverage verification of the actual photographed results after the inspection. After the inspection, the acceptance of coverage still relies on manual verification of component photographs against bridge drawings, which is inefficient and prone to omissions. On the other hand, existing defect detection algorithms assume that the input image already covers the area to be inspected, without considering the integrity of the input data itself. When critical areas are missed, the algorithm does not issue a prompt, and the omission of defects is directly passed on to maintenance decisions. Specifically, there is currently a lack of a technical solution that uses components as basic units to automatically identify which structural component surfaces have not been effectively photographed and provides guidance for supplementary inspection. To address this, we propose a rapid assessment system for road and bridge surface defects based on deep learning. Summary of the Invention

[0003] To address the aforementioned technical issues, a rapid assessment system for road and bridge surface defects based on deep learning is provided, including: Component reference table construction module, image coverage mapping module, blind spot recognition module, and supplementary inspection guidance generation module; The component reference table construction module is used to divide the bridge surface according to the visible surface as the basic unit based on the bridge design data, and generate a component reference table containing component category, span number, number, spatial coordinates and surface area; The image coverage mapping module is electrically connected to the component reference table construction module. It is used to perform component semantic segmentation on the inspection image to identify the category of each pixel, combine the image positioning information with the component reference table to determine the specific component number, and determine the coverage validity according to the effective identification distance threshold, observation angle threshold and minimum contrast corresponding to the component category, and map the effective coverage to the component reference table. The blind spot identification module is electrically connected to the image coverage mapping module. It is used to calculate the coverage rate according to the ratio of the effective coverage area to the surface area of ​​each component surface, mark the component surface below the corresponding coverage threshold as a blind spot component, and aggregate adjacent blind spot components in the same span into a blind spot region. The supplementary inspection guidance generation module is electrically connected to the blind spot identification module. It is used to output a supplementary inspection path sequence that includes a list of supplementary inspection components, suggested shooting directions and distance ranges, and spatial adjacency relationships, based on the spatial location and component type of each blind spot area.

[0004] Preferably, the component reference table construction module constructs the component reference table using the following method: Extract the number of spans, the arrangement of components in each span, and the geometric dimensions of the components from the bridge design drawings. Divide the bridge surface level by level according to the span number, component type, and visible surface direction, and use the single visible surface of a single component as the basic recording unit. For each recording unit, the surface area is calculated based on its geometric dimensions, and the three-dimensional coordinates of the center point are determined based on its position in the overall coordinate system of the bridge. The component category, span number, component number, center point coordinates and surface area are combined into a component surface record. Collect all component surface records and construct a component reference table.

[0005] Preferably, the component categories include the bottom surface of the bridge deck, the web of the T-beam, the side surface of the pier, the bottom and side surfaces of the cap beam, and the side surface of the stop block; The number of each component face is generated by combining the span number, component category identifier, and sequence number within the same span and category; The surface area is calculated based on the geometry of the corresponding component surface and the dimensional parameters marked on the design drawings.

[0006] Preferably, in the image overlay mapping module, the method for performing component semantic segmentation and component number matching is as follows: A semantic segmentation network for bridge components is used to classify inspection images pixel by pixel and label the component category to which each pixel belongs; Read the GPS coordinates and IMU attitude data attached to the image to determine the shooting location and camera orientation; Match the shooting location with the spatial coordinates of each component surface in the component reference table to lock the corresponding span number and candidate component number set of the image; By taking the intersection of pixel-level category labels and the candidate component number set, the specific component number corresponding to each pixel region is determined, thus completing the mapping from the image to the component reference table.

[0007] Preferably, in the image coverage mapping module, the method for determining coverage validity based on the threshold corresponding to the component category is as follows: The shooting distance is estimated based on the imaging proportion of the participating components in the image, the shooting angle is determined based on the angle between the image plane normal vector and the component surface normal vector, and the exposure sufficiency is measured based on the local contrast of the pixel grayscale of the component area. When the shooting distance does not exceed the effective recognition distance threshold corresponding to the component category, the shooting angle does not exceed the effective observation angle threshold, and the local contrast is not lower than the minimum recognition contrast, the image is determined to be an effective cover of the component surface; otherwise, it is determined to be an invalid cover and is not included in the coverage area of ​​the component surface.

[0008] Preferably, in the image coverage mapping module, the method for setting the effective recognition distance threshold, the effective observation angle threshold, and the minimum recognition contrast is as follows: The effective identification distance threshold is determined based on the minimum defect identification scale corresponding to the component category and the spatial resolution parameters of the camera. The effective observation angle threshold is set based on the geometric features of the component surface; the threshold for planar components is greater than that for curved components. The minimum recognition contrast is set based on the typical grayscale difference between surface defects and the background for this component category.

[0009] Preferably, the blind spot identification module uses the following method for coverage calculation and blind spot marking: Traverse the component reference table, and for each component surface, divide its recorded effective coverage area by the surface area of ​​that component surface to obtain the coverage rate of that component surface; The coverage rate of each component surface is compared with the coverage threshold corresponding to that component category. Component surfaces with a coverage rate lower than the coverage threshold are marked as blind zone components. The coverage threshold is set separately for each component category, with the coverage threshold for component categories prone to defects being higher than that for component categories with low incidence of defects.

[0010] Preferably, the blind spot recognition module aggregates blind spot areas using the following method: The marked blind zone components are grouped by span number. Within the same span, it is determined whether two blind zone components meet the spatial adjacency condition. The spatial adjacency condition is that the distance between the center points of the two component surfaces in the bridge coordinate system is less than a preset adjacency distance threshold or the two component surfaces share a structural boundary. Blind zone components that meet the spatial adjacency condition are grouped into the same blind zone area, and the spatial range of the area and the list of components contained therein are marked.

[0011] Preferably, the supplementary inspection guidance generation module generates the supplementary inspection path sequence using the following method: For each blind spot area, the recommended shooting direction and shooting distance range are determined according to the component type of each component surface it contains; Starting from the current position of the inspection personnel, the component surface with the closest spatial distance in the blind area is selected as the first target for supplementary inspection. Subsequently, the closest adjacent uninspected component surface of the current component surface is selected as the next target for supplementary inspection. This process is repeated for all component surfaces in the blind area to form a supplementary inspection path sequence.

[0012] Preferably, after the inspection personnel complete the supplementary shooting according to the supplementary inspection guide, the supplementary inspection guide generation module inputs the newly acquired image into the image coverage mapping module and the blind spot recognition module to re-execute the coverage mapping and blind spot recognition, and updates the coverage rate and blind spot mark of each component surface; The blind spot identification process terminates when the coverage of all component surfaces reaches the corresponding coverage threshold, or when the remaining non-compliant component surfaces are marked as physically inaccessible areas.

[0013] Compared with the prior art, the beneficial effects of the present invention are as follows: The road and bridge surface defect rapid assessment system proposed in this invention extends the verification of inspection integrity from pre-inspection route planning to the coverage acceptance stage after inspection. It fills the gap in existing technologies where it is impossible to automatically identify which component surfaces were not effectively photographed after inspection, providing a quantifiable means of verifying the integrity of inspection results. By introducing a shooting quality assessment to distinguish between "captured" and "captured" images, it can be used for defect identification. By incorporating shooting distance, observation angle, and exposure sufficiency into the coverage effectiveness judgment, the coverage statistics are closer to actual detection capabilities, avoiding overestimation of inspection integrity due to simply counting whether an image covers a certain area. The system locates blind spots using components as basic units and outputs supplementary inspection guidance, including a list of components to be inspected, suggested shooting directions and distance ranges, and supplementary inspection route sequences. This allows for direct supplementary inspection actions after blind spots are discovered, reducing the workload of manual secondary planning of supplementary inspection routes. The component reference table is constructed based on bridge design drawings, without relying on complete BIM models or high-precision 3D point clouds, and can also be implemented for old bridges without established digital models, making it widely applicable. Attached Figure Description

[0014] Figure 1 This is a block diagram of the rapid assessment system for surface defects of roads and bridges according to the present invention. Detailed Implementation

[0015] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.

[0016] Reference Figure 1 As shown, the rapid assessment system for surface defects of roads and bridges based on deep learning includes: a component reference table construction module, an image overlay mapping module, a blind spot identification module, and a supplementary inspection guidance generation module.

[0017] The component reference table construction module is used to divide the bridge surface according to the visible surface as the basic unit based on the bridge design data, and generate a component reference table containing component category, span number, number, spatial coordinates and surface area.

[0018] The component reference table construction module constructs the component reference table by extracting the number of spans, the arrangement of components in each span, and the geometric dimensions of the components from the bridge design drawings. The bridge surface is then divided into levels according to the span number, component category, and visible surface direction, with each visible surface of a single component serving as the basic recording unit.

[0019] For each record unit, calculate its surface area based on its geometric dimensions, determine the three-dimensional coordinates of its center point based on its position in the overall bridge coordinate system, and combine the component category, span number, component number, center point coordinates, and surface area into a single component surface record. Compile all component surface records to construct a component reference table.

[0020] The component categories include the bottom surface of the bridge deck, the web of the T-beam, the side surface of the pier, the bottom and side surfaces of the cap beam, and the side surface of the retaining block. Each component surface is numbered by combining the span number, the component category identifier, and the sequential number within the same span and category. For example, the left web of the second T-beam in the third span is designated K03-TB02-L. The surface area is calculated based on the geometry of the corresponding component surface and the dimensions indicated on the design drawings. Planar components are calculated directly by multiplying their length and width, while curved components are calculated based on their unfolded dimensions.

[0021] The image coverage mapping module is electrically connected to the component reference table construction module. It is used to perform component semantic segmentation on the inspection image to identify the category of each pixel, combine the image positioning information with the component reference table to determine the specific component number, and determine the coverage validity according to the effective identification distance threshold, observation angle threshold and minimum contrast corresponding to the component category, and map the effective coverage to the component reference table.

[0022] In the image overlay mapping module, the bridge component semantic segmentation network adopts an encoder-decoder architecture. The encoder part consists of several convolutional blocks and downsampling layers alternately, extracting multi-scale features of the image step by step. The decoder part restores the spatial resolution through upsampling and skip connections, and outputs a pixel-by-pixel category label map of the same size as the input image. The label categories cover seven categories: bottom surface of bridge deck, web of T-beam, side surface of pier, bottom and side surface of cap beam, side surface of abutment block, and background.

[0023] The network's training data comes from manually labeled bridge inspection images. The labeling method involves drawing polygonal outlines of the visible areas of each component in each image and assigning corresponding category labels. During training, a weighted cross-entropy loss function is used to address the issue of uneven pixel counts between categories. Component categories with fewer pixels are assigned higher loss weights, with the weight value inversely proportional to the proportion of pixels in that category in the training set. The network's optimizer uses a stochastic gradient descent method with momentum, and the initial learning rate decays at fixed intervals.

[0024] In the image overlay mapping module, the method for performing component semantic segmentation and component number matching is as follows: The aforementioned semantic segmentation network for bridge components is used to classify the inspection images pixel by pixel, labeling the component category to which each pixel belongs.

[0025] The image is accompanied by GPS coordinates and IMU attitude data to determine the shooting location and camera orientation. Since GPS coordinate positioning accuracy is typically on the order of meters, while the distance between adjacent bridge components may be less than one meter, directly matching the spatial coordinates of components with GPS coordinates is prone to mismatches. Therefore, the following method is used for positioning: Based on GPS coordinates, all component surfaces whose center point three-dimensional coordinates fall within a set radius around the GPS coordinates are retrieved from the component reference table as candidate sets. The radius value is set according to the nominal positioning accuracy of the GPS device used, and is twice the nominal positioning accuracy. Then, using the camera orientation data from the IMU attitude data, the ray direction emitted from the shooting position along the camera orientation is calculated. Components whose center point is located within a fan-shaped region directly in front of this ray direction are retained. The angle of this fan-shaped region is set according to the nominal error of the IMU heading angle. Other candidate component surfaces are excluded. The intersection of the pixel-level category label and the set of remaining candidate component numbers after direction filtering is taken to determine the specific component number corresponding to each pixel region, completing the mapping from image to component reference table. When the intersection is empty, the image is marked as an unmatched image and is not included in subsequent coverage statistics.

[0026] In the image overlay mapping module, the method for determining the overlay validity based on the threshold corresponding to the component category is as follows: the shooting distance is calculated based on the imaging proportion of the participating components in the image, the shooting angle is determined based on the angle between the image plane normal vector and the component surface normal vector, and the exposure sufficiency is measured based on the local contrast of the pixel grayscale of the component area.

[0027] The method for calculating the shooting distance is as follows: Let the camera focal length be... The unit is millimeters, and the pixel size is... The unit is millimeters per pixel, and the pixel area of ​​the corresponding region of the component surface in the image is... The unit is pixels, and the actual surface area of ​​the component is... The unit is square meters.

[0028] First, calculate the equivalent side length (in pixels) of the component in the image. Then calculate the actual size corresponding to the equivalent side length. Then the shooting distance The unit is meters.

[0029] When there are multiple component regions in the image, the shooting distance of each region is calculated independently based on the pixel area of ​​each region and the corresponding component surface area.

[0030] The method for calculating the shooting angle is as follows: extract the camera's pitch and roll angles from the IMU attitude data, and combine this with the camera's optical orientation to determine the normal vector of the image plane. Obtain the geometric parameters and center point coordinates of the component surface from the component reference table, and determine its surface normal vector according to the component category. The normal vector of the bridge deck bottom surface is vertically upward, the normal vector of the T-beam web is horizontally pointing outward, the normal vector of the pier side is horizontally pointing outward, the normal vector of the cap beam bottom surface is vertically upward, and the normal vectors of the cap beam side and the block side are horizontally pointing outward; the formula for calculating the shooting angle is: ; In the formula, The angle of the shot is in degrees.

[0031] Extract the pixel region corresponding to the component category from the image. Divide this region into non-overlapping squares with a side length of 32 pixels. Calculate the standard deviation of the grayscale values ​​of the pixels within each square. Then, take the arithmetic mean of the standard deviations of all squares as the local contrast index for that component region. When the area of ​​the component region is less than a complete square, use the standard deviation of the grayscale values ​​of all pixels in that region as the local contrast.

[0032] When the shooting distance does not exceed the effective recognition distance threshold corresponding to the component category, the shooting angle does not exceed the effective observation angle threshold, and the local contrast is not lower than the minimum recognition contrast, the image is determined to be an effective cover of the component surface; otherwise, it is determined to be an invalid cover and is not included in the coverage area of ​​the component surface.

[0033] In the image overlay mapping module, the effective recognition distance threshold, effective observation angle threshold, and minimum recognition contrast are set as follows: the effective recognition distance threshold is determined based on the minimum defect recognition scale corresponding to the component category and the camera's spatial resolution parameters. Specifically, the minimum defect width to be recognized for the component category is set as follows: The unit is millimeters, and the camera pixel size is... The unit is millimeters per pixel, and the focal length is... If the unit is millimeters, then the formula for calculating the effective recognition distance threshold is: ; in, The minimum number of pixels required for a defect to be reliably identified is set to 3, meaning that a defect must occupy at least 3 pixels to be reliably identified. The minimum width of cracks to be identified in the web of the T-beam is 0.2 mm, based on the crack width limit specified in the bridge maintenance code. The minimum size of spalling to be identified on the side of the pier is 1 mm, on the bottom surface of the bridge deck is 0.2 mm, on the bottom surface of the cap beam is 0.3 mm, and on the side of the stop block is 0.5 mm.

[0034] The effective observation angle threshold is set based on the geometric features of the component surface. For planar components, the surface normal vector is single and the projection distortion of the image in that direction is small, so the threshold is set to 60 degrees. For curved components, the surface normal vector varies with position, and the projection distortion is significant when shooting at large angles, so the threshold is set to 45 degrees. The minimum recognition contrast is set based on the typical gray-level difference between the surface defects and the background of the component category. The specific method is as follows: collect no less than 50 inspection image samples of the component category under normal lighting conditions, extract the gray-level mean of the manually marked defect area and the adjacent background area in the samples, calculate the gray-level difference between the defect and the background in each sample, and take the 10th percentile of the gray-level difference of all samples as the minimum recognition contrast of the component category. This value ensures that the contrast between the defect and the background is higher than this threshold in more than 90% of the samples under normal lighting.

[0035] In the image coverage mapping module, the effective coverage area is calculated as follows: for each image determined to be effectively covered, the pixels within its component area are converted into the actual area of ​​the component surface based on the shooting distance and camera parameters. The conversion method is as follows: Let the actual area corresponding to a single pixel be The unit is square meters; among which The shooting distance for this component area. For pixel size, Focal length; The effective coverage area of ​​the image on the surface of this component is: ,in This represents the number of pixels in the image for the component area. The same component surface may be covered by multiple images, and the effective coverage areas of each image may overlap on the component surface. The cumulative effective coverage area of ​​the component surface is calculated using a pixel-level union method, that is, each actual area unit on the component surface is counted only once, and is not repeatedly accumulated.

[0036] The blind spot identification module is electrically connected to the image coverage mapping module. It calculates the coverage rate based on the ratio of the effective coverage area to the surface area of ​​each component surface, marks component surfaces with coverage below the corresponding coverage threshold as blind spot components, and aggregates adjacent blind spot components within the same span into a blind spot region. The blind spot identification module calculates the coverage rate and marks blind spots by: traversing the component reference table, dividing the recorded cumulative effective coverage area of ​​each component surface by its surface area to obtain the coverage rate of that component surface.

[0037] The coverage rate of each component surface is compared with the coverage threshold corresponding to that component category. Component surfaces with a coverage rate lower than the coverage threshold are marked as blind spot components. The coverage threshold is set separately for each component category, based on the defect inspection level requirements of that component category in the bridge maintenance code and the historical defect detection rate of that category. The web of T-beams and the bottom surface of bridge decks are key inspection components in the code and have a high historical defect detection rate, so the coverage threshold is set at 80%; the defect detection rate of bridge pier sides is moderate, so the coverage threshold is set at 70%; the defect detection rate of the bottom and sides of cap beams is moderate, so the coverage threshold is set at 75%; and the defect detection rate of the sides of abutments is low, so the coverage threshold is set at 60%.

[0038] The blind spot identification module aggregates blind spot regions as follows: Marked blind spot components are grouped by span number. Within the same span, it is determined whether two blind spot components meet the spatial adjacency condition. The spatial adjacency condition is that the distance between the center points of the two component surfaces in the bridge coordinate system is less than a preset adjacency distance threshold, or the two component surfaces share a structural boundary. The adjacency distance threshold is set based on the typical spacing of adjacent components within the same span in the bridge design drawings, taking the minimum value of the center point spacing of similar components within the same span. Blind spot components that meet the spatial adjacency condition are grouped into the same blind spot region. The spatial extent of this region is marked as the smallest bounding rectangle of the center point coordinates of the contained component surfaces, and a list of the contained component surfaces is provided.

[0039] The supplementary inspection guidance generation module is electrically connected to the blind spot identification module. It outputs a list of components to be inspected, suggested shooting directions and distance ranges, and a sequence of supplementary inspection paths sorted by spatial adjacency, based on the spatial location and component type of each blind spot area. The method for determining the suggested shooting directions and distance ranges in the supplementary inspection guidance generation module is as follows: for each component surface to be inspected, the suggested shooting direction is determined based on the surface normal vector of its component category. The shooting direction is to shoot directly at the component surface along the direction of the component surface normal vector. When the component surface is located under the bridge, the suggested direction is upward shooting; when it is located on the side of the bridge pier, the suggested direction is horizontal shooting. The lower limit of the suggested shooting distance range is 0.5 times the effective identification distance threshold for that component category to avoid shooting too close and resulting in a narrow field of view. The upper limit is the effective identification distance threshold for that component category to ensure that the shooting quality meets the effective coverage judgment conditions.

[0040] The supplementary inspection guidance generation module generates the supplementary inspection path sequence as follows: For each blind zone, the suggested shooting direction and shooting distance range are determined according to the component type of each component surface contained therein. Starting from the current position of the inspection personnel, the component surface with the closest spatial distance among the component surfaces contained in the blind zone is selected as the first supplementary inspection target. Subsequently, each time, the closest adjacent uninspected component surface of the current component surface is selected as the next supplementary inspection target, and so on, traversing all component surfaces in the blind zone to form a supplementary inspection path sequence.

[0041] After the inspection personnel complete the supplementary shooting according to the supplementary inspection guidelines, the supplementary inspection guidance generation module inputs the newly acquired images into the image coverage mapping module and the blind spot recognition module to re-execute the coverage mapping and blind spot recognition, updating the coverage rate and blind spot markings of each component surface. When the coverage rate of all component surfaces reaches the corresponding coverage threshold, or when the remaining non-compliant component surfaces are marked as physically inaccessible areas, the blind spot recognition process terminates.

[0042] The criteria for determining physically inaccessible areas are as follows: the shortest distance between the three-dimensional coordinates of the center point of the component surface in the bridge coordinate system and the bridge's accessible inspection passage is greater than twice the upper limit of the recommended shooting distance range for that component category, and the component surface does not share a structural boundary with any covered component surface, indicating that inspection personnel cannot reach the effective shooting range of the component surface from the existing passage. Accessible inspection passages include bridge deck pedestrian walkways, under-bridge inspection walkways, and pier maintenance platforms, the locations of which are extracted from bridge design drawings or as-built data. Component surfaces marked as physically inaccessible are no longer included in the supplementary inspection guidance generation in the blind spot identification process, but their blind spot status is recorded for subsequent manual evaluation.

[0043] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the present invention is defined by the appended technical solutions and their equivalents.

Claims

1. A rapid assessment system for surface defects of roads and bridges based on deep learning, characterized in that, include: The component reference table construction module is used to divide the bridge surface according to the visible surface as the basic unit based on the bridge design data, and generate a component reference table containing component category, span number, number, spatial coordinates and surface area; The image coverage mapping module is used to perform semantic segmentation of the components in the inspection image to identify the category of each pixel, combine the image positioning information with the component reference table to determine the specific component number, and determine the coverage validity according to the effective identification distance threshold, observation angle threshold and minimum contrast corresponding to the component category, and map the effective coverage to the component reference table. The blind spot identification module is used to calculate the coverage rate based on the ratio of the effective coverage area to the surface area of ​​each component surface, mark the component surface below the corresponding coverage threshold as a blind spot component, and aggregate adjacent blind spot components in the same span into a blind spot region. The supplementary inspection guidance generation module is used to output a list of supplementary inspection components, suggested shooting directions and distance ranges, and a supplementary inspection path sequence sorted by spatial adjacency, based on the spatial location and component type of each blind spot area.

2. The rapid assessment system for road and bridge surface defects based on deep learning according to claim 1, characterized in that, The component reference table construction module constructs the component reference table using the following method: Extract the number of spans, the arrangement of components in each span, and the geometric dimensions of the components from the bridge design drawings. Divide the bridge surface level by level according to the span number, component type, and visible surface direction, and use the single visible surface of a single component as the basic recording unit. For each recording unit, the surface area is calculated based on its geometric dimensions, and the three-dimensional coordinates of the center point are determined based on its position in the overall coordinate system of the bridge. The component category, span number, component number, center point coordinates and surface area are combined into a component surface record. Collect all component surface records and construct a component reference table.

3. The rapid assessment system for road and bridge surface defects based on deep learning according to claim 2, characterized in that, The component categories include the bottom surface of the bridge deck, the web of the T-beam, the side surface of the pier, the bottom and side surfaces of the cap beam, and the side surface of the stop block; The number of each component face is generated by combining the span number, component category identifier, and sequence number within the same span and category; The surface area is calculated based on the geometry of the corresponding component surface and the dimensional parameters marked on the design drawings.

4. The rapid assessment system for road and bridge surface defects based on deep learning according to claim 1, characterized in that, In the image overlay mapping module, the method for performing component semantic segmentation and component number matching is as follows: A semantic segmentation network for bridge components is used to classify inspection images pixel by pixel and label the component category to which each pixel belongs; Read the GPS coordinates and IMU attitude data attached to the image to determine the shooting location and camera orientation; Match the shooting location with the spatial coordinates of each component surface in the component reference table to lock the corresponding span number and candidate component number set of the image; By taking the intersection of pixel-level category labels and the candidate component number set, the specific component number corresponding to each pixel region is determined, thus completing the mapping from the image to the component reference table.

5. The rapid assessment system for road and bridge surface defects based on deep learning according to claim 1, characterized in that, In the image overlay mapping module, the method for determining overlay validity based on the threshold corresponding to the component category is as follows: The shooting distance is estimated based on the imaging proportion of the participating components in the image, the shooting angle is determined based on the angle between the image plane normal vector and the component surface normal vector, and the exposure sufficiency is measured based on the local contrast of the pixel grayscale of the component area. When the shooting distance does not exceed the effective recognition distance threshold corresponding to the component category, the shooting angle does not exceed the effective observation angle threshold, and the local contrast is not lower than the minimum recognition contrast, the image is determined to be an effective cover of the component surface; otherwise, it is determined to be an invalid cover and is not included in the coverage area of ​​the component surface.

6. The rapid assessment system for road and bridge surface defects based on deep learning according to claim 1, characterized in that, In the image overlay mapping module, the methods for setting the effective recognition distance threshold, the effective observation angle threshold, and the minimum recognition contrast are as follows: The effective identification distance threshold is determined based on the minimum defect identification scale corresponding to the component category and the spatial resolution parameters of the camera. The effective observation angle threshold is set based on the geometric features of the component surface; the threshold for planar components is greater than that for curved components. The minimum recognition contrast is set based on the typical grayscale difference between surface defects and the background for this component category.

7. The rapid assessment system for road and bridge surface defects based on deep learning according to claim 1, characterized in that, The blind spot identification module uses the following methods for coverage calculation and blind spot marking: Traverse the component reference table, and for each component surface, divide its recorded effective coverage area by the surface area of ​​that component surface to obtain the coverage rate of that component surface; The coverage rate of each component surface is compared with the coverage threshold corresponding to that component category. Component surfaces with a coverage rate lower than the coverage threshold are marked as blind zone components. The coverage threshold is set separately for each component category, with the coverage threshold for component categories prone to defects being higher than that for component categories with low incidence of defects.

8. The rapid assessment system for road and bridge surface defects based on deep learning according to claim 1, characterized in that, The blind spot identification module aggregates blind spot regions using the following method: The marked blind zone components are grouped by span number. Within the same span, it is determined whether two blind zone components meet the spatial adjacency condition. The spatial adjacency condition is that the distance between the center points of the two component surfaces in the bridge coordinate system is less than a preset adjacency distance threshold or the two component surfaces share a structural boundary. Blind zone components that meet the spatial adjacency condition are grouped into the same blind zone area, and the spatial range of the area and the list of components contained therein are marked.

9. The rapid assessment system for road and bridge surface defects based on deep learning according to claim 1, characterized in that, The supplementary inspection guidance generation module generates the supplementary inspection path sequence using the following method: For each blind spot area, the recommended shooting direction and shooting distance range are determined according to the component type of each component surface it contains; Starting from the current position of the inspection personnel, the component surface with the closest spatial distance in the blind area is selected as the first target for supplementary inspection. Subsequently, the closest adjacent uninspected component surface of the current component surface is selected as the next target for supplementary inspection. This process is repeated for all component surfaces in the blind area to form a supplementary inspection path sequence.

10. The rapid assessment system for road and bridge surface defects based on deep learning according to claim 1, characterized in that, After the inspection personnel complete the supplementary shooting according to the supplementary inspection guide, the supplementary inspection guide generation module inputs the newly acquired image into the image coverage mapping module and the blind spot recognition module to re-execute the coverage mapping and blind spot recognition, and update the coverage rate and blind spot mark of each component surface. The blind spot identification process terminates when the coverage of all component surfaces reaches the corresponding coverage threshold, or when the remaining non-compliant component surfaces are marked as physically inaccessible areas.