Bridge engineering drawing elevation intelligent examination method and system based on deep learning

Abstract

The application discloses a kind of bridge engineering drawing elevation intelligent examination method and system based on deep learning.The method comprises the following steps: obtaining bridge engineering drawing file and converting into standardized drawing image;Using YOLOv8 elevation related subgraph detection model to infer drawing image, identify all elevation related subgraph and output subgraph positioning information;Extract elevation related parameters;The elevation parameters extracted from each type of subgraph are uniformly mapped to the global coordinate system of drawing, the cross-subgraph parameter correlation is completed, and the structured dataset indexed by stake number is generated;Calculate the theoretical elevation value of each bridge structure control point;Compare the theoretical elevation value with the original elevation standard value of drawing, and determine the abnormality in combination with the preset deviation threshold, and output the examination result with parameter tracing.The application can automatically complete the subgraph positioning, parameter extraction, cross-subgraph correlation, theoretical elevation calculation and examination result output in bridge engineering drawing, and improve the efficiency of bridge engineering drawing elevation examination.

Description

Technical Field

[0001] This invention relates to the fields of intelligent review of engineering drawings, computer vision, deep learning, and digital bridge engineering technologies, specifically to a method and system for intelligent review of elevation in bridge engineering drawings based on deep learning. Background Technology

[0002] Bridge engineering drawings are crucial technical guidelines for construction, quality control, and project acceptance. Elevations characterize the height relationship of various control points on the bridge relative to a reference plane, directly affecting the bridge deck alignment, structural clearance, component installation positions, and the connection between the bridge and surrounding roads, terrain, and ancillary facilities. Omissions in design elevations, structural dimensions, or control point elevations in the drawings may lead to rework, increased costs, or even structural safety hazards during the construction phase.

[0003] Existing bridge engineering drawings are mostly circulated in PDF format, with elevation-related information typically scattered across multiple pages and sub-drawings. For example, bridge layout drawings include pier outlines and design elevation tables, while general pier construction drawings include elevation views, cross-sections, and elevation tables. The title sheet also records the drawing name, number, and project information. Reviewers need to repeatedly search between different pages, manually read dimensions, table values, and stationing information, and perform point-by-point verification based on professional formulas.

[0004] Traditional manual review methods have the following shortcomings: First, elevation information is scattered and presented in heterogeneous forms, requiring different interpretation methods for different sub-drawings, resulting in low efficiency for manual location and extraction; Second, parameters such as station number, design elevation, cross-sectional dimensions, pier length, and pile length are often located in different areas, making mismatches prone to occur when linking across areas and drawings; Third, the elevation calculation process involves multiple structural parameters and formulas, resulting in a large workload for manual calculation and verification, making it difficult to conduct a comprehensive review of drawings for large bridge projects; Fourth, the review results lack structured statistics and sufficient visualization, which is not conducive to archiving and subsequent quality analysis.

[0005] In existing technologies, some solutions attempt to automate the parsing of bridge design drawings using image processing, object detection, text recognition, or table recognition techniques. Chinese patent document CN116978052A discloses a method for identifying sub-graph layouts in bridge design drawings based on an improved YOLOv5. This method primarily identifies sub-graph areas or page layouts in bridge design drawings through object detection, and can be used for drawing area location and classification. This solution focuses on sub-graph layout identification and has not yet established a closed loop for elevation review, including dimensional parameter extraction, station-to-design elevation matching, formula library calculation, and anomaly output. Chinese patent document CN118587733A discloses a method for bridge structure identification and parameter extraction in bridge PDF design drawings, primarily targeting bridge structure identification and key point or structural parameter extraction. This solution focuses on structure identification and local parameter extraction, but does not integrate cross-sectional views, elevation views, title blocks, pier outlines, design elevation tables, and elevation tables into the elevation review process, nor does it create a theoretical elevation calculation dataset indexed by station number. Chinese patent document CN119964188A discloses a method and system for extracting key information from bridge design drawings based on multi-source fusion. It mainly combines object detection, semantic segmentation, OCR, and logical constraints to extract key information from bridge design drawings. This solution focuses on key information extraction but does not include a formula library, manual verification, or parameter tracing output mechanism for elevation review.

[0006] Existing technologies typically fail to address the following issues simultaneously: they do not uniformly integrate various elevation-related sub-drawings, such as cross-sectional views, elevation views, title blocks, pier outlines, design elevation tables, and elevation tables, into the review process; they do not configure differentiated extraction methods for dimension annotations, complex tables, circular stationing, and vertical design elevations; they do not map stationing, design elevation, structural dimensions, pier length, pile length, and cross slope parameters extracted from different sub-drawings to a unified global coordinate system and perform cross-regional matching; and they do not establish a configurable elevation calculation formula library. Therefore, it remains necessary to provide an intelligent elevation review method and system for bridge engineering drawings that can integrate drawing positioning, parameter extraction, cross-sub-drawing association, formula calculation, calculation comparison, and result display. Summary of the Invention

[0007] To address the shortcomings of existing technologies, the present invention aims to provide a method and system for intelligent elevation review of bridge engineering drawings based on deep learning. This method and system can perform standardized image conversion on PDF format bridge engineering drawings, automatically locate elevation-related sub-drawings in the drawings, extract the parameters required for elevation review according to the sub-drawing type, map the station numbers, design elevations, structural dimensions, and table parameters scattered in different sub-drawings to a unified coordinate system and associate them, and after manual verification, call the elevation calculation formula library to complete the theoretical elevation calculation and numerical consistency judgment.

[0008] To achieve the above objectives, the technical solution of the present invention is as follows:

[0009] In a first aspect of the present invention, a method for intelligent review of elevation in bridge engineering drawings based on deep learning is disclosed, the method comprising the following steps:

[0010] S1. Obtain bridge engineering drawing files and convert them into standardized drawing images;

[0011] S2. The trained YOLOv8 elevation-related sub-map detection model is used to infer the drawing image, identify all elevation-related sub-maps in the drawing image and output the sub-map location information;

[0012] S3. Based on the image type of different elevation-related sub-maps, use an adaptation extraction method to extract elevation-related parameters;

[0013] S4. Map the elevation parameters extracted from various sub-maps to the global coordinate system of the drawing, complete the cross-sub-map parameter association, and generate a structured dataset indexed by station number;

[0014] S5. Based on the structured dataset, call the preset elevation calculation formula to calculate the theoretical elevation value of each bridge structure control point;

[0015] S6. Compare the theoretical elevation value with the original elevation standard value on the drawing, determine the anomaly based on the preset deviation threshold, and output the review result with parameter traceability.

[0016] As a further improvement to the above technical solution, the bridge engineering drawing file mentioned in step S1 is a PDF format drawing; the standardized drawing image conversion includes format verification, PDF rasterization, resolution detection, image enhancement and format unification conversion of the PDF format drawing to obtain a standardized drawing image.

[0017] As a further improvement to the above technical solution, in step S2, the drawing image is input into the trained YOLOv8 elevation-related sub-image detection model to detect and locate the elevation-related sub-images in the drawing image, and output the category, confidence level and bounding box coordinates of each elevation-related sub-image; the elevation-related sub-images include cross-sectional views, elevation views, title blocks, pier outline views, design elevation tables and elevation tables.

[0018] The YOLOv8 elevation correlation sub-map detection model uses the YOLOv8 basic detection network for supervised training without changing the YOLOv8 basic network structure. The training process of the model includes: acquiring the original bridge engineering drawing files and converting the PDF format drawings into image files; manually annotating the six types of elevation correlation sub-maps in the sample to be annotated using the LabelImg tool, with the annotation format conforming to the YOLO object detection format, and the annotation content including the sub-map category and bounding box position; the YOLOv8 elevation correlation sub-map detection model uses bridge drawing images annotated with the six types of elevation correlation sub-maps as training samples, and the training samples are data augmented by at least one of random rotation, scale transformation, Gaussian noise, and brightness perturbation; the data-augmented samples are divided into training set and validation set; the preprocessed full-page bridge engineering drawing image is used as the model input, and the manually annotated sub-map categories and bounding boxes are used as supervision signals to supervise the training of the YOLOv8 basic detection network; after training, the model outputs the bounding box coordinates, category, and confidence score of each candidate elevation correlation sub-map.

[0019] As a further improvement to the above technical solution, step S3 specifically includes:

[0020] S31. Cut out each elevation-related sub-image from the original drawing image according to the coordinates of the detection box;

[0021] S32. Extract the drawing name and title of each sub-drawing; the drawing name and title are used to determine the drawing type, table or structural part to which the sub-drawing belongs, and serve as a semantic index for cross-drawing and cross-sub-drawing parameter association;

[0022] S33. Extract the corresponding parameters according to the sub-map category: extract the local structural dimensions from the elevation and cross-section views in the general construction drawing of the pier; extract the pier length, pile length, cross slope and elevation of each control point from the elevation table; extract the station number and its position coordinates from the pier outline in the bridge layout drawing; extract the design elevation value corresponding to each pile position from the design elevation table.

[0023] As a further improvement to the above technical solution, in step S33, the process of extracting local structural dimensions from the elevation and section views is as follows:

[0024] The dimension annotation area is detected and cropped. After image preprocessing, horizontal dimension lines and diagonal short dashes are removed, and pure numerical regions are filtered out. The connected component labeling algorithm is used to extract the target geometric attributes. The ratio of the width of the bounding rectangle of the connected component to the total width of the image is calculated. When the ratio is greater than a set empirical threshold, the corresponding connected component is identified as a dimension line and filtered out. Then, combined with the area characteristics of the connected component and the vertical distribution constraints, the remaining diagonal short dash fragments after filtering are cleaned up again. OCR technology is used to identify the dimension numbers. According to the positional relationship between the dimension numbers and the structural outline, center line or annotation line, the corresponding structural dimension parameters are determined. The cross-sectional view output parameters include: distance WA from the route center line to the design elevation, pavement thickness HA, superstructure height HB, distance HC from the bottom of the beam to the top of the cap beam, and cap beam height HD. The elevation view output parameters include: total length of the cap beam WB, distance WC from the route center line to the edge of the cap beam, distance WD from the edge of the cap beam to the inner pier, and pier spacing parameters 1WE and 2WF.

[0025] In step S33, the process of extracting pier length, pile length, cross slope, and elevation of each control point from the elevation table includes table structure detection and text recognition, specifically:

[0026] A table structure recognition model is used to locate cell positions and reconstruct the table's row and column structure. OCR technology is used to recognize the text content within each cell. The text recognition results are associated with the cells through a text spatial relationship-based matching algorithm, and two-dimensional table data is generated through row and column alignment and structured reconstruction. Specifically, the table structure recognition uses the Cycle-CenterNet architecture to identify cell boundaries, and the table text recognition uses PaddleOCR to identify numbers, text, and technical symbols in the table.

[0027] As a further improvement to the above technical solution, the matching algorithm based on text spatial relationships specifically includes:

[0028] After completing table structure detection and text recognition, the text area ratio (IoT) is used as the matching criterion between the text box and the cell box. Let the text box be Bt and the cell box be Bc. The text area ratio (IoT) is the intersection area of ​​the text box and the cell box divided by the area of ​​the text box itself, as shown in the following formula:

[0029]

[0030] When the text area ratio of a text box to a cell box exceeds a preset threshold, the text content is determined to belong to the corresponding cell. When multiple OCR text blocks match the same cell, the text fragments are concatenated into a complete string according to the natural reading order from left to right and from top to bottom.

[0031] After matching text with cells, the detected cells are divided into rows and columns based on the principle of spatial clustering. Cells with similar vertical positions are clustered into the same row based on the distribution of the Y coordinate of the top edge of the cell, and columns are clustered based on the distribution of the X coordinate of the left edge. If the boundary of a cell covers multiple consecutive rows or columns, it is determined to be a row-merged or column-merged cell.

[0032] In step S33, the specific process of extracting the station number and its location coordinates from the pier outline drawing is as follows:

[0033] The circular numbers are located in the detected pier outline area, and the task of extracting the pile number is divided into two steps: circular number detection and circular number recognition. First, the YOLOv8 object detection model is used to locate and detect the circular numbers, and then the image classification model is used to recognize the detected circular numbers to obtain the pile number of each pile and its relative position coordinates in the pier outline. The image classification model is the ResNet-18 image classifier.

[0034] In step S33, the specific process of extracting the design elevation values ​​corresponding to each pile position from the design elevation table is as follows: image preprocessing is performed on the extracted design elevation table area. The image preprocessing includes using the LANCZOS interpolation algorithm to enlarge the image by 4 times, sharpening the image to enhance the contrast between the text strokes and the background, and using adaptive binarization to convert the image into a binary image. For vertically arranged design elevation values, the image is rotated 90° clockwise to make the vertical text horizontal, and PaddleOCR is used for recognition. After recognition, the coordinates of the recognition results are restored to the original sub-image space, and coordinate normalization is performed to obtain the design elevation values ​​and their position coordinates in the design elevation table.

[0035] As a further improvement to the above technical solution, step S4 maps the parameters extracted from each elevation-related sub-map from the local coordinate system of the sub-map to the global coordinate system of the original drawing, and combines the drawing name, title, station number and spatial position relationship to complete the cross-sub-map parameter association, generating a structured elevation review dataset indexed by station number.

[0036] After generating the structured elevation review dataset in step S4 and before calling the elevation calculation formula in step S5, the system executes a human-computer interaction verification process: the system visualizes the structured elevation review dataset in an editable form; responding to external input correction instructions, the system uses the corrected parameters to cover the corresponding automatic identification parameters, field attribution, or station number and design elevation matching relationship, automatically generates a verified structured elevation review dataset, and sends it to step S5 for subsequent calculations.

[0037] Step S4, involving coordinate mapping and cross-submap parameter association, includes: after detecting and identifying the circular numbers in the pier outline map and extracting values ​​from the design elevation table, matching the station number and design elevation information based on their spatial relationship; assuming the coordinates of the upper left corner of the submap region in the original large map are (x1, y1), and the local coordinates of a target within the region are (x0, y0), the following formula is used to add the local coordinates to the coordinates of the upper left corner of the submap to obtain the global coordinates of the target in the original large map:

[0038]

[0039] After completing the coordinate transformation, compare the global X coordinates corresponding to the station number and design elevation, pair the station number with the design elevation that is closest to the horizontal position, and establish the correspondence between the station number and the design elevation of each station.

[0040] As a further improvement to the above technical solution, step S5 specifically includes:

[0041] Based on the verified structured elevation review dataset, the section type and pier type corresponding to each station number are determined, and the pre-configured bridge elevation calculation formula library is called to calculate the theoretical elevation value corresponding to each bridge structural control point.

[0042] In step S5, the system establishes an elevation parameter calculation formula library to calculate the elevation values ​​of each control point based on the verified structural dimension parameters; the system maintains the mapping relationship between section type and station number, as well as the pier type classification relationship; the mapping relationship between section type and station number is used to determine the section dimension parameters used for each pile position, and the pier type classification relationship is used to distinguish between continuous piers and transition piers to match the corresponding elevation table parameters; the parameters configured in the calculation formula library include cross slope P, design elevation O, middle pier length h2, and pile length L.

[0043] As a further improvement to the above technical solution, step S6 specifically includes:

[0044] The calculated theoretical elevation value is compared point by point with the corresponding design elevation value or elevation label value obtained from the drawing. If the theoretical elevation value and the corresponding value in the drawing meet the preset deviation conditions, the data point is judged to pass. If the preset deviation conditions are exceeded, the review is judged to fail. For data points that do not need to be checked and calculated, the result of "cannot be judged" is output. The comparison results corresponding to all station numbers are summarized, a review report is generated, and the total number of data points, the number of passed, the number of failed, and the number of "cannot be judged" are output.

[0045] In a second aspect of the present invention, a deep learning-based intelligent elevation review system for bridge engineering drawings is disclosed, comprising:

[0046] The drawing preprocessing module is used to acquire the PDF format bridge engineering drawing files to be reviewed, and to perform format verification, PDF rasterization, resolution detection, image enhancement and format unification conversion on the PDF drawings to obtain standardized drawing images.

[0047] The sub-image detection module is used to input the drawing image into the trained YOLOv8 elevation-related sub-image detection model, detect and locate the cross-sectional view, elevation view, title block, pier outline view, design elevation table and elevation table in the drawing, output the bounding box coordinates, sub-image category and confidence score, and crop the corresponding sub-image from the original drawing image according to the detection box coordinates.

[0048] The parameter extraction module is used to extract structural dimension parameters, pier length, pile length, cross slope, station number, elevation or design elevation data according to the category of elevation-related sub-maps.

[0049] The parameter association module is used to map the parameters extracted from each elevation-related sub-map to the global coordinate system of the original drawing, and combine the drawing name, title, station number and spatial position relationship to complete the cross-sub-map parameter association, and generate a structured elevation review dataset indexed by station number;

[0050] The manual verification module is used to visualize the structured elevation review dataset in an editable manner; in response to external input correction instructions, it uses the corrected parameters to cover the corresponding automatically identified parameters, field attribution, or station number and design elevation matching relationship to form a verified structured elevation review dataset.

[0051] The theoretical elevation calculation module is used to determine the section type and pier type corresponding to each station number based on the verified structured elevation review dataset, call the pre-configured bridge elevation calculation formula library, and calculate the theoretical elevation value corresponding to each bridge structural control point.

[0052] The verification output module is used to compare the calculated theoretical elevation value with the corresponding design elevation value or elevation label value in the drawing point by point, summarize the comparison results of all station numbers and generate a review report, outputting the total number of data, the number of passed, the number of failed, and the number of data that cannot be determined.

[0053] Compared with the prior art, the advantages of the present invention are:

[0054] (1) This invention unifies the cross-sectional views, elevation views, title blocks, pier outline views, design elevation tables and elevation tables in bridge engineering drawings as elevation-related sub-maps for automatic positioning. It can decompose the entire page of PDF drawings into semantically clear areas to be processed, reduce the workload of manual flipping, looking up and selecting areas, and improve the front-end positioning efficiency of elevation review.

[0055] (2) This invention employs differentiated parameter extraction methods for sub-maps related to different elevations: for elevation and section views, length ratio filtering and connected component analysis are used to remove interference from dimension lines and oblique short dashes; for elevation tables, table structure recognition, OCR, IoT matching, and row and column clustering are used to reconstruct structured data; for pier outline views, circular number detection and image classification models are used to identify station numbers; for design elevation tables, vertical text rotation recognition and inverse coordinate transformation are used to extract design elevations. This improves the reliability of information parsing from heterogeneous drawings.

[0056] (3) This invention transforms the station numbers, design elevations, structural dimensions, and table parameters in different sub-maps to the global coordinate system of the original drawings through unified coordinate mapping, and completes the matching of station numbers and design elevations based on horizontal distance thresholds, recognition confidence, and page title consistency. This method can reduce cross-regional and cross-sub-map information mismatches, and sort multiple candidate cases according to horizontal distance and recognition confidence, thereby reducing cross-regional mismatches.

[0057] (4) The present invention establishes a configurable bridge elevation calculation formula library, which can automatically select the calculation formula according to the bridge structure type, pier type, section type and control point type, calculate the theoretical elevation value of each control point, and use the consistency of the values ​​as the judgment standard to quantitatively compare the theoretical elevation value with the value marked on the drawing, thereby transforming manual experience review into a verifiable and structured quantitative review process.

[0058] (5) This invention supports external verification and parameter correction. When there are OCR errors, field attribution errors or station number matching errors in the automatic recognition results, external correction instructions can be input to correct key parameters and retain correction records, thereby improving the applicability of the system in engineering review scenarios.

[0059] (6) The review results output by this invention include theoretical elevation values, drawing annotation values, deviations, judgment results, total data, and number of inconsistent values, which can help reviewers quickly grasp the overall quality status of the drawing elevation data. Attached Figure Description

[0060] Figure 1 This is a flowchart of the intelligent elevation review method for bridge engineering drawings based on deep learning in this invention.

[0061] Figure 2 This is a schematic diagram of the six types of elevation-related sub-graphs in this invention;

[0062] Figure 3 This is a schematic diagram showing a cut-off of the dimension annotation area in this invention;

[0063] Figure 4 This is a flowchart of the dimension annotation information recognition process in this invention;

[0064] Figure 5 This is a schematic diagram illustrating the preprocessing effect of dimension lines and oblique short slashes in this invention;

[0065] Figure 6 This is a schematic diagram of the elevation table in this invention;

[0066] Figure 7 This is a flowchart of the high-level table recognition and structured reconstruction process in this invention;

[0067] Figure 8 This is a schematic diagram illustrating the cell boundary recognition effect in this invention;

[0068] Figure 9 This is a schematic diagram illustrating the text recognition effect of the high-level table in this invention;

[0069] Figure 10 This is a schematic diagram illustrating the OCR text and cell matching effect in this invention;

[0070] Figure 11 This is a schematic diagram illustrating the row and column alignment grouping effect of cells in this invention;

[0071] Figure 12 This is a flowchart of the key parameter extraction process for the bridge layout diagram in this invention;

[0072] Figure 13 This is a schematic diagram of the outline of the bridge pier in this invention;

[0073] Figure 14 This is a schematic diagram of the elevation table design in this invention;

[0074] Figure 15 This is a schematic diagram of the verification and parameter correction interface in this invention;

[0075] Figure 16 This is a schematic diagram of the interface of the high elevation parameter summary table in this invention;

[0076] Figure 17 This is a schematic diagram of the review results display interface in this invention. Detailed Implementation

[0077] To provide a better understanding of the structural features and effects achieved by the present invention, a detailed description is provided below, accompanied by preferred embodiments and accompanying drawings:

[0078] Example 1

[0079] like Figure 1 The method shown is an intelligent elevation review method for bridge engineering drawings based on deep learning. The method includes:

[0080] S1. Obtain the bridge engineering drawing file to be reviewed. The bridge engineering drawing file is in PDF format. Perform format verification, PDF rasterization, resolution detection, image enhancement, and format unification conversion on the PDF drawing to obtain a drawing image that can be parsed by subsequent algorithms.

[0081] S2. Input the preprocessed drawing image into the trained YOLOv8 elevation correlation sub-map detection model to detect and locate the elevation correlation sub-maps in the drawing. The elevation correlation sub-maps include cross-sectional views, elevation views, title blocks, pier outlines, design elevation tables, and elevation tables, such as... Figure 2 As shown.

[0082] In this embodiment, the YOLOv8 elevation correlation submap detection model uses the YOLOv8 basic detection network for supervised training without changing the YOLOv8 basic network structure. The model training process includes:

[0083] (1) Collect the original files of bridge engineering drawings and convert the PDF format drawings into image files to form original drawing image samples for target detection training.

[0084] (2) The LabelImg tool was used to manually annotate the six types of elevation-related sub-maps in the sample to be annotated. The annotation format followed the YOLO target detection format, and the annotation content included the sub-map category and the bounding box position. The six types of elevation-related sub-maps are cross-section view, elevation view, title block, pier outline view, design elevation table and elevation table.

[0085] (3) Data augmentation processing is performed on the annotated drawing image samples. The data augmentation processing includes adding Gaussian noise, random rotation and scaling transformation to improve data diversity and model generalization ability. In one embodiment, 2000 sample images are obtained after data augmentation.

[0086] (4) Dataset partitioning. The augmented samples are divided into a training set and a validation set; the training set is used to update the model parameters of the YOLOv8 base detection network, and the validation set is used to evaluate the model's detection performance on sub-image categories and bounding box locations during training.

[0087] (5) Using the preprocessed full-page bridge engineering drawing image as the model input, and manually annotated sub-image categories and bounding boxes as supervision signals, the YOLOv8 basic detection network was trained under supervision. The training parameters can adopt the conventional parameter range of the object detection model, the input image size is 640×640, the number of iterations is 100, the batch size is 16, and the initial learning rate lr0 is 0.001; YOLOv8n uses CIoU Loss based on IoU for bounding box regression training.

[0088] (6) Model inference. After training, the full-page image of the bridge engineering drawing to be reviewed is input into the YOLOv8 elevation-related sub-map detection model. The model outputs the bounding box coordinates, sub-map category and confidence of each candidate elevation-related sub-map. Then, the corresponding sub-map is cropped from the original drawing image according to the detection box coordinates for use in the subsequent parameter extraction steps.

[0089] In the general structural drawings of bridge piers, the elevation-related sub-drawings obtained by detection mainly include elevation views, cross-sectional views, and elevation tables; in the bridge layout drawings, the elevation-related sub-drawings obtained by detection mainly include bridge pier outline drawings and design elevation tables.

[0090] After the inspection is completed, the relevant sub-images for each elevation are cropped from the original drawing image according to the coordinates of the inspection frame. The elevation and section views in the general construction drawing of the pier are used to extract local structural dimensions, and the elevation table is used to extract the pier length, pile length, cross slope, and control point elevation; the pier outline view in the bridge layout drawing is used to extract the station number, and the design elevation table is used to extract the design elevation corresponding to each pile position.

[0091] S3. Crop each elevation-related sub-drawing according to the bounding box coordinates, and extract the drawing name and title of each sub-drawing. The drawing name and title are used to determine the drawing type, table type, or structural part to which the sub-drawing belongs, and serve as a semantic index for subsequent cross-drawing and cross-sub-drawing parameter association.

[0092] S31. For the detected elevation and section views, first detect and crop the dimension annotation area. The cropped dimension annotation image typically includes dimension numbers, horizontal dimension lines, and diagonal dashes, such as... Figure 3 As shown.

[0093] Because the dimension figures are close to the dimension lines and slanted dashes, direct OCR recognition is prone to interference. Therefore, the dimension annotation image is preprocessed first. The dimension annotation information recognition process is as follows: Figure 4 As shown. First, the dimension annotation image is extracted using the bounding box coordinates; second, image preprocessing removes horizontal dimension lines and diagonal dashes as interference, and filters out pure numerical regions; finally, OCR technology is used to identify the dimension numbers. During preprocessing, the geometric properties of dimension lines penetrating the cropping region horizontally are utilized, and a connected component labeling algorithm is used to extract the geometric attributes of the target, calculating the ratio of the bounding rectangle width to the total image width. When this ratio exceeds a set empirical threshold, the corresponding connected component is identified as a dimension line and filtered out; for the remaining diagonal dash fragments after filtering, secondary cleaning is performed using connected component area features and vertical distribution constraints. The preprocessing results are shown below. Figure 5 As shown.

[0094] After OCR recognition, parameters such as the distance from the route centerline to the design elevation, pavement thickness, superstructure height, distance from the bottom of the beam to the top of the cap beam, and cap beam height are obtained from the cross-sectional view; parameters such as the total length of the cap beam, the distance from the route centerline to the edge of the cap beam, the distance from the edge of the cap beam to the inner pier, and pier spacing are obtained from the elevation view. To facilitate subsequent formula calculations, the above parameters are defined as the following fields:

[0095] Table 1. Parameter Names and Field Correspondence Table

[0096] Distance from the centerline of the route to the design elevation WA Pier spacing 2 WF Total length of the cap beam WB Pavement thickness HA Distance from the center line of the route to the edge of the cap beam WC Superstructure high HB Distance from the edge of the cap beam to the inner pier WD Distance from bottom of beam to top of cap beam HC Pier spacing 1 WE Gai Liang Gao HD

[0097] S32. The elevation table is used to record the length of the pier, the length of the pile, the cross slope, and the elevation values ​​of each control point. An example is shown below. Figure 6 As shown. This type of table contains multi-level headers, a large number of columns, different types of numerical values, and professional abbreviations, requiring the table image to be converted into structured data. The elevation table recognition process is as follows: Figure 7 As shown in the diagram. First, a table structure recognition model is used to locate cell positions and reconstruct the table's row and column structure. Then, OCR technology is used to recognize the text content within each cell. Finally, a spatial relationship-based matching algorithm is used to associate the text recognition results with the cells, and two-dimensional table data is generated through row and column alignment and structured reconstruction. In the table structure recognition stage, the Cycle-CenterNet architecture is used to identify cell boundaries, and the recognition effect is shown in the diagram. Figure 8 As shown in the figure. In the table text recognition stage, PaddleOCR is used to recognize numbers, text, and technical symbols in the table. The recognition results are as follows: Figure 9 As shown. After completing table structure detection and text recognition, the text area ratio (IoT) is used as the matching criterion between text boxes and cell boxes. Let the text box be B. t The cell box is B c The calculation formula is as follows:

[0098]

[0099] When the area of ​​a text box relative to a cell exceeds a preset threshold, the text content is considered to belong to the corresponding cell. When multiple OCR text blocks are matched into the same cell, the text fragments are concatenated into a complete string according to the natural reading order from left to right and top to bottom. The text and cell matching effect is as follows. Figure 10 As shown.

[0100] After matching text with cells, the detected cells are divided into rows and columns based on spatial clustering principles. Cells with similar vertical positions are clustered into the same row based on the distribution of their top Y-coordinates, and columns are clustered based on the distribution of their left X-coordinates. When the boundary of a cell covers multiple consecutive rows or columns, it is determined to be a row-merged or column-merged cell. The row alignment grouping effect is as follows: Figure 11 As shown above, the fields such as pier number, H1 to H7, pier length h2, pile length L, and cross slope P are extracted from the elevation table.

[0101] S33. Extract the station numbers and their coordinates from the pier outline drawing. The bridge layout drawing includes a bridge graphic area and a parameter table area. The bridge graphic area contains station numbers labeled with circular numbers, and the parameter table area records the design elevation values ​​corresponding to each station location. The process for extracting key parameters from the bridge layout drawing is as follows: Figure 12 As shown. For the detected bridge pier outline, first locate the circular number in this area. The bridge outline is as follows. Figure 13 As shown.

[0102] Due to the small size, dense distribution, and potential tilt of the circular numbering system, this embodiment divides the station number extraction task into two steps: circular number detection and circular number recognition. First, a YOLOv8 object detection model is used to locate and detect the circular numbers. Then, an image classification model is used to recognize the detected circular numbers, ultimately obtaining the station number of each pile and its relative position coordinates in the pier outline map. In one embodiment, the circular number recognition uses a ResNet-18 image classifier, covering 31 common station number categories from 0 to 30. Experiments were conducted based on 8000 circular number samples obtained from 300 bridge layout maps, achieving a station number recognition accuracy of 98.19%.

[0103] S34. Extract design elevations from the design elevation table. Example of a design elevation table: Figure 14 As shown in the table, the "Design Elevation (m)" data rows are distributed in a strip pattern, the header text is arranged horizontally, and the key numerical data within the cells are distributed vertically at 90°. For the design elevation table, the extracted area is first pre-processed. Due to the small size of the original table area, a 4x magnification is used, and the LANCZOS interpolation algorithm is applied to enlarge the image. LANCZOS interpolation is based on the windowed sinc function, which can effectively suppress aliasing while preserving edge details. Its kernel function is defined as:

[0104] When the window size parameter of the interpolation kernel is set to a=4, it considers the surrounding 4×4 pixel area. This algorithm can maintain the sharpness of digital edges even at magnifications up to 4x.

[0105] After magnification, sharpening is applied to enhance the contrast between text strokes and the background. The sharpening kernel is defined as follows:

[0106]

[0107] Subsequently, adaptive binarization is used to convert the image into a binary image, where the neighborhood size of the Gaussian adaptive thresholding algorithm is 31 and the constant is 10.

[0108] For vertically arranged design elevation values, the image is rotated 90° clockwise to make the vertical text horizontal, and PaddleOCR is used for recognition. After recognition, the coordinates of the recognition result are restored to the original sub-image image space. Since the recognition is performed on an image magnified 4 times, let the recognition coordinates in the rotated magnified image be... The scaled coordinates are:

[0109]

[0110] If the recognition is performed on an image rotated 90° clockwise, let the height of the original image be H, and the coordinates of a point in the rotated image be... Then the corresponding coordinates of this point in the original image are:

[0111]

[0112] After rotation restoration, coordinate normalization is performed. Let the original image width be W and the height be H, then the normalized coordinates are:

[0113]

[0114] Through the above processing, the design elevation value and its position coordinates in the design elevation table are obtained.

[0115] S4. Coordinate Mapping and Cross-Submap Parameter Association. After completing the detection and identification of circular numbers in the pier outline map and the extraction of values ​​from the design elevation table, the two types of information are matched according to their spatial position to establish the association between each station number and its corresponding design elevation. Since the station number and the design elevation are located in different submap areas, it is necessary to unify the two types of identification results to the original drawing coordinate system through coordinate mapping, and then match them based on the principle of horizontal proximity.

[0116] Let the coordinates of the top-left corner of the region in the original large image be (x1, y1), and the local coordinates of a target within the region be (x0, y0). Then the global coordinates of the target in the original large image are:

[0117]

[0118] After coordinate transformation, the x-coordinate values ​​of the station and the design elevation are compared, and the station with the closest horizontal position is paired with the design elevation. In this embodiment, the matching result of the station and the design elevation is as follows:

[0119] Table 2. Results of matching station number with design elevation

[0120] 0 71.279 1 71.579 2 71.879 3 72.179 4 72.490 5 72.822 6 73.176 7 73.551 8 73.948 9 74.366 10 74.805

[0121] Through the above matching, the correspondence between the pile number and the design elevation is established, providing a data basis for the subsequent review of elevation parameters.

[0122] Before calling the elevation calculation formula library, the system displays the structured elevation review dataset obtained in step S4 to the reviewers in an editable manner. Reviewers can verify the automatic recognition and matching results. When OCR recognition errors, table field assignment errors, station number and design elevation matching errors, or section type or pier type judgment errors are found, the corresponding parameter values, field assignments, or matching relationships can be directly corrected in the verification interface. The system saves the correction results and overwrites the corresponding automatic recognition results with the corrected parameters, forming a verified structured elevation review dataset, such as... Figure 15 As shown.

[0123] S5. Formula Library Call and Theoretical Elevation Calculation. The system establishes a formula library for calculating elevation parameters, used to deduce the elevation values ​​of each control point based on the verified structural dimensions. The system also manages the mapping relationship between section type and station number, as well as the pier type classification relationship: the mapping relationship between section type and station number is used to determine the section dimension parameters referenced for each pile location; the pier type classification relationship is used to distinguish between continuous piers and transition piers, so as to reference the corresponding elevation table parameters.

[0124] The main parameters involved in the system are as follows:

[0125] Table 3. Parameter Names and Field Comparison Table

[0126] Distance from the centerline of the route to the design elevation WA Pier spacing 2 WF Total length of the cap beam WB Pavement thickness HA Distance from the center line of the route to the edge of the cap beam WC Superstructure high HB Distance from the edge of the cap beam to the inner pier WD Distance from bottom of beam to top of cap beam HC Pier spacing 1 WE Gai Liang Gao HD cross slope P Design elevation O Middle pier length h2 Pile length L

[0127] Taking this bridge drawing as an example, the calculation formulas for H1 to H7 are defined as follows:

[0128]

[0129]

[0130]

[0131]

[0132]

[0133]

[0134]

[0135] In this system, O and P are global design parameters extracted from the drawings; WA to WF and HA to HD are structural dimensions extracted from the cross-sectional and elevation views; h2 and L are pier parameters extracted from the elevation table; and H1 to H7 are the calculated elevation values ​​for each control point. The formula library supports configuration and expansion according to project type and design specifications, and administrators can add or modify calculation formulas based on specific bridge designs.

[0136] S6. Numerical Consistency Judgment and Review Result Output. The theoretical elevation values ​​of each control point calculated in step S5 are compared point-by-point with the corresponding design elevation values ​​on the drawings. If the calculated value is completely consistent with the design value, the data point is judged as "passed"; if any inconsistency exists, it is judged as "failed". Finally, the system summarizes the comparison results for all station numbers, generates a review report, and outputs the total number of data points, the number of passed values, and the number of failed values.

[0137] Taking pier No. 1 on the left span as an example, the system confirms that the span is left span, the station number is 1, the section type is I-I, and references the corresponding dimensional parameters; the system determines that its pier type is a continuous pier, and extracts the corresponding h2 and L values ​​from the elevation table; then it extracts the design value 71.579 from the design elevation table, substitutes it into the formula to calculate the theoretical values ​​of H1 to H7, and finally compares the calculated values ​​with the corresponding H1-H7 labeled values ​​in the elevation table on the drawing. If it is less than or equal to the preset deviation threshold, it is marked as "passed"; otherwise, it is marked as "failed".

[0138] System interface as follows Figures 16 to 17 As shown. Figure 16 It is an elevation summary table that links together the elevation parameters extracted from different sub-maps; Figure 17 The review results show that 60 elevation data points were checked, with 38 passing, 18 failing, and 4 unable to be determined. The failing data was found to be due to incorrect values ​​on the drawings, and the 4 unable-to-determine data points were because the data for abutments 0 and 29 on both the left and right spans did not require checking or calculation.

[0139] Example 2

[0140] This invention also discloses a deep learning-based intelligent elevation review system for bridge engineering drawings, comprising:

[0141] The drawing preprocessing module is used to acquire the PDF format bridge engineering drawing files to be reviewed, and to perform format verification, PDF rasterization, resolution detection, image enhancement and format unification conversion on the PDF drawings to obtain standardized drawing images.

[0142] The sub-image detection module is used to input the drawing image into the trained YOLOv8 elevation-related sub-image detection model, detect and locate the cross-sectional view, elevation view, title block, pier outline view, design elevation table and elevation table in the drawing, output the bounding box coordinates, sub-image category and confidence score, and crop the corresponding sub-image from the original drawing image according to the detection box coordinates.

[0143] The parameter extraction module is used to extract structural dimension parameters, pier length, pile length, cross slope, station number, elevation or design elevation data according to the category of elevation-related sub-maps.

[0144] The parameter association module is used to map the parameters extracted from each elevation-related sub-map to the global coordinate system of the original drawing, and combine the drawing name, title, station number and spatial position relationship to complete the cross-sub-map parameter association, and generate a structured elevation review dataset indexed by station number;

[0145] The manual verification module is used to visualize the structured elevation review dataset in an editable manner; in response to external input correction instructions, it uses the corrected parameters to cover the corresponding automatically identified parameters, field attribution, or station number and design elevation matching relationship to form a verified structured elevation review dataset.

[0146] The theoretical elevation calculation module is used to determine the section type and pier type corresponding to each station number based on the verified structured elevation review dataset, call the pre-configured bridge elevation calculation formula library, and calculate the theoretical elevation value corresponding to each bridge structural control point.

[0147] The verification output module is used to compare the calculated theoretical elevation value with the corresponding design elevation value or elevation label value in the drawing point by point, summarize the comparison results of all station numbers and generate a review report, outputting the total number of data, the number of passed, the number of failed, and the number of data that cannot be determined.

[0148] 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 appended claims and their equivalents is defined.

Claims

1. A method for intelligently reviewing the elevation of bridge engineering drawings based on deep learning, characterized in that, Includes the following steps: S1. Obtain bridge engineering drawing files and convert them into standardized drawing images; S2. The trained YOLOv8 elevation-related sub-map detection model is used to infer the drawing image, identify all elevation-related sub-maps in the drawing image and output the sub-map location information; S3. Based on the image type of different elevation-related sub-maps, use an adaptation extraction method to extract elevation-related parameters; S4. Map the elevation parameters extracted from various sub-maps to the global coordinate system of the drawing, complete the cross-sub-map parameter association, and generate a structured dataset indexed by station number; S5. Based on the structured dataset, call the preset elevation calculation formula to calculate the theoretical elevation value of each bridge structure control point; S6. Compare the theoretical elevation value with the original elevation standard value on the drawing, determine the anomaly based on the preset deviation threshold, and output the review result with parameter traceability.

2. The intelligent elevation review method for bridge engineering drawings based on deep learning according to claim 1, characterized in that, In step S1, the bridge engineering drawing file is a PDF format drawing; the standardized drawing image conversion includes format verification, PDF rasterization, resolution detection, image enhancement and format unification conversion of the PDF format drawing to obtain a standardized drawing image.

3. The intelligent elevation review method for bridge engineering drawings based on deep learning according to claim 1, characterized in that, In step S2, the drawing image is input into the trained YOLOv8 elevation correlation sub-map detection model to detect and locate the elevation correlation sub-maps in the drawing image, and output the category, confidence level and bounding box coordinates of each elevation correlation sub-map; the elevation correlation sub-maps include cross-sectional views, elevation views, title blocks, pier outline views, design elevation tables and elevation tables; The YOLOv8 elevation correlation subgraph detection model is trained under supervision using the YOLOv8 basic detection network without changing the YOLOv8 basic network structure. The training process of this model includes: acquiring the original files of bridge engineering drawings and converting the PDF format drawings into image files; manually annotating the six types of elevation-related sub-maps in the sample to be annotated using the LabelImg tool, with the annotation format conforming to the YOLO object detection format, and the annotation content including the sub-map category and bounding box position; the YOLOv8 elevation-related sub-map detection model uses bridge drawing images annotated with the six types of elevation-related sub-maps as training samples, and the training samples undergo data augmentation processing through at least one of random rotation, scale transformation, Gaussian noise, and brightness perturbation; the data-augmented samples are divided into training set and validation set; the preprocessed full-page bridge engineering drawing image is used as the model input, and the manually annotated sub-map categories and bounding boxes are used as supervision signals to supervise the training of the YOLOv8 basic detection network; After training, the model outputs the bounding box coordinates, category, and confidence score of each candidate elevation-related sub-map.

4. The intelligent elevation review method for bridge engineering drawings based on deep learning according to claim 1, characterized in that, Step S3 specifically includes: S31. Cut out each elevation-related sub-image from the original drawing image according to the coordinates of the detection box; S32. Extract the drawing name and title of each sub-drawing; the drawing name and title are used to determine the drawing type, table or structural part to which the sub-drawing belongs, and serve as a semantic index for cross-drawing and cross-sub-drawing parameter association; S33. Extract the corresponding parameters according to the sub-map category: extract the local structural dimensions from the elevation and cross-section views in the general construction drawing of the pier; extract the pier length, pile length, cross slope and elevation of each control point from the elevation table; extract the station number and its position coordinates from the pier outline in the bridge layout drawing; extract the design elevation value corresponding to each pile position from the design elevation table.

5. The intelligent elevation review method for bridge engineering drawings based on deep learning according to claim 4, characterized in that, In step S33, the process of extracting local structural dimensions from the elevation and section views is as follows: The dimension annotation area is detected and cropped. Image preprocessing removes horizontal dimension lines and diagonal short dashes, and filters out pure numerical regions. A connected component labeling algorithm is used to extract the target geometric attributes. The ratio of the width of the bounding rectangle of the connected component to the total width of the image is calculated. When this ratio is greater than a set empirical threshold, the corresponding connected component is identified as a dimension line and filtered out. Then, combined with the area features of the connected component and the vertical distribution constraints, the remaining diagonal short dash fragments after filtering are cleaned up again. OCR technology is used to identify the dimension numbers. Based on the positional relationship between the dimension numbers and the structural outline, center line, or annotation line, the corresponding structural dimension parameters are determined. The cross-sectional view output parameters include: distance WA from the route centerline to the design elevation, pavement thickness HA, superstructure height HB, distance HC from the bottom of the beam to the top of the cap beam, and cap beam height HD; the elevation view output parameters include: total length of the cap beam WB, distance WC from the route centerline to the edge of the cap beam, distance WD from the edge of the cap beam to the inner pier, and pier spacing parameters 1WE and 2WF. In step S33, the process of extracting pier length, pile length, cross slope, and elevation of each control point from the elevation table includes table structure detection and text recognition, specifically: A table structure recognition model is used to locate cell positions and reconstruct the table's row and column structure. OCR technology is used to recognize the text content within each cell. The text recognition results are associated with the cells through a text spatial relationship-based matching algorithm, and two-dimensional table data is generated through row and column alignment and structured reconstruction. Specifically, the table structure recognition uses the Cycle-CenterNet architecture to identify cell boundaries, and the table text recognition uses PaddleOCR to identify numbers, text, and technical symbols in the table.

6. The intelligent elevation review method for bridge engineering drawings based on deep learning according to claim 5, characterized in that, The matching algorithm based on text spatial relationships specifically includes: After completing table structure detection and text recognition, the text area ratio (IoT) is used as the matching criterion between the text box and the cell box. Let the text box be Bt and the cell box be Bc. The text area ratio (IoT) is the intersection area of ​​the text box and the cell box divided by the area of ​​the text box itself, as shown in the following formula: ； When the text area ratio of a text box to a cell box exceeds a preset threshold, the text content is determined to belong to the corresponding cell; when multiple OCR text blocks match the same cell, the text fragments are concatenated into a complete string according to the natural reading order from left to right and from top to bottom. After completing the text and cell matching, the detected cells are divided into rows and columns based on the principle of spatial clustering. Cells with similar vertical positions are clustered into the same row based on the distribution of the Y coordinate of the top edge of the cell, and columns are clustered based on the distribution of the X coordinate of the left edge. If the boundary of a cell covers multiple consecutive rows or columns, it is determined to be a row-merged or column-merged cell. In step S33, the specific process of extracting the station number and its location coordinates from the pier outline drawing is as follows: The circular numbers are located in the detected pier outline area, and the task of extracting the pile number is divided into two steps: circular number detection and circular number recognition. First, the YOLOv8 object detection model is used to locate and detect the circular numbers, and then the image classification model is used to recognize the detected circular numbers to obtain the pile number of each pile and its relative position coordinates in the pier outline. The image classification model is the ResNet-18 image classifier. In step S33, the specific process of extracting the design elevation values ​​corresponding to each pile position from the design elevation table is as follows: image preprocessing is performed on the extracted design elevation table area. The image preprocessing includes using the LANCZOS interpolation algorithm to enlarge the image by 4 times, sharpening the image to enhance the contrast between the text strokes and the background, and using adaptive binarization to convert the image into a binary image. For vertically arranged design elevation values, the image is rotated 90° clockwise to make the vertical text horizontal, and PaddleOCR is used for recognition. After recognition, the coordinates of the recognition results are restored to the original sub-image space, and coordinate normalization is performed to obtain the design elevation values ​​and their position coordinates in the design elevation table.

7. The intelligent elevation review method for bridge engineering drawings based on deep learning according to claim 4, characterized in that, Step S4 maps the parameters extracted from each elevation-related sub-map from the sub-map's local coordinate system to the original drawing's global coordinate system. It then combines the drawing name, title, station number, and spatial relationship to complete the cross-sub-map parameter association, generating a structured elevation review dataset indexed by station number. After generating the structured elevation review dataset in step S4 and before calling the elevation calculation formula in step S5, the system executes a human-computer interaction verification process: the system visualizes the structured elevation review dataset in an editable form; responding to external input correction instructions, the system uses the corrected parameters to cover the corresponding automatic identification parameters, field attribution, or station number and design elevation matching relationship, automatically generates a verified structured elevation review dataset and sends it to step S5 for subsequent calculations; Step S4, involving coordinate mapping and cross-submap parameter association, includes: after detecting and identifying the circular numbers in the pier outline map and extracting values ​​from the design elevation table, matching the station number and design elevation information based on their spatial relationship; assuming the coordinates of the upper left corner of the submap region in the original large map are (x1, y1), and the local coordinates of a target within the region are (x0, y0), the following formula is used to add the local coordinates to the coordinates of the upper left corner of the submap to obtain the global coordinates of the target in the original large map: ； After completing the coordinate transformation, compare the global X coordinates corresponding to the station number and design elevation, pair the station number with the design elevation that is closest to the horizontal position, and establish the correspondence between the station number and the design elevation of each station.

8. The intelligent elevation review method for bridge engineering drawings based on deep learning according to claim 7, characterized in that, Step S5 specifically includes: Based on the verified structured elevation review dataset, the section type and pier type corresponding to each station number are determined, and the pre-configured bridge elevation calculation formula library is called to calculate the theoretical elevation value corresponding to each bridge structural control point. In step S5, the system establishes an elevation parameter calculation formula library to calculate the elevation values ​​of each control point based on the verified structural dimension parameters; the system maintains the mapping relationship between section type and station number, as well as the pier type classification relationship; the mapping relationship between section type and station number is used to determine the section dimension parameters used for each pile position, and the pier type classification relationship is used to distinguish between continuous piers and transition piers to match the corresponding elevation table parameters; the parameters configured in the calculation formula library include cross slope P, design elevation O, middle pier length h2, and pile length L.

9. The intelligent elevation review method for bridge engineering drawings based on deep learning according to claim 1, characterized in that, Step S6 specifically includes: The calculated theoretical elevation value is compared point by point with the corresponding design elevation value or elevation label value obtained from the drawing. If the theoretical elevation value and the corresponding value in the drawing meet the preset deviation conditions, the data point is judged to pass. If the preset deviation conditions are exceeded, the review is judged to fail. For data points that do not need to be checked and calculated, the result of "cannot be judged" is output. The comparison results corresponding to all station numbers are summarized, a review report is generated, and the total number of data points, the number of passed, the number of failed, and the number of "cannot be judged" are output.

10. A deep learning-based intelligent elevation review system for bridge engineering drawings, characterized in that, include: The drawing preprocessing module is used to acquire the PDF format bridge engineering drawing files to be reviewed, and to perform format verification, PDF rasterization, resolution detection, image enhancement and format unification conversion on the PDF drawings to obtain standardized drawing images; The sub-image detection module is used to input the drawing image into the trained YOLOv8 elevation-related sub-image detection model, detect and locate the cross-sectional view, elevation view, title block, pier outline view, design elevation table and elevation table in the drawing, output the bounding box coordinates, sub-image category and confidence score, and crop the corresponding sub-image from the original drawing image according to the detection box coordinates. The parameter extraction module is used to extract structural dimension parameters, pier length, pile length, cross slope, station number, elevation or design elevation data according to the category of elevation-related sub-maps. The parameter association module is used to map the parameters extracted from each elevation-related sub-map to the global coordinate system of the original drawing, and combine the drawing name, title, station number and spatial position relationship to complete the cross-sub-map parameter association, and generate a structured elevation review dataset indexed by station number; The manual verification module is used to visualize the structured elevation review dataset in an editable manner; In response to external correction commands, the corrected parameters are used to cover the corresponding automatically identified parameters, field attribution, or station number and design elevation matching relationship, forming a verified structured elevation review dataset; The theoretical elevation calculation module is used to determine the section type and pier type corresponding to each station number based on the verified structured elevation review dataset, call the pre-configured bridge elevation calculation formula library, and calculate the theoretical elevation value corresponding to each bridge structural control point. The verification output module is used to compare the calculated theoretical elevation value with the corresponding design elevation value or elevation label value in the drawing point by point, summarize the comparison results of all station numbers and generate a review report, outputting the total number of data, the number of passed, the number of failed, and the number of data that cannot be determined.

Images