Bridge structure identification and parameter extraction method for bridge PDF design drawing

Abstract

The invention relates to a bridge structure recognition and parameter extraction method for a bridge PDF design drawing. The method comprises the steps that the bridge design drawing stored in a PDF format is acquired and preprocessed; constructing a bridge structure detection model for training; inputting the PDF design drawing of the bridge to be detected into the PaddleOCR model for parameter identification; and inputting the preprocessed PDF design drawing of the to-be-detected bridge into the trained bridge structure detection model, and identifying the bridge structure to obtain the bridge structures of different views. According to the method, the bridge structure in the drawing is detected by combining the improved YOLOv10 model and the K-Net semantic segmentation algorithm, so that the contour of the bridge structure is drawn; carrying out text detection on a bridge design drawing by utilizing a PaddleOCR model, and extracting parameters of a bridge structure; through a series of logic constraints, the structure is matched with the parameters, finally structure identification and parameter extraction of the bridge in the bridge design drawing are realized, and different elements in the drawing can be effectively detected.

Description

Technical Field

[0001] The invention relates to the technical field of PDF image recognition, in particular to a bridge structure recognition and parameter extraction method for bridge PDF design drawings. Background Art

[0002] With the rapid development of transportation design and construction, bridge design has become the most technically difficult and safety-conscious link in the entire transportation engineering design. In the bridge construction drawing design stage, designers usually refer to the design drawings of existing bridge projects, learn from the parameter information therein, reduce the trial calculation process, and pass on design experience. Traditional methods usually retrieve information by manually reviewing drawings and documents, but these drawings and documents are very complex and prone to errors. In the field of architecture, especially in the recognition of bridge drawings, the current research status at home and abroad shows the dual characteristics of interdisciplinary integration and technological innovation. The rapid development of computer vision and artificial intelligence provides strong technical support for the automatic recognition of bridge drawings. The geometric and semantic information of bridge structures is crucial for the reconstruction of building information models (BIM). However, many bridge design drawings stored in PDF format lack BIM, and manual reconstruction of BIM is very cumbersome and time-consuming. Therefore, in order to more effectively manage the structured data in architectural drawings, an efficient method is needed to extract structures and parameters from bridge design drawings.

[0003] In practical applications, bridge design drawings in PDF format usually exist in the form of images or vector graphics, but these drawings often lack key semantic information, making it challenging for computers to understand the exact meaning of the drawing content. First, how to accurately extract key parameters in bridge drawings and perform effective rule matching is a technical challenge. Although OCR technology can recognize text and position information in drawings, this information often lacks deep semantic meaning, which makes it difficult to directly apply the identified parameters and establish connections between drawings in different views. Secondly, multi-view parameter matching is a key step in the interpretation of bridge architectural drawings. Bridge drawings usually include plan views, elevation views, and sections, each of which shows the structural details of the bridge from different angles. These views together constitute a complete image of the bridge. For example, the plan view provides a horizontal view of the bridge, while the elevation view and section view provide vertical and cross-sectional views, respectively. Although these views are projected from different angles of the same entity, the conversion and association between them are not always intuitive. For a complete bridge structure, its parameters are scattered in multiple views. Parameters obtained from only one view cannot support the information reconstruction of the entire bridge structure. Therefore, different views must be associated according to the characteristics of the drawings to obtain complete bridge structure parameters. Summary of the invention

[0004] In order to solve the defects that target detection technology is difficult to capture the detailed structure of bridge structure drawings and difficult to extract parameters, the purpose of the present invention is to provide a bridge structure recognition and parameter extraction method for bridge PDF design drawings that can realize bridge structure recognition and parameter extraction in bridge design drawings and effectively detect different elements in the drawings.

[0005] To achieve the above object, the present invention adopts the following technical solution: a bridge structure identification and parameter extraction method for bridge PDF design drawings, the method comprising the following steps in order:

[0006] (1) Obtain bridge design drawings saved in PDF format, preprocess them, and obtain preprocessed bridge design drawings to form a data set;

[0007] (2) Combine the improved YOLOv10 model and contour extraction module to build a bridge structure detection model;

[0008] (3) Using the data set to train the bridge structure detection model to obtain a trained bridge structure detection model;

[0009] (4) Obtain the PDF design drawing of the bridge to be inspected, input the PDF design drawing of the bridge to be inspected into the PaddleOCR model for parameter recognition, and obtain the necessary parameters and their coordinate information;

[0010] (5) Preprocess the PDF design drawings of the bridge to be inspected, input the preprocessed PDF design drawings of the bridge to be inspected into the trained bridge structure detection model, identify the bridge structure, and obtain the bridge structure in different views.

[0011] In step (1), the preprocessing refers to grayscale processing to obtain a binary image, and then the binary image is labeled and data expanded. The expanded bridge design drawings constitute a data set.

[0012] Step (2) specifically refers to: improving the YOLOv10 model, wherein the YOLOv10 model includes a backbone network, a neck network, and a head network, adding a CGA attention mechanism module to a layer behind the C2f module in the backbone network, and adding the CGA attention mechanism module to a layer behind the C2f module in the neck network, to obtain an improved YOLOv10 model;

[0013] The output of the improved YOLOv10 model is used as the input of the contour extraction module, and the output of the contour extraction module is used as the output of the bridge structure detection model. The improved YOLOv10 model and the contour extraction module together constitute the bridge structure detection model;

[0014] The contour extraction module uses the K-Net semantic segmentation algorithm to obtain the foreground mask of the structure to be detected, and then uses the edge pixel detection method and polygon fitting in the image processing method to finally obtain the specific contour of the structure to be detected;

[0015] The edge pixel detection method is used to detect the contour of the foreground in the mask. The edge pixel detection method uses the Canny detection operator. The Canny detection includes four steps: noise elimination, calculation of gradient amplitude and direction, non-maximum suppression and hysteresis threshold. The two-dimensional Gaussian formula for noise elimination is:

[0016]

[0017] Where G(x,y) represents the value of the Gaussian function at the point (x,y), that is, the probability density at the point (x,y) in two-dimensional space; x and y are coordinate variables in two-dimensional space; σ is the standard deviation.

[0018] Step (3) specifically includes the following steps in order:

[0019] (3a) Set the training parameters, i.e. the training parameters of the improved YOLOv10 model, including the number of iterations, batch size, and initial learning rate. The number of iterations epoch is 300, the batch size batch-size is 8, and the initial learning rate lr0 of the model is set to 0.01;

[0020] (3b) Input the dataset into the improved YOLOv10 model;

[0021] The data set is sent to the backbone network, which extracts and outputs the feature information of the training image;

[0022] Input feature information into the neck network, perform multi-scale feature fusion, enhance feature representation capability, and output fused feature information;

[0023] Input the fused feature information into the head network and output the key information of the target;

[0024] (3c) The image output by the improved YOLOv10 model is sent to the contour extraction module, and the cropped bridge structure image is used to identify the structural foreground, extract the mask and fit the polygon. The obtained polygon is the bridge structure contour map.

[0025] Step (4) specifically includes the following steps in order:

[0026] (4a) Input the PDF design drawing of the bridge to be inspected into the PaddleOCR model, set the direction of manual reading of the drawing as the standard direction, extract the text in the standard direction normally, and for the text perpendicular to the standard direction, first rotate the image 90° and then identify the vertical text, then perform a one-step coordinate conversion, normalize and save the identified text, and output the detected text information and position information;

[0027] (4b) Setting logical constraints: Using the text information detected by the PaddleOCR model, combined with the annotation area identified by the improved YOLOv10 model and the structure identified by the bridge structure detection model, the text and coordinate information of the necessary parameters are jointly screened out;

[0028] At this time, the required parameters obtained are not assigned attributes. The required parameters are then matched with the obtained bridge structure. Specifically, the Euclidean distance between the coordinates of the required parameters and the coordinates of the midpoint of the bridge edge is calculated. When the distance is the smallest, the bridge edge and the required parameters are matched together. At this time, the parameter and coordinate matching of a view is completed. After that, based on the inherent relationship between different views, multiple parameters between multiple views are summarized into an Excel table to achieve the final parameter extraction.

[0029] It can be seen from the above technical scheme that the beneficial effects of the present invention are: first, compared with the prior art, the present invention realizes the detection of bridge structures in drawings by combining the improved YOLOv10 model and the K-Net semantic segmentation algorithm, thereby drawing the outline of the bridge structure; uses the PaddleOCR model to perform text detection on bridge design drawings and extracts the parameters of the bridge structure; matches the structure with the parameters through a series of logical constraints, and finally realizes the structural recognition and parameter extraction of the bridge in the bridge design drawings; second, the present invention adds the CGA attention mechanism module to the YOLOv10 model, combines it with the K-net semantic segmentation algorithm, and solves the problem of inaccurate outline detection in bridge design drawings; at the same time, in view of the problem that the text information detected by the PaddleOCR model lacks semantic information, semantic information is given to the text and parameters in the drawings; third, experimental results show that the recall rate of the bridge structure detection model can reach 96.1%, and the average precision reaches 96.2%, which are 14.6% and 13.1% higher than the original YOLOv10 model respectively, and can effectively detect different elements in the drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 is a method sequence diagram of the present invention;

[0031] Figure 2 It is a network structure diagram of the improved YOLOv10 model in the present invention;

[0032] Figure 3This is a network diagram of the CGA attention mechanism involved in the present invention;

[0033] Figure 4 It is a schematic diagram of the cross section of a cast-in-place box girder;

[0034] Figure 5 This is a schematic diagram of the elevation of the cast-in-place box girder;

[0035] Figure 6 This is a schematic diagram of the top surface of the cast-in-place box girder;

[0036] Figure 7 This is a schematic diagram of the bottom surface of the cast-in-place box girder;

[0037] Figure 8 PDF design drawings of the bridge to be inspected;

[0038] Fig. 9 Schematic diagram of semantic segmentation results;

[0039] Fig.10 is the contour fitting result;

[0040] Fig.11 This is a diagram of parameter extraction results involved in the present invention. DETAILED DESCRIPTION

[0041] like Figure 1 As shown, a bridge structure identification and parameter extraction method for bridge PDF design drawings includes the following steps in order:

[0042] (1) Obtain bridge design drawings saved in PDF format, preprocess them, and obtain preprocessed bridge design drawings to form a data set;

[0043] (2) Combine the improved YOLOv10 model and contour extraction module to build a bridge structure detection model;

[0044] (3) Using the data set to train the bridge structure detection model to obtain a trained bridge structure detection model;

[0045] (4) Obtain the PDF design drawings of the bridge to be inspected, such as Figure 8 As shown in the figure, the PDF design drawing of the bridge to be inspected is input into the PaddleOCR model for parameter recognition to obtain the necessary parameters and their coordinate information;

[0046] (5) Preprocess the PDF design drawings of the bridge to be inspected, input the preprocessed PDF design drawings of the bridge to be inspected into the trained bridge structure detection model, identify the bridge structure, and obtain the bridge structure in different views.

[0047] In step (1), the preprocessing refers to grayscale processing to obtain a binary image, and then the binary image is labeled and data expanded. The expanded bridge design drawings constitute a data set.

[0048] Step (2) specifically refers to: Figure 2 As shown, the YOLOv10 model is improved, and the YOLOv10 model includes a backbone network, a neck network and a head network, such as Figure 3 As shown in the figure, adding the CGA attention mechanism module to the back layer of the C2f module in the backbone network and adding the CGA attention mechanism module to the back layer of the C2f module in the neck network can improve the model's detection ability for bridge structures; an improved YOLOv10 model is obtained;

[0049] The output of the improved YOLOv10 model is used as the input of the contour extraction module, and the output of the contour extraction module is used as the output of the bridge structure detection model. The improved YOLOv10 model and the contour extraction module together constitute the bridge structure detection model;

[0050] The contour extraction module uses the K-Net semantic segmentation algorithm to obtain the semantic segmentation result, i.e., the foreground mask of the structure to be detected, such as Fig. 9 As shown, the black part is the background and the red part is the foreground.

[0051] Then, the edge pixel detection method and polygon fitting in the image processing method are used to finally obtain the specific outline of the structure to be detected;

[0052] The edge pixel detection method is used to detect the contour of the foreground in the mask. The edge pixel detection method uses the Canny detection operator. The Canny detection operator is an edge detection operator that integrates filtering, enhancement, detection and other stages. Its goal is to find an optimal edge contour. The Canny detection operator is used to fuse their detection results to improve the accuracy of the detection results. The Canny detection includes four steps: eliminating noise, calculating gradient amplitude and direction, non-maximum suppression and hysteresis threshold. The two-dimensional Gaussian formula for eliminating noise is:

[0053]

[0054] Where G(x,y) represents the value of the Gaussian function at the point (x,y), that is, the probability density at the point (x,y) in two-dimensional space; x and y are coordinate variables in two-dimensional space; σ is the standard deviation.

[0055] Step (3) specifically includes the following steps in order:

[0056] (3a) Set the training parameters, i.e. the training parameters of the improved YOLOv10 model, including the number of iterations, batch size, and initial learning rate. The number of iterations epoch is 300, the batch size batch-size is 8, and the initial learning rate lr0 of the model is set to 0.01;

[0057] (3b) Input the dataset into the improved YOLOv10 model;

[0058] The data set is sent to the backbone network, which extracts and outputs the feature information of the training image;

[0059] Input feature information into the neck network, perform multi-scale feature fusion, enhance feature representation capability, and output fused feature information;

[0060] Input the fused feature information into the head network and output the key information of the target;

[0061] (3c) The image output by the improved YOLOv10 model is sent to the contour extraction module, and the cropped bridge structure image is used to identify the structural foreground, extract the mask and fit the polygon. The obtained polygon is the bridge structure contour map.

[0062] Step (4) specifically includes the following steps in order:

[0063] (4a) Input the PDF design drawing of the bridge to be inspected into the PaddleOCR model, set the direction of manual reading of the drawing as the standard direction, extract the text in the standard direction normally, and for the text perpendicular to the standard direction, first rotate the image 90° and then identify the vertical text, then perform a one-step coordinate conversion, normalize and save the identified text, and output the detected text information and position information;

[0064] (4b) Setting logical constraints: Using the text information detected by the PaddleOCR model, combined with the annotation area identified by the improved YOLOv10 model and the structure identified by the bridge structure detection model, the text and coordinate information of the necessary parameters are jointly screened out;

[0065] At this time, the required parameters obtained are not assigned attributes. The required parameters are then matched with the obtained bridge structure. Specifically, the Euclidean distance between the coordinates of the required parameters and the coordinates of the midpoint of the bridge edge is calculated. When the distance is the smallest, the bridge edge and the required parameters are matched together. At this time, the parameter and coordinate matching of a view is completed. After that, based on the inherent relationship between different views, multiple parameters between multiple views are summarized into an Excel table to achieve the final parameter extraction.

[0066] Figure 4Several required parameters are marked in the figure, among which "8-8" represents the serial number, "160" represents the beam height, "20" represents the thickness of the cantilever end, "1686.8" represents the width of the top surface of the bridge, and "186.8" represents the width of the bottom surface of the bridge.

[0067] Figure 5 Several required parameters are marked in the figure, among which "2495" represents the span length, "|8" and "|11" represent the sectioning sequence numbers, and "Z12" and "Z13" represent the span positioning numbers.

[0068] Figure 6 Several required parameters are marked in the figure, among which "2495" represents the span length, "1675" represents the width of the bridge top surface, and "Z12" and "Z13" represent the span location numbers.

[0069] Figure 7 Several required parameters are marked in the figure, among which "2495" represents the span length, "1275" represents the width of the bridge bottom surface, and "Z12" and "Z13" represent the span location numbers.

[0070] like Fig.10 As shown, the bridge structure detection model can accurately extract the outline structure of the bridge from the bridge design drawings.

[0071] like Fig.11 As shown, the present invention can detect the type of bridge structure in the bridge design drawings, obtain necessary parameter information, and classify all structural dimensions into a library.

[0072] In summary, the present invention realizes the detection of bridge structures in drawings by combining the improved YOLOv10 model and the K-Net semantic segmentation algorithm, thereby drawing the outline of the bridge structure; uses the PaddleOCR model to perform text detection on bridge design drawings and extracts the parameters of the bridge structure; matches the structure with the parameters through a series of logical constraints, and finally realizes the structural recognition and parameter extraction of the bridge in the bridge design drawings; the present invention adds the CGA attention mechanism module to the YOLOv10 model, combines it with the K-net semantic segmentation algorithm, and solves the problem of inaccurate outline detection in bridge design drawings; at the same time, in view of the problem that the text information detected by the PaddleOCR model lacks semantic information, semantic information is given to the text and parameters in the drawings; experimental results show that the recall rate of the bridge structure detection model can reach 96.1%, and the average precision reaches 96.2%, which are 14.6% and 13.1% higher than the original YOLOv10 model respectively, and can effectively detect different elements in the drawings.

Claims

1. A bridge structure identification and parameter extraction method for bridge PDF design drawings, characterized by: The method comprises the following steps in sequence: (1) Obtain bridge design drawings saved in PDF format, preprocess them, and obtain preprocessed bridge design drawings to form a data set; (2) Combining the improved YOLOv10 model and contour extraction module, a bridge structure detection model is constructed; (3) Using the data set to train the bridge structure detection model, and obtaining the trained bridge structure detection model; (4) Obtain the PDF design drawing of the bridge to be inspected, input the PDF design drawing of the bridge to be inspected into the PaddleOCR model for parameter recognition, and obtain the necessary parameters and their coordinate information; (5) Preprocess the PDF design drawings of the bridge to be inspected, input the preprocessed PDF design drawings of the bridge to be inspected into the trained bridge structure detection model, identify the bridge structure, and obtain the bridge structure in different views.

2. The bridge structure identification and parameter extraction method for bridge PDF design drawings according to claim 1 is characterized by: In step (1), the preprocessing refers to grayscale processing to obtain a binary image, and then the binary image is labeled and data expanded. The expanded bridge design drawings constitute a data set.

3. The bridge structure identification and parameter extraction method for bridge PDF design drawings according to claim 1 is characterized in that: Step (2) specifically refers to: improving the YOLOv10 model, wherein the YOLOv10 model includes a backbone network, a neck network, and a head network, adding the CGA attention mechanism module to the layer after the C2f module in the backbone network, and adding the CGA attention mechanism module to the layer after the C2f module in the neck network, to obtain an improved YOLOv10 model; The output of the improved YOLOv10 model is used as the input of the contour extraction module, and the output of the contour extraction module is used as the output of the bridge structure detection model. The improved YOLOv10 model and the contour extraction module together constitute the bridge structure detection model; The contour extraction module uses the K-Net semantic segmentation algorithm to obtain the foreground mask of the structure to be detected, and then uses the edge pixel detection method and polygon fitting in the image processing method to finally obtain the specific contour of the structure to be detected; The edge pixel detection method is used to detect the outline of the foreground in the mask. The edge pixel detection method uses the Canny detection operator. The Canny detection includes four steps: noise elimination, calculation of gradient amplitude and direction, non-maximum suppression, and hysteresis threshold. The two-dimensional Gaussian formula for noise elimination is: Where G(x,y) represents the value of the Gaussian function at the point (x,y), that is, the probability density at the point (x,y) in two-dimensional space; x and y are coordinate variables in two-dimensional space; σ is the standard deviation.

4. The bridge structure identification and parameter extraction method for bridge PDF design drawings according to claim 1 is characterized in that: Step (3) specifically includes the following steps in order: (3a) Set the training parameters, i.e., the training parameters of the improved YOLOv10 model. The training parameters include the number of iterations, batch size, and initial learning rate. The number of iterations is 300, the batch size is 8, and the initial learning rate lr0 of the model is set to 0.

01. (3b) Input the dataset into the improved YOLOv10 model; The dataset is fed into the backbone network, which extracts and outputs the feature information of the training image; Input feature information into the neck network, perform multi-scale feature fusion, enhance feature representation capability, and output the fused feature information; Input the fused feature information into the head network and output the key information of the target; (3c) The image output by the improved YOLOv10 model is fed into the contour extraction module. The cropped bridge structure image is used to identify the structural foreground, extract the mask, and fit the polygon. The resulting polygon is the bridge structure contour map.

5. The bridge structure identification and parameter extraction method for bridge PDF design drawings according to claim 1 is characterized in that: Step (4) specifically includes the following steps in order: (4a) Input the PDF design drawing of the bridge to be inspected into the PaddleOCR model, set the direction of manual reading of the drawing to the standard direction, and extract the text in the standard direction normally. For the text perpendicular to the standard direction, first rotate the image 90° and then identify the vertical text. Then, through a one-step coordinate transformation, the recognized text is normalized and saved, and the detected text information and position information are output; (4b) Setting logical constraints: Using the text information detected by the PaddleOCR model, combined with the annotation area identified by the improved YOLOv10 model and the structure identified by the bridge structure detection model, the text and coordinate information of the necessary parameters are filtered out; At this time, the required parameters obtained are not assigned attributes. Then the required parameters are matched with the obtained bridge structure. Specifically, the Euclidean distance between the required parameter coordinates and the coordinates of the midpoint of the bridge edge is calculated. When the distance is minimum, the bridge edge and the required parameters are matched together. At this time, the parameter and coordinate matching of a view is completed. Then, based on the inherent relationship between different views, multiple parameters between multiple views are summarized into an Excel table to achieve the final parameter extraction.

Images