Defect automatic positioning method based on BIM virtual image

Abstract

The invention relates to the technical field of computer vision processing, in particular to an automatic defect positioning method based on a BIM virtual image. The method comprises the following steps: establishing a high-rise building model, and making a data set integrating an illumination condition and a full view angle; calculating camera parameters; performing cross-modal image retrieval: performing targeted optimization on the basis of a classical ResNet architecture, and constructing a backbone network structure suitable for a building image retrieval task; initializing position attitude estimation based on matching; and correcting the camera position posture. According to the method, a defect fixed frame based on a building BIM virtual image is provided, a building image data set BIM-Vision based on Revit is constructed, rich visual angles and illumination condition setting are achieved, accurate camera position postures and 3D labels between beam columns are provided, high-quality basic data support is provided for building visual research, and the method has the advantages of being high in practicability and high in practicability. And during inspection, accurate positioning of defect positions and component association can be completed only by shooting a field image, so that the field operation process is greatly simplified.

Description

Technical Field

[0001] This invention relates to the field of computer vision processing technology, and more specifically, to an automatic defect location method based on BIM virtual images. Background Technology

[0002] In architectural settings, accurately locating defects is a crucial yet challenging task. Traditional assessment methods typically rely on manual inspections, where inspectors meticulously examine the building structure based on their experience and effort. While this method has some practicality, the inspection process is not only extremely time-consuming but also highly dependent on the inspector's experience and effort, easily leading to missed or false detections, thus severely impacting the accuracy and reliability of the assessment results. With technological advancements, image-based retrieval methods, due to their strong perception capabilities in complex defect environments, are increasingly recognized as a promising solution. These methods not only reduce the time and effort costs of manual inspections but also improve the accuracy and precision of defect location through big data analysis. Currently, research and applications in building defect retrieval mainly focus on two types of methods: end-to-end deep learning methods and geometric optimization methods.

[0003] End-to-end deep learning methods typically extract features directly from raw images using deep learning techniques such as convolutional neural networks (CNNs) for automatic defect identification and classification. This approach boasts strong adaptability, achieving good performance in complex scenes and dynamic environments, particularly when handling large-scale datasets. However, a significant drawback of deep learning methods is their reliance on large amounts of labeled data, the acquisition of which is extremely costly. Furthermore, the diversity and accuracy of the data directly impact model performance; therefore, in some scenarios, the model may fail to adapt to unknown defect types or new building environments.

[0004] In contrast, geometric optimization methods primarily extract geometric features from building structures and combine them with geometric constraints to locate defects. Geometric optimization methods have certain advantages when handling low-texture or highly repetitive building scenes, especially in scenarios where traditional methods struggle, providing relatively robust detection results. However, geometric optimization methods also have limitations, particularly in complex and varied architectural scenes, where their performance is often restricted. Because this method relies heavily on the geometric features of the structure, it cannot effectively cope with environmental changes and diverse defect types; therefore, its applicability and robustness need improvement. Summary of the Invention

[0005] The purpose of this invention is to provide an automatic defect location method based on BIM virtual images to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, an automatic defect location method based on BIM virtual images is provided, comprising the following steps:

[0007] S1. Build a high-rise building model and create a dataset that integrates lighting conditions and full viewing angles, and train the dataset.

[0008] S2. Calculate camera parameters;

[0009] S3, Cross-modal image retrieval: Based on the classic ResNet architecture, targeted optimizations are made to construct a backbone network structure suitable for architectural image retrieval tasks;

[0010] S4. Matching-based initial position and pose estimation;

[0011] S5. Correct the camera position and attitude.

[0012] As a further improvement to this technical solution, step S1 includes the following steps:

[0013] S1.1 Under the BIM technology framework, construct a parametric model of a high-rise building based on the Autodesk Revit 2024 platform;

[0014] S1.2 Utilize Revit and Dynamo to form an efficient and automated viewpoint generation and data acquisition process;

[0015] S1.3 Automated rendering of the generated observation points;

[0016] S1.4 Finally, the rendered image and the camera's geometric feature data are generated.

[0017] As a further improvement to this technical solution, the process steps in S1.2 are as follows:

[0018] ① First, manually preset 50 key camera positions in the Revit model. The coordinates of these camera positions come from the measurement point (0, 0, 0) in the building model. Input these positions into Dynamo, and Dynamo will perform a uniform transformation of the measurement points to the standard viewpoint height.

[0019] ② Construct the camera path matrix at the camera observation point using a polar coordinate algorithm, and finally generate a perspective view, which is then saved in the Revit 3D view.

[0020] As a further improvement to this technical solution, the specific steps of S2 are as follows:

[0021] S2.1 Define a world coordinate system OX w Y w Zw Select a point in the 3D model in the world coordinate system and define this point as P. w (X w Y w Z w The origin of this coordinate system corresponds to the model measurement point set in Revit;

[0022] S2.2, Set the camera coordinate system OX c Y c Z c By transforming the model point P w Transform to the camera coordinate system;

[0023] S2.3 In the camera coordinate system, the 3D model points are mapped to the image coordinate system through perspective projection model to obtain the image coordinates of the 3D model points;

[0024] S2.4 Finally, by combining the projection relationship between the 3D model points and their corresponding pixels, the camera intrinsic parameter matrix K is obtained.

[0025] As a further improvement to this technical solution, the step of obtaining the image coordinates of the three-dimensional model points in S2.3 is as follows:

[0026] ① Define the image coordinate system O-xy as a virtual plane located at a focal length f in front of the camera;

[0027] ② The virtual plane is perpendicular to the Z-axis of the camera coordinate system, and its origin is the foot of the perpendicular from the Z-axis, which is defined as the image center;

[0028] ③ Set a 3D point (X) with camera coordinates. c Y c Z c Based on the principle of similar triangles, the camera coordinates are projected onto the image coordinate system to obtain the image coordinates (x, y).

[0029] As a further improvement to this technical solution, in step S2.3, after calculating the image coordinates (x, y), they are converted into pixel coordinates (u, v). The steps are as follows:

[0030] ① Perform a scale transformation to convert physical units (feet) into pixel units;

[0031] ② Flip the Y-axis to match the rule of the Y-axis pointing downwards in the pixel coordinate system;

[0032] ③ Perform origin translation, moving the origin of the coordinate system from the center of the image to the top left corner of the image, thereby achieving accurate mapping.

[0033] As a further improvement to this technical solution, the specific steps of S4 are as follows:

[0034] Based on the data in S3, we focus on the correspondence between the Top-K images and the test images, and use the ORB algorithm to extract two-dimensional features from the Top-K images and the test images, and calculate the corresponding descriptors.

[0035] As a further improvement to this technical solution, the specific process of two-dimensional feature extraction in S4 is as follows:

[0036] ① First, use Oriented FAST to quickly find key corner points of different scales with orientation in the image;

[0037] ② Then, use Rotated BRIEF to create a rotation-insensitive descriptor for each corner point;

[0038] ③ Perform preliminary feature matching based on Hamming distance using the Brute-Force matcher, sort the matching results according to the distance, and select the optimal matching point pairs;

[0039] ④ The homography matrix between the two images is estimated using the RANSAC algorithm to obtain stable two-dimensional pixel matching points. Combined with the information of the three-dimensional spatial points and two-dimensional pixel points obtained in S1, the camera position and pose are estimated using the EPnP algorithm.

[0040] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0041] 1. This method for automatic defect localization based on BIM virtual images proposes a defect localization framework based on building BIM virtual images and constructs a Revit-based building image dataset, BIM-Vision. Addressing the shortcomings of existing research in handling the complexity of the building environment, multi-view variations, and lighting interference, it designs rich viewpoint and lighting condition settings and provides precise camera position and pose, as well as 3D annotations between beams and columns. This provides high-quality basic data support for building vision research. During inspections, only one on-site image needs to be captured to accurately locate the defect and associate it with components, greatly simplifying on-site operation procedures and improving the efficiency and intelligence level of defect management.

[0042] 2. In this automatic defect localization method based on BIM virtual images, a structure-aware image retrieval model is proposed. By introducing spatial distribution guidance and geometric structure enhancement modules, the ability to perceive differences in the layout of building components is strengthened, significantly improving image discrimination and retrieval accuracy in multi-view and low-texture scenes. A cross-modal image feature alignment method is also proposed. By introducing multi-scale feature fusion and adaptive mapping mechanisms, the differences between on-site images and projected images in terms of lighting, viewing angle, etc., are overcome, thereby improving the accuracy and robustness of cross-modal retrieval. Attached Figure Description

[0043] Figure 1 The structured framework design diagram of the present invention

[0044] Figure 2 Example diagram of cross-view image retrieval with different perturbations according to the present invention

[0045] Figure 3 Example diagram of the initial matching point of the present invention;

[0046] Figure 4 This is a flowchart of the steps in this solution. Detailed Implementation

[0047] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0048] Example 1

[0049] Please see Figures 1-4 As shown, the purpose of this embodiment is to provide an automatic defect location method based on BIM virtual images. This method starts by establishing a building database, and through a three-stage process of "retrieval-matching-solving," obtains the corrected camera position and pose. Then, based on the image background information, it determines the defect location, achieving automated, high-precision component-level defect location. Specifically, it includes the following steps:

[0050] S1. Build a high-rise building model and create a dataset that integrates lighting conditions and full viewing angles, then train the model on the dataset:

[0051] S1.1 Under the BIM technology framework, a parametric model of a high-rise building is constructed based on the Autodesk Revit 2024 platform, covering major components such as concrete columns, concrete beams, basic walls and floor slabs;

[0052] S1.2 Utilize Revit and Dynamo to form an efficient automated viewpoint generation and data acquisition workflow; this workflow adopts a strategy that combines manually preset key viewpoints with batch generation by algorithms, and the specific steps are as follows:

[0053] ① First, manually preset 50 key camera positions in the Revit model. The coordinates of these camera positions come from the measurement point (0, 0, 0) in the building model. Input these positions into Dynamo. Dynamo will perform a uniform transformation of the measurement points to a standard viewpoint height (in this study, a standard viewpoint height of 5 feet is used, which matches human height. By setting a reasonable observation point height that is consistent with human height, the actual images captured can be better matched with the images in the dataset when they are taken later). Finally, reasonable camera observation points are generated.

[0054] ② At the location of the camera observation point, a camera path matrix is ​​constructed using a polar coordinate algorithm. The azimuth angle α is discretely sampled in steps of 30° within the range of 0° to 360°, and the pitch angle β is set in steps of 5° within the range of -5° to 5°. Finally, a perspective view is generated and saved in the 3D view of Revit. Considering the data compatibility between Revit and Dynamo environments, feet are used as the unit of length.

[0055] S1.3 Automated rendering of the generated observation points;

[0056] The architectural model was visualized using Revit's built-in rendering engine. The image quality was set to medium, and different lighting scenarios were adjusted to generate high-quality architectural representation images. To realistically simulate diverse lighting environments, enhance the richness of the image database, and reduce the interference of lighting differences in subsequent comparisons with real-world images, this study set up three typical lighting scenarios: 10:00 AM during the summer solstice, 4:00 PM during the summer solstice, and artificial light only. Simultaneously, suitable concrete texture images were downloaded from the texture website and applied to Revit's built-in material manager. This allowed all component surfaces to be covered with standard concrete textures with different roughness, color, and texture details, enhancing the realism and material diversity of the images. To ensure consistency in material mapping, the UVW map ratio was uniformly set to 65×65. This setting, verified through multiple rounds of testing, balanced image clarity, rendering efficiency, and surface perception quality.

[0057] S1.4 Finally, generate the rendered image and the camera's geometric feature data:

[0058] ① A 927×729 pixel JPG format rendered image;

[0059] ② The geometric coordinates and rotation angle of the camera in Revit, as well as the geometric coordinates of its target point; the camera coordinates are manually selected coordinate points (x, y, z), and the rotation angle is set to 30 degrees; the target point coordinates are calculated on a horizontal plane using cosine and sinine functions with the camera coordinates (x, y) as the center and radius as the radius, while the target point itself is only a spatial coordinate without a rotation angle; at the same time, Revit can directly read CSV files containing 3D point information such as beam-column intersections and wall feature lines in the world coordinate system, as well as the corresponding 2D projected coordinates in the pixel coordinate system. There is a one-to-one correspondence between the 3D spatial points and the positions of the 2D pixels obtained after projection onto a specific image, which is used for subsequent position and pose calculation and error evaluation.

[0060] S2. Calculate camera parameters;

[0061] To achieve a precise correspondence between the rendered image and the 3D model data, the camera intrinsic parameter matrix K will be calculated by inversely using the spatial transformation relationship between the coordinate systems.

[0062] S2.1 Define a world coordinate system OX w Y w Z w Select a point in the 3D model in the world coordinate system and define this point as P. w (X w Y w Z w The origin of this coordinate system corresponds to the model measurement point set in Revit;

[0063] S2.2, Set the camera coordinate system OX c Y c Z c The camera coordinate system is a reference system centered on the camera, with its origin at the camera's optical center and the Z-axis typically pointing in the camera's viewing direction. Through rigid body transformation (i.e., continuously rotating and translating the world coordinate system), the model point P is transformed... w Transform to the camera coordinate system; at this point, the coordinates of the model points become (X... c Y c Z c This establishes the relative relationship between the world coordinate system and the camera coordinate system.

[0064] S2.3. In the camera coordinate system, the 3D model points are mapped to the image coordinate system through perspective projection to obtain the image coordinates of the 3D model points. The specific steps are as follows:

[0065] ① Define the image coordinate system O-xy as a virtual plane located at a focal length f in front of the camera;

[0066] ② The virtual plane is perpendicular to the Z-axis of the camera coordinate system, and its origin (principal point) is the foot of the perpendicular from the Z-axis, which is defined as the image center;

[0067] ③ Set a 3D point (X) with camera coordinates. c Y c Z c Based on the principle of similar triangles, the camera coordinates are projected onto the image coordinate system to obtain the image coordinates (x, y).

[0068] Furthermore, after calculating the image coordinates (x, y), in order to align the calculated image coordinates (x, y) with the actual digital image pixel grid into the image coordinate system O-xy, they need to be converted into pixel coordinates (u, v). This transformation process is completed through three core steps:

[0069] ① Perform a scale transformation to convert physical units (feet) into pixel units;

[0070] ② Flip the Y-axis to match the rule of the Y-axis pointing downwards in the pixel coordinate system;

[0071] ③ Perform origin translation, moving the origin of the coordinate system from the center of the image (origin) to the top left corner of the image, thereby achieving accurate mapping.

[0072] S2.4 Finally, by combining the projection relationship between the 3D model points and their corresponding pixels, the camera intrinsic parameter matrix K is obtained.

[0073] The above refers to the three-dimensional model point P in the world coordinate system. w (X w ,Y w Z w The formula for transforming a pixel (u, v) into pixel coordinates is as follows: The overall formula is as follows:

[0074]

[0075] z c R represents the depth value of the model point in the camera coordinate system, R represents the rotation matrix in the world coordinate system, and t represents the translation vector in the world coordinate system.

[0076] The final camera intrinsic parameter matrix K is:

[0077]

[0078] (f x f y ) represents the equivalent focal length in the horizontal and vertical directions.

[0079] The specific operating procedure is as follows:

[0080] Input a BIM-based virtual test image. First, a 3×3 convolutional layer (3 input channels, 64 output channels, stride=1, padding=1) is used to extract preliminary texture features. Then, the nonlinear expressive power is enhanced by a BatchNorm layer and the PReLU activation function.

[0081] Subsequently, the feature map passes through four cascaded residual modules, which progressively deepen semantic understanding, expand the channels from 64 to 512, and halve the resolution step by step, thereby effectively capturing multi-scale architectural features from local details to overall outlines. Each residual module is composed of multiple Bottleneck modules stacked in the order of "normalization-convolution-activation".

[0082] To address the significant differences between real and virtual images in terms of lighting, viewpoint, etc., the model introduces a spatial distribution guidance and geometric enhancement module after the residual module.

[0083] At the end of the network, the AdaptiveAvgPool2d layer unifies the feature map size to 7×7 to adapt to different input resolutions. Then, it passes through BatchNorm2d, Dropout, and Flatten operations in sequence. Finally, a fully connected layer efficiently compresses the 512×7×7 feature map into a 512-dimensional compact feature vector, which is used to retrieve and return the Top-K most similar virtual images and their similarity scores.

[0084] The test image contains three perturbation scenarios:

[0085] (1) Translational disturbance, angle remains unchanged;

[0086] (2) Rotational disturbance, position unchanged;

[0087] (3) Combined translation and rotation disturbances;

[0088] Examples of perturbations for different virtual test images are as follows: Figure 2 As shown, Figure 2 In the diagram, a represents translational perturbation, b represents rotational perturbation, and c represents a combination of random translational and rotational perturbation.

[0089] S3, Cross-modal image retrieval:

[0090] Based on the classic ResNet architecture, a backbone network structure suitable for architectural image retrieval tasks is constructed through targeted optimizations. This network supports multiple depth configurations (50, 100, and 152 layers), allowing for a flexible trade-off between accuracy and efficiency according to actual needs. Furthermore, the network achieves a dynamic balance between expressive power and computational cost by enabling modular switching between IR and IR-SE architectures.

[0091] The input image is encoded using a Backbone network to represent its high-dimensional features, which are then mapped to a 512-dimensional normalized feature vector via a head network.

[0092]

[0093] q represents the query image; d i The candidate images are shown in the formula. This formula is a ternary loss function, which can transform image recognition from a "classification problem" into a "metric learning problem". It is often used in the field of image retrieval. The normalization method is used to standardize the feature vector to a unit length, so as to lay the foundation for subsequent image matching using the cosine similarity method.

[0094] The system retrieves the top five candidate images with the highest similarity based on similarity (Top-K) to evaluate the accuracy and robustness of the backbone network structure in cross-view retrieval tasks and to provide support for subsequent position and pose estimation.

[0095] The specific operating steps are as follows:

[0096] Select the top-1 image that is most similar to the test image, and use the ORB algorithm to extract key points from the image (such as...). Figure 3 (As shown), then the BFMatcher brute-force matcher and RANSAC algorithm are combined to eliminate incorrect matches and ensure accurate pairing of key points; next, by establishing the correspondence between 3D model points and 2D pixels in the test image, and between 3D model points and 2D pixels in the Top-1 image, the Epnp algorithm is used to calculate the initial position and pose of the camera.

[0097] S4. Matching-based initial position and pose estimation:

[0098] Based on the returned Top-K result images and similarity data, we focus on the correspondence between the Top-K images and the test images. We use the ORB (Oriented Fast and Rotated BRIEF) algorithm to extract two-dimensional features from both the Top-K and test images and calculate the corresponding descriptors. (A descriptor can be imagined as a unique "digital fingerprint" or "identity card" for a feature point. Its core function is to compare the descriptors of each feature point in the two images to identify which points match, thus achieving image alignment and matching.) The specific process of two-dimensional feature extraction is as follows:

[0099] ① First, use Oriented FAST to quickly find key corner points of different scales with orientation in the image;

[0100] ② Then, Rotated BRIEF is used to create a rotation-insensitive descriptor for each corner point, thereby achieving fast and robust feature matching;

[0101] ③ Perform preliminary feature matching based on Hamming distance using the Brute-Force matcher, sort the matching results according to the distance, and select the optimal matching point pairs;

[0102] ④ The RANSAC algorithm is used to estimate the homography matrix between the two images to obtain stable two-dimensional pixel matching points. Combined with the information of the three-dimensional spatial points and two-dimensional pixel points obtained in S1, the EPnP (Efficient Perspective-n-Point) algorithm is used to estimate the camera position and pose.

[0103] The RANSAC algorithm consists of three steps: hypothesis-verification-iteration. This process can eliminate incorrect matching results and ensure the geometric consistency of correct matching results to obtain a more accurate homography matrix.

[0104] This process not only achieves spatial alignment between images, but also provides an accurate initial correspondence for subsequent position and pose estimation.

[0105] Based on the initial estimation of the EPnP algorithm, the position and attitude are finely adjusted by combining nonlinear optimization, which effectively reduces the accumulated error and finally obtains the optimal camera position and attitude. Based on this position and attitude, the location of the target point can be determined. In the final test, the actual captured image is input, and the target point is the defect in the image. Then, the component where the defect is located can be determined by the BIM model, so as to achieve accurate positioning of the defect.

[0106] S5. Correct camera position and attitude:

[0107] To further correct the position and attitude, this study constructs the following optimization function with the goal of minimizing the reprojection error:

[0108]

[0109] Where E(r,t) represents the total error, which is a function of the camera position and pose R and t. The goal of this formula is to minimize the value of E; (c x c y ) represents the coordinates of the principal point, and π(·) represents the projection function from the point in three-dimensional space to the image plane:

[0110]

[0111] Where [X, Y, Z] are the 3D point coordinates input to the projection function, relative to the camera coordinate system; [u, v] refer to the 2D pixel coordinates output by the projection function; (f x ,f y () represents the focal length.

[0112] This study solves this nonlinear optimization problem using the Gauss-Newton method, ultimately obtaining the optimal position and attitude parameters (r). ＊ ,t ＊ Furthermore, the position C of the camera in the world coordinate system is derived. ＊ and towards R ＊ This enables precise attitude positioning of the camera in three-dimensional space.

[0113] The specific embodiments of the present invention have been described in detail through the above examples and steps. This method requires no additional economic investment and can achieve automated, high-precision component-level positioning of defects using only a single image.

[0114] 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 preferred examples and are not intended to limit 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 present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for automatic defect localization based on BIM virtual images, characterized in that: Includes the following steps: S1. Build a high-rise building model and create a dataset that integrates lighting conditions and full viewing angles, and train the dataset. S2. Calculate camera parameters; S3, Cross-modal image retrieval: Based on the classic ResNet architecture, targeted optimizations are made to construct a backbone network structure suitable for architectural image retrieval tasks; S4. Matching-based initial position and pose estimation; S5. Correct the camera position and attitude.

2. The automatic defect location method based on BIM virtual images according to claim 1, characterized in that: S1 includes the following steps: S1.1 Under the BIM technology framework, construct a parametric model of a high-rise building based on the Autodesk Revit 2024 platform; S1.2 Utilize Revit and Dynamo to form an efficient and automated viewpoint generation and data acquisition process; S1.3 Automated rendering of the generated observation points; S1.4 Finally, the rendered image and the camera's geometric feature data are generated.

3. The automatic defect location method based on BIM virtual images according to claim 2, characterized in that: The process steps in S1.2 are as follows: ① First, manually preset 50 key camera positions in the Revit model. The coordinates of these camera positions come from the measurement point (0, 0, 0) in the building model. Input these positions into Dynamo, and Dynamo will perform a uniform transformation of the measurement points to the standard viewpoint height. ② Construct the camera path matrix at the camera observation point using a polar coordinate algorithm, and finally generate a perspective view, which is then saved in the Revit 3D view.

4. The automatic defect location method based on BIM virtual images according to claim 2, characterized in that: The specific steps of S2 are as follows: S2.1 Define a world coordinate system OX w Y w Z w Select a point in the 3D model in the world coordinate system and define this point as P. w (X w Y w Z w The origin of this coordinate system corresponds to the model measurement point set in Revit; S2.2, Set the camera coordinate system OX c Y c Z c By transforming the model point P w Transform to the camera coordinate system; S2.3 In the camera coordinate system, the 3D model points are mapped to the image coordinate system through perspective projection model to obtain the image coordinates of the 3D model points; S2.4 Finally, by combining the projection relationship between the 3D model points and their corresponding pixels, the camera intrinsic parameter matrix K is obtained.

5. The automatic defect location method based on BIM virtual images according to claim 1, characterized in that: The steps in S2.3 to obtain the image coordinates of the three-dimensional model points are as follows: ① Define the image coordinate system O-xy as a virtual plane located at a focal length f in front of the camera; ② The virtual plane is perpendicular to the Z-axis of the camera coordinate system, and its origin is the foot of the perpendicular from the Z-axis, which is defined as the image center; ③ Set a 3D point (X) with camera coordinates. c Y c Z c Based on the principle of similar triangles, the camera coordinates are projected onto the image coordinate system to obtain the image coordinates (x, y).

6. The automatic defect location method based on BIM virtual images according to claim 1, characterized in that: In step S2.3, after calculating the image coordinates (x, y), they are converted into pixel coordinates (u, v). The steps are as follows: ① Perform a scale transformation to convert physical units into pixel units; ② Flip the Y-axis to match the rule of the Y-axis pointing downwards in the pixel coordinate system; ③ Perform origin translation, moving the origin of the coordinate system from the center of the image to the top left corner of the image, thereby achieving accurate mapping.

7. The automatic defect location method based on BIM virtual images according to claim 1, characterized in that: The specific steps of S4 are as follows: Based on the data in S3, we focus on the correspondence between the Top-K images and the test images, and use the ORB algorithm to extract two-dimensional features from the Top-K images and the test images, and calculate the corresponding descriptors.

8. The automatic defect location method based on BIM virtual images according to claim 7, characterized in that: The specific process of two-dimensional feature extraction in S4 is as follows: ① First, use Oriented FAST to quickly find key corner points of different scales with orientation in the image; ② Then, use Rotated BRIEF to create a rotation-insensitive descriptor for each corner point; ③ Perform preliminary feature matching based on Hamming distance using the Brute-Force matcher, sort the matching results according to the distance, and select the optimal matching point pairs; ④ The homography matrix between the two images is estimated using the RANSAC algorithm to obtain stable two-dimensional pixel matching points. Combined with the information of the three-dimensional spatial points and two-dimensional pixel points obtained in S1, the camera position and pose are estimated using the EPnP algorithm.

Images