A method for detecting construction quality defects of an external wall of a building

Abstract

The application discloses a building outer wall construction quality defect detection method, and relates to the field of building outer wall quality detection and image processing information technology. The scheme collects overlapping images of the outer wall through a multi-angle image acquisition device, generates a first feature image set through feature extraction, constructs a dense point cloud based on a dense matching algorithm and maps to generate an initial spatial coordinate set, obtains a three-dimensional coordinate point set through projection transformation of two-dimensional pixel information of defects, and corrects the coordinate offset through error correction. A three-dimensional grid model of defects is constructed based on the corrected coordinates and defect recognition results, a crack extension direction vector is solved by extracting the model contour, the depth range of defects is quantified through deep fusion combined with three-dimensional grid data, and finally, repair range data is accurately generated based on the depth and direction parameters, and the real spatial position and geometric shape of defects are output. The application improves the outer wall defect detection precision and the scientific nature of the later repair construction, and is suitable for automatic quality inspection operation of high-rise building outer walls.

Description

Technical Field

[0001] This invention relates to the fields of image processing and building construction quality inspection technology, and in particular to a method for detecting defects in the construction quality of building exterior walls. Background Technology

[0002] Building exterior wall defect detection is directly related to building structural safety, service life, facade maintenance costs, and overall construction quality. With the continuous growth of the stock of high-rise and super high-rise buildings in China, defects such as exterior wall cracks and peeling paint, if not accurately detected and addressed, can easily lead to safety hazards such as falling objects from heights. Therefore, the development of efficient and accurate automated defect detection technology is of practical necessity.

[0003] Currently, the quality inspection of building exterior walls generally relies on manual visual inspection or single-view two-dimensional image acquisition. Due to limitations imposed by factors such as the varied shapes of building facades and the obstruction of images by certain components, these traditional inspection methods can only capture two-dimensional images of the defect surface, failing to accurately represent the spatial morphology and three-dimensional coordinates of the defect entity. Because two-dimensional defect images lack depth dimension information, they are difficult to register and integrate with the three-dimensional spatial data of the building entity, resulting in significant deviations in the calculated defect size and boundary range. Furthermore, the calculation of repair construction materials and the delineation of repair areas lack reliable three-dimensional data support.

[0004] Oblique photogrammetry can acquire multi-view overlapping images and generate dense point clouds of walls, but existing technologies struggle to achieve high-precision spatial registration between 2D defect identification results and 3D point clouds. 3D point clouds carry complete spatial geometric coordinate information, while AI-generated defect annotations only provide image pixel information without depth parameters. Inaccurate mapping between the two leads to distortion in the 3D defect reconstruction model, further distorting the spatial orientation of cracks and the measurement of damage depth, making it impossible to accurately define the scope of repair work.

[0005] Therefore, how to accurately integrate two-dimensional defect pixel information with three-dimensional wall point cloud data, and accurately restore the true three-dimensional shape and spatial location of defects, is a technical bottleneck that urgently needs to be overcome in the field of external wall defect detection. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a method for detecting construction quality defects in building exterior walls. It features intelligent processing throughout the entire process, from multi-angle image acquisition, feature extraction, point cloud generation, coordinate correction, 3D modeling, depth calculation to automatic delineation of repair areas. It accurately outputs data on the 3D morphology, spatial location, crack direction, and damage depth of defects, solving the industry pain points of insufficient data dimensions and inadequate accuracy in traditional detection methods.

[0007] To achieve the above objectives, the technical solution of the present invention is as follows: A method for detecting construction quality defects in building exterior walls, the method comprising: Collect multiple sets of overlapping image data of the building's exterior walls and input them into the feature extraction module to generate the first feature image set; The first feature image set is processed using a dense matching algorithm to generate dense point cloud data, and the initial spatial coordinate set is obtained by the coordinate mapping module. Based on the initial spatial coordinate set, extract the two-dimensional pixel coordinate information corresponding to the defect identification result, and generate a projected three-dimensional coordinate point set through the projection transformation module; Calculate the spatial distance between the projected three-dimensional coordinate point set and the wall reference structural surface. When the spatial distance exceeds a preset threshold, correct the coordinate offset parameters through the error correction module and generate a corrected three-dimensional coordinate point set. Based on the corrected three-dimensional coordinate point set and the defect identification results, surface fitting is performed to construct a three-dimensional mesh model of the defect. The boundary contour information of the defect 3D mesh model is extracted and sent to the direction calculation module to solve for the crack propagation direction vector; The original depth data is extracted by combining the crack extension direction vector with the defect three-dimensional mesh model, and the defect depth range parameters are determined by the depth fusion module. Based on the defect depth range parameters and crack extension direction vector, repair range calibration data is generated, and the geometric shape and spatial location parameters of the defect in the real-world space are output.

[0008] Furthermore, the step of collecting multiple sets of overlapping image data of the building's exterior walls and inputting them into the feature extraction module to generate a first feature image set includes: A drone equipped with a multi-lens camera was used to collect multi-view overlapping image source data of building exterior walls through gridded flight. The sequence integrity of the multi-view overlapping image source data was verified, and valid sequence images were selected. The first feature image set was obtained by extracting features from the valid sequence images through a feature extraction module.

[0009] Further, the step of processing the first feature image set using a dense matching algorithm to generate dense point cloud data, and obtaining an initial spatial coordinate set via a coordinate mapping module, includes: The first feature image set is retrieved, and pixel-level disparity is calculated using a dense matching algorithm to generate dense point cloud data. When the dense point cloud index meets the preset judgment threshold, the dense point cloud data is imported into the coordinate mapping module. The coordinate mapping module establishes a coordinate system with the reference point of the building's exterior wall as the origin and generates an initial set of spatial coordinates.

[0010] Furthermore, the generation of the initial spatial coordinate set further includes: Image acquisition distance and pitch angle acquisition parameters are extracted from the initial spatial coordinate set. The correlation between the acquisition parameters and the spatial density of the point cloud is analyzed by linear regression algorithm. When the image acquisition parameters and the spatial density distribution of the point cloud match, the mapping function parameters are solved by spatial coordinate transformation algorithm. The mapping function parameters are used to complete the point coordinate accuracy constraint, control the absolute coordinate error of the point to be ≤0.02m, and standardize the output coordinate set.

[0011] Further, based on the initial spatial coordinate set, the two-dimensional pixel coordinate information corresponding to the defect identification result is extracted, and a projected three-dimensional coordinate point set is generated by mapping through the projection transformation module, including: The two-dimensional pixel coordinate information corresponding to the defect identification result is retrieved from the initial spatial coordinate set. When the effective index of the two-dimensional pixel coordinate reaches the preset judgment threshold, the projection transformation module is activated to perform spatial transformation on the two-dimensional pixel coordinate, and the projection three-dimensional coordinate point set is obtained by solving the projection transformation. The spatial topological distribution characteristics of the projected 3D coordinate point set are analyzed; based on the spatial topological distribution characteristics, it is determined whether the projected 3D coordinate point set belongs to the same defect region. If it belongs to the same defect region, the boundary point set of the defect region is determined, and the final 3D defect position coordinates are obtained from the boundary point set of the defect region.

[0012] Further, the spatial distance between the projected three-dimensional coordinate point set and the wall reference structural surface is calculated. When the spatial distance exceeds a preset threshold, the coordinate offset parameters are corrected through an error correction module to generate a corrected three-dimensional coordinate point set, including: Extract the projected 3D coordinate point set and the reconstructed mesh data of the wall reference structural surface respectively; use the nearest neighbor algorithm to calculate the vertical spatial distance from the coordinate point to the wall reference surface point by point; if the distance calculation result exceeds the preset error threshold, generate a correction trigger flag; The error correction module calculates the coordinate offset correction amount for each dimension based on the trigger identifier; the offset correction amount is used to iteratively update the projected three-dimensional coordinate point set to obtain the corrected three-dimensional coordinate point set; the corrected point cloud is used to refresh the wall surface reconstruction model data.

[0013] Furthermore, based on the corrected three-dimensional coordinate point set and the defect identification results, surface fitting is performed to construct a three-dimensional mesh model of the defect, including: Obtain the set of calibrated 3D coordinate points and extract the corresponding 2D region of the defect from the defect identification results; establish the spatial relationship between the defect region and the 3D point cloud and delineate the defect boundary point set; perform surface fitting with the boundary point set as a constraint to obtain the initial surface model; statistically analyze the vertex discrete distribution index of the initial surface model to determine the vertex density state; perform mesh subdivision processing on the initial surface model based on the vertex density state; update the mesh patch normal vector parameters to generate the final 3D mesh model of the defect.

[0014] Further, the boundary contour information of the three-dimensional mesh model of the defect is extracted and fed into the direction calculation module to solve for the crack propagation direction vector, including: Obtain the vertex coordinate set of the 3D mesh model of the defect; filter the mesh contour edges based on the vertex coordinate set and stitch them together to form the defect boundary contour line; perform closure verification on the boundary contour line and filter the effective closed contour sequence; obtain the vertex normal vector of the closed contour from the closed contour sequence and project it onto the reference plane to construct the projection vector group; send the projection vector group into the direction calculation module to solve the crack extension direction vector; calculate the spatial angle between the crack extension direction vector and the crack principal axis; if the angle is less than the preset angle threshold, output the crack extension trend judgment conclusion.

[0015] Furthermore, the original depth data is extracted by combining the crack extension direction vector with the defect 3D mesh model, and the defect depth range parameters are determined by the depth fusion module, including: Vector decomposition is performed on the crack extension direction vector to extract the depth component; raw depth data is collected by combining the depth component with the 3D mesh model of the defect; the raw depth data is filtered and fused by the depth fusion module to generate a continuous depth numerical sequence; the difference between adjacent values ​​in the sequence is compared segment by segment, and if the difference exceeds a preset threshold, a depth boundary judgment point is marked; a set of valid depth values ​​is selected from the depth numerical sequence based on the depth boundary judgment points; the upper and lower limit intervals of the defect depth are statistically determined based on the set of valid depth values, and the final defect depth interval parameters are output.

[0016] Furthermore, based on the defect depth range parameters and crack extension direction vector, repair range calibration data is generated, and the geometric shape and spatial location parameters of the defect in the real-world space are output, including: The crack extension direction vector and defect depth range parameters are separated to extract three-dimensional spatial coordinate components. These three-dimensional spatial coordinate components are then mapped onto the three-dimensional mesh surface of the defect to generate point mapping data. The mapped points are then smoothed using a deep fusion module to generate a spatial point sequence. The spatial difference between adjacent points in the sequence is compared with a preset threshold. Points exceeding the threshold are marked as shape boundary judgment points. Valid shape value sets are filtered using these shape boundary judgment points. The final repair range delineation data is generated based on the shape value set. The spatial dimensions and boundary parameters corresponding to the repair range are then output.

[0017] The technical effects and advantages of this invention are as follows: 1. The method for detecting construction quality defects in building exterior walls provided in this application integrates a multi-algorithm system, including multi-angle oblique photography, dense point cloud matching, projection transformation error correction, three-dimensional surface fitting, vector analysis, and depth data fusion, to establish a data link between two-dimensional defect images and three-dimensional building entities, thus overcoming the shortcomings of traditional two-dimensional detection, which lacks depth and spatial dimensions. 2. The method for detecting construction quality defects of building exterior walls provided in this application solves the problems of two-dimensional pixel projection misalignment and model distortion by adding a coordinate error correction step and using the distance threshold judgment between point cloud and wall reference model + iterative registration to correct coordinate offset. The accuracy of three-dimensional defect modeling is significantly improved. 3. The method for detecting defects in the construction quality of building exterior walls provided in this application automatically generates repair boundary delineation data based on three-dimensional spatial parameters. It can be directly used for construction material cutting, repair range layout, accurate calculation of repair material usage, reduction of repair losses, and improvement of the standardization of exterior wall maintenance. 4. The method for detecting construction quality defects in building exterior walls provided in this application can automatically analyze the crack extension direction vector and the defect depth range, realize the quantitative value of damage depth and extension direction, and get rid of the subjective error of manual estimation; 5. The method for detecting construction quality defects in building exterior walls provided in this application uses non-contact data collection by drones throughout the process, eliminating the need for high-altitude suspended platform operations. This significantly reduces the labor costs and safety risks associated with high-altitude construction for the inspection of exterior walls of high-rise buildings, and is suitable for routine inspections of the exterior walls of large numbers of residential communities and super high-rise buildings. Attached Figure Description

[0018] Figure 1 is a schematic diagram of the overall process steps of the present invention; Figure 2 is a schematic diagram of the defect depth calculation process in step S107 of the embodiment of the present invention; Figure 3 is a schematic diagram of the defect three-dimensional mesh modeling process in step S105 of the embodiment of the present invention. Detailed Implementation

[0019] The present invention will be further described in detail below with reference to the embodiments given in the accompanying drawings.

[0020] See Figures 1-3 As shown, a method for detecting construction quality defects in building exterior walls includes the following steps: Step S101: Acquire multiple sets of overlapping image data of the building's exterior wall surface through a multi-angle image acquisition device and input them into a pre-established feature extraction module to generate the first feature image set.

[0021] The specific execution process of this step is as follows: Collect multiple sets of overlapping image data of the building's exterior wall surface; after confirming that the image data source comes from a multi-angle acquisition device, verify the integrity of the image sequence and generate compliant sequence image data; import the sequence image data into a preset feature extraction module to complete the computational processing and obtain a first feature image set; based on the first feature image set, determine the feature correspondence of the overlapping areas of multiple frames of images and generate a matching feature point set; based on the matching feature point set, calculate and generate a three-dimensional point cloud of the building's exterior wall surface to obtain the original surface point cloud data of the exterior wall; perform geometric configuration analysis on the acquired surface point cloud of the exterior wall to construct a geometric structure model of the wall; perform spatial mapping matching between the geometric structure model and the first feature image set, and finally output the distribution results of defects on the exterior wall surface.

[0022] In the specific implementation process, the overlapping images of the exterior walls were acquired by a drone equipped with a multi-lens industrial camera. The drone maintained a height of 5 meters above the wall surface and used a grid-based cruising flight path to capture the entire facade. To ensure the integrity of the image sequence, a phase correlation algorithm was used to calculate the translation amount and image overlap rate of adjacent images frame by frame. The formula for calculating the adjacent image overlap rate is as follows: In the formula: This represents the overlap ratio between two adjacent images. The area of ​​overlapping pixels between two images; This represents the total pixel area of ​​a single image. If the overlap rate between adjacent images is less than 70%, it is determined that the current image sequence has missing data. The drone is then controlled to selectively capture images of the missing areas, ultimately obtaining a complete sequence of image data that meets the overlap requirements. By verifying image integrity, the problem of model holes in subsequent 3D wall reconstruction can be avoided, improving the reliability of the entire set of detection data from the data source level.

[0023] In this embodiment, the feature extraction module incorporates a Scale Invariant Feature Transform (SIFT) algorithm. After the sequential image data is input into the module, the algorithm retrieves image extrema points in a multi-scale space, simultaneously calculates the local gradient direction histogram of pixels, and outputs a first feature image set containing information about image corners and edge details. When matching corresponding feature points based on the first feature image set, a fast nearest neighbor search algorithm is used to calculate the Euclidean distance between different image feature descriptors. A distance ratio threshold of 0.6 is set to eliminate cross-frame mismatched feature points, resulting in a high-precision matching feature point set. This processing method can accurately locate the corresponding pixel position of the same entity point on the wall in different shooting images, providing reliable basic data for subsequent three-dimensional spatial coordinate calculation.

[0024] Furthermore, based on the matching feature point set, the Structure for Motion Restoration (SfM) algorithm is used to solve the 3D point cloud of the exterior wall. The algorithm takes the matching feature point set and the intrinsic parameter matrix of the camera as input, and uses bundle adjustment to simultaneously optimize the camera's spatial pose parameters and 3D point coordinates to generate a high-density exterior wall surface point cloud. For the generated exterior wall surface point cloud, the Poisson surface reconstruction algorithm is used to fit a continuous geometric structure model of the wall. This algorithm uses discrete point clouds as input conditions and constructs a continuous and smooth solid geometric model by solving the Poisson equation. After the geometric structure model is completed, the 3D spatial coordinates of the model are spatially mapped one-to-one with the pixels of the first feature image set. Combined with the Support Vector Machine (SVM) classification algorithm, exterior wall defects are identified: wall cracks with a width ≥ 2mm, surface peeling with a single area > 5cm², and efflorescence defects are classified and labeled. The output is the distribution results of exterior wall defects with defect type and 3D spatial coordinate parameters, realizing the quantitative statistics and spatial positioning of building exterior wall defects.

[0025] Step S102: For the first feature image set, a dense matching algorithm is used to generate corresponding dense point cloud data and output to the coordinate mapping module to obtain the initial spatial coordinate set.

[0026] The specific execution process of this step is as follows: retrieve the generated first feature image set; apply a dense matching algorithm to the first feature image set to generate dense point cloud data; under the condition that the various indicators of the dense point cloud meet the preset judgment threshold, transmit the dense point cloud data to the coordinate mapping module; after receiving the point cloud data, the coordinate mapping module establishes an initial spatial coordinate set; extract the device image acquisition parameters from the initial spatial coordinate set and analyze the correlation between the acquisition parameters and the spatial density of the point cloud; when the image acquisition parameters and the point cloud density distribution pattern match, select a spatial coordinate transformation algorithm to solve the mapping function parameters; use the mapping function to complete the point coordinate accuracy constraint and encapsulate the coordinate results according to the preset output format.

[0027] In practice, after retrieving the first feature image set, a semi-global matching (SGM) algorithm is used to solve the disparity pixel by pixel, and pixel-level dense point cloud data of the exterior wall is generated based on the disparity results. The point cloud density quality control threshold is preset to no less than 500 spatial points per square meter. When the density of the dense point cloud generated in a local wall area meets the standard, the point cloud data of that area is pushed to the coordinate mapping module for coordinate conversion. After receiving the dense point cloud, the coordinate mapping module selects a fixed reference corner point on the exterior wall as the origin of the spatial coordinate system, constructs a global coordinate system, and generates an initial set of spatial coordinates. The device acquisition parameters such as the shooting distance of 12.5m and the camera tilt angle of 45° are extracted from the acquired data. The negative correlation between the shooting distance and the local point cloud density is quantified by a linear regression algorithm. If the measured point cloud density change at a shooting distance of 12.5m conforms to the preset distance attenuation model, the acquisition parameters are deemed to be effectively matched with the point cloud density distribution. After confirming the matching relationship, a seven-parameter coordinate transformation model is used to perform global coordinate conversion. The coordinate mapping function parameters are obtained by solving three translation parameters, three rotation parameters, and one scaling factor. Based on the mapping function, global registration is performed on all point clouds to control the spatial absolute position error of key feature points of the exterior wall within 0.02m, thus completing coordinate accuracy control. Finally, all coordinate data is packaged according to a preset standardized format to achieve standardized output of exterior wall spatial point data.

[0028] Step S103: Obtain the two-dimensional pixel position information corresponding to the defect identification result from the initial spatial coordinate set, and input it into the projection transformation module to obtain the projected three-dimensional coordinate point set.

[0029] The specific execution process of this step is as follows: Extract the two-dimensional pixel position information corresponding to the defect identification result from the initial spatial coordinate set; when the two-dimensional pixel information index reaches a preset threshold, send the pixel data to the projection transformation module for calculation; the projection transformation module converts the data into a projected three-dimensional coordinate point set; analyze the spatial distribution characteristics of the projected three-dimensional coordinate point set; determine whether the current point set belongs to the same defect area based on the spatial distribution characteristics; if it is determined to be the same defect area, extract the defect area boundary point set; and calculate the final three-dimensional positioning coordinates of the defect based on the boundary point set.

[0030] In practice, relying on the pre-defined mapping index relationship between 3D points and image pixels, the pixel row and column coordinates of the defect identification box area in the original high-definition image are retrieved in reverse. Taking a crack defect as an example, the effective pixel range corresponding to this defect is 1500.5~1520.0 on the horizontal coordinate and 2000.0~2050.5 on the vertical coordinate. The effective defect pixel area threshold is preset to 50 pixel units. When the total number of pixels corresponding to the defect reaches 120 pixels, the threshold requirement is met, and the projection transformation module is activated. The projection transformation module is based on the camera perspective projection mathematical model, combined with the camera calibration focal length and image principal point coordinate parameters, to transform the 2D pixel coordinates to the camera space coordinate system. In the example, all pixels are uniformly projected onto a reference space plane with a depth of 800.0mm to generate an initial projected 3D coordinate point set. Then, a spatial neighborhood search algorithm is used to count the neighborhood information of the points. With a search radius of 5.0mm, the number of neighboring points around each single point is counted one by one. The preset spatial connectivity coefficient threshold is 0.85. If the overall connectivity coefficient of the point set is >0.85, it is determined that all points belong to the same entity defect. After confirming the attribution, a 3D spatial edge extraction operator is used to retrieve points where the grid curvature changes abruptly. These abrupt points are then combined to form a defect boundary point set. Taking a pit defect as an example, its contour edge contains a total of 200 valid spatial points, which together constitute the boundary set. By calculating the spatial centroid of the boundary point set, the global precise coordinates of the defect can be obtained. Example defect location coordinates: X=450.2mm, Y=600.8mm, Z=120.5mm, completing the precise conversion of the defect from planar pixels to a 3D solid position.

[0031] Step S104: If the distance between the projected three-dimensional coordinate point set and the actual structural surface of the wall exceeds a preset threshold, the coordinate offset is adjusted by the error correction module to obtain the corrected three-dimensional coordinate point set.

[0032] The specific execution process of this step is as follows: retrieve the projected three-dimensional coordinate point set; extract the surface reconstruction mesh data corresponding to the wall entity structure; calculate the vertical distance from the three-dimensional coordinate point to the reconstructed wall surface point by point and generate the distance calculation result; when the distance result exceeds the preset threshold, generate an out-of-standard comparison mark; the error correction module calculates the three-axis coordinate offset adjustment amount according to the comparison mark; use the offset adjustment amount to correct the original three-dimensional coordinates item by item and generate the corrected three-dimensional coordinate point set; use the corrected point cloud to refresh the wall surface reconstruction data.

[0033] During implementation, the projected 3D coordinate point set and the reconstructed model data of the wall reference surface are retrieved simultaneously. The nearest neighbor search algorithm is used to calculate the vertical distance of a single point from the wall grid surface point by point. The formula for the vertical distance from a single point to the wall plane is: In the formula, This represents the perpendicular distance from a single point in three-dimensional space to the fitted plane of the wall, in millimeters (mm). For plane normal vector ( The three components of the equation together define the orientation and tilt angle of the wall plane. This is a plane constant term that controls the spatial position offset of the wall plane in the three-dimensional coordinate system. To fit the plane equation to the local area of ​​the wall; The magnitude of the normal vector is used to normalize the distance; the distance threshold is set as follows. When the measured maximum vertical distance of a single point When the deviation exceeds a threshold, a coordinate deviation exceeding the standard is generated, triggering the correction process. Upon receiving the exceeding standard indicator, the error correction module uses the Iterative Closest Point (ICP) registration algorithm, with the wall reference reconstruction model as the reference, to solve for the translation vectors and rotation matrices in the horizontal, vertical, and depth directions. After iterative optimization, the offset parameters are obtained: horizontal translation +1.2mm, vertical translation -0.8mm, and depth translation +2.5mm. These translation and rotation parameters are substituted into the original coordinates to complete the numerical correction, obtaining the corrected 3D coordinate point set. Finally, based on the corrected point cloud, the wall mesh model is reconstructed again using the Poisson surface reconstruction algorithm, updating the wall surface reconstruction data and achieving a closed-loop correction of point correction and model update.

[0034] Step S105: Perform surface fitting calculations based on the corrected three-dimensional coordinate point set and the defect identification results to obtain a three-dimensional mesh model of the defect.

[0035] For detailed instructions on this step, please refer to [link / reference]. Figure 3 As shown: Obtain the corrected 3D coordinate point set; extract the surface region corresponding to the defect from the defect identification result; verify the spatial correspondence between the defect region and the corrected 3D coordinate point set, and extract the defect boundary point set if they match; perform surface fitting operation based on the defect boundary point set to generate an initial surface model; statistically analyze the density distribution of the initial surface vertices and output the vertex distribution judgment result; perform mesh subdivision processing on the initial surface based on the vertex distribution result to generate a defect 3D mesh model; update the surface normal vectors of each vertex of the mesh model to obtain the final formed defect 3D mesh model.

[0036] During on-site implementation, the three-dimensional coordinate point set for correction and the defect identification results containing information on defects such as cracks and surface peeling are imported. The two-dimensional defect area in the identification results is mapped to three-dimensional space, and the spatially overlapping points are matched and screened. After the two are spatially matched, the rolling ball boundary extraction algorithm is used to delineate the defect boundary. The rolling ball radius of the algorithm is set to 0.05m. Based on the calculation results, a complete defect boundary point set is generated to accurately lock the damaged outline.

[0037] After obtaining the boundary point set, a non-uniform rational B-spline (NURBS) surface fitting algorithm is used to generate an initial defect surface model using the 3D coordinates of the boundary points as control points. The average distance between adjacent vertices of the initial surface is calculated. If the average distance is greater than 0.02m, the vertex distribution of the model is determined to be sparse. For sparse surfaces, a butterfly mesh subdivision algorithm is used to iteratively refine the mesh, inserting new vertices at the edges of existing triangles to improve the fineness of the patches and generate a refined 3D defect mesh model. Finally, principal component analysis is used to traverse the neighborhood of each vertex in the mesh, solve for the covariance matrix, eigenvalues, and eigenvectors, and refresh the surface normal vectors based on the calculation results to obtain the final 3D defect mesh model that accurately reflects the concave and convex morphology of the defect.

[0038] Step S106: Extract the boundary contour lines of the defect 3D mesh model and input them into the direction calculation module to obtain the crack propagation direction vector.

[0039] The specific execution process of this step is as follows: Summarize all vertex coordinate data of the defect 3D mesh model; filter mesh contour edges based on the vertex coordinate set and stitch them together to form the defect boundary contour line; perform closure verification on the boundary contour line and filter valid closed contour sequences; extract the vertex normal vectors of the closed contour and project them onto the reference plane to form a projection vector group; send the projection vector group into the direction calculation module to solve for the crack extension direction vector; calculate the spatial angle between the extension direction vector and the crack main axis and obtain the angle value; when the angle is less than a preset threshold, output the crack extension trend judgment conclusion.

[0040] During implementation, the vertex coordinates corresponding to all faces of the defect's 3D mesh are fully collected. Using a half-edge data structure algorithm, contour edges attached to only a single triangular face are selected and connected end-to-end according to spatial topological order to form continuous boundary contours. A depth-first search algorithm is used to trace the beginning and end endpoints of the contours. If the spatial Euclidean distance between two points is less than 0.5 mm, the contour segment is considered closed. All closed contours are selected to form a closed contour sequence, and invalid pseudo-contours generated by scanning noise are discarded.

[0041] Using the local fitting plane of the crack as the projection reference plane, the spatial normal vectors of all vertices on the closed contour are orthogonally projected onto the reference plane, and the two-dimensional components of the projection are extracted to construct a projection vector set. The direction calculation module uses principal component analysis to reduce the dimensionality of the vector set, and selects the eigenvector corresponding to the maximum variance of the data as the local extension direction vector of the crack. The spatial angle between the extension vector and the overall principal axis of the crack is solved by vector dot product operation, with a preset angle threshold of 15°. When the measured angle is less than 15°, it is determined that the crack has a single linear extension trend and there is no bifurcation or network cracking phenomenon. This judgment result provides a quantitative basis for the assessment of the safety of building structures.

[0042] Step S107: Obtain depth information from the crack extension direction vector and the 3D mesh model of the defect, and determine the numerical value of the defect depth range through the depth fusion module.

[0043] For detailed instructions on this step, please refer to [link / reference]. Figure 2 As shown: Vector decomposition is performed on the crack extension direction vector to extract the depth component; the original depth data is collected by combining the depth component with the 3D mesh model of the defect to generate depth information acquisition results; the depth fusion module performs fusion operation on the original depth data to generate a continuous depth numerical sequence; the difference between adjacent values ​​in the sequence is compared segment by segment, and the depth boundary judgment point is marked if the difference exceeds the preset threshold; the effective depth numerical set is selected from the full depth sequence based on the boundary judgment point; the upper and lower limit intervals of the defect depth are statistically determined based on the effective set, and the final depth range data is output.

[0044] In practice, the crack extension direction vector is split, and the depth component pointing towards the interior of the wall is extracted. For example, the projection coefficient of this component on the vertical axis is 0.85. Using this direction component as a reference, a ray casting algorithm is used to sample within the 3D mesh to obtain the original depth data of the crack from the wall surface to the bottom. The depth fusion module uses median filtering and Gaussian weighting algorithms to smooth the discrete original depths, eliminating single-point noise interference and generating a continuous depth numerical sequence. For example, the discrete depth values ​​of 0.45, 0.48, and 0.62 at the crack bifurcation are transformed into a continuous ordered sequence after fusion.

[0045] A threshold of 0.03 is preset for the difference between adjacent depth values. When the difference between adjacent data in a sequence exceeds the limit, the corresponding point is marked as a depth boundary judgment point. The boundary point is used to distinguish the interval between sudden changes in crack depth. With the starting boundary of 0.15 and the ending boundary of 0.75 as the limit, all effective depth data within the interval are extracted to form a depth value set. Combining the extreme value data within the set with the local curvature correction error of the grid, the final depth range of the crack is determined. The example effective depth range is 0.15m to 0.75m, which quantifies the longitudinal dimension of the defect.

[0046] Step S108: Generate the final repair range delineation data based on the defect depth range value and the crack extension direction vector to output the shape and location of the defect in real space.

[0047] The specific execution process of this step is as follows: decompose the crack extension direction vector and defect depth range parameters, and extract the three-dimensional spatial coordinate components; map the spatial coordinate components onto the surface of the defect three-dimensional mesh to generate point location mapping results; the deep fusion module performs weighted operations on the mapped points to generate a continuous spatial sequence set; compare the Euclidean distance difference between adjacent points in the spatial sequence, and mark the shape boundary judgment points for points exceeding the threshold; filter the effective shape value set through the boundary judgment points; generate the final repair range delineation data based on the shape value set; and output the spatial dimensions and boundary parameters corresponding to the repair range.

[0048] During the implementation phase, the crack extension vector and defect depth parameters are correlated, and the three-dimensional spatial coordinate components are obtained by using a vector decomposition algorithm. In the example, the horizontal projection coefficient of the crack extension vector is 0.75, and the depth range is 3.5mm to 8.2mm, from which the spatial trajectory coordinate components are obtained. The coordinate components are projected onto the surface of the defect's three-dimensional mesh using a ray projection algorithm, and the intersection points of the components and triangular facets are solved. The normal vectors and curvature parameters of the points are recorded simultaneously to complete the position mapping.

[0049] The deep fusion module employs a multi-point weighted fusion algorithm to smoothly fit discrete mapping points. It optimizes coordinate values ​​based on single-point confidence weights (e.g., 0.85, 0.92) to generate a continuous spatial point sequence. A threshold of 0.45 for the Euclidean distance difference between sequence points is set, and the coordinate abrupt change difference at crack inflection points is increased from 0.1 to 0.6. Points exceeding the limit are marked as shape boundary judgment points. A shape value set is formed by summing all boundary judgment points. A convex hull algorithm is used to perform envelope operations on the set of points, connecting the outer key coordinate points to form a closed polygon repair area. An example shows a 25.4 mm² repair area enclosed by 15 boundary coordinate points, generating standardized repair delineation data. This data can be directly imported into automated repair equipment to guide on-site construction layout and material calculation, fully realizing the output of all parameters related to defect spatial morphology, location, and repair scope.

[0050] This invention acquires multiple sets of overlapping image data of the building's exterior wall surface using a multi-angle image acquisition device and inputs them into a feature extraction module to obtain a first feature image set. A dense matching algorithm is then used to generate corresponding dense point cloud data for the first feature image set, which is output to a coordinate mapping module to obtain an initial spatial coordinate set. Two-dimensional pixel position information corresponding to the defect identification result is obtained from the initial spatial coordinate set and input into a projection transformation module to obtain a projected three-dimensional coordinate point set. If the distance between the projected three-dimensional coordinate point set and the actual structural surface of the wall exceeds a preset threshold, the coordinate offset is adjusted through an error correction module to obtain a corrected three-dimensional coordinate point set. A surface fitting operation is performed based on the corrected three-dimensional coordinate point set and the defect identification result to obtain a three-dimensional mesh model of the defect. The boundary contour lines of the three-dimensional mesh model are extracted and input into a direction calculation module to obtain a crack extension direction vector. Depth information is obtained from the crack extension direction vector and the three-dimensional mesh model of the defect, and the defect depth range value is determined through a depth fusion module. Finally, based on the defect depth range value and the crack extension direction vector, final repair range delineation data is generated to output the shape and location of the defect in real space. This invention organically integrates image dense matching, projection transformation error correction, three-dimensional mesh fitting, crack direction calculation and depth fusion technology, solving the core business problems in traditional defect detection such as three-dimensional reconstruction distortion caused by coordinate deviation, difficulty in quantifying crack depth and inability to accurately define the repair range. It realizes accurate reconstruction and intelligent judgment of building exterior wall defects from two-dimensional images to the three-dimensional shape, position and depth range in real space, greatly improving the scientific nature and work efficiency of exterior wall repair.

[0051] The above description is only a preferred embodiment of the present invention. It should be noted that those skilled in the art can make several modifications and improvements without departing from the inventive concept of the present invention, and these all fall within the protection scope of the present invention.

Claims

1. A method for detecting construction quality defects in building exterior walls, characterized in that, The method includes: Collect multiple sets of overlapping image data of the building's exterior walls and input them into the feature extraction module to generate the first feature image set; The first feature image set is processed using a dense matching algorithm to generate dense point cloud data, and the initial spatial coordinate set is obtained by the coordinate mapping module. Based on the initial spatial coordinate set, extract the two-dimensional pixel coordinate information corresponding to the defect identification result, and generate a projected three-dimensional coordinate point set through the projection transformation module; Calculate the spatial distance between the projected three-dimensional coordinate point set and the wall reference structural surface. When the spatial distance exceeds a preset threshold, correct the coordinate offset parameters through the error correction module and generate a corrected three-dimensional coordinate point set. Based on the corrected three-dimensional coordinate point set and the defect identification results, surface fitting is performed to construct a three-dimensional mesh model of the defect. The boundary contour information of the defect 3D mesh model is extracted and sent to the direction calculation module to solve for the crack propagation direction vector; The original depth data is extracted by combining the crack extension direction vector with the defect three-dimensional mesh model, and the defect depth range parameters are determined by the depth fusion module. Based on the defect depth range parameters and crack extension direction vector, repair range calibration data is generated, and the geometric shape and spatial location parameters of the defect in the real-world space are output.

2. The method for detecting construction quality defects in building exterior walls according to claim 1, characterized in that, The process of collecting multiple sets of overlapping image data of the building's exterior walls and inputting them into the feature extraction module to generate a first feature image set includes: A drone equipped with a multi-lens camera was used to collect multi-view overlapping image source data of building exterior walls through gridded flight. The sequence integrity of the multi-view overlapping image source data was verified, and valid sequence images were selected. The first feature image set was obtained by extracting features from the valid sequence images through a feature extraction module.

3. The method for detecting construction quality defects in building exterior walls according to claim 1, characterized in that, The process of generating dense point cloud data by processing the first feature image set using a dense matching algorithm, and obtaining an initial spatial coordinate set through a coordinate mapping module, includes: The first feature image set is retrieved, and pixel-level disparity is calculated using a dense matching algorithm to generate dense point cloud data. When the dense point cloud index meets the preset judgment threshold, the dense point cloud data is imported into the coordinate mapping module. The coordinate mapping module establishes a coordinate system with the reference point of the building's exterior wall as the origin and generates an initial set of spatial coordinates.

4. The method for detecting construction quality defects in building exterior walls according to claim 3, characterized in that, The generation of the initial spatial coordinate set further includes: Image acquisition distance and pitch angle acquisition parameters are extracted from the initial spatial coordinate set. The correlation between the acquisition parameters and the spatial density of the point cloud is analyzed by linear regression algorithm. When the image acquisition parameters and the spatial density distribution of the point cloud match, the mapping function parameters are solved by spatial coordinate transformation algorithm. The mapping function parameters are used to complete the point coordinate accuracy constraint, control the absolute coordinate error of the point to be ≤0.02m, and standardize the output coordinate set.

5. The method for detecting construction quality defects in building exterior walls according to claim 1, characterized in that, Based on the initial spatial coordinate set, the two-dimensional pixel coordinate information corresponding to the defect identification result is extracted, and then mapped by the projection transformation module to generate a projected three-dimensional coordinate point set, including: The two-dimensional pixel coordinate information corresponding to the defect identification result is retrieved from the initial spatial coordinate set. When the effective index of the two-dimensional pixel coordinate reaches the preset judgment threshold, the projection transformation module is activated to perform spatial transformation on the two-dimensional pixel coordinate, and the projection three-dimensional coordinate point set is obtained by solving the projection transformation. The spatial topological distribution characteristics of the projected 3D coordinate point set are analyzed; based on the spatial topological distribution characteristics, it is determined whether the projected 3D coordinate point set belongs to the same defect region. If it belongs to the same defect region, the boundary point set of the defect region is determined, and the final 3D defect position coordinates are obtained from the boundary point set of the defect region.

6. The method for detecting construction quality defects in building exterior walls according to claim 1, characterized in that, Calculate the spatial distance between the projected three-dimensional coordinate point set and the wall reference structural surface. When the spatial distance exceeds a preset threshold, correct the coordinate offset parameters through an error correction module to generate a corrected three-dimensional coordinate point set, including: Extract the projected 3D coordinate point set and the reconstructed mesh data of the wall reference structural surface respectively; use the nearest neighbor algorithm to calculate the vertical spatial distance from the coordinate point to the wall reference surface point by point; if the distance calculation result exceeds the preset error threshold, generate a correction trigger flag; The error correction module calculates the coordinate offset correction amount for each dimension based on the trigger identifier; the offset correction amount is used to iteratively update the projected three-dimensional coordinate point set to obtain the corrected three-dimensional coordinate point set; the corrected point cloud is used to refresh the wall surface reconstruction model data.

7. The method for detecting construction quality defects in building exterior walls according to claim 1, characterized in that, Based on the corrected three-dimensional coordinate point set and the defect identification results, surface fitting is performed to construct a three-dimensional mesh model of the defect, including: Obtain the set of calibrated 3D coordinate points and extract the corresponding 2D region of the defect from the defect identification results; establish the spatial relationship between the defect region and the 3D point cloud and delineate the defect boundary point set; perform surface fitting with the boundary point set as a constraint to obtain the initial surface model; statistically analyze the vertex discrete distribution index of the initial surface model to determine the vertex density state; perform mesh subdivision processing on the initial surface model based on the vertex density state; update the mesh patch normal vector parameters to generate the final 3D mesh model of the defect.

8. The method for detecting construction quality defects in building exterior walls according to claim 1, characterized in that, Extract the boundary contour information of the three-dimensional mesh model of the defect and send it to the direction calculation module to solve for the crack propagation direction vector, including: Obtain the vertex coordinate set of the 3D mesh model of the defect; filter the mesh contour edges based on the vertex coordinate set and stitch them together to form the defect boundary contour line; perform closure verification on the boundary contour line and filter the effective closed contour sequence; obtain the vertex normal vector of the closed contour from the closed contour sequence and project it onto the reference plane to construct the projection vector group; send the projection vector group into the direction calculation module to solve the crack extension direction vector; calculate the spatial angle between the crack extension direction vector and the crack principal axis; if the angle is less than the preset angle threshold, output the crack extension trend judgment conclusion.

9. The method for detecting construction quality defects in building exterior walls according to claim 1, characterized in that, The original depth data is extracted by combining the crack extension direction vector with the defect 3D mesh model, and the defect depth range parameters are determined by the depth fusion module, including: Vector decomposition is performed on the crack extension direction vector to extract the depth component; raw depth data is collected by combining the depth component with the 3D mesh model of the defect; the raw depth data is filtered and fused by the depth fusion module to generate a continuous depth numerical sequence; the difference between adjacent values ​​in the sequence is compared segment by segment, and if the difference exceeds a preset threshold, a depth boundary judgment point is marked; a set of valid depth values ​​is selected from the depth numerical sequence based on the depth boundary judgment points; the upper and lower limit intervals of the defect depth are statistically determined based on the set of valid depth values, and the final defect depth interval parameters are output.

10. The method for detecting construction quality defects in building exterior walls according to claim 1, characterized in that, Based on the defect depth range parameters and crack propagation direction vector, repair range calibration data is generated, and the geometric shape and spatial location parameters of the defect in the real-world space are output, including: The crack extension direction vector and defect depth range parameters are separated to extract three-dimensional spatial coordinate components. These three-dimensional spatial coordinate components are then mapped onto the three-dimensional mesh surface of the defect to generate point mapping data. The mapped points are then smoothed using a deep fusion module to generate a spatial point sequence. The spatial difference between adjacent points in the sequence is compared with a preset threshold. Points exceeding the threshold are marked as shape boundary judgment points. Valid shape value sets are filtered using these shape boundary judgment points. The final repair range delineation data is generated based on the shape value set. The spatial dimensions and boundary parameters corresponding to the repair range are then output.

Images