A method for identifying diseases of ancient building wood components based on three-dimensional point cloud semantic segmentation

Abstract

The application discloses a kind of ancient building woodwork disease identification method based on three-dimensional point cloud semantic segmentation, belong to ancient building protection and three-dimensional point cloud processing technical field.This method completes multi-view point cloud registration and records acquisition attribute, establishes component structure coordinate system, and demarcates normal structure constraint area with type identification;Stable point is screened, and local disease-free reference surface is iteratively fitted, and surface directed deviation is calculated;According to the reliability of wood grain direction, an adaptive multi-scale neighborhood is constructed, and a disease-sensitive feature vector is extracted;The initial semantic label is output by the multi-channel point cloud semantic segmentation model containing the gate fusion unit, and the label is corrected by combining the geodesic adjacency graph and the structural prior rule to obtain the disease point cloud region.The application can effectively distinguish between the inherent structure of the component and the real disease, adapt to uneven point cloud quality, and improve the accuracy and stability of complex wood component disease identification.

Description

Technical Field

[0001] This invention relates to the field of ancient building conservation and three-dimensional point cloud processing technology, and in particular to a method for identifying defects in wooden components of ancient buildings based on three-dimensional point cloud semantic segmentation. Technical Background

[0002] During their long service life, wooden components of ancient buildings are prone to damage such as cracks, insect infestation, decay, surface defects, and overall deformation due to environmental changes in temperature and humidity, biological erosion, load-bearing effects, and material aging. Three-dimensional point cloud detection can record the surface morphology of components in a non-contact manner and is increasingly being used for routine inspections and conservation assessments of ancient buildings.

[0003] Existing methods for identifying defects often rely on local geometric changes or apparent differences to determine abnormal areas. However, ancient architectural wooden components typically retain traces of traditional processing, assembly connections, and irregular shapes. These inherent forms and defects such as depressions, missing parts, or breaks may exhibit similar spatial characteristics in point clouds, easily leading to confusion between normal and defective areas. Natural wood grain and age-related deformation can also cause similar defects to exhibit different directions and scales of extension in different components and locations, making it difficult to simultaneously account for both minor defects and continuous defects within a fixed analysis range. Furthermore, limitations imposed by site space, occlusion relationships, and scanning conditions result in inconsistent data completeness and measurement quality across different areas of the same component. Local missing parts, noise, and registration deviations further disrupt defect boundaries and continuous forms. Therefore, existing technologies struggle to reliably distinguish between inherent structural changes and actual defects when component shapes are complex and point cloud quality is uneven, leading to misjudgments, missed detections, and fragmented regional identification results.

[0004] Therefore, how to improve the accuracy and stability of three-dimensional point cloud disease identification of ancient building wooden components is a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0005] In view of this, embodiments of the present invention provide a method for identifying defects in ancient building wooden components based on three-dimensional point cloud semantic segmentation, in order to solve or alleviate the technical problems existing in the prior art.

[0006] The technical solution of this invention is implemented as follows: A method for identifying defects in ancient architectural wooden components based on three-dimensional point cloud semantic segmentation, comprising the following steps:

[0007] S1. Acquire and register multi-view point clouds of wooden components, and record the scanning position and direction, incident angle, number of effective viewpoints and registration residuals;

[0008] S2. Extract the axis, cross-sectional profile and normal, establish the structural coordinate system, identify normal construction and form a normal structural constraint region with type identification;

[0009] S3. Select stable points outside the normal structural constraint area where the number of effective viewpoints reaches the threshold and the registration residual and normal dispersion are lower than the threshold. Fit the reference surface and iteratively remove residual abnormal points to obtain a local disease-free reference surface. Calculate the directional deviation and record the fitting residual.

[0010] S4. Determine the wood grain direction and confidence level. When the confidence level reaches the threshold, establish an anisotropic multi-scale neighborhood with a larger scale along the wood grain direction. When the confidence level is lower than the threshold, establish the neighborhood with the principal curvature direction. When the direction is unstable, establish an equal-scale neighborhood. Extract geometric, appearance and structural location features to form a disease-sensitive feature vector.

[0011] S5. Generate the point cloud acquisition reliability based on point cloud density, scanning incident angle, number of effective viewing angles and fitting residuals. Input this vector into the point cloud semantic segmentation model that includes geometric defects, material appearance, structural prior channels and gated fusion units. The gated fusion unit generates point-level weights and fuses channel features based on acquisition reliability and wood grain direction confidence, and outputs category probabilities and initial semantic labels.

[0012] S6. Establish an adjacency graph based on surface geodesy and correct the connected regions of the initial semantic labels; according to the structural type identifier and the parameter statistics of the corresponding normal structure, adjust the regions that conform to the normal structure parameter range to be labeled as disease-free, retain the regions that conform to the disease rules in terms of connected morphology, directional deviation changes and wood grain direction relationship as disease labels, and mark the remaining regions as regions to be reviewed, thus obtaining the disease point cloud region.

[0013] Compared with the prior art, the present invention has the following advantages:

[0014] This invention establishes a structural coordinate system based on the axis, cross-sectional profile, normal, and boundary of the wooden component. It also identifies geometrically abrupt regions that repeat from multiple perspectives, extend regularly, and maintain a stable spatial relationship with the wooden component structure as normal structural constraint regions. This allows changes in mortise and tenon boundaries, component edges, end profiles, and regular cross-sections to be distinguished from abnormal surface changes before defects are identified, reducing the chances of normal structures being misidentified as cracks, insect holes, or surface erosion.

[0015] This invention selects stable points outside the normal structural constraint area that meet the requirements for quality and surface condition. By iteratively eliminating abnormal points with residuals, it fits a local disease-free reference surface and calculates the directional deviation of the actual surface of the wooden component relative to the reference surface. This transforms disease identification from simply judging local curvature or surface roughness to judging depressions, material loss, and continuous deformation by combining the normal surface morphology of the wooden component. This helps to reduce the interference of natural surface undulations and age-related deformation on disease feature extraction.

[0016] This invention selects anisotropic multi-scale neighborhoods set along the wood grain direction, anisotropic multi-scale neighborhoods set along the principal curvature direction, or equal-scale multi-scale neighborhoods based on the reliability of the wood grain direction. Geometric, apparent, and structural position features are extracted within the corresponding neighborhoods, so that the feature extraction range can adapt to changes in wood grain direction, local surface orientation, and disease scale. It retains the extension features of narrow diseases such as cracks, while also taking into account disease areas with different spatial forms such as insect holes, decay, and surface erosion.

[0017] This invention generates point cloud acquisition reliability based on point cloud density, scanning incident angle, number of effective viewing angles, and fitting residuals. It also performs gated fusion of point-level features in geometric disease channels, material appearance channels, and structural prior channels by combining wood grain direction credibility. After obtaining initial semantic labels, it performs regional correction using surface geodesic adjacency relations, normal structural parameter ranges, and disease morphology rules. This prevents local abnormal features generated by low-quality point clouds from participating in classification with fixed weights and reduces the fragmentation of initial semantic labels and the erroneous merging of adjacent disease areas. As a result, it improves the accuracy and stability of disease identification results under complex component morphology and non-uniform scanning conditions. Attached Figure Description

[0018] The accompanying drawings are provided to further understand the present invention and form part of the specification. They are used together with the embodiments of the present invention to explain the invention and do not constitute a limitation thereof.

[0019] Figure 1 This is an overall flowchart of the method for identifying defects in wooden components of ancient buildings according to the present invention;

[0020] Figure 2 This is a flowchart of the multi-view point cloud acquisition, registration, and normal structure constraint region formation process of the present invention.

[0021] Figure 3 This is a flowchart of the stable point screening and local disease-free reference surface fitting process of the present invention;

[0022] Figure 4 This is a flowchart of the adaptive multi-scale neighborhood formation process in this invention;

[0023] Figure 5 This is a schematic diagram of the multi-channel gated semantic segmentation model in this invention;

[0024] Figure 6 This is a flowchart of the surface geodesic adjacency map and label correction process in this invention;

[0025] Figure 7 This is a flowchart of the supplementary data collection and local re-identification of low-confidence disease areas in this invention. Detailed Implementation

[0026] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0027] This invention uses multi-view point clouds of ancient architectural wooden components and their corresponding scanning poses and acquisition attributes as raw inputs. First, point clouds acquired from different scanning positions are registered, duplicate points are merged, and outliers are removed to form a unified coordinate system for the wooden component point cloud. Then, based on the type of wooden component, the axis, cross-sectional contour, and normal are extracted to establish a structural coordinate system, identifying normal structures and forming a normal structural constraint region with type identifiers. Subsequently, stable points are selected outside the normal structural constraint region, and a local, disease-free reference surface is fitted. The directional deviation of each point relative to the local, disease-free reference surface is calculated, and the fitting residual is recorded. Anisotropic multi-scale or equal-scale neighborhoods are established based on the wood grain direction reliability, and geometric, apparent, and structural positional features are extracted to form a disease-sensitive feature vector. Point cloud acquisition reliability is generated based on point cloud density, scanning incident angle, number of effective viewing angles, and fitting residuals. Point-level fusion of different feature channels is controlled by the point cloud acquisition reliability and wood grain direction reliability, outputting category probabilities and initial semantic labels. Finally, the initial semantic labels are corrected based on surface geodesy relations, the range of normal structural parameters, and disease rules to obtain the diseased point cloud region.

[0028] The ancient architectural wooden components include beams, purlins, rafters, columns, brackets, and irregularly shaped wooden components with irregular surfaces. Each point in the multi-view point cloud includes at least three-dimensional coordinates; when the acquisition device has the capability to acquire surface data, each point also includes reflection intensity, grayscale, or color. After registration is completed at different scanning perspectives, the three-dimensional coordinates, acquisition attributes, structural coordinates, disease sensitivity features, category probability, and semantic labels of each point are stored in association according to a unified point index.

[0029] Example 1

[0030] This embodiment provides a method for identifying defects in ancient building wooden components based on three-dimensional point cloud semantic segmentation, including the following steps.

[0031] S1. Obtain and register multi-view point clouds of wooden components:

[0032] Multiple point clouds of the wooden components to be identified are collected from multiple scanning positions, and the scanning position, scanning direction and calibration information of the acquisition equipment are recorded.

[0033] Based on the calibration points, stable structural feature points, and overlapping point clouds of adjacent scanning areas, the initial correspondence between point clouds from different scanning perspectives is established. The initial rotation matrix and translation vector of each scanning perspective relative to the reference coordinate system are calculated to complete the coarse registration of multi-view point clouds.

[0034] Based on the coarse registration results, the corresponding points within the overlapping area are iteratively updated using the point-to-plane distance between overlapping point clouds as the registration target. The rotation matrix and translation vector are then corrected based on the updated corresponding points. Iteration stops when the pose change and registration error change between two adjacent iterations are lower than the corresponding thresholds, and the point clouds from each scanning viewpoint are transformed to a unified coordinate system.

[0035] The scanning position and direction of each point are determined based on the location of the acquisition device and the scanning direction. The scanning incident angle is determined based on the relationship between the scanning light direction and the point normal. The number of scanning angles that are effectively observed at the same surface position is counted to obtain the number of effective angles. The registration residual is recorded based on the distance between the corresponding points after registration.

[0036] The registration point cloud is processed by merging duplicate points and removing outliers. Duplicate points with a spatial distance below the merging threshold and the same normal direction are merged according to a unified point index, while retaining the acquisition attributes corresponding to different scanning perspectives. Isolated points with a distance exceeding the threshold from neighboring points and not repeatedly observed by other scanning perspectives are removed from the registration point cloud.

[0037] Step S1 takes the original point clouds obtained from different scanning positions as input and outputs a multi-view point cloud of the wooden component in a unified coordinate system, as well as the scanning position and direction associated with each point, the scanning incident angle, the number of effective viewing angles, and the registration residual. The above output serves as the input for the structural analysis in step S2 and the stable point selection in step S3.

[0038] S2. Establish the structural coordinate system and form the normal structural constraint region:

[0039] Read the registration point cloud output in step S1, calculate the local normal of each point, and extract the component axis and cross-sectional profile based on the overall extension direction and cross-sectional shape of the wooden component.

[0040] For beams, purlins, and rafters, the main extension direction of the component is taken as the axis, and multiple cross-sectional profiles are extracted along the axis. For column components, the column centerline is fitted according to the cross-sectional center at different height positions, and the column centerline is taken as the axis. For brackets and irregular components, the surface is divided into segments according to the branch direction of the point cloud skeleton and the local surface orientation, and a segmented structural coordinate system is established for each segment.

[0041] The structural coordinates of each point are determined based on the component's axis, cross-sectional center, outer surface, ends, edges, and tenon / mortise boundaries. These structural coordinates include the axial position, the position relative to the cross-sectional center, the normal position relative to the outer surface of the component, and the distance to the ends, edges, and tenon / mortise boundaries.

[0042] Geometric abrupt change regions are detected in the registered point cloud, and their spatial location, extension direction, and relationship with the component axis or cross-sectional profile are compared in different scanning views. Geometric abrupt change regions that recur in multiple scanning views, extend regularly, and have a fixed spatial relationship with the component axis or cross-sectional profile are identified as normal structures.

[0043] A normal structural constraint region is formed based on the point index and spatial extent covered by the normal structure, and a structural type identifier is assigned based on the spatial location and geometry of the normal structure. The structural type identifier includes at least one of the following: component edge, end boundary, tenon and mortise boundary, component joint, carved outline, and regular cross-sectional variation.

[0044] Step S2 takes the registered point cloud and its acquired attributes as input and outputs the component axis, cross-sectional profile, point normal, structural coordinates, and a normal structural constraint region with type identification. The normal structural constraint region is used for stable point screening and reference surface fitting in step S3, and for normal structural parameter statistics and label correction in step S6.

[0045] S3. Fit a locally disease-free reference surface and calculate the directional deviation:

[0046] Read the normal structural constraint region output in step S2, and divide the surface of the wooden component outside the normal structural constraint region into overlapping point cloud blocks.

[0047] For each point cloud block, the effective number of viewpoints, registration residuals, local normals, local curvature, and surface roughness are read. Points with an effective number of viewpoints reaching a threshold, registration residuals below a threshold, normal dispersion below a threshold, and local curvature and surface roughness within a stable range are identified as stable points, forming a set of stable points for the corresponding point cloud block.

[0048] The fitting constraints of the reference surface are selected according to the type of wooden component. For beams and purlins, planar or regular section constraints are applied to the reference surface; for columns, cylindrical or elliptical cylindrical constraints are applied to the reference surface; for brackets and irregularly shaped components, segmented surface fitting is performed based on the surface segments formed in step S2.

[0049] Within the overlapping range of adjacent point cloud blocks, positional continuity and normal continuity constraints are set to ensure that the reference surface formed by fitting adjacent point cloud blocks is continuously connected in spatial position and surface direction.

[0050] After completing one reference surface fitting, the fitting residual of the stable points relative to the current reference surface is calculated. Stable points whose fitting residuals exceed a threshold are removed from the stable point set, and the reference surface is refitted using the remaining stable points. This process of calculating fitting residuals, removing abnormal residual points, and refitting the reference surface is repeated until the change in the stable point set and the change in fitting residuals are both below the corresponding thresholds, resulting in a locally defect-free reference surface.

[0051] For each point in the point cloud, determine its corresponding position on a locally defect-free reference surface, and calculate the directional deviation of that point relative to the reference surface based on the reference surface normal. When the outer normal of the reference surface is taken as the positive direction, deviations towards the interior of the component are recorded as inward deviations, and deviations towards the exterior of the component are recorded as outward deviations. Simultaneously, record the fitting residuals of each point relative to the locally defect-free reference surface.

[0052] Step S3 takes the point cloud and acquired attributes outside the normal structural constraint region as input, and outputs the local disease-free reference surface, the directed deviation of each point, and the fitting residual. These outputs serve as inputs for step S4 (disease-sensitive feature extraction) and step S5 (point cloud acquisition reliability generation).

[0053] S4. Determine the wood grain direction and form a disease-sensitive feature vector:

[0054] The color texture gradient direction, reflection intensity gradient direction, and local geometric principal direction are calculated in the surface tangent plane at each point, and the directional consistency of each candidate direction in the neighborhood is analyzed.

[0055] Based on the completeness of color or reflection intensity information, the stability of local geometric orientation, and the neighborhood consistency of candidate orientations, the wood grain orientation and the corresponding wood grain orientation confidence level of each point are determined.

[0056] When the confidence level of the wood grain direction reaches the threshold, an anisotropic multi-scale neighborhood with a large scale along the wood grain direction is established, with the wood grain direction as the major axis, the surface tangent perpendicular to the wood grain direction as the minor axis, and the surface normal as the thickness direction.

[0057] When the confidence level of the wood grain direction is lower than a threshold, the principal curvature direction of the corresponding point is calculated, and an anisotropic multi-scale neighborhood is established using the principal curvature direction as the major axis direction. When the dispersion of the candidate principal curvature direction in the neighborhood exceeds a threshold, the direction is determined as an unstable direction, and an equal-scale multi-scale neighborhood is established.

[0058] Within small-scale, medium-scale, and large-scale neighborhoods, calculate the directional deviation and its gradient, the angle between the point normal and the normal of the local undamaged reference surface, the dispersion of the neighborhood normal, curvature, surface roughness, concavity depth continuity, reflection intensity or color difference, distance to the normal structural constraint region, and the angle between the neighborhood main extension direction and the wood grain direction.

[0059] By combining the calculation results at each scale with the wood grain direction reliability using a unified point index, a disease-sensitive feature vector is formed for each point. This disease-sensitive feature vector includes geometric features, material appearance features, and structural location features.

[0060] Step S4 takes a local disease-free reference surface, directional deviation, point cloud appearance data, and structural coordinates as input, and outputs the wood grain direction, wood grain direction confidence, multi-scale neighborhood, and disease-sensitive feature vector. The disease-sensitive feature vector serves as the input to the point cloud semantic segmentation model in step S5.

[0061] S5. Generate point cloud acquisition reliability and output initial semantic labels:

[0062] Read the point cloud density, scan incidence angle, number of effective viewpoints and fitting residuals of each point, and convert the above-mentioned collected attributes into data of a uniform scale.

[0063] The reliability of point cloud acquisition for each point is generated based on point cloud density, scanning incident angle, number of effective viewing angles, and fitting residuals. Higher point cloud acquisition reliability is generated when point cloud density is high, scanning incident angle is small, number of effective viewing angles is large, and fitting residuals are low; lower point cloud acquisition reliability is generated when there is occlusion, single-view coverage, large scanning incident angle, or high fitting residuals.

[0064] The defect-sensitive feature vector formed in step S4 is input into the point cloud semantic segmentation model. The point cloud semantic segmentation model includes a geometric defect channel, a material appearance channel, a structural prior channel, and a gated fusion unit.

[0065] Geometric defects channel coding includes directional deviation, normal variation, curvature, surface roughness, and indentation depth continuity; material apparent channel coding includes reflection intensity, grayscale, or color; structural a priori channel coding includes structural coordinates, normal structural constraint area identification, and distance to component boundaries.

[0066] The gated fusion unit reads the reliability of point cloud acquisition and the credibility of wood grain direction, maps the two into point-level weights corresponding to geometric defects channel, material appearance channel and structural prior channel respectively, and normalizes the point-level weights.

[0067] Based on the normalized point-level weights, the point-level features output from the three channels are weighted and fused to form a fused feature. The fused feature is then propagated through neighborhood features, and the propagated features are restored to the original point cloud resolution and original point indices.

[0068] Based on the restored point-level features, output the category probability of each point belonging to the categories of no disease, crack, insect-eaten hole, decay, surface erosion and deformation, and determine the category with the highest category probability as the initial semantic label of the corresponding point.

[0069] Step S5 takes the disease-sensitive feature vector, point cloud acquisition reliability, and wood grain direction confidence as input, and outputs the category probability and initial semantic label for each point. These outputs serve as input for step S6 to establish the initial label connected regions and perform label correction.

[0070] S6. Correct the initial semantic labels and obtain the disease point cloud region:

[0071] An adjacency graph is constructed based on the continuous connectivity of the wooden component surface. The nodes in the adjacency graph correspond one-to-one with the points in the point cloud, and the connecting edges are used to represent the connection relationship between adjacent points on the surface.

[0072] For candidate adjacent points, read the surface geodesic distance, normal difference, directional deviation difference, and wood grain direction consistency between them, and determine whether the surface connection path between candidate adjacent points crosses the normal structural constraint area. Based on this, calculate the connection weight of the adjacency graph.

[0073] Based on the adjacency graph connection weights, points with the same or similar initial semantic labels and continuous surface connections are aggregated into initially labeled connected regions. For candidate adjacent points whose connection paths cross normal structural constraint regions, their connection weights are reduced to limit erroneous aggregation of points on both sides of the normal construction boundary.

[0074] According to the structural type identifier, the width, main extension direction and directional deviation of the normal structural constraint region are statistically analyzed to form the normal structural parameter range corresponding to each structural type.

[0075] The spatial location, width, main extension direction, and directional deviation of each initial label's connected region are compared with the corresponding normal structural parameter range. For regions that conform to the normal structural parameter range, their initial semantic labels are adjusted to "no disease" labels.

[0076] For areas that do not conform to the normal structural parameter range, read their connectivity, directional deviation, boundary morphology, surface roughness, normal dispersion, and the relationship between the main extension direction and the wood grain direction, and make a judgment according to the corresponding disease rules.

[0077] Areas with a narrow, connected shape, local inward deviation, and an angle between the main extension direction and the wood grain direction within a preset range are retained as crack areas; areas with closed boundaries and concentrated inward deviation are retained as insect-eaten hole areas; areas with increased roughness and normal dispersion and shallow inward deviation are retained as decay areas; areas with flaky defects and irregular boundaries are retained as surface erosion areas; and areas with continuous deviation within a preset space are retained as deformation areas.

[0078] Areas that do not conform to the normal structural parameter range and do not stably meet the corresponding disease rules are marked as areas to be reviewed.

[0079] After completing the initial semantic label correction, the points with labels such as cracks, insect holes, decay, surface erosion or deformation, and their spatial connections are output as the disease point cloud region.

[0080] Example 2

[0081] This embodiment is based on the above embodiment. In step S1, the process of acquiring, registering and cleaning the multi-view point cloud of the wooden component is described.

[0082] At least one of three-dimensional laser scanning equipment, structured light scanning equipment, and depth vision acquisition equipment is used to acquire multi-view point clouds of wooden components from multiple scanning positions. The scanning positions are set according to the size of the wooden components, surface orientation, and on-site occlusion conditions, and overlapping scanning areas that can establish point cloud correspondences are preserved between adjacent scanning views.

[0083] Let the first The source point in each scanning perspective is Its corresponding target point is The target point normal is The rotation matrix and translation vector are respectively and The registration target is the point-to-plane distance between overlapping point clouds.

[0084]

[0085] in, The number of overlapping point pairs, The corresponding point weights are determined based on the normal consistency of the point pair, local curvature differences, and the measurement quality of the corresponding scanning angle.

[0086] The source point position is updated using the obtained rotation matrix and translation vector. The corresponding point in the overlapping point cloud is searched again and the iteration continues. When the changes in the rotation matrix, translation vector, and average point-to-plane distance are all below the corresponding thresholds, the final rotation matrix and translation vector are output.

[0087] Based on the final rotation matrix and translation vector, the point clouds from different scanning perspectives are transformed to a unified coordinate system. For observation points that are registered and located in the same surface neighborhood and meet the requirements of effective ranging range, incident angle range, and visibility, the number of their respective scanning perspectives is counted to obtain the number of effective perspectives; the point-to-plane distances of the corresponding points are counted to obtain the registration residuals of the corresponding points or local areas.

[0088] Points with a spatial distance below the repetition point merging threshold and consistent normals are merged, and the scanning incident angle, registration residual, and scanning pose of each scanning view are retained in the acquisition attribute set of the merged points.

[0089] For candidate isolated points whose distance from neighboring points exceeds a threshold, query whether there is a corresponding observation point in other scanning views. If there is no corresponding point in other views, delete the candidate isolated point; if there is a corresponding point in other views, retain the point and its corresponding acquisition attributes to avoid deleting crack edges, hole edges, or sharp edges of components as outliers.

[0090] This yields the registration point cloud of the wooden components and its acquisition attributes in a unified coordinate system, which are then input into steps S2 and S3.

[0091] Example 3

[0092] This embodiment is based on the above embodiment. In step S2, the structural coordinate system and the formation process of the normal structural constraint region of different types of wooden components are explained.

[0093] For beams, purlins, and rafters, the overall spatial distribution of the registered point cloud is analyzed by principal direction, and the direction of maximum spatial extension is determined as the component's axis. Point cloud sections are extracted along the component's axis at set intervals, section profiles are fitted to each section, and the component's axis is corrected based on the center positions of multiple section profiles.

[0094] For column-type components, multiple cross-sectional point sets are extracted along the column's extension direction, and the center of each cross-section is determined. The column's centerline is then fitted based on the sequence of cross-sectional center points. When the cross-section is circular or approximately circular, the cross-sectional radius is recorded; when the cross-section is elliptical, the major axis, minor axis, and cross-sectional direction are recorded.

[0095] For bracket sets and irregularly shaped components, the point cloud skeleton is extracted, and the point cloud is divided into multiple surface segments according to the branch direction of the skeleton and the orientation of the local surface. Local axial, tangential and normal directions are established in each surface segment.

[0096] For point Its structural coordinates can be expressed as:

[0097]

[0098] in, This refers to the axial position. The position relative to the center of the cross section. The normal position relative to the outer surface of the component. , and These are the distances to the end of the component, the edge, and the mortise and tenon boundary, respectively.

[0099] In a unified coordinate system, candidate regions of geometrical abrupt changes are extracted, and the number of times the candidate regions recur, their main extension direction, width variation, and distance relationship with the component axis, cross-sectional profile, and component boundary are calculated in different scanning views.

[0100] For a candidate region of geometric abrupt change, multiple region sections are extracted along its main extension direction at a set interval, and the center position, region width, and local main extension direction of the candidate region are calculated in each region section.

[0101] The region width is determined by projecting points within the region's cross-section onto a direction perpendicular to the main extension direction, and based on the effective distribution range of the projected points. To reduce the influence of a small number of outliers, points located within a preset ratio at both ends of the projected coordinates can be excluded, and the region width is then determined based on the difference between the maximum and minimum values ​​of the remaining projected points.

[0102] Calculate the angle between the main extension directions of adjacent cross sections, the change in the width of the region, and the change in the distance from the center of the region to the component axis, the cross section profile, the end of the component, the edge, or the tenon and mortise boundary.

[0103] When a candidate region is observed in at least two scanning perspectives, the included angle of the main extension direction between adjacent region sections is lower than the direction change threshold, the change in region width is lower than the width change threshold, and the change in distance from the region center to the corresponding axis, cross-sectional profile, or component boundary is lower than the distance fluctuation threshold, the candidate region is determined as a region with regular extension and fixed spatial relationship.

[0104] For candidate regions that simultaneously meet multiple structure type determination criteria, the distance fluctuation between the region and the corresponding spatial relationship of each structure type is calculated. The structure type with the smallest distance fluctuation is identified as the structure type identifier for the candidate region; when the distance fluctuations are the same, the structure type determination result with more repeated observations from multiple perspectives is given priority.

[0105] For geometric mutation candidate regions that do not reach the minimum number of effective viewpoints, or whose direction change, width change, or distance fluctuation exceeds the corresponding threshold, normal structural constraint regions are not formed, and their original point indexes are retained for subsequent disease feature extraction and semantic segmentation processing.

[0106] Candidate regions that repeatedly appear in multiple scanning perspectives, extend regularly along a fixed direction, and maintain a fixed spatial relationship with the component axis, cross-sectional profile, component ends, edges, or tenon-and-mortise boundaries are identified as normal structures, forming normal structural constraint regions.

[0107] Structural type identifiers are assigned based on the spatial location and structural relationships of the normal structure. For example, an edge type identifier is assigned to a region located at the inflection point of the cross-section profile and extending continuously along the axial direction of the component; an end boundary type identifier is assigned to a region located at the end of the component and consistent with the end face profile; and a tenon-mortise boundary type identifier is assigned to a region with regular connection boundaries and corresponding to the assembly position of adjacent components.

[0108] Normal structural constraint regions are recorded using point index sets, spatial boundaries, and structural type identifiers, serving as inputs for step S3 stable point screening, step S4 structural location feature extraction, and step S6 label correction.

[0109] Example 4

[0110] This embodiment is based on the above embodiment. In step S3, the process of stable point selection, reference surface iterative fitting and directional deviation calculation is explained.

[0111] Setting points The neighborhood contains Each point corresponds to a unit normal. Then point The degree of normal dispersion is:

[0112]

[0113] When the normal directions of neighboring points are consistent, the dispersion of the normal direction is small; when there are crack edges, hole edges or surface damage in the neighborhood, the dispersion of the normal direction increases accordingly.

[0114] Points that have reached a threshold in the number of effective viewpoints, have a registration residual below a threshold, have a normal dispersion below a threshold, and whose local curvature and surface roughness are within a stable range are added to the stable point set.

[0115] Let the first The set of stable points for each point cloud block is The reference surface to be fitted is The reference surface is determined by minimizing the distance from the stable point to the reference surface and the corresponding constraint terms:

[0116]

[0117] in, The distance from the stationary point to the reference surface. For the stable point weights, This includes constraints on component type, as well as positional and normal continuity constraints for adjacent point cloud blocks. This is the constraint coefficient.

[0118] The weights of stable points are determined based on the number of effective viewpoints, registration residuals, and normal dispersion. Stable points with a larger number of effective viewpoints, lower registration residuals, and lower normal dispersion have larger fitting weights.

[0119] After completing one fitting, stable points whose fitting residuals exceed a threshold are deleted, and the reference surface is refitted using the remaining stable points. Iteration stops when the percentage change in the set of stable points and the percentage change in the average fitting residuals of two consecutive fittings are both below the corresponding thresholds.

[0120] Setting points The projection point on the locally defect-free reference surface is The unit normal of the reference surface at the projection point is Then the directed deviation is:

[0121]

[0122] When the outward normal of the reference surface is defined as the positive direction, the deviation towards the interior of the component corresponds to a negative value, and the deviation towards the exterior of the component corresponds to a positive value. The fitting residual is the shortest distance from the point to the locally defect-free reference surface.

[0123] Associate the local disease-free reference surface, the directional deviation, and the fitting residual with the origin index, and input them into steps S4 and S5.

[0124] Example 5

[0125] This embodiment is based on the above embodiment. In step S4, the formation process of wood grain direction, anisotropic multi-scale neighborhood and disease-sensitive feature vector is explained.

[0126] The color texture gradient, reflection intensity gradient, and local geometric principal direction are projected onto the surface tangent plane of each point to obtain the corresponding candidate directions.

[0127] For the Candidate directions for class, let the candidate unit direction for the center point be... The direction of the candidate neighboring point is Then the neighborhood consistency of the candidate direction is:

[0128]

[0129] Candidate direction weights are set based on the completeness of color texture data, reflection intensity data, and local geometric data. The candidate direction with the highest weighted neighborhood consistency is determined as the wood grain direction, and the corresponding weighted neighborhood consistency is normalized and determined as the wood grain direction confidence.

[0130] When the confidence level of the wood grain direction reaches a threshold, let the wood grain direction be... The surface tangent perpendicular to the wood grain direction is The normal direction of the point is , No. The anisotropic multiscale neighborhood at each scale is:

[0131]

[0132] in, For the major axis search scale, For the short axis search scale, and , The normal thickness.

[0133] When the confidence level of the wood grain direction is lower than the threshold, the anisotropic multi-scale neighborhood is established by replacing the wood grain direction with the principal curvature direction; when the dispersion of the neighborhood of the candidate principal curvature direction exceeds the threshold, the major axis scale and the minor axis scale are set to be the same to form an equal-scale multi-scale neighborhood.

[0134] Calculate the mean, extreme values, gradient, and continuity of directional deviations within the neighborhood at each scale. Calculate the angle between the point normal and the normal of the local undamaged reference surface, the dispersion of the neighborhood normal, curvature, surface roughness, reflection intensity or color difference, and the angle between the neighborhood main extension direction and the wood grain direction.

[0135] For point The disease-sensitive feature vector is represented as:

[0136]

[0137] in, For multi-scale geometric features, For material appearance characteristics, For structural location features, The reliability of the wood grain direction.

[0138] The disease-sensitive feature vector is input into the corresponding feature channel in step S5 according to the origin index.

[0139] Example 6

[0140] This embodiment is based on the above embodiment. In step S5, the reliability of point cloud acquisition, channel coding, gating fusion and category probability output process are explained.

[0141] The point cloud density, scan incidence angle, effective viewing angle number, and fitting residual at each point are normalized. Let the normalized point cloud density quality, scan incidence angle quality, effective viewing angle quality, and fitting quality be respectively... , , and Then the reliability of point cloud acquisition is:

[0142]

[0143] in:

[0144]

[0145] The smaller the incident angle, the higher the quality of the scanning incident angle; the smaller the fitting residual, the higher the fitting quality.

[0146] The geometric defect channel, material appearance channel, and structural prior channel each perform point-level encoding on their corresponding input features to obtain the geometric defect channel features. Material surface channel characteristics and structural prior channel features .

[0147] In this embodiment, the geometric defect channel, the material appearance channel, and the structural prior channel each include a point-level feature encoding layer, a neighborhood feature aggregation layer, and a channel output layer. The three channels use the same point index and sampling level, but the encoding parameters of each channel are independent of each other.

[0148] The point-level feature encoding layer normalizes and maps the dimension of the continuous numerical features and identifier features input from each channel. The geometric defects channel maps directional deviation, normal variation, curvature, surface roughness, and indentation depth continuity to geometric point-level features; the material appearance channel maps reflection intensity, grayscale, or color to material appearance point-level features; and the structural a priori channel maps structural coordinates, normal structural constraint region identifiers, and distances to component boundaries to structural a priori point-level features.

[0149] The neighborhood feature aggregation layer establishes local neighborhoods based on the spatial distance between points, and inputs the center point features, neighborhood point features, relative coordinates, and normal differences into the neighborhood aggregation process. For each center point, the aggregation result that can characterize the local maximum response and average change is extracted from the neighborhood point features and connected with the center point features to form the updated point-level features.

[0150] Gated fusion units improve the reliability of point cloud acquisition Reliability of wood grain direction The gated input is constructed, and the point-level weights of the three channels are obtained through parameter mapping and normalization:

[0151]

[0152] in, For gating mapping parameters, The gating bias parameter is the sum of the point-level weights of the three channels, which is 1.

[0153] The features of the three channels are weighted and fused according to point-level weights:

[0154]

[0155] Neighborhood feature propagation and resolution restoration are performed on the fused features to obtain the classification values ​​corresponding to the original point indices, and the class probabilities are output through a normalized exponential function:

[0156]

[0157] in, For point Corresponding category The classification value, The total number of categories. The category with the highest probability is selected as the initial semantic label.

[0158] The point cloud semantic segmentation model is trained using point clouds of wooden components that have been labeled with categories such as no disease, cracks, insect holes, decay, surface erosion, and deformation. The training samples use the same structural coordinates, feature composition, and category definition as the point cloud to be identified.

[0159] This embodiment employs a multi-level point cloud encoding structure. Between adjacent encoding layers, representative points are selected based on the spatial distribution of the point cloud to form a downsampled point set, and the original point index and neighborhood point index corresponding to each downsampled point are recorded. The three channels use the same downsampled point set, ensuring that the point-level features output by different channels have a one-to-one spatial correspondence and the same feature dimension.

[0160] After the three channels output point-level features of the same dimension, the gated fusion unit generates point-level weights for the three channels based on the reliability of point cloud acquisition and the credibility of wood grain direction. It then performs a weighted sum of the point-level features output by the three channels to obtain the fused features. The mapping parameters in the gated fusion unit, along with the encoding parameters in the three channels, serve as the training parameters for the point cloud semantic segmentation model.

[0161] When propagating neighborhood features to the fused features, the point index relationships recorded in the encoding stage are used to propagate the fused features from the low-resolution point set to the point set of the next layer. For points that do not directly correspond to low-resolution representative points, the interpolation weight is determined based on the spatial distance between the point and its neighboring representative points. The fused features of the neighboring representative points are then interpolated to the point and connected with the point-level features of the same level in the encoding stage. After completing the feature propagation at each level, the fused features are restored to the original point cloud resolution and the original point index.

[0162] The training samples for the point cloud semantic segmentation model are divided into training and validation sets according to the wooden component numbers. Point cloud blocks generated from the same wooden component are only included in one of the datasets to avoid adjacent regions of the same component being included in both the training and validation sets simultaneously. The ratio of the number of components in the training and validation sets can be set to eight to two.

[0163] Each point in the training point cloud is labeled with a category such as no disease, crack, insect-eaten hole, decay, surface erosion, or deformation. Points whose category cannot be determined based on the scan results are not used as supervision points for model training. Normal construction points in the normal structural constraint region are labeled as no disease, and the corresponding structural type identifier is retained as input for the structural prior channel.

[0164] The model training uses multi-class cross-entropy loss to calculate the difference between the predicted class probability and the labeled class. Class weights are set according to the number of points for each disease category in the training set, with higher weights assigned to disease categories with fewer samples. An upper limit is set on the class weights to avoid a small number of outlier labels having an excessive impact on model parameter updates.

[0165] After each training round, the validation loss and the identification results for each disease category are calculated using the validation set. Training stops when the validation loss does not decrease for a preset number of consecutive rounds, or when the number of training rounds reaches a preset upper limit. The channel encoding parameters, gating mapping parameters, neighborhood feature propagation parameters, and classification parameters corresponding to the lowest validation loss are retained.

[0166] The defect-sensitive feature vectors of the wooden components to be identified are input into the trained point cloud semantic segmentation model using the same feature order, normalization parameters, and point index organization method as the training samples, and the output is the category probability and initial semantic label of each point.

[0167] Input the category probability and initial semantic label of each point into step S6 according to the origin index.

[0168] Example 7

[0169] This embodiment is based on the above embodiment. In step S6, the process of adjacency graph establishment, connection weight calculation and initial semantic label correction is described.

[0170] For points adjacent to the surface geodesy and Let the differences in surface geodesic distance, normal direction, directional deviation, and wood grain direction between the two be respectively... , , and The adjacency graph connection weights are:

[0171]

[0172] Among them, when point With point When the surface connection path between them crosses a normal structural constraint region. A connection penalty value less than 1 is used; when the normal structural constraint region is not crossed. Take 1.

[0173] Connected regions are formed based on connection weights and initial semantic labels, and the spatial location, width, length, main extension direction, boundary closure degree, directional deviation distribution, normal dispersion, surface roughness, and the angle between the main extension direction and the wood grain direction of each connected region are calculated.

[0174] According to the structural type, the width, main extension direction, and directional deviation of the constraint region of the normal structure are statistically analyzed to form the corresponding normal structural parameter range.

[0175] The region parameters of the connected regions are compared with the corresponding normal structural parameter ranges. Regions that conform to the normal structural parameter ranges are adjusted to be labeled as disease-free.

[0176] Areas with a narrow, connected shape, local inward deviation, and an angle between the main extension direction and the wood grain direction within a preset range are retained as crack areas; areas with closed boundaries and concentrated inward deviation are retained as insect-eaten hole areas; areas with increased roughness and normal dispersion and shallow inward deviation are retained as decay areas; areas with flaky defects and irregular boundaries are retained as surface erosion areas; and areas with continuous deviation within a preset space are retained as deformation areas.

[0177] Regions that cannot stably conform to the normal structural parameter range or any disease rule are marked as regions to be reviewed. The final output includes the disease point cloud region, disease category, and regions to be reviewed.

[0178] Example 8

[0179] This embodiment is based on the above embodiment, and describes the process of quantizing the size of the disease point cloud region after step S6.

[0180] For the cracked area, a centerline is extracted along the main extension direction of the crack, and cross-sections perpendicular to the centerline are established at multiple locations along the centerline. The crack length is obtained by accumulating the surface geodesic distances between adjacent center points of the centerline; the crack width is determined by the distance between the two boundaries of the crack in each cross-section; and the measurable depth is determined by the maximum inward directional deviation of a point within the cracked area relative to a local undamaged reference surface.

[0181] For areas with insect-eaten holes, the hole boundaries are projected onto a local, unaffected reference surface to obtain a projected profile. Let the area of ​​the projected profile be... The equivalent aperture is:

[0182]

[0183] The measurable depth is determined by the maximum inward directional deviation of the visible point inside the hole relative to the local undamaged reference surface.

[0184] For areas of decay or surface erosion, the point cloud of the diseased area is projected onto a local, unaffected reference surface, and the coverage area is calculated based on the projection range. Let the first... The area of ​​each projection unit is The corresponding directional deviation is The material loss volume is:

[0185]

[0186] For the deformation region, a reference axis is fitted based on the center of the stable section that is not marked as deformed in the structural coordinate system, and a measured axis is fitted based on the center of the section of the deformation region; the distance between the two axes at the corresponding axial positions is compared to obtain the bending amount; the torsional angle is calculated based on the changes in different cross-sectional directions.

[0187] The dimensions of the defects are associated with and stored in relation to the defect point cloud region, defect category, and structural coordinates to form the defect identification results for wooden components.

[0188] Example 9

[0189] This embodiment is based on the above embodiment. After step S6, it explains the regional comprehensive reliability, the area to be supplemented for collection, and the supplementary scanning process.

[0190] For the disease point cloud area The system reads the disease category probability, point cloud acquisition reliability, and fitting residual for each point within the region. Let the average category probability for the region be... The average point cloud acquisition reliability is The normalized average fitting residual is The overall regional credibility is:

[0191]

[0192] in:

[0193]

[0194] Areas with disease point cloud that have an overall credibility below the threshold and are subject to occlusion, single-view coverage, or scanning incident angle exceeding the threshold are identified as areas to be supplemented for data collection.

[0195] Based on the missing boundary of the local point cloud in the area to be supplemented, and combined with the visibility of the area from the existing scan pose, the main occlusion direction is determined. Candidate supplementary scan poses are generated based on the surface normal of the area to be supplemented, the occlusion direction, and the existing scan pose.

[0196] Select a supplementary scan pose from the candidate supplementary scan poses that can cover the missing area of ​​the point cloud, meet the requirements for the scan incident angle, and are different from the existing scan poses.

[0197] Supplementary point clouds are obtained according to the supplementary scan pose, registered to the original coordinate system, and merged with the original point cloud for duplicate points and fused with collected attributes.

[0198] The scanning incident angle, effective number of viewing angles, registration residuals, and point cloud acquisition reliability of the area to be supplemented and its adjacent areas are recalculated. Steps S3 to S6 are then re-executed for the area and its adjacent areas to regenerate the local disease-free reference surface, directional deviation, disease-sensitive feature vector, category probability, initial semantic label, and disease point cloud area.

[0199] For other areas not located within the supplementary data collection area and its adjacent areas, retain the original processing results. After completing the supplementary data collection and local re-identification, update the disease point cloud area, disease type, disease size, and overall area reliability, and retain the areas that still require manual review.

[0200] In the above specific implementation, the point cloud block size, multi-scale neighborhood range, effective viewpoint number threshold, registration residual threshold, normal dispersion threshold, wood grain direction confidence threshold, connection weight threshold, and region comprehensive confidence threshold are determined based on the scanning equipment accuracy, wooden component size, average point cloud spacing, and minimum identification size of the target disease. The above parameter settings do not alter the data processing relationship formed by multi-view point cloud input, structural coordinate establishment, fitting of local disease-free reference surfaces, extraction of disease-sensitive features, point cloud semantic segmentation, initial semantic label correction, and output of the disease point cloud region.

[0201] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various variations or substitutions within the technical scope disclosed in the present invention, and these should all be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for identifying defects in ancient architectural wooden components based on 3D point cloud semantic segmentation, characterized in that, Includes the following steps: S1. Acquire and register multi-view point clouds of wooden components, and record the scanning position and direction, incident angle, number of effective viewpoints and registration residuals; S2. Extract the axis, cross-sectional profile and normal, establish the structural coordinate system, identify normal construction and form a normal structural constraint region with type identification; S3. Select stable points outside the normal structural constraint area where the number of effective viewpoints reaches the threshold and the registration residual and normal dispersion are lower than the threshold. Fit the reference surface and iteratively remove residual abnormal points to obtain a local disease-free reference surface. Calculate the directional deviation and record the fitting residual. S4. Determine the wood grain direction and confidence level. When the confidence level reaches the threshold, establish an anisotropic multi-scale neighborhood with a larger scale along the wood grain direction. When the confidence level is lower than the threshold, establish the neighborhood with the principal curvature direction. When the direction is unstable, establish an equal-scale neighborhood. Extract geometric, appearance and structural location features to form a disease-sensitive feature vector. S5. Generate the point cloud acquisition reliability based on point cloud density, scanning incident angle, number of effective viewing angles and fitting residuals. Input the vector into the point cloud semantic segmentation model containing geometric defects, material appearance, structural prior channels and gated fusion units. The gated fusion unit generates point-level weights and fuses channel features based on acquisition reliability and wood grain direction confidence, and outputs category probability and initial semantic label. S6. Establish an adjacency graph based on surface geodesy and correct the connected regions of the initial semantic labels; according to the structural type identifier and the parameter statistics of the corresponding normal structure, adjust the regions that conform to the normal structure parameter range to be labeled as disease-free, retain the regions that conform to the disease rules in terms of connected morphology, directional deviation changes and wood grain direction relationship as disease labels, and mark the remaining regions as regions to be reviewed, thus obtaining the disease point cloud region.

2. The method according to claim 1, characterized in that, In step S1, at least one of a three-dimensional laser scanning device, a structured light scanning device, and a depth vision acquisition device is used to acquire multi-view point clouds of the wooden component from multiple scanning positions; Coarse registration is performed based on calibration points, stable structural feature points, and overlapping point clouds of adjacent scanning areas. The rotation matrix and translation vector are iteratively corrected with the goal of minimizing the point-to-plane distance between overlapping point clouds. The registered point cloud is processed by merging duplicate points and removing outliers, and isolated points that are more than a threshold away from their neighbors and have not been observed repeatedly from other perspectives are deleted.

3. The method according to claim 1, characterized in that, In step S2, for beams, purlins, and rafters, the main extension direction of the component is taken as the axis; for columns, the center line of the column is taken as the axis; for brackets and irregular components, the surface is divided into segments and a segmented structural coordinate system is established according to the branch direction of the point cloud skeleton and the local surface orientation. The structural coordinates include axial position, position relative to the center of the cross section, normal position relative to the outer surface of the component, and distance to the end, edge, and tenon / mortise boundary of the component; Geometrically abrupt regions that repeat in multiple scanning views, extend regularly, and have a fixed spatial relationship with the axis or cross-sectional profile are marked as normal structural constraint regions and assigned a structural type identifier.

4. The method according to claim 1, characterized in that, In step S3, the surface of the wooden component is divided into overlapping point cloud blocks, and points with local curvature and surface roughness within a stable range are selected from each point cloud block as stable points. Apply planar or regular section constraints to the reference curved surfaces of beams and purlins, apply cylindrical or elliptical cylindrical constraints to column components, and use piecewise surface fitting for brackets and irregular components, and set positional and normal continuity constraints for adjacent point cloud blocks. After each fitting, stable points whose fitting residuals exceed the threshold are removed, and the remaining stable points are used to refit until the change in the set of stable points and the change in the fitting residuals are lower than the corresponding thresholds.

5. The method according to claim 1, characterized in that, In step S4, the color texture gradient direction, reflection intensity gradient direction, and local geometric principal direction are calculated in the surface tangent plane at each point. The wood grain direction and the confidence level of the wood grain direction are determined based on the neighborhood consistency of the candidate directions. The anisotropic multi-scale neighborhood has the wood grain direction or principal curvature direction as the major axis, the direction perpendicular to the major axis as the minor axis, and the surface normal as the thickness direction, and the major axis search scale is larger than the minor axis search scale. When the neighborhood dispersion of the candidate principal curvature direction exceeds the threshold, an equal-scale multi-scale neighborhood is established.

6. The method according to claim 1, characterized in that, In step S4, the directional deviation and its gradient, the angle between the point normal and the local undamaged reference surface normal, the dispersion of the neighborhood normal, curvature, surface roughness, the continuity of the indentation depth, the reflection intensity or color difference, the distance to the normal structural constraint area, and the angle between the neighborhood main extension direction and the wood grain direction are calculated in the small-scale, medium-scale, and large-scale neighborhoods, respectively. The disease-sensitive feature vector is formed by combining the calculation results of each scale according to the point index and the confidence level of the wood grain direction.

7. The method according to claim 1, characterized in that, In step S5, the geometric defect channel coding includes directional deviation, normal variation, curvature, surface roughness, and indentation depth continuity; the material apparent channel coding includes reflection intensity, grayscale, or color; and the structural prior channel coding includes structural coordinates, normal structural constraint area identification, and distance to the component boundary. The gated fusion unit maps the reliability of point cloud acquisition and the confidence level of wood grain direction to point-level weights corresponding to three channels and normalized, and performs a weighted summation of the point-level features output by the three channels according to the weights. The obtained fused features are then subjected to neighborhood feature propagation and resolution restoration to output the category probabilities of no disease, cracks, insect holes, decay, surface erosion and deformation.

8. The method according to claim 1, characterized in that, In step S6, the connection weights of the adjacency graph are calculated based on the surface geodesic distance, normal difference, directional deviation difference, wood grain direction consistency, and whether it crosses the normal structural constraint area. According to the structural type, the width, main extension direction and directional deviation of the constraint region of the normal structure are statistically analyzed to form the corresponding normal structure parameter range; Areas with a narrow, connected shape, local inward deviation, and an angle between the main extension direction and the wood grain direction within a preset range are retained as crack areas; areas with closed boundaries and concentrated inward deviation are retained as wormhole areas; areas with increased roughness and normal dispersion and shallow inward deviation are retained as decay areas; areas with flaky defects and irregular boundaries are retained as surface erosion areas; and areas with continuous deviation within a preset space are retained as deformation areas.

9. The method according to claim 1, characterized in that, After step S6, the centerline is extracted along the main extension direction of the crack region and a cross section is established to calculate the crack length, width, and measurable depth based on the directional deviation. The equivalent aperture is calculated based on the projection profile of the insect-eaten hole boundary onto a local undamaged reference surface, and the measurable depth is calculated based on the maximum inward deviation of the visible point inside the hole. The decayed or eroded area is projected onto a local, undamaged reference surface to calculate the coverage area, and the directional deviation is spatially integrated to calculate the material loss volume. The bending amount and torsional angle are calculated by fitting a reference axis to the center of the stable section in the structural coordinate system, fitting the measured axis to the center of the section in the deformation region, and then calculating the bending amount and torsional angle.

10. The method according to claim 1, characterized in that, After step S6, the overall credibility of the region is determined based on the category probability of each point within the disease point cloud region, the reliability of point cloud acquisition, and the fitting residual. Areas with a comprehensive confidence level below the threshold and that are obstructed, covered by a single viewpoint, or have an incident angle exceeding the threshold are identified as areas to be supplemented for data collection. The occlusion direction is determined based on the missing boundary of the local point cloud and the visibility of each scanning viewpoint. The supplementary scanning pose is determined based on the surface normal of the area to be supplemented, the occlusion direction, and the existing scanning pose. The supplementary point cloud is registered to the original coordinate system and merged with the original point cloud. The acquisition attributes and point cloud acquisition reliability of the region and its adjacent regions are recalculated. Steps S3 to S6 are re-executed for the region and its adjacent regions, while the processing results of other regions are retained.

Images