A prefabricated part steel bar quality detection method and system

Abstract

The application discloses a prefabricated steel bar quality detection method and system, generates an illumination robust image feature set; obtains initial dense three-dimensional point cloud data and corresponding reconstruction confidence maps; obtains confidence screening three-dimensional point cloud data, and converts the same into steel bar mesh topological graph structure data to obtain topological correction three-dimensional point cloud data; obtains a steel bar mesh macroscopic skeleton curved surface model; obtains a flexible registration alignment point cloud model; calculates single steel bar three-dimensional shape deviation vectors, steel bar spacing deviations and overall mesh flatness deviations to obtain multi-dimensional steel bar shape deviation data; and generates a steel bar shape deviation detection report. The application significantly improves the three-dimensional point cloud completeness rate under strong reflection and multi-layer shielding environment, obviously reduces the local fracture rate, can maintain the spatial integrity of the continuous cylindrical structure of the steel bar, and improves the stability of millimeter-level measurement from the source.

Description

Technical Field

[0001] This invention relates to the field of steel reinforcement quality testing technology, and in particular to a method and system for testing the quality of precast steel reinforcement. Background Technology

[0002] With the rapid development of prefabricated buildings and industrialized precast components, precast steel mesh, as a core load-bearing unit in composite slabs, shear walls, and floor slabs, directly affects structural safety performance due to its spatial arrangement accuracy, spacing consistency, and node binding quality. In actual production, steel mesh is usually large in size and complex in structure, and the production environment is located on assembly lines or construction sites, which present complex conditions such as unstable lighting, strong metal reflection, crisscrossing binding wires, and multiple layers of steel reinforcement obstructing the view.

[0003] In 3D detection, existing multi-view geometric methods rely on precise camera calibration and feature point matching, establishing cross-view correspondences through feature descriptors. However, the cylindrical metal structure of steel bars is prone to high light reflection, and the texture of threaded steel bars is highly repetitive, which easily leads to mismatches or matching degradation in the feature matching stage. This results in point cloud collapse, local missing parts, or even topological errors such as broken steel bars during 3D reconstruction.

[0004] In recent years, end-to-end deep learning-based 3D reconstruction methods have been increasingly applied to complex scene modeling. However, general network models typically rely on data-driven methods for depth regression, lacking physical prior constraints for continuous cylindrical structures with reinforced concrete and the topology of intersecting nodes. In areas with metallic reflections, oil stains, or partial occlusion, network regression results are prone to producing depth outliers, disrupting the connectivity within the point cloud. Since 3D point clouds are essentially discrete and unordered sets of points, without topology preservation mechanisms, the reconstruction results are difficult to automatically repair when local breaks or distortions occur, thus affecting the reliability of subsequent dimensional measurements.

[0005] On the other hand, precast steel mesh panels may experience overall flexible sagging or slight bending during handling, hoisting, or uneven support. This macroscopic, non-rigid deformation is a physical phenomenon within acceptable limits. However, existing 3D point cloud alignment methods often employ rigid registration, rigidly transforming and aligning the measured point cloud with a standard model. This fails to effectively distinguish between overall flexible deformation and actual installation deviations. When the entire mesh panel deflects, rigid alignment often misjudges macroscopic deformation as large-scale out-of-tolerance, leading to a high false alarm rate. Summary of the Invention

[0006] One objective of this invention is to propose a method for inspecting the quality of reinforcing steel bars in precast components. This invention significantly improves the integrity rate of three-dimensional point clouds under strong reflective light and multi-layer shading environments, and significantly reduces the local fracture rate. It can maintain the spatial integrity of continuous cylindrical steel bar structures and improve the stability of millimeter-level measurements from the source.

[0007] A method for inspecting the quality of reinforcing steel bars in precast components according to an embodiment of the present invention includes:

[0008] Unlabeled multi-view RGB image datasets of precast steel mesh are collected and subjected to illumination-robust precoding to generate illumination-robust image feature sets.

[0009] The unlabeled multi-view RGB image dataset and the illumination-robust image feature set are input into the improved DUSt3R end-to-end 3D reconstruction network to obtain the initial dense 3D point cloud data and the corresponding reconstruction confidence map.

[0010] Based on the reconstructed confidence map, low-confidence points are removed from the initial dense 3D point cloud data to obtain confidence-filtered 3D point cloud data, which is then converted into steel mesh topology data. Based on the steel mesh topology data, topology-preserving correction is performed to obtain topology-corrected 3D point cloud data.

[0011] A flowing surface skeleton was fitted to the topology-corrected 3D point cloud data to obtain a macroscopic skeleton surface model of the steel mesh.

[0012] The topology-corrected 3D point cloud data and the macroscopic skeleton surface model of the steel mesh are flexibly registered with the 3D design standard model to obtain a flexibly registered aligned point cloud model.

[0013] Based on the spatial difference between the flexible registration and alignment point cloud model and the three-dimensional design standard model, the three-dimensional morphological deviation vector of a single steel bar, the steel bar spacing deviation, and the overall mesh flatness deviation are calculated to obtain multi-dimensional steel bar morphological deviation data.

[0014] The multidimensional rebar morphology deviation data are graded and judged to generate a rebar morphology deviation detection report.

[0015] Optionally, the process of acquiring unlabeled multi-view RGB image datasets of precast steel mesh and performing illumination-robust precoding includes:

[0016] Multiple views of the same precast steel mesh were simultaneously acquired to form an uncalibrated multi-view RGB image dataset.

[0017] Perform illumination consistency normalization on each RGB image in the unlabeled multi-view RGB image dataset to obtain an illumination-normalized RGB image;

[0018] Based on illumination-normalized RGB images, construct illumination-invariant color representations;

[0019] Multi-scale illumination-robust precoding is performed on illumination-normalized RGB images to form illumination-robust texture representations;

[0020] The illumination-invariant color representation and the illumination-robust texture representation are concatenated at the pixel level along the channel dimension to obtain an illumination-robust precoded feature map. The illumination-robust precoded feature maps corresponding to all viewpoints are then collected in viewpoint order to form an illumination-robust image feature set.

[0021] Optionally, the step of inputting the unlabeled multi-view RGB image dataset and the illumination-robust image feature set into the improved DUSt3R end-to-end 3D reconstruction network includes:

[0022] Based on each RGB image in the unlabeled multi-view RGB image dataset and the corresponding illumination robust pre-coded feature map, the steel bar reflectivity map and the repeating texture ambiguity map are calculated respectively.

[0023] While keeping the pixel spatial position unchanged, the RGB image and the corresponding illumination robust pre-coding feature map are concatenated in the channel dimension to obtain the pixel-level fusion feature. The guiding gate coefficient is calculated based on the steel bar reflectivity map and the repeating texture ambiguity map. The guiding gate coefficient and the pixel-level fusion feature are multiplied element-wise at the corresponding pixel position to obtain the input feature after gate modulation.

[0024] The gated and modulated input features are used to generate view guidance feature sequences through block embedding operators and position encoding operators. All view guidance feature sequences corresponding to all viewpoints are collected in viewpoint index order to form a multi-view guidance feature sequence set.

[0025] The multi-view guided feature sequence set is input into the cross-view global attention backbone of the improved DUSt3R end-to-end 3D reconstruction network. During the cross-view attention calculation process, a reflectivity-sensitive bias term and a texture ambiguity bias term are introduced for each pixel position, and the multi-view dense 3D coordinate prediction tensor and reconstruction confidence prediction tensor are output.

[0026] Initial dense 3D point cloud data is generated based on the multi-view dense 3D coordinate prediction tensor, and a reconstruction confidence map is generated based on the reconstruction confidence prediction tensor.

[0027] The improved DUSt3R end-to-end 3D reconstruction network outputs absolute spatial scale parameters, and performs absolute scale assignment on the 3D point coordinates of the initial dense 3D point cloud data based on the absolute spatial scale parameters to obtain the initial dense 3D point cloud data carrying absolute spatial scale information.

[0028] Optionally, the topology-preserving correction processing performed on the structural data based on the steel mesh topology diagram includes:

[0029] Based on the reconstructed confidence map, low-confidence points are removed from the initial dense 3D point cloud data to obtain confidence-filtered 3D point cloud data;

[0030] Convert confidence-based filtering of 3D point cloud data into steel mesh topology structure data;

[0031] Topology-preserving correction processing is performed on the structural data of the steel mesh topology diagram to obtain topology-corrected 3D point cloud data.

[0032] Optionally, the step of performing streaming surface skeleton fitting on the topology-corrected 3D point cloud data includes:

[0033] According to the acquisition sequence of the topology-corrected 3D point cloud data, the topology-corrected 3D point cloud data is segmented and divided into continuously arriving topology-corrected 3D point cloud segments.

[0034] Within each topology-corrected 3D point cloud fragment, a set of representative points for skeleton fitting is constructed, and the set of representative points is accumulated in chronological order to form a streaming set of representative points.

[0035] Based on the representative point set of the flow cytometry, a macroscopic skeleton surface model of the steel mesh is defined;

[0036] A flowing surface skeleton fitting is performed on the macroscopic skeleton surface model of the steel mesh, and the control point set is incrementally updated based on the representative point set at the current time to obtain the macroscopic skeleton surface model of the steel mesh at the current time.

[0037] Optionally, the step of flexibly registering the topology-corrected 3D point cloud data and the macroscopic skeleton surface model of the reinforcing mesh with the 3D design standard model includes:

[0038] A macroscopic deformation field is established based on the current moment's macroscopic skeleton surface model of the steel mesh. The topology-corrected three-dimensional point cloud data is mapped onto the reference plane of the macroscopic deformation field. Macroscopic deformation displacement stripping is performed to obtain a macroscopic deformation stripped point cloud containing only high-frequency local morphological deviations.

[0039] The macroscopic deformation stripping point cloud is subjected to non-rigid registration and solution with the 3D design standard model. The overall rigid pose alignment and local non-rigid deviation are decoupled, and a flexible registration and alignment point cloud model with accurate topological matching relationship and deviation attribute is output.

[0040] Optionally, the calculation of the three-dimensional morphological deviation vector of a single rebar, the rebar spacing deviation, and the overall mesh flatness deviation based on the spatial difference between the flexible registration and alignment point cloud model and the three-dimensional design standard model includes:

[0041] Extract the deformation attributes encapsulated in the flexible registration and alignment point cloud model, and establish a set of point-level spatial difference vectors;

[0042] Based on the point-level spatial difference vector set, calculate the three-dimensional morphological deviation vector of a single steel bar;

[0043] Based on the flexible registration and alignment point cloud model and the three-dimensional design standard model, the deviation of the rebar spacing is calculated.

[0044] Based on the flexible registration and alignment point cloud model, the overall mesh flatness deviation is calculated;

[0045] The three-dimensional morphological deviation vector of a single rebar, the rebar spacing deviation, and the overall mesh flatness deviation are structurally aggregated to obtain multi-dimensional rebar morphological deviation data.

[0046] Optionally, the step of classifying and determining the multidimensional rebar morphology deviation data includes:

[0047] Establish a grading standard for multidimensional steel bar morphological deviation data;

[0048] Based on the hierarchical judgment criteria, abnormal node binding information is identified, forming a set of abnormal node binding information.

[0049] Based on the grading criteria, determine the overall quality level information;

[0050] A rebar morphology deviation detection report is generated based on the set of node binding anomaly identification information and the overall quality grade information.

[0051] Optionally, the overall quality level information is determined according to the following rules:

[0052] When the number of abnormal deviations in the three-dimensional shape of the reinforcing bars is zero, the number of abnormal deviations in the spacing of the reinforcing bars is zero, and the overall flatness deviation of the mesh is less than the overall flatness deviation threshold, it is judged to be of qualified grade.

[0053] When at least one of the abnormal number of three-dimensional shape deviations of steel bars or abnormal number of steel bar spacing deviations is greater than zero and does not exceed a preset proportional threshold, it is determined to be a limit level.

[0054] When the number of abnormal deviations in the three-dimensional shape of the reinforcing bars or the number of abnormal deviations in the spacing of the reinforcing bars exceeds the preset proportional threshold, or when the overall flatness deviation of the mesh is greater than or equal to the overall flatness deviation threshold, it is judged as unqualified.

[0055] Optionally, a precast steel reinforcement quality inspection system includes:

[0056] The image acquisition module acquires uncalibrated multi-view RGB image datasets of precast steel mesh and performs illumination-robust precoding processing to generate an illumination-robust image feature set.

[0057] The 3D reconstruction module inputs the unlabeled multi-view RGB image dataset and the illumination-robust image feature set into the improved DUSt3R end-to-end 3D reconstruction network to obtain the initial dense 3D point cloud data and the corresponding reconstruction confidence map.

[0058] The topology correction module removes low-confidence points from the initial dense 3D point cloud data based on the reconstructed confidence map and performs correction processing to obtain topology-corrected 3D point cloud data.

[0059] The skeleton fitting module performs flow surface skeleton fitting on topology-corrected 3D point cloud data to obtain a macroscopic skeleton surface model of steel mesh.

[0060] The flexible registration module flexibly registers the topology-corrected 3D point cloud data and the macroscopic skeleton surface model of the steel mesh with the 3D design standard model to obtain a flexibly registered aligned point cloud model.

[0061] The deviation calculation module obtains multi-dimensional rebar morphology deviation data based on the spatial difference between the flexible registration and alignment point cloud model and the three-dimensional design standard model.

[0062] The grading and judgment module grades and judges the multi-dimensional rebar morphology deviation data and generates a rebar morphology deviation detection report.

[0063] The beneficial effects of this invention are:

[0064] This invention introduces a steel reinforcement reflectivity map and a repeating texture ambiguity map as guiding bias terms for cross-view attention into the improved DUSt3R end-to-end 3D reconstruction network, realizing a structured suppression mechanism for metal steel reinforcement scenes. By negatively modulating the similarity scores of reflective and repeating texture regions during the cross-view global attention calculation stage, the network automatically reduces the weight of unreliable regions during feature aggregation, suppressing erroneous cross-view matching. Compared with the original network without the guiding bias, the integrity rate of 3D point clouds is significantly improved in strong reflectivity and multi-layer occlusion environments, and the local breakage rate is significantly reduced. It can maintain the spatial integrity of the continuous cylindrical steel reinforcement structure and improve the stability of millimeter-level measurements from the source.

[0065] This invention converts the confidence-filtered 3D point cloud into a steel mesh topology data and performs topology-preserving correction based on graph Laplace constraints. This achieves the transformation from data-driven results to structurally constrained results. While ensuring that the corrected 3D coordinates do not deviate too far from the original point cloud, it forces adjacent nodes to maintain geometric continuity. This automatically pulls breakpoints caused by reflection or occlusion back onto the continuous cylindrical trajectory. It can suppress local drift while maintaining the overall topological connectivity structure without breakage, significantly improving the reconstruction stability of the intersection node area and the steel mesh overlapping area.

[0066] This invention introduces a deformation field stripping mechanism based on a macroscopic skeleton surface model during the flexible registration stage. This achieves layered decoupling between macroscopic low-frequency flexible sag and local high-frequency manufacturing deviations. The topology-corrected 3D point cloud is mapped to the macroscopic skeleton surface model of the steel mesh, and the macroscopic deformation displacement vector is calculated by the difference vector between the point cloud and the ideal reference plane model. Macroscopic deformation displacement stripping is then performed on the original point cloud, retaining only the high-frequency local deviation components. On the stripped point cloud, overall rigid pose alignment and local non-rigid displacement field joint optimization are performed, eliminating overall angular tilt and positional offset. The actual steel bar arrangement error is retained and structurally bound in the form of local displacement vectors. This allows for accurate identification of single steel bar offset, steel bar spacing deviation, and node binding anomalies even in the presence of large-scale flexible deflection, reducing the misjudgment rate caused by macroscopic deformation and improving the reliability of the overall mesh quality level determination. Attached Figure Description

[0067] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0068] Figure 1 This is a flowchart of a method for inspecting the quality of reinforcing steel bars in precast components proposed in this invention;

[0069] Figure 2 This is a schematic diagram of the improved DUSt3R end-to-end three-dimensional reconstruction network structure in the precast steel reinforcement quality inspection method and system proposed in this invention. Detailed Implementation

[0070] Example 1: Reference Figures 1-2 A method for inspecting the quality of reinforcing steel bars in precast components, comprising:

[0071] Unlabeled multi-view RGB image datasets of precast steel mesh are collected and subjected to illumination-robust precoding to generate illumination-robust image feature sets.

[0072] In this embodiment, an unlabeled multi-view RGB image dataset of the precast steel mesh is collected and subjected to illumination-robust precoding processing, including:

[0073] Multiple views of the same precast steel mesh were simultaneously acquired to form an uncalibrated multi-view RGB image dataset.

[0074] In Example 1, at least two ordinary RGB cameras are used to simultaneously acquire the steel mesh of the same precast component from multiple perspectives to form an uncalibrated multi-view RGB image dataset. The uncalibrated multi-view RGB image dataset consists of RGB images acquired from multiple different perspectives, and each RGB image is a normalized pixel intensity value matrix.

[0075] Perform illumination consistency normalization on each RGB image in the unlabeled multi-view RGB image dataset to obtain an illumination-normalized RGB image;

[0076] Based on illumination-normalized RGB images, construct illumination-invariant color representations;

[0077] In Example 1, for each pixel position and each color channel in the illumination-normalized RGB image, the natural logarithm of the sum of the pixel intensity and stability term in that color channel is calculated. The average of the natural logarithms of the sum of the pixel intensity and stability term in the three color channels is calculated. The natural logarithm of the pixel in the current color channel is subtracted from the average of the natural logarithms of the pixel in the three color channels to obtain the illumination-invariant color representation of the pixel in the current color channel.

[0078] Multi-scale illumination-robust precoding is performed on illumination-normalized RGB images to form illumination-robust texture representations;

[0079] In Example 1, the normalized luminance component is calculated from the illumination-normalized RGB image. Multiple scale layers are set, and a corresponding scale parameter is set for each scale layer. For each scale layer, the normalized luminance component is convolved with a two-dimensional Gaussian kernel of the corresponding scale parameter to obtain the smooth response at the corresponding scale. The smooth response at the small scale is subtracted from the smooth response at the large scale, and the difference results of all scale layers are accumulated to obtain the illumination-robust texture response corresponding to the pixel position, thus forming an illumination-robust texture representation.

[0080] The illumination-invariant color representation and the illumination-robust texture representation are concatenated at the pixel level along the channel dimension to obtain an illumination-robust precoded feature map. The illumination-robust precoded feature maps corresponding to all viewpoints are then collected in viewpoint order to form an illumination-robust image feature set.

[0081] In Example 1, the channel dimension splicing method is as follows: while keeping the pixel spatial position unchanged, the illumination-invariant color component and the illumination-robust texture response at the same pixel position are arranged sequentially according to the preset channel arrangement order to form a multi-channel feature tensor containing the sum of the original number of color channels and the number of texture channels, thereby obtaining the illumination-robust pre-coded feature map of the corresponding viewpoint.

[0082] All the illumination robust precoded feature maps corresponding to the viewpoints are arranged in order according to the viewpoint acquisition order in the unlabeled multi-view RGB image dataset, and stacked with the viewpoint index as the first dimension to form a three-dimensional structure data containing multiple viewpoint feature tensors, thus obtaining the illumination robust image feature set. The illumination robust image feature set corresponds one-to-one with the unlabeled multi-view RGB image dataset.

[0083] The unlabeled multi-view RGB image dataset and the illumination-robust image feature set are input into the improved DUSt3R end-to-end 3D reconstruction network to obtain the initial dense 3D point cloud data and the corresponding reconstruction confidence map.

[0084] In this embodiment, an unlabeled multi-view RGB image dataset and an illumination-robust image feature set are input into the improved DUSt3R end-to-end 3D reconstruction network, including:

[0085] Based on each RGB image in the unlabeled multi-view RGB image dataset and the corresponding illumination robust pre-coded feature map, the steel bar reflectivity map and the repeating texture ambiguity map are calculated respectively.

[0086] In Example 1, the steel bar reflectivity map is obtained by taking the difference between the maximum and minimum values ​​of the pixel intensity of the three color channels at the same pixel location, dividing it by the sum of the maximum value of the pixel intensity of the three color channels at the pixel location and the stable term, and cropping the result to the range of zero to one. The steel bar reflectivity map is used to measure the degree of interference of the high-gloss reflection area on the surface of the steel bar cylinder on the three-dimensional coordinate regression.

[0087] ;

[0088] in, Represents pixel coordinates, Indicates the color channel index. Indicates a stable term. This indicates that the result will be cropped to a range. The operator, This represents a graph showing the reflectivity of the reinforcing steel bars. This represents the RGB image acquired from the k-th viewpoint. The repeating texture ambiguity map is obtained by weighted aggregation of the illumination robust pre-coded feature map along the channel dimension to obtain the single-channel texture response. The mean of the single-channel texture response is calculated in the spatial dimension. Then, the absolute value of the difference between the single-channel texture response and the mean at each pixel location is calculated. This absolute value is divided by the sum of the average value and a stable term of the absolute value of the difference between the single-channel texture response and the mean in the spatial dimension. The repeating texture ambiguity map is used to measure the spatial distribution of cross-viewpoint matching ambiguity caused by repeating textures in rebar.

[0089] ;

[0090] in, Represents a precoding feature map robust to illumination. The single-channel texture response operator is obtained by weighted aggregation along the channel dimension. This represents the mean of the single-channel texture response in the spatial dimension. This represents a repeating texture ambiguity map.

[0091] While keeping the pixel spatial position unchanged, the RGB image and the corresponding illumination robust pre-coding feature map are concatenated in the channel dimension to obtain the pixel-level fusion feature. The guiding gate coefficient is calculated based on the steel bar reflectivity map and the repeating texture ambiguity map. The guiding gate coefficient and the pixel-level fusion feature are multiplied element-wise at the corresponding pixel position to obtain the input feature after gate modulation.

[0092] In Example 1, the guiding gating coefficient is obtained by multiplying the reflective sensitive image of the steel bar and the ambiguous image of the repeating texture by the corresponding suppression coefficient, summing them, taking the negative value of the summation result, and then performing an exponential operation. The guiding gating coefficient has a value range of greater than zero and not greater than one.

[0093] The gated and modulated input features are used to generate view guidance feature sequences through block embedding operators and position encoding operators. All view guidance feature sequences corresponding to all viewpoints are collected in viewpoint index order to form a multi-view guidance feature sequence set.

[0094] The multi-view guided feature sequence set is input into the cross-view global attention backbone of the improved DUSt3R end-to-end 3D reconstruction network. During the cross-view attention calculation process, a reflectivity-sensitive bias term and a texture ambiguity bias term are introduced for each pixel position, and the multi-view dense 3D coordinate prediction tensor and reconstruction confidence prediction tensor are output.

[0095] In Example 1, query features, key features, and value features are generated for each viewpoint guided feature sequence in the cross-view global attention backbone network. When calculating the similarity matrix between the query features and key features, a reflection-sensitive bias term and a texture ambiguity bias term are superimposed on the similarity score corresponding to each pixel position to form a modified similarity matrix. The reflection-sensitive bias term is obtained by multiplying the reflective image of the rebar with the reflection-sensitive bias coefficient pixel by pixel. The texture ambiguity bias term is obtained by multiplying the repeated texture ambiguity image with the texture ambiguity bias coefficient pixel by pixel. Both the reflection-sensitive bias term and the texture ambiguity bias term are superimposed on the similarity matrix in the form of negative values, so that the similarity score corresponding to the reflective area of ​​the rebar and the repeated texture area is reduced when matching features across views, thus suppressing erroneous cross-view aggregation.

[0096] Normalization is performed on the modified similarity matrix to generate a cross-view attention weight matrix. The cross-view attention weight matrix is ​​then used to perform a weighted summation of the value features to obtain a cross-view fusion feature tensor. The cross-view fusion feature tensor is then used to generate a multi-view dense 3D coordinate prediction tensor through a 3D coordinate regression head. The 3D coordinate regression head is a coordinate mapping network composed of multiple linear mappings and nonlinear activation functions, and the output is the three coordinate components corresponding to each pixel position.

[0097] The cross-view fusion feature tensor is used to generate a reconstructed confidence prediction tensor through a confidence prediction head. The confidence prediction head is a probability mapping network composed of multiple linear mappings and normalized activation functions. The output is a confidence value between zero and one corresponding to each pixel position. The multi-view dense 3D coordinate prediction tensor and the reconstructed confidence prediction tensor maintain a one-to-one correspondence in the view index and pixel coordinate dimensions.

[0098] Initial dense 3D point cloud data is generated based on the multi-view dense 3D coordinate prediction tensor, and a reconstruction confidence map is generated based on the reconstruction confidence prediction tensor.

[0099] In Example 1, the multi-view dense three-dimensional coordinate prediction tensor is a four-dimensional data structure consisting of a view index, pixel x-coordinate and pixel y-coordinate as three-dimensional index dimensions, and three coordinate components as channel dimensions.

[0100] The process of generating initial dense 3D point cloud data is as follows: For each view index and the corresponding x-coordinate and y-coordinate of each pixel under the view, traverse the three coordinate components of the multi-view dense 3D coordinate prediction tensor at the corresponding index position, combine the three coordinate components into a 3D point coordinate, and collect all 3D point coordinates in the order of view index to form an initial dense 3D point set containing the 3D points corresponding to all pixels, thus obtaining the initial dense 3D point cloud data.

[0101] The three-dimensional point coordinates are located in the preset world coordinate system. Each three-dimensional point coordinate contains the horizontal coordinate component, the vertical coordinate component, and the vertical coordinate component in the preset world coordinate system.

[0102] The process of generating the reconstruction confidence map is as follows: read the confidence values ​​of the reconstruction confidence prediction tensor at each view index and the corresponding pixel horizontal and vertical coordinate positions, and arrange them according to the view index and pixel spatial position to form a set of confidence matrices that are completely consistent with the initial dense 3D point cloud data in the index dimension.

[0103] Each confidence value in the confidence matrix set maintains a strict one-to-one correspondence with the corresponding 3D point coordinates in terms of view index and pixel coordinate dimensions. This ensures that when removing low-confidence points from the initial dense 3D point cloud data, the 3D point coordinates are filtered point by point based on the corresponding confidence value.

[0104] The improved DUSt3R end-to-end 3D reconstruction network outputs absolute spatial scale parameters, and performs absolute scale assignment on the 3D point coordinates of the initial dense 3D point cloud data based on the absolute spatial scale parameters to obtain the initial dense 3D point cloud data carrying absolute spatial scale information.

[0105] In Example 1, a three-dimensional spatial reference object with a known real physical length is pre-set in the acquisition scenario of the precast steel mesh. Local reference feature point cloud corresponding to the three-dimensional spatial reference object is extracted from the initial dense three-dimensional point cloud data. The absolute spatial scale parameter is calculated and output. Based on the absolute spatial scale parameter, the three-dimensional point coordinates of the initial dense three-dimensional point cloud data are assigned an absolute scale value to obtain the initial dense three-dimensional point cloud data carrying the absolute spatial scale information.

[0106] The local reference feature point cloud includes the first relative three-dimensional coordinates and the second relative three-dimensional coordinates of two reference endpoints set on a three-dimensional spatial reference object in the initial dense three-dimensional point cloud data.

[0107] The Euclidean distance between the first relative 3D coordinates and the second relative 3D coordinates is calculated to obtain the relative reconstruction length of the 3D spatial reference object in the initial dense 3D point cloud data. The known true physical length of the 3D spatial reference object is divided by the relative reconstruction length to calculate the absolute spatial scale parameter.

[0108] ;

[0109] ;

[0110] in, and These represent the first relative three-dimensional coordinates and the second relative three-dimensional coordinates, respectively. Represents the L2 norm operator, Indicates the relative reconstruction length. The known true physical length of a reference object in three-dimensional space. This represents the absolute spatial scale parameter.

[0111] The three-dimensional vectors of all three-dimensional point coordinates in the initial dense three-dimensional point cloud data are multiplied by the absolute spatial scale parameter to achieve global scale scaling and uniformly convert them into absolute three-dimensional coordinates in meters or millimeters.

[0112] ;

[0113] in, This represents the initial dense 3D point cloud data. Represents the first element in the initial dense 3D point cloud data. Three-dimensional point coordinates, This represents the initial dense 3D point cloud data carrying absolute spatial scale information. This represents the absolute three-dimensional coordinates after the scale is assigned.

[0114] Based on the reconstructed confidence map, low-confidence points are removed from the initial dense 3D point cloud data to obtain confidence-filtered 3D point cloud data, which is then converted into steel mesh topology data. Based on the steel mesh topology data, topology-preserving correction is performed to obtain topology-corrected 3D point cloud data.

[0115] In this embodiment, topology-preserving correction processing is performed based on the structural data of the steel mesh topology diagram, including:

[0116] Based on the reconstructed confidence map, low-confidence points are removed from the initial dense 3D point cloud data to obtain confidence-filtered 3D point cloud data;

[0117] In Example 1, the initial dense 3D point cloud data and the reconstructed confidence map are made to have a strict one-to-one correspondence in terms of view index and pixel coordinate dimension.

[0118] For each view index and the corresponding pixel x and y coordinates, the 3D point coordinates of the initial dense 3D point cloud data at the corresponding index position are read, and the reconstruction confidence value of the reconstruction confidence map at the corresponding index position is read.

[0119] When the reconstruction confidence value at the corresponding location is greater than or equal to the confidence threshold, the corresponding 3D point coordinates are retained; when the reconstruction confidence value at the corresponding location is less than the confidence threshold, the corresponding 3D point coordinates are discarded. All retained 3D point coordinates are then aggregated in the order of view index to form confidence-filtered 3D point cloud data.

[0120] Convert confidence-based filtering of 3D point cloud data into steel mesh topology structure data;

[0121] In Example 1, the confidence-filtered 3D point cloud data is represented as a node set V, where each node in the node set corresponds to the coordinates of a 3D point in the confidence-filtered 3D point cloud data.

[0122] The topology data of the steel mesh is constructed using a set of nodes. The topology data of the steel mesh includes a set of nodes, a set of edges, and a set of edge weights. For any two different nodes in the set of nodes, the Euclidean distance between the corresponding three-dimensional point coordinates of the two nodes is calculated. When the Euclidean distance between two nodes is less than or equal to the neighborhood radius threshold, an undirected edge connecting the two nodes is established in the set of edges.

[0123] For each node in the node set, select a set of neighboring nodes within the neighborhood radius threshold range. Calculate the local normal vector of the node by performing least-squares plane fitting on the neighboring node set. For each established undirected edge, calculate the dot product between the local normal vectors of the nodes at both ends of the undirected edge. When the dot product is greater than or equal to the normal consistency threshold, the undirected edge is retained as a continuous reinforcement edge; when the dot product is less than the normal consistency threshold, the undirected edge is deleted.

[0124] For each undirected edge retained as a continuous rebar edge, the angle deviation between the square of the Euclidean distance between the two ends of the undirected edge and the local normal vector is calculated. The edge weight of the undirected edge is calculated based on the square of the distance and the angle deviation. All nodes, retained continuous rebar edges and their corresponding edge weights are collected to form the rebar mesh topology structure data.

[0125] Topology-preserving correction processing is performed on the structural data of the steel mesh topology diagram to obtain topology-corrected 3D point cloud data.

[0126] In Example 1, the node coordinates of the confidence-filtered 3D point cloud data are stacked by index to form a coordinate matrix. With the constraints of maintaining the continuity of the steel reinforcement structure and suppressing local drift caused by reflection and occlusion, topology-preserving correction is performed on the coordinate matrix to obtain the topology-corrected coordinate matrix. Each row in the topology-corrected coordinate matrix corresponds to the corrected 3D coordinate vector of a node. The topology-corrected coordinate matrix is ​​restored to the topology-corrected 3D point cloud data according to the index order of the node set, and the topology-corrected 3D point cloud data and the node set maintain a strict one-to-one correspondence in the node index dimension.

[0127] ;

[0128] in, Represents the topological correction coordinate matrix. Describe the Frobenius norm. This represents the coordinate consistency term under the constraints of the steel mesh topology structure data, used to pull the nodes corresponding to the continuous edges of the steel bars back to the spatial trajectory of the continuous cylindrical shape. To ensure that the coordinates after forced correction do not deviate too far from the original 3D point cloud, To ensure geometric continuity between adjacent nodes, if a steel bar breaks due to reflection or obstruction, this item will pull the break back onto the continuous curve trajectory.

[0129] A flowing surface skeleton was fitted to the topology-corrected 3D point cloud data to obtain a macroscopic skeleton surface model of the steel mesh.

[0130] In this embodiment, the topology-corrected 3D point cloud data is fitted with a flowing surface skeleton, including:

[0131] According to the acquisition sequence of the topology-corrected 3D point cloud data, the topology-corrected 3D point cloud data is segmented and divided into continuously arriving topology-corrected 3D point cloud segments.

[0132] In Example 1, each topology-corrected 3D point cloud segment represents a subset of topology-corrected 3D point cloud data arriving within the same time window. The point cloud segments are arranged sequentially in time to form a time-ordered data structure for fitting the flowing surface skeleton.

[0133] Within each topology-corrected 3D point cloud fragment, a set of representative points for skeleton fitting is constructed, and the set of representative points is accumulated in chronological order to form a streaming set of representative points.

[0134] In Example 1, the topology-corrected 3D point cloud segment is divided into spatial grids according to the preset spatial grid side length. Within each grid cell, the arithmetic mean of the horizontal, vertical, and longitudinal coordinate components of all 3D point coordinates within the grid cell is calculated. The obtained average values ​​are combined as the centroid coordinates of the grid cell, and the centroid coordinates are used as a representative point. The representative points corresponding to all grid cells constitute the representative point set at the current time.

[0135] The current set of representative points is combined with the historical set of representative points in chronological order to form a streaming set of representative points from the start time to the current time. The streaming set of representative points is used to depict the overall spatial distribution trend of the steel mesh on a macro scale.

[0136] Based on the representative point set of the flow cytometry, a macroscopic skeleton surface model of the steel mesh is defined;

[0137] In Example 1, the macroscopic skeleton surface model of the steel mesh is represented as a tensor product of the control point grid and the basis functions.

[0138] Establish a control point grid, which consists of multiple control points. Each control point is a 3D control point with coordinates in a preset world coordinate system. Define a set of basis functions in two parameter directions, namely the first parameter direction and the second parameter direction.

[0139] The macroscopic skeleton surface model of the steel mesh is obtained by multiplying all control points by the basis function values ​​in the corresponding parameter directions and then performing a weighted summation. The output of the macroscopic skeleton surface model of the steel mesh is the coordinates of the three-dimensional surface points in the preset world coordinate system. The macroscopic skeleton surface model of the steel mesh is used to describe the overall flexible deformation trend of the steel mesh.

[0140] ;

[0141] in, and Represents surface parameters, Indicates the control point grid in Upper bound of the direction control point index Indicates the control point grid in Upper bound of the direction control point index express Direction Secondary basis functions express Direction Secondary basis functions Indicates control points, This is a macroscopic skeleton surface model of the steel mesh.

[0142] A flowing surface skeleton fitting is performed on the macroscopic skeleton surface model of the steel mesh, and the control point set is incrementally updated based on the representative point set at the current time to obtain the macroscopic skeleton surface model of the steel mesh at the current time.

[0143] In Example 1, for each representative point in the set of representative points at the current time, a corresponding first parameter direction parameter value and a second parameter direction parameter value are assigned to the representative point. The first parameter direction parameter value and the second parameter direction parameter value are obtained by normalizing the projection of the representative point within the bounding box of the control point grid.

[0144] Construct a streaming fitting objective function. The first part is the sum of squared Euclidean distances between the representative point and the surface point. The second part is the second-order difference smoothing term of the surface in the parameter domain, which is used to limit the curvature of the surface and ensure the continuity of the surface. The incremental update of the control points is obtained by minimizing the streaming fitting objective function.

[0145] ;

[0146] in, Indicates the first At the [time]th moment The fitting weights for each representative point are dimensionless. This represents the Euclidean distance from a point to a point on the surface. Represents the surface smoothing regularity coefficient. This represents the result of the second-order difference operator of the surface with respect to the parameter domain. Indicates the first The moment of the first One representative point, This represents the objective function for flow cytometry fitting.

[0147] A streaming forgetting factor is introduced to preserve historical control points, and the current control point update is linearly combined with the historical control points according to a preset ratio to obtain the control point coordinates at the current moment. The updated control point coordinate set is then substituted into the expression of the macroscopic skeleton surface model of the steel mesh to obtain the macroscopic skeleton surface model of the steel mesh at the current moment.

[0148] ;

[0149] in, Represents the streaming forgetting factor. Indicates time control points, Indicates time control points, Indicated by minimizing The moment of gain Control point increments.

[0150] The topology-corrected 3D point cloud data and the macroscopic skeleton surface model of the steel mesh are flexibly registered with the 3D design standard model to obtain a flexibly registered aligned point cloud model.

[0151] In this embodiment, the topology-corrected 3D point cloud data and the macroscopic skeleton surface model of the steel mesh are flexibly registered with the 3D design standard model, including:

[0152] A macroscopic deformation field is established based on the current moment's macroscopic skeleton surface model of the steel mesh. The topology-corrected three-dimensional point cloud data is mapped onto the reference plane of the macroscopic deformation field. Macroscopic deformation displacement stripping is performed to obtain a macroscopic deformation stripped point cloud containing only high-frequency local morphological deviations.

[0153] In Example 1, for each 3D point coordinate in the topology-corrected 3D point cloud data, the coordinates are orthogonally projected onto the macroscopic skeleton surface model of the reinforcing mesh by minimizing the Euclidean distance, and the optimal surface parameters are solved. The 3D surface point coordinates of the macroscopic skeleton surface model of the reinforcing mesh are calculated at the optimal surface parameters. The difference vector between the 3D point coordinates and the 3D surface point coordinates is defined as the macroscopic deformation displacement vector. :

[0154] ;

[0155] in, This represents an ideal reference plane model that is not affected by gravitational deflection.

[0156] Displacement compensation is performed on the coordinates of each three-dimensional point along the opposite direction of the macroscopic deformation displacement vector to obtain the macroscopic deformation stripping point cloud.

[0157] The macroscopic deformation stripping point cloud is subjected to non-rigid registration and solution with the 3D design standard model. The overall rigid pose alignment and local non-rigid deviation are decoupled, and a flexible registration and alignment point cloud model with accurate topological matching relationship and deviation attribute is output.

[0158] In Example 1, the macroscopic deformation stripping point cloud is used as the source point set to be registered, and the surface sampling point cloud of the three-dimensional design standard model is used as the target point set to construct a joint registration optimization target that includes the overall rigid transformation parameters and the local non-rigid displacement field.

[0159] The optimal global rotation matrix and optimal global translation vector for eliminating global spatial coordinate system misalignment between the source point set and the target point set are solved by alternating iterative calculations, and the optimal local non-rigid displacement vector attached to the coordinates of each three-dimensional point in the source point set is solved simultaneously.

[0160] After the joint registration optimization objective converges, the optimal global rotation matrix, optimal global translation vector, and optimal local non-rigid displacement vector are applied sequentially to the coordinates of each 3D point in the macroscopic deformation stripped point cloud. The 3D target coordinates aligned with the 3D design standard model in spatial position are calculated. Each 3D target coordinate and its corresponding optimal local non-rigid displacement vector are structurally bound and combined into a 3D data tuple with spatial deformation attributes. All 3D data tuples are collected according to the original node index order of the point cloud to generate a flexible registration aligned point cloud model.

[0161] The flexible registration and alignment point cloud model achieves flexible fitting and alignment with the three-dimensional design standard model in terms of spatial geometry, and the optimal local non-rigid displacement vector bound inside its data tuples completely preserves the actual local physical deviation caused by steel bar arrangement errors, binding misalignment or dimensional deviation.

[0162] The optimal global rotation matrix is The matrix represents the rotation angles (pitch, yaw, roll) of an object around the X, Y, and Z axes in three-dimensional space. The optimal overall rotation matrix is ​​used in rebar detection to eliminate overall angular tilt.

[0163] Assuming the standard 3D design model is placed due south and due north, but the steel mesh on the production line is slightly misaligned on the conveyor belt or the camera is installed at a slightly skewed angle, the role of the optimal overall rotation matrix is ​​to treat the entire measured steel mesh as an absolutely rigid whole and rotate and reset it in 3D digital space so that it is completely parallel to the standard 3D design model in the macroscopic direction.

[0164] The optimal global translation vector is a vector containing three elements (dx, dy, dz), representing the straight-line movement distance in three-dimensional space. In rebar detection, it is used to eliminate the overall positional offset.

[0165] The center point of the 3D design standard model is at the origin (0,0,0) of the coordinate system, but the center of the steel mesh captured by the camera is actually at (2 meters, 1 meter, 0.5 meters). The role of the optimal overall translation vector is to treat the entire steel mesh as a whole and translate it back to the origin so that it coincides with the approximate outline of the 3D design standard model.

[0166] The optimal local non-rigid displacement vector is not a single value, but a vector field. Each measured 3D point i has a unique displacement vector. It is used in rebar inspection to quantify the actual local manufacturing / tying errors.

[0167] After aligning the steel mesh with the two parameters mentioned above, we find that due to improper binding by the workers, some steel bars are not perfectly aligned. For example, the second longitudinal bar should be at the 20 cm position, but it is actually bound at the 22 cm position. In order to map this incorrect measured point to the target benchmark nonlinear space onto the perfect 3D design standard model, we need to apply an additional local spatial displacement compensation to this point. The torque and direction of the additional local spatial displacement compensation are the local non-rigid displacement vector. If a steel bar is bound perfectly, its local displacement vector is 0. If the steel bar is 2 cm off, the length of its local displacement vector is 2 cm, and the direction is pointing to its correct position.

[0168] This implementation constructs a flexible registration and alignment point cloud model with spatial deformation properties, accurately isolates the rigid placement posture error of the entire mesh, and precisely quantifies the actual arrangement misalignment and binding abnormality into independent local non-rigid displacement vectors. In uncontrolled industrial scenarios, it achieves a fully automatic, interference-resistant, and highly physically interpretable millimeter-level high-precision detection closed loop for the three-dimensional morphology of steel bars.

[0169] Based on the spatial difference between the flexible registration and alignment point cloud model and the three-dimensional design standard model, the three-dimensional morphological deviation vector of a single steel bar, the steel bar spacing deviation, and the overall mesh flatness deviation are calculated to obtain multi-dimensional steel bar morphological deviation data.

[0170] In this embodiment, based on the spatial difference between the flexible registration and alignment point cloud model and the three-dimensional design standard model, the three-dimensional shape deviation vector of a single rebar, the rebar spacing deviation, and the overall mesh flatness deviation are calculated, including:

[0171] Extract the deformation attributes encapsulated in the flexible registration and alignment point cloud model, and establish a set of point-level spatial difference vectors;

[0172] In Example 1, each three-dimensional data tuple in the flexible registration and alignment point cloud model is analyzed, the optimal local non-rigid displacement vector bound in the three-dimensional data tuple is directly extracted, and its negative value is taken, which is defined as the point-level spatial difference vector. All point-level spatial difference vectors constitute the point-level spatial difference vector set.

[0173] Based on the point-level spatial difference vector set, calculate the three-dimensional morphological deviation vector of a single steel bar;

[0174] In Example 1, the topology data of the steel mesh corresponding to the flexible registration and alignment point cloud model is used as the basis for grouping steel bars. Each steel bar in the topology data of the steel mesh corresponds to a set of node indices.

[0175] For each point-level spatial difference vector in the node index set corresponding to each rebar, multiply it by the corresponding aggregation weight and then sum it component by component. Divide the sum of component by the sum of all aggregation weights to obtain the three-dimensional morphological deviation vector corresponding to the rebar. Collect all the three-dimensional morphological deviation vectors corresponding to the rebars in the order of rebar number to represent the average offset direction and average offset of each rebar in the overall space.

[0176] Based on the flexible registration and alignment point cloud model and the three-dimensional design standard model, the deviation of the rebar spacing is calculated.

[0177] In Example 1, for each rebar in the flexible registration and alignment point cloud model, the coordinates of its corresponding three-dimensional points are sorted along the rebar's direction, and samples are taken at equal arc length intervals to obtain the set of centerline points of the rebar. The same equal arc length sampling is performed on the standard rebar corresponding to the rebar in the three-dimensional design standard model to obtain the set of standard centerline points.

[0178] For any two adjacent reinforcing bars, calculate the Euclidean distance between the centerline points of the two reinforcing bars at the same sampling number position. Take the arithmetic mean of the Euclidean distances at all sampling number positions to obtain the actual spacing between the two reinforcing bars. Calculate the standard spacing between the corresponding standard centerline point sets in the same way. Subtract the standard spacing from the actual spacing to obtain the reinforcing bar spacing deviation, which is used to represent the deviation of the reinforcing bar spacing from the three-dimensional design standard model.

[0179] Based on the flexible registration and alignment point cloud model, the overall mesh flatness deviation is calculated;

[0180] In Example 1, least squares plane fitting is performed on the flexible registration and alignment point cloud model to obtain a flatness reference plane. For each three-dimensional point coordinate in the flexible registration and alignment point cloud model, the signed distance from the three-dimensional point coordinate to the flatness reference plane is calculated. The maximum and minimum values ​​are taken among all signed distances, and the maximum value is subtracted from the minimum value to obtain the overall mesh flatness deviation. The overall mesh flatness deviation is used to represent the remaining high-frequency spatial fluctuation amplitude of the flexible registration and alignment point cloud model after stripping macroscopic deformation.

[0181] The three-dimensional morphological deviation vector of a single rebar, the rebar spacing deviation, and the overall mesh flatness deviation are structurally aggregated to obtain multi-dimensional rebar morphological deviation data.

[0182] The multidimensional rebar morphology deviation data are graded and judged to generate a rebar morphology deviation detection report.

[0183] In this embodiment, the multidimensional rebar morphology deviation data is graded and judged, including:

[0184] Establish a grading standard for multidimensional steel bar morphological deviation data;

[0185] Based on the hierarchical judgment criteria, abnormal node binding information is identified, forming a set of abnormal node binding information.

[0186] In Example 1, for each rebar intersection node in the rebar mesh topology data, the coordinates of the three-dimensional points associated with the rebar intersection node are extracted, and the point-level spatial difference vector corresponding to each three-dimensional point coordinate associated with the rebar intersection node is statistically analyzed.

[0187] When the amplitude of the three-dimensional shape deviation of the rebars corresponding to multiple three-dimensional points associated with the same rebar intersection node is greater than or equal to the node abnormality threshold, it is determined that there is a binding abnormality at the corresponding rebar intersection node.

[0188] When the absolute value of the vertical difference component of the three-dimensional point associated with the same rebar intersection node is greater than or equal to the vertical offset threshold, it is determined that the corresponding rebar intersection node has loose binding or abnormal intersection height.

[0189] All rebar intersections that are identified as abnormal are collected in order of their node index to form a node binding anomaly identification information set. This set is used to indicate the location of nodes with binding offset, loose binding, or abnormal intersection height, and their corresponding deviations.

[0190] The amplitude of the three-dimensional shape deviation of the reinforcing bar is obtained by squaring the transverse, longitudinal, and vertical difference components of the reinforcing bar, summing the squared results, and taking the square root of the summation result. The amplitude of the three-dimensional shape deviation of the reinforcing bar is used to represent the comprehensive offset of the reinforcing bar in three-dimensional space. The transverse, longitudinal, and vertical difference components of the reinforcing bar are the coordinates of the three-dimensional shape deviation vector of a single reinforcing bar.

[0191] Based on the grading criteria, determine the overall quality level information;

[0192] In Example 1, the number of steel bars with a three-dimensional shape deviation amplitude greater than or equal to the threshold value of the three-dimensional shape deviation amplitude of the steel bars is counted, the number of adjacent steel bar pairs with a steel bar spacing deviation greater than or equal to the steel bar spacing deviation threshold value is counted, and it is determined whether the overall mesh flatness deviation is greater than or equal to the overall mesh flatness deviation threshold value. Based on the relationship between the number of abnormal steel bar three-dimensional shape deviations, the number of abnormal steel bar spacing deviations, and the overall mesh flatness deviation and the corresponding threshold values, the overall quality grade information is defined.

[0193] In this embodiment, the overall quality level information is determined according to the following rules:

[0194] When the number of abnormal deviations in the three-dimensional shape of the reinforcing bars is zero, the number of abnormal deviations in the spacing of the reinforcing bars is zero, and the overall flatness deviation of the mesh is less than the overall flatness deviation threshold, it is judged to be of qualified grade.

[0195] When at least one of the abnormal number of three-dimensional shape deviations of steel bars or abnormal number of steel bar spacing deviations is greater than zero and does not exceed a preset proportional threshold, it is determined to be a limit level.

[0196] When the number of abnormal deviations in the three-dimensional shape of the reinforcing bars or the number of abnormal deviations in the spacing of the reinforcing bars exceeds the preset proportional threshold, or when the overall flatness deviation of the mesh is greater than or equal to the overall flatness deviation threshold, it is judged as unqualified.

[0197] A rebar morphology deviation detection report is generated based on the set of node binding anomaly identification information and the overall quality grade information.

[0198] In this embodiment, a method for inspecting the quality of reinforcing steel bars in precast components includes:

[0199] The image acquisition module acquires uncalibrated multi-view RGB image datasets of precast steel mesh and performs illumination-robust precoding processing to generate an illumination-robust image feature set.

[0200] The 3D reconstruction module inputs the unlabeled multi-view RGB image dataset and the illumination-robust image feature set into the improved DUSt3R end-to-end 3D reconstruction network to obtain the initial dense 3D point cloud data and the corresponding reconstruction confidence map.

[0201] The topology correction module removes low-confidence points from the initial dense 3D point cloud data based on the reconstructed confidence map and performs correction processing to obtain topology-corrected 3D point cloud data.

[0202] The skeleton fitting module performs flow surface skeleton fitting on topology-corrected 3D point cloud data to obtain a macroscopic skeleton surface model of steel mesh.

[0203] The flexible registration module flexibly registers the topology-corrected 3D point cloud data and the macroscopic skeleton surface model of the steel mesh with the 3D design standard model to obtain a flexibly registered aligned point cloud model.

[0204] The deviation calculation module obtains multi-dimensional rebar morphology deviation data based on the spatial difference between the flexible registration and alignment point cloud model and the three-dimensional design standard model.

[0205] The grading and judgment module grades and judges the multi-dimensional rebar morphology deviation data and generates a rebar morphology deviation detection report.

[0206] Example 2: During a continuous inspection process, the system performs multi-view acquisition on a precast component's steel mesh, obtaining an uncalibrated multi-view RGB image dataset. The selected image is the first... An RGB image from multiple perspectives As an example, the system in pixel coordinates at this viewpoint The pixel intensities of the three color channels read at the location are respectively The pixel is located at the edge of the highlight strip on the surface of the steel bar.

[0207] System settings stability items (Pixel intensity dimensionless stable term), calculate the reflectivity map of reinforcing bars. :

[0208] ;

[0209] From the same viewpoint, the system focuses on another highlight center pixel. Read channel strength Substituting, we can get .

[0210] Based on this, the system obtains a lower reflectivity value in the highlight center area and a higher reflectivity value at the highlight edge, forming a reflectivity spatial distribution. In the calculation of the repeating texture ambiguity map, the system... Viewpoint-based illumination robust precoding feature map Perform channel-weighted aggregation to obtain single-channel texture response. The system statistically obtains the spatial mean of the texture response from this viewpoint within a region of repetitive texture on a section of rebar. Furthermore, the spatial mean absolute deviation was statistically obtained. Stable terms are taken For pixels Read at [location] :

[0211] ;

[0212] The system obtains a repeating texture ambiguity map saturation value of 1 at this pixel, indicating that the texture response at this location deviates significantly from the mean and belongs to a highly ambiguous region for cross-view matching.

[0213] The system calculates the guiding gating coefficient for pixel positions, assuming the reflection suppression coefficient is... Texture ambiguity suppression coefficient is The gating coefficient is defined as follows: .

[0214] The system multiplies the gating coefficients element-wise with the pixel-level fused features, causing the amplitude of the input feature of each pixel to be attenuated to approximately [a fraction of its original value]. Entering the confidence screening stage, the system outputs a reconstructed confidence map using the improved DUSt3R, and reads the reconstructed confidence values ​​for the same pixel location from the same viewpoint. And set a confidence threshold. .because The system removes the 3D point coordinates corresponding to the corresponding pixel and then processes the other pixel. Read ,because The system retains the 3D coordinates of the corresponding pixel. The system calculates the initial number of points from this viewpoint as follows: Points, the number of points removed is The elimination rate was 10.7%.

[0215] In the topology-preserving correction phase, the system stacks the node coordinates of the confidence-filtered 3D point cloud data into a coordinate matrix. and set To provide verifiable numerical examples, the system selects five consecutive nodes on the same rebar to form a local subgraph, with a local coordinate matrix. for:

[0216] ;

[0217] The vertical coordinate of the third node (0.020m) exhibits an anomalous jump of 0.008m relative to its adjacent node (0.012m). The system constructs the corresponding graph Laplacian matrix on this local subgraph. (Chained connection, endpoint degree is 1, intermediate degree is 2):

[0218] ;

[0219] The system is solved, and the corrected local solution is obtained. for:

[0220] ;

[0221] The implementers found that the vertical anomaly of the third node was reduced from 0.020m to 0.014m, and the anomaly jump decreased from 0.008m to 0.002m. The system also recorded two costs for this local subgraph:

[0222] ;

[0223] Substitution Obtain the objective function value This serves as a verifiable numerical record for the local topology correction.

[0224] Entering the macroscopic deformation stripping and flexible registration stage, the system obtains the optimal surface parameters from the coordinates of a certain 3D point in the topology-corrected 3D point cloud data. The system calculates the output points of the macroscopic skeleton surface model of the steel mesh at the corresponding parameters. Ideal reference plane model output points The system calculates the macroscopic deformation displacement vector. .

[0225] The system compensates the three-dimensional point coordinates by 0.006m in the opposite direction of the macroscopic deformation displacement vector, so that the three-dimensional point returns to the reference position on the reference plane after the macroscopic deflection is removed.

[0226] Entering the deviation quantization stage, the system analyzes the 3D data tuple corresponding to a node in the flexible registration and alignment point cloud model, directly extracts the optimal local non-rigid displacement vector (-0.015, 0.002, -0.001)m bound in the 3D data tuple, and calculates the point-level spatial difference vector as follows: ;

[0227] The magnitude of the vector is .

[0228] The system aggregates 20 points from the node index set of the same rebar, assuming the aggregation weight is 1. And obtain the three-dimensional morphological deviation vector of the steel bar. The amplitude is .

[0229] In the calculation of rebar spacing deviation, the system obtains 10 Euclidean distance samples at the same sampling sequence position for two adjacent rebars:

[0230] ;

[0231] The arithmetic mean yields the actual spacing. The corresponding standard spacing is The deviation of the rebar spacing is .

[0232] In the same round of detection, after the system generates the flexible registration and alignment point cloud model, it enters the deviation quantification and hierarchical judgment stage. The system establishes a nearest neighbor correspondence for the coordinates of each 3D point in the flexible registration and alignment point cloud model, forming a point-level correspondence mapping, and calculates the point-level spatial difference vector set point by point. The system records in the log that the average nearest neighbor distance for this round of point matching is 0.0016m, the maximum nearest neighbor distance is 0.0069m, and the point matching success rate (the percentage of points whose nearest neighbor distance is no more than 0.01m) is [missing information]. .

[0233] The system then aggregates the point-level spatial difference vector set according to the topological connectivity of the rebar mesh, grouping and aggregating the rebars, and outputs a set of three-dimensional morphological deviation vectors for individual rebars. The mesh contains... The system provides an over-limit warning for the magnitude of the three-dimensional morphological deviation vector of three of the reinforcing bars.

[0234] The system sets the threshold value for the three-dimensional shape deviation of the reinforcing bars to be [value]. The threshold value for the deviation of rebar spacing is The overall mesh flatness deviation threshold is The node anomaly threshold is The vertical offset threshold is 0.006m.

[0235] The system extracts the first from the set of three-dimensional shape deviation vectors of the reinforcing bars. The polymerization results of the rebar were used as a record sample to obtain... And calculate the amplitude according to the following formula. ,because , No. The rebar does not trigger the three-dimensional shape deviation exceeding the limit.

[0236] The system in the same round for the first Rebar output The amplitude is ,because The system recorded that rebar number 19 had an excessive three-dimensional morphological deviation of 0.00036m and entered the defect attribution statistics queue.

[0237] The system for the first Rebar output Amplitude ,because The system recorded that rebar number 27 had an excessive three-dimensional morphological deviation of 0.00161m.

[0238] The system for the first Rebar output The amplitude is ,because The system recorded that rebar number 33 had an excessive three-dimensional morphological deviation of 0.00306m.

[0239] In the calculation of rebar spacing deviation, the system obtains the actual spacing samples at 10 sampling sequence positions of a pair of adjacent rebars (numbered 19 and 20) with equal arc lengths: Its average actual spacing The corresponding standard spacing is Therefore, the spacing deviation is ,because The system recorded that the spacing deviation of adjacent reinforcing bars (19, 20) exceeded the limit by 0.0009m.

[0240] The system obtained the average actual spacing for another pair of adjacent reinforcing bars (numbered 27 and 28). The corresponding standard spacing is 0.2000m, and the spacing deviation is... ,because Adjacent steel bars do not trigger over-limit events.

[0241] The system satisfies the requirement for centering of adjacent reinforcing bars across the entire network. The number of overlimit pairs is .

[0242] In the calculation of overall mesh flatness deviation, the system performs least-squares plane fitting on the flexible registration and alignment point cloud model to obtain a flatness reference plane. The system calculates the signed distance from all points to this plane and records the maximum signed distance. The minimum signed distance is Therefore, the overall flatness deviation of the mesh is ,because The overall flatness deviation of the mesh does not exceed the limit, and the system marks the flatness as "passed" in the log.

[0243] In the identification of node binding anomalies, the system establishes a local node index set for rebar intersections and extracts point-level spatial difference vectors within the neighborhood of each node. At a given intersection, the system statistically analyzes the point-level spatial difference vector amplitudes of eight points within the node's neighborhood: 0.0132, 0.0128, 0.0141, 0.0135, 0.0122, 0.0130, 0.0126, and 0.0138. The system then calculates the mean amplitude of these neighborhood amplitudes. ,because The system determines that a node has a risk of binding anomalies. Simultaneously, the system checks the set of vertically dissimilar components in the node's neighborhood and records the largest vertical absolute value among them. .because The system labels the node anomaly as cross height anomaly / loose binding risk and writes it into the node binding anomaly identification information set.

[0244] The system counted the number of nodes that met the anomaly criteria across the entire network. And record three of the nodes that simultaneously satisfy and Under these two conditions, the system generates multi-dimensional rebar morphology deviation data and proceeds to the overall quality grade determination. The system then counts the number of rebars exceeding the three-dimensional morphology deviation limit. The system sets a proportional threshold. And calculate the excess ratio. ,because ,and Not zero but small in quantity, at the same time The system will determine the overall quality level as the limit level and write "Overall Quality Level: Limit Level" in the report header.

[0245] The system ultimately generates a rebar morphology deviation detection report, which is output in structured text fields and includes information on node binding anomalies and overall quality level information.

[0246] To compare the output of traditional methods, the system ran the traditional multi-view geometry method and the unmodified end-to-end reconstruction method in parallel on the same mesh, using the same 3D design standard model as a reference. The traditional multi-view geometry method failed to match in the reflective area, resulting in local point cloud loss. The system recorded that the effective point ratio in the area corresponding to rebar number 33 was only 61% of that of the method in this invention, and a continuous missing section of approximately 0.028m appeared, causing the centerline sampling of this rebar to fail and the spacing deviation to be output unstably. While the unmodified end-to-end reconstruction method could output centerline sampling, under the condition that macroscopic deflection was not stripped, the system calculated the overall mesh flatness deviation to be 0.0129m, exceeding... 0.008m, and the overall quality grade was determined to be unqualified, while the method of this invention obtains after macroscopic deformation peeling. It is determined to be at the limit level, and the difference between the two allows the implementer to observe that macroscopic downward misjudgment has been eliminated.

[0247] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for inspecting the quality of reinforcing steel bars in precast components, characterized in that, include: Unlabeled multi-view RGB image datasets of precast steel mesh are collected and subjected to illumination-robust precoding to generate illumination-robust image feature sets. The unlabeled multi-view RGB image dataset and the illumination-robust image feature set are input into the improved DUSt3R end-to-end 3D reconstruction network to obtain the initial dense 3D point cloud data and the corresponding reconstruction confidence map. Based on the reconstructed confidence map, low-confidence points are removed from the initial dense 3D point cloud data to obtain confidence-filtered 3D point cloud data, which is then converted into steel mesh topology data. Based on the steel mesh topology data, topology-preserving correction is performed to obtain topology-corrected 3D point cloud data. A flowing surface skeleton was fitted to the topology-corrected 3D point cloud data to obtain a macroscopic skeleton surface model of the steel mesh. The topology-corrected 3D point cloud data and the macroscopic skeleton surface model of the steel mesh are flexibly registered with the 3D design standard model to obtain a flexibly registered aligned point cloud model. Based on the spatial difference between the flexible registration and alignment point cloud model and the three-dimensional design standard model, the three-dimensional morphological deviation vector of a single steel bar, the steel bar spacing deviation, and the overall mesh flatness deviation are calculated to obtain multi-dimensional steel bar morphological deviation data. The multidimensional rebar morphology deviation data are graded and judged to generate a rebar morphology deviation detection report.

2. The method for quality inspection of precast steel reinforcement according to claim 1, characterized in that, The process of acquiring unlabeled multi-view RGB image datasets of precast steel mesh and performing illumination-robust precoding includes: Multiple views of the same precast steel mesh were simultaneously acquired to form an uncalibrated multi-view RGB image dataset. Perform illumination consistency normalization on each RGB image in the unlabeled multi-view RGB image dataset to obtain an illumination-normalized RGB image; Based on illumination-normalized RGB images, construct illumination-invariant color representations; Multi-scale illumination-robust precoding is performed on illumination-normalized RGB images to form illumination-robust texture representations; The illumination-invariant color representation and the illumination-robust texture representation are concatenated at the pixel level along the channel dimension to obtain an illumination-robust precoded feature map. The illumination-robust precoded feature maps corresponding to all viewpoints are then collected in viewpoint order to form an illumination-robust image feature set.

3. The method for quality inspection of precast steel reinforcement according to claim 1, characterized in that, The process of inputting an unlabeled multi-view RGB image dataset and an illumination-robust image feature set into the improved DUSt3R end-to-end 3D reconstruction network includes: Based on each RGB image in the unlabeled multi-view RGB image dataset and the corresponding illumination robust pre-coded feature map, the steel bar reflectivity map and the repeating texture ambiguity map are calculated respectively. While keeping the pixel spatial position unchanged, the RGB image and the corresponding illumination robust pre-coding feature map are concatenated in the channel dimension to obtain the pixel-level fusion feature. The guiding gate coefficient is calculated based on the steel bar reflectivity map and the repeating texture ambiguity map. The guiding gate coefficient and the pixel-level fusion feature are multiplied element-wise at the corresponding pixel position to obtain the input feature after gate modulation. The gated and modulated input features are used to generate view guidance feature sequences through block embedding operators and position encoding operators. All view guidance feature sequences corresponding to all viewpoints are collected in viewpoint index order to form a multi-view guidance feature sequence set. The multi-view guided feature sequence set is input into the cross-view global attention backbone of the improved DUSt3R end-to-end 3D reconstruction network. During the cross-view attention calculation process, a reflectivity-sensitive bias term and a texture ambiguity bias term are introduced for each pixel position, and the multi-view dense 3D coordinate prediction tensor and reconstruction confidence prediction tensor are output. Initial dense 3D point cloud data is generated based on the multi-view dense 3D coordinate prediction tensor, and a reconstruction confidence map is generated based on the reconstruction confidence prediction tensor. The improved DUSt3R end-to-end 3D reconstruction network outputs absolute spatial scale parameters, and performs absolute scale assignment on the 3D point coordinates of the initial dense 3D point cloud data based on the absolute spatial scale parameters to obtain the initial dense 3D point cloud data carrying absolute spatial scale information.

4. The method for quality inspection of precast steel reinforcement according to claim 1, characterized in that, The topology-preserving correction process performed on the structural data based on the steel mesh topology diagram includes: Based on the reconstructed confidence map, low-confidence points are removed from the initial dense 3D point cloud data to obtain confidence-filtered 3D point cloud data; Convert confidence-based filtering of 3D point cloud data into steel mesh topology structure data; Topology-preserving correction processing is performed on the structural data of the steel mesh topology diagram to obtain topology-corrected 3D point cloud data.

5. The method for quality inspection of precast steel reinforcement according to claim 1, characterized in that, The process of fitting a flowing surface skeleton to the topology-corrected 3D point cloud data includes: According to the acquisition sequence of the topology-corrected 3D point cloud data, the topology-corrected 3D point cloud data is segmented and divided into continuously arriving topology-corrected 3D point cloud segments. Within each topology-corrected 3D point cloud fragment, a set of representative points for skeleton fitting is constructed, and the set of representative points is accumulated in chronological order to form a streaming set of representative points. Based on the representative point set of the flow cytometry, a macroscopic skeleton surface model of the steel mesh is defined; A flowing surface skeleton fitting is performed on the macroscopic skeleton surface model of the steel mesh, and the control point set is incrementally updated based on the representative point set at the current time to obtain the macroscopic skeleton surface model of the steel mesh at the current time.

6. The method for quality inspection of precast steel reinforcement according to claim 1, characterized in that, The process of flexibly registering the topology-corrected 3D point cloud data and the macroscopic skeleton surface model of the steel mesh with the 3D design standard model includes: A macroscopic deformation field is established based on the current moment's macroscopic skeleton surface model of the steel mesh. The topology-corrected three-dimensional point cloud data is mapped onto the reference plane of the macroscopic deformation field. Macroscopic deformation displacement stripping is performed to obtain a macroscopic deformation stripped point cloud containing only high-frequency local morphological deviations. The macroscopic deformation stripping point cloud is subjected to non-rigid registration and solution with the 3D design standard model. The overall rigid pose alignment and local non-rigid deviation are decoupled, and a flexible registration and alignment point cloud model with accurate topological matching relationship and deviation attribute is output.

7. The method for quality inspection of precast steel reinforcement according to claim 1, characterized in that, The calculation of the spatial difference between the flexible registration and alignment point cloud model and the three-dimensional design standard model, including the calculation of the three-dimensional shape deviation vector of a single rebar, the rebar spacing deviation, and the overall mesh flatness deviation, includes: Extract the deformation attributes encapsulated in the flexible registration and alignment point cloud model, and establish a set of point-level spatial difference vectors; Based on the point-level spatial difference vector set, calculate the three-dimensional morphological deviation vector of a single steel bar; Based on the flexible registration and alignment point cloud model and the three-dimensional design standard model, the deviation of the rebar spacing is calculated. Based on the flexible registration and alignment point cloud model, the overall mesh flatness deviation is calculated; The three-dimensional morphological deviation vector of a single rebar, the rebar spacing deviation, and the overall mesh flatness deviation are structurally aggregated to obtain multi-dimensional rebar morphological deviation data.

8. The method for quality inspection of precast steel reinforcement according to claim 1, characterized in that, The classification and determination of multidimensional steel bar morphology deviation data includes: Establish a grading standard for multidimensional steel bar morphological deviation data; Based on the hierarchical judgment criteria, abnormal node binding information is identified, forming a set of abnormal node binding information. Based on the grading criteria, determine the overall quality level information; A rebar morphology deviation detection report is generated based on the set of node binding anomaly identification information and the overall quality grade information.

9. A method for inspecting the quality of reinforcing steel bars in precast components according to claim 8, characterized in that, The overall quality level information is determined according to the following rules: When the number of abnormal deviations in the three-dimensional shape of the reinforcing bars is zero, the number of abnormal deviations in the spacing of the reinforcing bars is zero, and the overall flatness deviation of the mesh is less than the overall flatness deviation threshold, it is judged to be of qualified grade. When at least one of the abnormal number of three-dimensional shape deviations of steel bars or abnormal number of steel bar spacing deviations is greater than zero and does not exceed a preset proportional threshold, it is determined to be a limit level. When the number of abnormal deviations in the three-dimensional shape of the reinforcing bars or the number of abnormal deviations in the spacing of the reinforcing bars exceeds the preset proportional threshold, or when the overall flatness deviation of the mesh is greater than or equal to the overall flatness deviation threshold, it is judged as unqualified.

10. A precast steel reinforcement quality inspection system, used to execute the precast steel reinforcement quality inspection method according to any one of claims 1-9, characterized in that, include: The image acquisition module acquires uncalibrated multi-view RGB image datasets of precast steel mesh and performs illumination-robust precoding processing to generate an illumination-robust image feature set. The 3D reconstruction module inputs the unlabeled multi-view RGB image dataset and the illumination-robust image feature set into the improved DUSt3R end-to-end 3D reconstruction network to obtain the initial dense 3D point cloud data and the corresponding reconstruction confidence map. The topology correction module removes low-confidence points from the initial dense 3D point cloud data based on the reconstructed confidence map and performs correction processing to obtain topology-corrected 3D point cloud data. The skeleton fitting module performs flow surface skeleton fitting on topology-corrected 3D point cloud data to obtain a macroscopic skeleton surface model of steel mesh. The flexible registration module flexibly registers the topology-corrected 3D point cloud data and the macroscopic skeleton surface model of the steel mesh with the 3D design standard model to obtain a flexibly registered aligned point cloud model. The deviation calculation module obtains multi-dimensional rebar morphology deviation data based on the spatial difference between the flexible registration and alignment point cloud model and the three-dimensional design standard model. The grading and judgment module grades and judges the multi-dimensional rebar morphology deviation data and generates a rebar morphology deviation detection report.

Images