Point cloud self-supervised quality evaluation and completion method

Abstract

The application is a point cloud self-supervised quality evaluation and completion method, relating to the technical field of point cloud processing and computer vision, comprising: normal data set construction and refined defect simulation; double-branch network training and normality feature deep learning; to-be-detected point cloud processing and completion reasoning; residual error calculation and multi-dimensional abnormality positioning analysis. In the application, two-dimensional analysis of point-by-point residual error and voxelized residual error comprehensively depicts the single-point abnormality intensity and the density and position difference of the local area, combines with the optimized threshold to screen the candidate abnormal points, and effectively reduces the false detection rate; further, through DBSCAN clustering, a continuous abnormal area is obtained, an abnormality index is calculated by fusing the residual error coefficient and the two-dimensional volume coefficient, the accurate grading of the abnormality is realized, and the problem of inaccurate grading of different types of abnormality in the traditional method is solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of point cloud processing and computer vision technology, and in particular to a point cloud self-supervised quality evaluation and completion method. BACKGROUND

[0002] Point cloud data is widely used in industrial detection, civil engineering and other fields because it can accurately represent the structure of three-dimensional space. Current point cloud anomaly detection and completion methods mostly rely on supervised learning, which requires a large number of manually labeled normal-anomaly sample pairs, resulting in high labeling costs and poor generalization. At the same time, most methods separate the completion task from the quality evaluation task, and the completion results lack targeted quality feedback. Anomaly detection only relies on single-dimensional residual error judgment, resulting in high false detection rate and inaccurate anomaly classification.

[0003] For example, traditional supervised learning methods require professionals to label abnormal areas point by point, which is extremely inefficient for large-scale point cloud data. Some self-supervised methods only focus on completion effects and do not establish a correlation between completion quality and anomaly degree. Anomaly evaluation only uses point-by-point residual error, ignoring the spatial aggregation and overall size of the anomaly, resulting in inaccurate classification of different types of anomalies such as scattered and concentrated anomalies.

[0004] Therefore, a point cloud self-supervised quality evaluation and completion method is proposed to address the above-mentioned problems. SUMMARY

[0005] The purpose of the present application is to solve the above problems by proposing a point cloud self-supervised quality evaluation and completion method.

[0006] To achieve the above purpose, the present application adopts the following technical solutions:

[0007] A point cloud self-supervised quality evaluation and completion method includes:

[0008] Normal data set construction and refined defect simulation: After collecting and screening normal point clouds that meet the quality standards and preprocessing, simulate multiple types of defects through parameterized design, generate defect and normal training pairs, and construct a data set;

[0009] Dual-branch network training and normality feature deep learning: build a dual-branch network, design a composite loss function and set reasonable hyperparameters for training, and learn the deep geometric features of normal point clouds;

[0010] Processing and completion inference of the point cloud to be detected: standardize the unknown point cloud to be detected, complete the inference and optimize the results through the trained model, and obtain the repaired normal point cloud benchmark;

[0011] Residual calculation and multi-dimensional anomaly positioning analysis: calculate the multi-dimensional residual of the point cloud to be detected and the completed point cloud, screen the abnormal points after determining the threshold, cluster the discrete abnormal points to obtain continuous abnormal areas, analyze the abnormal areas to obtain an abnormal index, and match the corresponding abnormal level based on the abnormal index.

[0012] Preferably, the normal data set construction and fine defect simulation specifically includes:

[0013] Two types of point cloud data are selected, one is the point cloud generated by the CAD / BIM ideal model, and the other is the real scene scanning data; and the data is screened and preprocessed;

[0014] To ensure the generality of the model, different intensity levels are set for each defect, and the specific parameters are as follows:

[0015] Random holes: hole radius range 0.5-5 cm, 1-3 holes randomly generated for each point cloud;

[0016] Local point cloud deletion: the deletion area is a rectangle or an irregular polygon, and the area ratio is 5%-20%, simulating occlusion or data loss, and the deletion boundary uses smooth transition;

[0017] Geometric distortion: non-rigid transformation is achieved by thin plate spline interpolation, the distortion area ratio is 3%-15%, the distortion displacement is 0.1-2 cm, and the stretching / bending angle is ≤15°;

[0018] Noise injection: add Gaussian noise disturbance to part of the point positions, and add 0.5%-2% outliers;

[0019] Each normal point cloud Randomly select 2-3 types of defects to generate 1 defect point cloud , and finally construct a preset number of training sample pairs and divide the training set and the validation set.

[0020] Preferably, the double-branch network training and normality feature deep learning specifically includes:

[0021] The double-branch network training includes a shared encoder and a completion decoder;

[0022] Based on the global feature vector output by the encoder, through two fully connected layers and a Dropout layer, a global quality score in the interval of 0-1 is finally output ;

[0023] Loss function design and training hyperparameter setting:

[0024] Preset completion loss , consistency loss , regularization loss After applying the weighting factors, the losses will be compensated accordingly. Consistency loss Regularization loss The total loss function is obtained by weighting and summing the corresponding weight factors. ;

[0025] Simultaneously calculate the completed point cloud To normal point cloud Distance; normal point cloud To complete the point cloud The distance;

[0026] After assigning weight factors to the two distances respectively, a weighted summation is performed to obtain... ;

[0027] Using mean squared error loss, based on The derived expected quality score and the actual quality score output by the model Compare the two and minimize the error between them;

[0028] And add L2 regularization terms to all trainable weights in the network.

[0029] Preferably, the point cloud processing and completion inference specifically includes:

[0030] The original point cloud was converted into a tensor format required by the model input, which was consistent with the training data. Gaussian filtering, uniform downsampling, and coordinate normalization were performed to ensure that the model input distribution was consistent with the training data.

[0031] Determine the inference environment configuration and set the inference completion parameters;

[0032] For the model output Post-processing is performed, including outlier removal and smoothing, to prevent noise generated by model inference from affecting subsequent analysis.

[0033] Preferably, the residual calculation and multi-dimensional anomaly localization analysis specifically include:

[0034] Preprocessed point cloud to be detected Each point in In the optimized The algorithm uses the K-nearest neighbors algorithm to find neighboring points and calculates... Euclidean distance to the mean of neighboring points , as the outlier score for that point;

[0035] The 3D space is divided into a uniform voxel grid, and the content within each voxel is calculated. and Point density difference and center offset ;

[0036] respectively, and then the point density difference , the center offset are weighted and summed to obtain the voxel residual error ;

[0037] The point-by-point residual error and the voxel residual error are both normalized to the interval [0, 1].

[0038] Preferably, the method further comprises:

[0039] Based on the residual error distribution of the training set normal point cloud, a threshold is determined, the point-by-point residual error of all normal point clouds in the training set after completion is calculated, the 99.5th percentile is taken as the initial threshold, and the threshold is adjusted according to the detection recall rate of the simulated defects in the validation set to determine the optimal threshold ;

[0040] The points in that satisfy > are marked as candidate abnormal points, and all points within the voxel residual error >0.2 are combined to supplement the missed abnormal points;

[0041] The DBSCAN algorithm is used to cluster the candidate abnormal points, the neighborhood radius =2×voxel edge length is set, and the minimum cluster point number is preset based on the point cloud density, so that discrete abnormal points are aggregated into continuous abnormal regions, and isolated noise points are filtered out;

[0042] The residual error coefficient and the volume coefficient of the abnormal region are obtained respectively, and the abnormal index is obtained after weighted and summed calculation.

[0043] Preferably, the residual error coefficient is obtained as follows:

[0044] The abnormal region is divided into sub-regions with a preset spatial size, and a plurality of abnormal sub-regions with the same size are obtained;

[0045] The candidate abnormal points in each abnormal sub-region are obtained respectively, for each candidate abnormal point, the point-by-point residual error thereof is calculated, the point-by-point residual errors of all points in the region are summed, and then divided by the number of points to obtain the average residual error;

[0046] The number of candidate abnormal points in each abnormal sub-region is multiplied by the average residual error corresponding to the sub-region to obtain the total residual error;

[0047] ​​The various abnormal sub-regions are arranged in descending order according to their respective corresponding residual total amount, and the three abnormal sub-regions with the largest residual total amount are extracted and marked as a first region, a second region and a third region respectively;

[0048] The region centers of the first region, the second region and the third region are extracted respectively, and the three region centers are connected by a straight line to form a closed triangle, the area of the triangle is calculated, and the reciprocal of the area is taken to obtain a residual coefficient.

[0049] Preferably, the volume coefficient is obtained as follows:

[0050] Based on the abnormal point sets corresponding to the first region, the second region and the third region, a minimum bounding box is fitted to obtain the spatial occupation volumes corresponding to the three regions respectively, and the three obtained spatial occupation volumes are accumulated to obtain an accumulated volume;

[0051] The total volume of all abnormal regions is obtained, denoted as an object volume;

[0052] The accumulated volume is divided by the object volume to obtain a volume coefficient;

[0053] The volume of the complete point cloud of the object to be detected is obtained, and the total volume of the abnormal region is divided by the volume of the complete point cloud of the object to be detected to obtain a proportion volume coefficient;

[0054] The volume coefficient is obtained by weighted sum calculation of the volume coefficient and the proportion volume coefficient;

[0055] The weight factors of the volume coefficient and the proportion volume coefficient are preset, and the volume coefficient and the proportion volume coefficient are multiplied by the corresponding weight factors to obtain the volume coefficient.

[0056] Preferably, the method further comprises:

[0057] Anomaly verification and comprehensive detection report generation: through artificial review and cross-method verification, the detection results are corrected to generate a detection report containing basic information, quality evaluation and abnormal details, and the data is archived for model iteration.

[0058] As described above, due to the adoption of the above technical solutions, the present application has the following advantages:

[0059] 1. The present application comprehensively describes the single-point abnormal intensity and the density and position difference of the local region through the two-dimensional analysis of the point-by-point residual and the voxel residual, optimizes the threshold to screen the candidate abnormal points, effectively reduces the false detection rate, further obtains the continuous abnormal region through DBSCAN clustering, and fuses the residual coefficient and the two-dimensional volume coefficient to calculate the abnormal index, so as to realize the accurate grading of the anomaly and solve the problem of inaccurate grading of different types of anomalies in the traditional method.

[0060] 2. The application simulates multiple types and intensities of defects such as random holes and geometric distortion through parameterized design, generates large-scale normal-defect training pairs based on normal point clouds, and greatly reduces labeling costs and technical thresholds; at the same time, a double-branch network including a shared encoder, a completion decoder and a quality evaluation branch is built, and the completion and quality evaluation are co-trained by combining a compound loss function, the quality score can intuitively reflect the normality of the point cloud, and the problem of disconnection between completion and quality evaluation in the prior art is solved, and the generalization is stronger, and can adapt to point cloud processing needs in different scenes such as bridges and mechanical parts. BRIEF DESCRIPTION OF DRAWINGS

[0061] In the following description of the exemplary embodiments in conjunction with the accompanying drawings, more details, features and advantages of the present application are disclosed, in which:

[0062] Figure 1 The method flowchart of the present application. DETAILED DESCRIPTION

[0063] Several embodiments of the present application will be described in more detail below with reference to the accompanying drawings so as to enable those skilled in the art to implement the present application. The present application can be embodied in many different forms and purposes and should not be limited to the embodiments described herein. These embodiments are provided to make the present application comprehensive and complete, and to fully convey the scope of the present application to those skilled in the art. The embodiments do not limit the present application.

[0064] Unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. It will be further understood that terms such as those defined in commonly used dictionaries should be interpreted as having a meaning that is consistent with their meaning in the context of the relevant art and / or the present specification, and should not be interpreted in an idealized or overly formal sense unless expressly so defined herein.

[0065] Embodiment 1

[0066] The detailed description is combined with the accompanying drawings Figure 1 are described in detail.

[0067] The accompanying drawings Figure 1 A point cloud self-supervised quality evaluation and completion method flowchart is provided for the embodiments of the present application, which shows the complete steps from normal data set construction and fine defect simulation to abnormal verification and comprehensive detection report generation.

[0068] In this embodiment, it includes:

[0069] Normal dataset construction and refined defect simulation: After collecting and screening normal point clouds that meet the quality standards and performing preprocessing, multiple types of defects are simulated through parametric design to generate defect and normal training pairs and construct the dataset;

[0070] Specifically, it includes:

[0071] Normal point cloud data acquisition and filtering:

[0072] Two types of point cloud data are preferred: one is point clouds generated from ideal CAD / BIM models (ensuring absolutely no defects, with a point cloud density ≥ 1024 points / ...). Secondly, real-world scene scanning data (such as bridge and mechanical parts scanning) needs to be manually screened to remove samples containing cracks, corrosion, and deformation.

[0073] The screening criteria include: point cloud noise intensity ≤ 0.05mm (judged by calculating the average distance from the point cloud to the fitting plane), point cloud integrity ≥ 95% (no large areas of missing data), and point cloud density fluctuation ≤ 10% (to avoid local sparsity leading to model misjudgment).

[0074] The normal point cloud after screening is subjected to uniform preprocessing, including Gaussian filtering for noise reduction (σ=0.02), uniform downsampling (unifying the number of point cloud points to 2048 / 4096 while retaining key geometric features), and coordinate normalization (aligning the center of the point cloud to the origin and scaling it to the unit sphere).

[0075] The above point cloud preprocessing process is a direct reference to existing technology and will not be elaborated here;

[0076] Multi-type defect simulation and training pair generation:

[0077] To ensure the model's generalizability, different intensity levels are set for each type of defect, with specific parameters as follows:

[0078] Random holes: The radius of the holes ranges from 0.5 to 5 cm (adjusted according to the size of the point cloud object, such as using a smaller radius for part point clouds and a larger radius for bridge point clouds). 1 to 3 holes are randomly generated for each point cloud, and the point deletion rate in the hole area is 100%.

[0079] Local point cloud deletion: The deletion area is a rectangle or irregular polygon, accounting for 5% to 20% of the area, simulating occlusion or missing data. The deletion boundary adopts a smooth transition (to avoid sharp edges that cause the model to overfit).

[0080] Geometric Twist: Non-rigid transformation is achieved using thin plate spline interpolation (TPS). The twisted region accounts for 3% to 15%, the twist displacement is 0.1 to 2 cm, and the tension / bending angle is ≤15° to simulate slight deformation defects.

[0081] Noise injection: Gaussian noise (mean 0, variance 0.01-0.05) is added to disturb some point locations, and 0.5% to 2% outliers (≥3 times the noise standard deviation from the normal surface) are added to simulate scanning noise or small impurities;

[0082] Each normal point cloud Randomly select 2-3 defect type combinations to generate one defect point cloud. Finally, a preset number (≥10,000) of training sample pairs are constructed, and the training set and validation set are divided in an 8:2 ratio.

[0083] Dual-branch network training and normality feature deep learning: Build a dual-branch network, design a composite loss function and set reasonable hyperparameters for training, and learn the deep geometric features of normal point clouds;

[0084] Specifically, it includes:

[0085] Detailed design of the dual-branch network structure, including:

[0086] Shared encoder: It adopts an improved version of PointNet++ as the basic architecture, sets up 4 sampling layers (each layer has 1024, 512, 256 and 128 sampling points respectively), and each sampling layer is followed by a multilayer perceptron (MLP) and batch normalization (BN). The activation function is ReLU. Finally, it outputs a 1024-dimensional global feature vector and a pointwise 512-dimensional local feature vector, which takes into account both global structure and local details.

[0087] The completion decoder employs a deconvolution and residual connection structure. It takes the feature vector output from the encoder as input and gradually restores the point cloud dimension through three deconvolution layers, ultimately outputting a completed point cloud with the same number of points as the input. The last layer of the decoder adds a Sigmoid activation function to constrain the range of point cloud coordinates;

[0088] Quality assessment branch: Based on the global feature vector output by the encoder, through two fully connected layers (hidden layer dimensions 512 and 256) and a Dropout layer (dropout rate 0.3), it finally outputs a global quality score in the 0-1 range. The closer the score is to 1, the more normal the point cloud is;

[0089] Loss function design and training hyperparameter settings:

[0090] Preset compensation loss Consistency loss Regularization loss After applying the weighting factors, the losses will be compensated accordingly. Consistency loss Regularization loss The total loss function is obtained by weighting and summing the corresponding weight factors. ;

[0091] Total loss function: It is the weighted sum of the three sub-losses;

[0092] Completing the loss is the core function, with the highest weight, to ensure that the model can accurately repair defective point clouds;

[0093] Consistency loss is used to correlate the completion effect with the quality score, making the model's judgment of normality more reasonable.

[0094] Regularization loss is used to prevent overfitting and ensure the model's ability to generalize on new data.

[0095] The 0.7, 0.2, and 0.1 are respectively the compensation loss. Consistency loss Regularization loss Weighting factors.

[0096] ChamferDistance is a classic metric for measuring the similarity between two point clouds. Here, we improve upon it by performing bidirectional calculation, and the specific logic is as follows:

[0097] Make up for the loss Improved bidirectional constraints of ChamferDistance:

[0098] The limitation of unidirectional ChamferDistance: If only the point cloud is completed... To true normal point cloud The model may overemphasize the distance. Close , and ignore arrive The matching degree can lead to shape shifts in the completion result (such as overall scaling or partial missing points in the point cloud).

[0099] Improved bidirectional computation: Simultaneous computation to complete point clouds To normal point cloud Distance (calculate to complete point cloud) From each point in the cloud to the true normal point cloud The sum of the distances to the nearest neighbors (which measures whether the points in the completed point cloud can be matched in the real point cloud), normal point cloud. To complete the point cloud Distance (for calculating true normal point clouds) To complete the point cloud from each point in the middle The sum of the distances to the nearest neighbors in the complete point cloud. It measures whether a point in the real point cloud can be matched in the completed point cloud.

[0100] Calculating only one direction can lead to shape or density deviations in the completion results;

[0101] After assigning weight factors to the two distances respectively, a weighted sum is performed to obtain... The formula is: , , These are the weighting factors corresponding to the two distances in the formula;

[0102] Yes, the completed point cloud's shape, density, and spatial distribution are consistent with a normal point cloud. Highly consistent;

[0103] Consistency loss : Link the completion effect to the quality score:

[0104] The core of this loss is to establish a quality score where the smaller the completion loss (the better the completion effect). The higher the logical correlation (the more normal the point cloud), the more specific the implementation is as follows:

[0105] If the model represents the defect point cloud The closer the completion ( The smaller (the smaller), the better. The more normal it is, the higher the quality score. The higher the value, the better.

[0106] Using mean squared error loss, based on The derived expected quality score and the actual quality score output by the model Compare the two and minimize the error between them;

[0107] For example, through function mapping (such as...) )Will Convert into expected score, then with Calculate the mean square error;

[0108] Regularization loss : And add an L2 regularization term to all trainable weights of the network, the formula is as follows , It is the weight decay coefficient;

[0109] This forces the weight parameters to be as small as possible, preventing the model from becoming too complex in order to fit the noise in the training data, thereby improving the model's generalization ability on unknown point clouds (inference stage);

[0110] Training hyperparameters: The Adam optimizer is used, with an initial learning rate of 0.001, which decays to 0.8 every 10 epochs; the batch size is set to 32 (adjusted according to GPU memory, 64 if GPU memory ≥ 16G); the total number of training epochs is 100, and an early stopping mechanism is triggered when the validation set loss does not decrease for 15 consecutive epochs.

[0111] Point cloud processing and completion inference: Standardization processing is performed on the unknown point cloud to be detected, and completion inference is completed and the results are optimized through the trained model to obtain the normal point cloud baseline after repair;

[0112] Specifically, it includes:

[0113] The original point cloud was converted into a tensor format required by the model input, which was consistent with the training data. Gaussian filtering (σ=0.02), uniform downsampling (unified to 4096 points; if the number of points in the original point cloud was insufficient, interpolation was used to complete it), and coordinate normalization were all performed to ensure that the model input distribution was consistent with the training data.

[0114] Initial data quality check: Calculate the noise intensity and point cloud density of the point cloud to be detected. If the noise intensity is greater than 0.1 mm or the point cloud density fluctuation is greater than 30%, output a data quality warning, indicating that it may affect the detection accuracy.

[0115] Complete reasoning execution and result optimization:

[0116] Determine the inference environment configuration: It is recommended to use a GPU (NVIDIA RTX 3090 or above), and the inference framework should be PyTorch 1.10+. Set the inference mode (model.eval()) and turn off the Dropout layer to ensure the stability of the results.

[0117] Set the completion inference parameters: Set the batch size to 1 (single sample inference to ensure accuracy), and retain the original coordinate mapping relationship of the point cloud during inference to facilitate the calculation of the true location of the abnormal area in the subsequent process;

[0118] Optimization of completion results: Optimization of model output Post-processing is performed, including removing outliers (by statistically analyzing local point densities and deleting points with densities below 1 / 3 of the mean) and smoothing (using moving least squares with a window size of 20 neighboring points) to avoid noise generated by model inference affecting subsequent analysis.

[0119] Residual calculation and multi-dimensional anomaly localization analysis: Calculate the multi-dimensional residual between the point cloud to be detected and the completed point cloud, determine the threshold and filter out anomaly points, and cluster discrete anomaly points to obtain continuous anomaly regions. After analyzing the anomaly regions, obtain the anomaly index, and match the corresponding anomaly level based on the anomaly index to generate a corresponding anomaly heatmap. The heatmap uses Jet color mapping, mapping different colors based on the anomaly index, and generates a point cloud heatmap with color annotations. The output supports 3D formats such as PLY and OBJ (which can be viewed directly in software such as CloudCompare) and PNG 2D renderings (captured from multiple perspectives for quick browsing). The heatmap marks the number, level, and core location of the anomaly regions.

[0120] Specifically, it includes:

[0121] Point-by-point residual calculation: For the preprocessed point cloud to be detected Each point in In the optimized The algorithm uses the K-nearest neighbor algorithm (K=5) to find neighboring points and calculates... Euclidean distance to the mean of neighboring points The outlier score for that point is calculated using the following formula: , , , They are respectively Zhongyu The average coordinates of the 5 nearest points, that is, the average of the x, y, and z coordinates of these 5 neighboring points, represent the reference coordinates of this position under normal conditions; , , for Its own three-dimensional coordinates;

[0122] Voxelized Residual Analysis: The 3D space is divided into a uniform voxel mesh (the voxel size is set according to the size of the point cloud object, such as 10cm×10cm×10cm for bridge point cloud and 1cm×1cm×1cm for part point cloud), and the residual within each voxel is calculated. and Point density difference ( and center offset ( , , , for The coordinates of the center of the point cloud within this voxel. , , for (coordinates of the center of the point cloud within the same voxel).

[0123] Preset point density differences Center offset After applying the weighting factor, the point density difference will be... Center offset The voxel residuals are obtained by performing a weighted summation calculation. ;

[0124] Point-by-point residuals and voxel residuals All values ​​are normalized to the [0,1] interval to facilitate unified threshold judgment.

[0125] The threshold is determined based on the residual distribution of normal point clouds in the training set. The point-by-point residuals after completing all normal point clouds in the training set are calculated, and the 99.5 quantile is taken as the initial threshold. The threshold is then adjusted according to the detection recall of simulated defects in the validation set to determine the optimal threshold. ;

[0126] Will China satisfies > Points are marked as candidate outliers, and voxel residuals are also considered. All points within voxels with a value greater than 0.2 are used to supplement and filter out any missed outliers, thus avoiding misjudgment based on a single dimension.

[0127] The DBSCAN algorithm is used to cluster candidate outliers, and a neighborhood radius is set. =2×voxel side length, based on the point cloud density, preset the minimum number of cluster points, aggregate discrete outliers into continuous outlier regions, and filter out isolated noise points (cluster points < 5 are considered noise).

[0128] The residual coefficients and volume coefficients of the abnormal regions are obtained separately, and the anomaly index is obtained by weighted summation.

[0129] The weighted summation process is as follows: preset the weighting factors of the residual coefficient and the volume coefficient, multiply the residual coefficient and the volume coefficient by their corresponding weighting factors, and then sum them to obtain the anomaly index.

[0130] The process of obtaining the residual coefficients is as follows:

[0131] The abnormal region is divided into sub-regions with a preset spatial size, resulting in multiple abnormal sub-regions of the same size; (wherein, the size of the sub-region is adaptively adjusted according to the maximum dimension of the abnormal region (e.g., the side length of the sub-region = 1 / 8 of the maximum x / y / z dimension of the abnormal region) to reduce size sensitivity issues).

[0132] Candidate anomaly points are obtained in each anomaly sub-region. For each candidate anomaly point, its point-by-point residual (i.e., the Euclidean distance between the point cloud to be detected and the corresponding point in the completed point cloud, which has been described above and will not be repeated here) is calculated. The point-by-point residuals of all points in the region are summed and then divided by the number of points to obtain the average residual.

[0133] Multiply the number of candidate outliers in each outlier sub-region by the average residual corresponding to that sub-region to obtain the total residual.

[0134] The abnormal sub-regions are sorted in descending order according to the total residual of each region, and the three abnormal sub-regions with the largest total residual are extracted and labeled as the first region, the second region, and the third region, respectively.

[0135] Extract the center of the first, second, and third regions respectively, and connect the three center regions with straight lines to form a closed triangle. Calculate the area of ​​the triangle and take its reciprocal to obtain the residual coefficient.

[0136] The process of obtaining the volume factor is as follows:

[0137] Based on the outlier sets corresponding to the first, second, and third regions, their minimum bounding boxes are fitted to obtain the spatial occupancy volumes corresponding to the three regions respectively. The three spatial occupancy volumes are then accumulated to obtain the cumulative volume.

[0138] The process of fitting the minimum axis-aligned bounding box is as follows:

[0139] For the set of outliers obtained by clustering, find the minimum and maximum coordinate values ​​of each point on the x, y, and z axes. This forms a cuboid aligned with the coordinate axes. Calculate the volume of this cuboid to obtain the volume of the minimum axis-aligned bounding box.

[0140] Get the total volume of all abnormal regions, and denote it as the object volume;

[0141] Divide the cumulative volume by the object volume to obtain the volume coefficient;

[0142] Obtain the volume of the complete point cloud of the object to be detected, and divide the total volume of the abnormal region by the volume of the complete point cloud of the object to be detected to obtain the volume ratio coefficient.

[0143] The volume coefficient is obtained by weighted summation of the volume coefficient and the proportion volume coefficient;

[0144] The weighting factors for the volume coefficient and the percentage volume coefficient are preset. The volume coefficient is obtained by multiplying the volume coefficient and the percentage volume coefficient with their corresponding weighting factors and then summing them.

[0145] A dual-dimensional screening of candidate anomalies is adopted using pointwise residuals and voxel residuals. This approach captures the intensity of single-point anomalies through pointwise residuals and identifies regional anomaly features through voxel residuals (density difference, center shift). This avoids the omission of low single-point residual but overall shift anomalies or the misjudgment of high single-point residual but regional normal noise by a single dimension (such as looking only at single-point residuals).

[0146] Subsequent DBSCAN clustering further aggregates discrete outliers into physically meaningful continuous regions by using an adaptive neighborhood radius (2 × voxel side length) and density threshold (filtering isolated clusters with <5 points). This effectively eliminates scanning noise interference, ensures that the analyzed objects are real outlier regions, and lays a reliable foundation for subsequent evaluation.

[0147] The residual coefficient and volume coefficient are designed to characterize anomalies from two dimensions: core clustering, intensity, and scale characteristics. A comprehensive evaluation is then achieved through weighted fusion.

[0148] The residual coefficient filters the core area by adaptive sub-region division (matching the anomaly scale) and total residual (number of points × average residual). It combines the inverse of the triangle area to reflect the spatial clustering of the core area. The denser the clustering (the smaller the area), the larger the coefficient, which accurately captures the risk characteristics of local dense high-intensity anomalies.

[0149] The volume factor integrates the volume coefficient (the proportion of the core area to the total volume of the anomaly) and the percentage volume factor (the proportion of the anomaly to the total volume of the object), which reflects both the concentration of the anomaly energy and its absolute influence on the overall object, thus avoiding the one-sidedness of looking only at the local or the whole.

[0150] The anomaly index obtained by weighting the two can comprehensively distinguish different types of anomalies, such as dense, small-scale high-risk anomalies and dispersed, large-scale low-risk anomalies, providing a quantitative basis for hierarchical decision-making and significantly improving the scientific nature and practicality of the assessment.

[0151] Anomaly Validation and Comprehensive Test Report Generation: Test results are corrected through manual review and cross-method validation to generate a test report containing basic information, quality assessment, and anomaly details. Data is then archived before model iteration, specifically including:

[0152] Verification and correction of abnormal areas: Manual sampling review: For severely high-risk abnormal areas, manual review is adopted (by viewing the overlay effect of the original point cloud, the completed point cloud and the heat map through 3D point cloud visualization software) to confirm whether it is a real abnormality (such as excluding false positives caused by point cloud scanning errors).

[0153] Cross-method verification: For critical objects (such as key load-bearing structures of bridges and precision mechanical parts), the authenticity of abnormal areas can be verified by combining traditional detection methods (such as laser ranging and ultrasonic testing), and the detection results can be corrected (such as adjusting the boundaries of abnormal areas and correcting the anomaly level).

[0154] Results correction rules: If a false positive is found during manual review, the residual weight of the corresponding region is reduced, and the anomaly score is recalculated; if a missed anomaly is found, the threshold is adjusted according to a preset scale. The analysis process of re-executing residual calculation and multi-dimensional anomaly location analysis steps is then carried out.

[0155] The report's core elements include:

[0156] Basic information: Name of the object to be detected, point cloud acquisition time / device, preprocessing parameters, and detection time.

[0157] Overall Quality Assessment: Global Quality Score (≥0.8 is normal, 0.6-0.8 is slightly abnormal, <0.6 is severely abnormal), total number of abnormal areas, percentage of abnormal areas at each level, and overall risk level (low / medium / high).

[0158] Detailed information on abnormal areas: number, location (3D coordinate range), size (length × width × height / volume) of each abnormal area, abnormality level, average residual, mean value of voxel residual, and abnormality type determination (cracks / pits / corrosion / deformation, etc., based on the morphological characteristics of the area).

[0159] Visualization results: anomaly heatmap (multi-view), comparison of original point cloud and completed point cloud, and magnified view of anomaly area.

[0160] Decision recommendations: Provide targeted recommendations based on the level of anomaly (mild anomaly: regular monitoring; moderate anomaly: partial repair; severe anomaly: emergency handling and cessation of use), recommendations for subsequent testing cycles, and recommendations for data archiving.

[0161] Archive the point cloud to be tested, the completed point cloud, the residual data, and the test report in a unified manner, and name them according to the object name and the test time to establish a test database.

[0162] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0163] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

[0164] It should be noted that, in this document, the use of relational terms such as "first" and "second" is merely for distinguishing one entity or operation from another, and does not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0165] It should be understood that in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0166] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0167] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0168] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0169] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0170] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0171] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

Claims

1. A self-supervised quality evaluation and completion method for point clouds, characterized in that, include: Normal dataset construction and refined defect simulation: After collecting and screening normal point clouds that meet the quality standards and performing preprocessing, multiple types of defects are simulated through parametric design to generate defect and normal training pairs and construct the dataset; Dual-branch network training and normality feature deep learning: Build a dual-branch network, design a composite loss function and set reasonable hyperparameters for training, and learn the deep geometric features of normal point clouds; Dual-branch network training and normality feature deep learning, specifically including: Two-branch network training includes a shared encoder and a completion decoder; Based on the global feature vector output by the encoder, the final output is a global quality score in the 0-1 range, processed through two fully connected layers and a Dropout layer. ; Loss function design and training hyperparameter settings: Preset compensation loss Consistency loss Regularization loss After applying the weighting factors, the losses will be compensated accordingly. Consistency loss Regularization loss The total loss function is obtained by weighting and summing the corresponding weight factors. ; Simultaneously calculate the completed point cloud To normal point cloud Distance; normal point cloud To complete the point cloud The distance; After assigning weight factors to the two distances respectively, a weighted summation is performed to obtain... ; Using mean squared error loss, based on The derived expected quality score and the actual quality score output by the model Compare the two and minimize the error between them; And add L2 regularization terms to all trainable weights in the network; Point cloud processing and completion inference: Standardization processing is performed on the unknown point cloud to be detected, and completion inference is completed and the results are optimized through the trained model to obtain the normal point cloud baseline after repair; Residual calculation and multi-dimensional anomaly localization analysis: Calculate the multi-dimensional residual between the point cloud to be detected and the completed point cloud, determine the threshold and screen out anomaly points, and cluster the discrete anomaly points to obtain continuous anomaly regions. After analyzing the anomaly regions, obtain the anomaly index and match the corresponding anomaly level based on the anomaly index. Specifically, it includes: Preprocessed point cloud to be detected Each point in For the optimized The K-nearest neighbors algorithm is used to find neighboring points and calculate... Euclidean distance to the mean of neighboring points , as the outlier score for that point; The 3D space is divided into a uniform voxel grid, and the content within each voxel is calculated. and Point density difference and center offset ; Preset point density differences Center offset After applying the weighting factor, the point density difference will be... Center offset We perform a weighted summation calculation to obtain the voxel residuals. ; Point-by-point residuals and voxel residuals All values ​​are normalized to the [0,1] interval; The threshold is determined based on the residual distribution of normal point clouds in the training set. The point-by-point residuals after completing all normal point clouds in the training set are calculated, and the 99.5 quantile is taken as the initial threshold. The threshold is then adjusted according to the detection recall of simulated defects in the validation set to determine the optimal threshold. ; Will China satisfies > Points are marked as candidate outliers, and voxel residuals are also considered. For all points within a voxel with a value greater than 0.2, supplement any missed outliers during the screening process; The DBSCAN algorithm is used to cluster candidate outliers, and a neighborhood radius is set. =2×voxel side length, based on the minimum number of cluster points preset by the point cloud density, to aggregate discrete outliers into continuous outlier regions and filter out isolated noise points; The residual coefficients and volume coefficients of the abnormal regions are obtained separately, and the anomaly index is obtained by weighted summation. The process of obtaining the residual coefficients is as follows: The abnormal region is divided into sub-regions with preset spatial dimensions to obtain multiple abnormal sub-regions of the same size. Candidate outliers in each outlier sub-region are obtained. For each candidate outlier, its point-by-point residual is calculated. The point-by-point residuals of all points in the region are summed and then divided by the number of points to obtain the average residual. Multiply the number of candidate outliers in each outlier sub-region by the average residual corresponding to that sub-region to obtain the total residual. The abnormal sub-regions are sorted in descending order according to the total residual of each region, and the three abnormal sub-regions with the largest total residual are extracted and labeled as the first region, the second region, and the third region, respectively. Extract the center of the first, second, and third regions respectively, and connect the three center regions with straight lines to form a closed triangle. Calculate the area of ​​the triangle and take its reciprocal to obtain the residual coefficient. The process of obtaining the volume factor is as follows: Based on the outlier sets corresponding to the first, second, and third regions, their minimum bounding boxes are fitted to obtain the spatial occupancy volumes corresponding to the three regions respectively. The three spatial occupancy volumes are then accumulated to obtain the cumulative volume. Get the total volume of all abnormal regions, and denote it as the object volume; Divide the cumulative volume by the object volume to obtain the volume coefficient; Obtain the volume of the complete point cloud of the object to be detected, and divide the total volume of the abnormal region by the volume of the complete point cloud of the object to be detected to obtain the volume ratio coefficient. The volume coefficient is obtained by weighted summation of the volume coefficient and the proportion volume coefficient; The weighting factors for the volume coefficient and the percentage volume coefficient are preset. The volume coefficient is obtained by multiplying the volume coefficient and the percentage volume coefficient with their corresponding weighting factors and then summing them.

2. The point cloud self-supervised quality evaluation and completion method according to claim 1, characterized in that, Normal dataset construction and refined defect simulation, specifically including: Two types of point cloud data were selected: one was the point cloud generated from the CAD / BIM ideal model, and the other was the scanned data from the real scene; and the data were filtered and preprocessed. To ensure the model's generalizability, different intensity levels are set for each type of defect, with specific parameters as follows: Random holes: The radius of the holes ranges from 0.5 to 5 cm, and 1 to 3 holes are randomly generated for each point cloud; Local point cloud deletion: The deletion area is a rectangle or irregular polygon, accounting for 5% to 20% of the area, simulating occlusion or data loss, and the deletion boundary adopts a smooth transition; Geometric distortion: Non-rigid transformation is achieved by thin plate spline interpolation, with the distortion area accounting for 3% to 15%, the distortion displacement being 0.1 to 2 cm, and the tension / bending angle being ≤15°; Noise injection: Add Gaussian noise to disturb some points, and add 0.5% to 2% outliers; Each normal point cloud Randomly select 2-3 defect type combinations to generate one defect point cloud. Finally, a predetermined number of training sample pairs are constructed, and the training set and validation set are divided.

3. The point cloud self-supervised quality evaluation and completion method according to claim 1, characterized in that, Point cloud processing and completion inference to be detected, specifically including: The original point cloud was converted into a tensor format required by the model input, which was consistent with the training data. Gaussian filtering, uniform downsampling, and coordinate normalization were performed to ensure that the model input distribution was consistent with the training data. Determine the inference environment configuration and set the inference completion parameters; For the model output Post-processing is performed, including outlier removal and smoothing, to prevent noise generated by model inference from affecting subsequent analysis.

4. The point cloud self-supervised quality evaluation and completion method according to claim 1, characterized in that, Also includes: Anomaly Validation and Comprehensive Test Report Generation: Through manual review and cross-method validation, test results are corrected to generate a test report containing basic information, quality assessment, and anomaly details. The data is then archived for model iteration.

Images