Multi-view blade point cloud quick registration method based on node guide virtual correspondence

Abstract

The invention discloses a multi-view blade point cloud rapid registration method based on node guide virtual correspondence, and belongs to the technical field of blade contour reconstruction. The method comprises the following steps: firstly, performing feature extraction and down-sampling on multi-view point cloud to obtain super points and features thereof; then screening out reliable overpoint correspondence through an overlapping perception mechanism; then, based on super-point corresponding guidance, generating a confidence-weighted virtual corresponding point in the target point cloud; and finally, a bidirectional local-global strategy is adopted to efficiently estimate an optimal transformation parameter, and synchronous and accurate registration of the multi-view point clouds is realized. The method effectively solves the registration problem of complex curved surfaces such as blades caused by low overlapping and fuzzy characteristics in three-dimensional reconstruction, has high precision, high efficiency and strong robustness, and is suitable for high-precision industrial optical measurement.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of blade profile reconstruction, and particularly relates to a multi-view blade point cloud fast registration method based on node-guided virtual correspondence. BACKGROUND

[0002] The blade of an aero-engine is a core component, and its contour geometric precision directly determines the overall performance and reliability of the engine. Developing high-precision and high-efficiency blade profile measurement technology is crucial to ensuring the manufacturing quality of such key components.

[0003] Traditional blade profile measurement widely adopts a coordinate measuring machine (CMM). The CMM has the advantages of high precision and repeatability by using a stylus contact point. However, this method has inherent limitations such as low throughput, physical contact that may cause deformation of the blade surface, and slow measurement speed. These shortcomings have prompted the development of faster, non-contact optical measurement methods such as laser triangulation and structured light imaging.

[0004] However, due to the occlusion problem, a single optical scan cannot capture the complete blade profile. Therefore, it is necessary to align multiple scan views to a unified coordinate system through multi-view acquisition and registration technology to achieve comprehensive and accurate reconstruction of the blade geometry. In the registration technology, the Iterative Closest Point (ICP) algorithm and its variants are the most widely used point cloud alignment method. However, the ICP algorithm and its variants usually rely on good initial pose, and the convergence speed may be slow, and they are very sensitive to noise and outliers, which are common in blade optical measurement.

[0005] In order to overcome the limitations of "hard correspondence" relationship, the registration method adopts a "soft correspondence" strategy to enhance flexibility and robustness, especially under partial overlap and noisy measurement conditions. Representative methods include probabilistic frameworks and deep learning models such as RPM-Net, which learn feature embeddings and predict soft correspondence relationships, and optimize rigid transformations end-to-end.

[0006] Turbine blade reconstruction is a highly challenging application scenario, and its point cloud usually has the characteristics of low overlap, smooth, and thin-walled free-form surface. For this low overlap situation, methods that explicitly use overlapping regions are developed. For example, Predator guides registration by predicting overlapping regions, while Xu et al. combine automatic overlap discrimination with a hybrid weighting scheme based on feature distance.

[0007] Despite the aforementioned progress, global matching ambiguity remains a persistent challenge in low-overlap point cloud registration, often resulting in a low inlier ratio in the inferred correspondences. A typical coarse-to-fine registration paradigm adopts the following procedure: first, downsample the point clouds with a KPConv network and construct their feature descriptors; then, establish coarse correspondences at the superpoint level based on these descriptors; finally, further refine the matching in the Euclidean space neighborhood at the point level under the guidance of the coarse matching. However, due to the thin-walled nature of turbine blades and the small overlap region between views, real inliers are extremely sparse; even many nearest neighbors do not belong to the inlier set. Therefore, these methods usually assign high confidence to both real inliers and incorrect matches near them, ultimately limiting the reconstruction accuracy. To this end, the prior art alleviates this problem by estimating correspondence confidence and computing virtual target points weighted by these scores.

[0008] In addition, most existing methods rely on rigid pairwise point cloud registration (RPPCR). When the number of views increases, this method leads to cumulative alignment errors. Since reconstructing a complete blade usually requires at least three scans, the sequential pairwise registration chain superimposes the tiny misregistration problems in each step, often failing to achieve accurate reconstruction. To alleviate this problem, Zhu et al. proposed a learning framework that iteratively updates the translation estimation through a recurrent network, achieving accurate multi-view alignment. However, this iterative recursive scheme incurs a huge computational overhead.

[0009] In summary, accurate and efficient multi-view blade contour reconstruction still faces significant challenges. The main obstacles include low overlap rates between adjacent views, non-obvious feature characteristics in the overlap region, and error drift accumulated by sequential pairwise registration. Therefore, there is an urgent need in the art for a new multi-view point cloud registration method that can simultaneously address low-overlap matching ambiguity, suppress error accumulation, and have high computational efficiency. SUMMARY

[0010] The present application aims to overcome the deficiencies of the prior art and provide a multi-view blade point cloud fast registration method based on node-guided virtual correspondence (NGVC-Net), which realizes simultaneous, accurate, robust, and efficient multi-view point cloud registration.

[0011] To achieve the above-mentioned purpose, the present application adopts the following technical solutions: A multi-view blade point cloud fast registration method based on node-guided virtual correspondence, comprising the following steps: S1: performing feature extraction on the input source point cloud and target point cloud to generate a superpoint set, superpoint features, and point-level features; S2: performing overlap-aware matching based on the superpoint features to predict the overlap score of the superpoint and filter non-overlapping superpoints to obtain a reliable superpoint correspondence set; S3: generating node-guided, confidence-weighted virtual corresponding points for points in the source point cloud in the target point cloud based on the set of superpoint correspondences and the point-level features, forming a set of virtual corresponding point pairs; S4: estimating optimal rigid transformation parameters based on the set of virtual corresponding point pairs through a bidirectional local-to-global hypothesis verification strategy, completing multi-view point cloud synchronous registration.

[0012] Further, in S1, the KPConv backbone network is used to hierarchically downsample the source point cloud and the target point cloud to generate the superpoint set and the superpoint features thereof, and to upsample the superpoint features to obtain the point-level features.

[0013] Further, a local patch is constructed for each superpoint in the superpoint set in its original point cloud. ; wherein, , are local patches of superpoints and , is a point in the source point cloud , is a point in the target point cloud , is a point in the superpoint cloud , is a point in the superpoint cloud .

[0014] Further, in S2, the matching based on the superpoint features is performed to predict the overlap scores of the superpoints and filter the non-overlapping superpoints to obtain a reliable set of superpoint correspondences, specifically including: S21: extracting enhanced features in the superpoint features based on a context encoder of a GeoTransformer; S22: predicting overlap confidence scores of each superpoint feature through a fully connected layer and an activation function, and binarizing the overlap confidence scores to generate an overlap score matrix; S23: calculating a Gaussian correlation matrix between the enhanced features; S24: using the overlap score matrix to mask entries in the Gaussian correlation matrix corresponding to non-overlapping superpoint correspondences; S25: selecting Top-K entries to form the reliable set of superpoint correspondences based on a double soft-max operation to suppress ambiguous matches.

[0015] Further, S3 generates node-guided, confidence-weighted virtual corresponding points for points in the source point cloud in the target point cloud based on the super point correspondence set and the point-level features, and forms a virtual corresponding point pair set, specifically comprising: S31: calculate a point-by-point cost matrix between the source point cloud and the target point cloud based on the point-level features, and solve a soft correspondence confidence matrix by applying a Sinkhorn algorithm; S32: under the guidance of the super point correspondence set, select a bidirectional top-ranked point-by-point match from the soft correspondence confidence matrix as a potential corresponding point; S33: for a point in the source point cloud, calculate its virtual corresponding point in the target point cloud as a confidence-weighted average position of the potential corresponding points, thereby forming the virtual corresponding point pair set.

[0016] Further, S4 estimates optimal rigid transformation parameters based on the virtual corresponding point pair set through a bidirectional local-to-global hypothesis verification strategy, specifically comprising: S41: in the local stage, assign weights to each virtual corresponding point pair based on the current motion hypothesis to eliminate outliers, and calculate weighted motion parameters; S42: in the global stage, generate multiple candidate rotation centers, and select a candidate rotation center that can make the most number of virtual corresponding point pairs satisfy the inlier condition as the optimal posterior rotation center; S43: solve the final rigid transformation based on the optimal posterior rotation center.

[0017] Compared with the prior art, the present application has the following beneficial effects: (1) Through the overlap perception mechanism and the node-guided virtual corresponding generation, the matching ambiguity and the extremely low inlier rate problem in the low overlap scene are effectively solved, the method is not sensitive to noise and outliers, and the registration accuracy is significantly better than that of the traditional method.

[0018] (2) The synchronous multi-view registration paradigm is adopted to directly estimate the transformation parameters in the global coordinate system, which fundamentally eliminates the error chain propagation and accumulation problem in sequential registration.

[0019] (3) The bidirectional local-to-global solver does not need time-consuming RANSAC iteration, and the entire registration process only needs one forward propagation, so the inference speed is much faster than that of the existing method based on recursion or iteration, and the real-time requirement of industrial online detection is met.

[0020] (4) By introducing the Chamfer distance as a self-supervised loss, the domain gap between the synthetic training data and the real scanning data is effectively reduced, the generalization performance of the model in the real industrial scene is improved, and the data labeling cost is reduced. BRIEF DESCRIPTION OF DRAWINGS

[0021] Figure 1 is the overall flowchart of the NGVC-Net network of the present application.

[0022] Figure 2 is the flowchart of the overlapping perception superpoint matching module and the node-guided virtual corresponding point generation module of the present application.

[0023] Figure 3 is the schematic diagram of three representative turbine blades and their target measurement cross sections.

[0024] Figure 4 is the registration result visualization comparison diagram of different registration methods on the blade profile of the test set.

[0025] Figure 5 is the registration result visualization comparison diagram of different registration methods on the real-world blade profile.

[0026] Figure 6 is the blade target measurement cross section deviation chromatogram of the registration result of the method of the present application. DETAILED DESCRIPTION

[0027] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments.

[0028] The present embodiment provides a multi-view blade point cloud fast registration method based on node-guided virtual correspondence, which comprises the following steps: Step 1: generating multi-level superpoint features based on a feature extraction module and , and point-level features and .

[0029] Two fields of view with overlapping areas are selected, which are source point cloud and target point cloud , KPConv is used as the backbone network to downsample the source point cloud and the target point cloud to generate superpoint sets and , and corresponding multi-level superpoint features and ; Local patches and are constructed for each superpoint in the superpoint sets and , and local geometric information is extracted; ; wherein, is a point in a point cloud , is a point in a point cloud , is a point in a super point cloud , is a point in a super point cloud .

[0030] and up-sampling by NN-upsampling to generate point-level features and .

[0031] Step 2: Filter out reliable super point pairs in the overlapping region based on the overlapping-aware super point matching module.

[0032] Extract multi-level super point features and context information based on GeoTransformer to obtain fusion features and ; then predict the overlap score of each super point and binarize to filter out non-overlapping points; then calculate the Gaussian correlation matrix and use the double soft-max operation to suppress fuzzy matching; finally select the Top-k reliable super point pairs.

[0033] This embodiment connects two fully connected layers after GeoTransformer; use ReLU and sigmoid nonlinear activation functions to map the output to the interval [0, 1] for predicting the overlap score of each super point in the overlapping region, which is then binarized to generate the overlap score matrix and , the overlap score of the i-th corresponding super point wherein, is a threshold.

[0034] Calculate a Gaussian correlation matrix from the fusion features and , point wherein, is the fusion feature of the i-th super point in the fusion feature , is the fusion feature of the j-th super point in the fusion feature .

[0035] Use the double soft-max operation to suppress fuzzy matching: ; wherein, , Let G be the i-th row and j-th column of the Gaussian correlation matrix. Let be the i-th row and j-th column of the correlation matrix G of the Gaussian matrix after a softmax operation on the i-th row.

[0036] use and Masking Gaussian correlation matrix Generative matrix corresponding to non-overlapping points Select the top-k items to form the set of superpoints. .

[0037] Step 3: The node-guided virtual correspondence point generation module generates confidence-weighted virtual correspondence points in the target point cloud.

[0038] Based on the aforementioned point-level features, a point-by-point cost matrix between the source and target point clouds is calculated, and a confidence matrix is ​​obtained through point-by-point matching using the Sinkhorn algorithm. Bidirectional Top-k matching is selected as the potential corresponding points. For each source point, a virtual corresponding point is generated in the target point cloud as a weighted average of its neighboring points (weighted by confidence). A set of virtual corresponding point pairs is constructed for subsequent registration.

[0039] The cost matrix , Point-level features The features of the i-th point, Point-level features The features of the j-th point, For feature dimensions.

[0040] In this embodiment, the Sinkhorn algorithm is applied to the obtained cost matrix to obtain the soft correspondence confidence matrix. Select the top k matches in both rows and columns to form potential correspondences. Potential correspondence , in, , For the i-th group, there is a corresponding local patch, that is, the superpoints within the patch have a corresponding relationship; express The Middle The j-th point and the patch The j-th point, This represents the soft correspondence confidence level between the two points.

[0041] From a geometric perspective, this process aims to extract points located in the overlapping region of two local patches. For each The virtual corresponding points are calculated as a confidence-weighted average: ,in express and The corresponding confidence level.

[0042] Then, place the point Pairing with its virtual counterpart To form a virtual set of corresponding points Calculate in the same way and Aggregates into a virtual corresponding point set. and .

[0043] Step 4: Estimate the optimal transformation vector based on the dual fusion local-to-global strategy.

[0044] Local phase: Assign weights to each virtual corresponding point pair and remove outliers.

[0045] Based on virtual corresponding point set This embodiment is the first one. A weight is assigned to each pair of corresponding points in each local patch. , , Represents the Iverson symbol. It is the acceptance radius.

[0046] Transform vector calculated based on weights .

[0047] Similarly, obtain the transformation vector. .

[0048] Global phase: Select the rotation center that yields the most interior points among all virtual corresponding points as the optimal solution.

[0049] In this embodiment, the center of rotation is selected at all virtual corresponding points. To obtain the most interior points, use them as the optimal rotation center, and merge. The optimal transformation vector is calculated as follows: ,in Transformation vector Alignment operations, To accept the radius.

[0050] To evaluate the practical applicability of the proposed method, experiments were conducted using real-world data obtained through a high-precision four-axis motion platform. A three-dimensional solid model of the blade was measured by aligning the laser plane of the line laser scanner (LSS) with a reference A. The cross-sections (TMCSs) measured by the target were then analyzed. The vertical spacing is extracted and down-sampled by farthest point sampling (FPS) to meet the computational constraints. For accuracy evaluation, the measurement data of specified TMCSs from a coordinate measuring machine (CMM) are used as ground truth, following the blade manufacturing quality standards. This setup enables direct benchmarking with traditional high-precision measurement methods.

[0051] To evaluate the reliability of NGVC-Net, three representative turbine blade profiles are selected as experimental objects, as shown in Figure 3 NGVC-Net is compared with five state-of-the-art algorithms: a probabilistic-based method (CPD), a soft correspondence method RPM-Net, a coarse-to-fine method GeoTransformer (GT), and CSPN and RMCR-Net.

[0052] Evaluation on the theoretical dataset (contains 20000 samples, of which 16000 are used for training, 2000 for validation, and 2000 for testing): As shown in Table 1, all learning models are trained on the training set of the theoretical dataset and evaluated on the noise-free test set. Compared with RMCR-Net, NGVC-Net has higher reconstruction accuracy on Blade 1 and Blade 2, while its performance on Blade 3 is slightly lower than that of RMCR-Net.

[0053] CPD performs poorly, mainly due to its reliance on accurate initial pose estimation to avoid converging to local minima, which makes it very sensitive to poor initialization, a common problem in fast blade scanning. Although CSPN, RPM-Net, and GT have made significant progress in registration accuracy, they inherently rely on a rigid registration method (RPPCR), leading to a cascade of error propagation, where registration errors between view pairs accumulate. RPM-Net and CSPN heavily rely on fine point correspondences, which, for blades, can lead to global misalignment between parts due to similar features in different TMCSs.

[0054] Table 1 Results on the noise-free test set

[0055] In the superpoint stage, GT often matches wrong superpoint pairs, leading to wrong correspondences and unstable performance on the test set. Moreover, its accuracy is limited by the tendency to assign high confidence to real inliers and nearby but incorrect matches, thus limiting the reconstruction accuracy. When selecting the correspondence set using the Top-k strategy, it is difficult to ensure accurate matches, i.e., a larger k value usually introduces irrelevant points. In contrast, the centralized registration paradigm of NGVC-Net eliminates the error chain propagation by enforcing the simultaneous multi-view constraint. On Blade 3, compared to the three FOVs of Blade 1 and Blade 2, Blade 3 has five FOVs, and the overlap between adjacent FOVs decreases, leading to a decrease in real superpoint pairs. This decrease reduces the effectiveness of BLGS in the global stage, thus limiting the potential of NGVC-Net in further improving accuracy.

[0056] NGVC-Net solves the global ambiguous matching problem through OASM, and further filters out incorrect superpoint correspondences using the overlap score. To cope with low overlap rates, it implements NGVC, which reformulates the problem as generating confidence-weighted virtual correspondences consistent with the correspondence space. BLGS adopts a two-stage outlier suppression mechanism to reduce the distortion of the center of rotation transformation vector, while the center-aware registration mechanism fundamentally solves the error propagation chain in RPPCR. Therefore, compared to other methods, NGVC-Net is more likely to achieve accurate reconstruction results.

[0057] RPM-Net, CSPN, and RMCR require multiple iterations of recursive forward propagation to improve performance, which greatly increases the inference time. In contrast, the OASM and BLGS of NGVC-Net effectively filter out outliers, thus suppressing their impact on registration and improving overall accuracy. In addition, the RANSAC-free BLGS registration strategy eliminates the time-consuming RANSAC process in traditional correspondence-based methods. This registration process only requires point cloud data from two viewpoints to align all viewpoints, rather than computing registration for each pair of viewpoints, which further reduces the forward propagation time. Therefore, NGVC-Net achieves the shortest processing time among all comparison methods. It is worth noting that for RPM-Net and CSPN, reaching optimal performance requires calling the Open3D library to estimate the point normal vector for each sample - an additional cost not included in the current timing test, but very important in real-world engineering applications.

[0058] Evaluation in real-world applications: The performance of the above methods on real-world data is further evaluated through qualitative and quantitative analysis.

[0059] Qualitative and quantitative analysis of reconstructed target measurement cross-section: To quantify the registration accuracy of NGVC-Net and other methods, the TMCS was imported into Geomagic Control 2014 and compared with the blade profile measured by CMM as a reference. Specifically, in Geomagic, the registration result (as test) was first aligned with the CMM measured TMCS (as reference), and then the distance of each point to the nearest point in the reference was calculated as error.

[0060] The non-learning algorithm (CPD) showed unacceptable results, as it performed poorly even on clean theoretical datasets, and its registration performance further declined when there was noise in real-world data. For deep learning methods, domain shift was a significant challenge. Therefore, RPM-Net, CSPN, and GT did not perform as well as NGVC-Net on real-world data. As shown in Table 2, it is worth noting that GT incorrectly evaluated the superpoint correspondences, which were located outside the overlapping area, resulting in a large number of outliers and ultimately leading to registration failure. In contrast, NGVC-Net successfully resolved the global ambiguous matching using a more rigorous OBSM and reduced the incorrect superpoint correspondences. NGVC-Net and RMCR-Net benefited from the self-supervised loss function, which helped to alleviate the domain difference, enabling them to achieve the best overall performance on real-world data.

[0061] Table 2. Results on real data sets

[0062] On both theoretical and real blade datasets, NGVC-Net achieved higher reconstruction accuracy with shorter inference time than the state-of-the-art baseline methods, thanks to the spatial consistency constraint on virtual correspondences and the elimination of iterative outlier rejection in the pose estimation process. In addition, the registration process only aligns all views using two viewpoints, avoiding tedious pairwise matching and ensuring the shortest processing time among all comparison methods—this feature has clear practical value for high-throughput metrology.

[0063] NGVC-Net provides a fast, robust, and practically deployable solution for precise blade profile reconstruction. Its centralized multi-view constraint, outlier-resistant correspondence modeling, and efficient RANSAC-free estimation provide a significant improvement over existing methods, with immediate industrial inspection application significance where accuracy, throughput, and reliability must be achieved simultaneously. Future work will focus on fully self-supervised training to further enhance generalization and reduce dependence on synthetic data.

[0064] The above merely describes preferred specific embodiments of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art, according to the technical solution and inventive concept of the present application, makes equivalent replacement or change within the technical range disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1. A method for rapid registration of multi-view leaf point clouds based on node-guided virtual correspondence, characterized in that, Includes the following steps: S1: Extract features from the input source point cloud and target point cloud to generate superpoint sets, superpoint features, and point-level features; S2: Based on the superpoint features, perform overlap-aware matching, predict the superpoint overlap score, and filter out non-overlapping superpoints to obtain a reliable set of superpoint correspondences; S3: Based on the superpoint correspondence set and the point-level features, generate node-guided, confidence-weighted virtual correspondence points for points in the source point cloud in the target point cloud, forming a set of virtual correspondence point pairs; S4: Based on the set of virtual corresponding point pairs, the optimal rigid body transformation parameters are estimated through a two-way local to global hypothesis verification strategy, and the rotation center is solved to achieve multi-view synchronous registration.

2. The method for rapid registration of multi-view leaf point clouds based on node-guided virtual correspondence according to claim 1, characterized in that, In S1, the KPConv backbone network is used to perform hierarchical downsampling on the source point cloud and the target point cloud to generate the superpoint set and its superpoint features, and the superpoint features are upsampled to obtain the point-level features.

3. The method for rapid registration of multi-view leaf point clouds based on node-guided virtual correspondence according to claim 2, characterized in that, For each superpoint in the superpoint set, a local patch is constructed in its original point cloud; ； in, , Superpoints and Super Point Local patch, Source Point Cloud The point in the middle, For target point cloud The point in the middle, For super point cloud The point in the middle, For super point cloud point.

4. The method for rapid registration of multi-view leaf point clouds based on node-guided virtual correspondence according to claim 3, characterized in that, S2 performs overlap-aware matching based on the superpoint features, predicts the superpoint overlap score, and filters out non-overlapping superpoints to obtain a reliable set of superpoint correspondences, specifically including: S21: Extract enhanced features from the superpoint features using a context encoder based on GeoTransformer; S22: Predict the overlap confidence score of each superpoint feature through a fully connected layer and an activation function, and binarize the overlap confidence score to generate an overlap score matrix; S23: Calculate the Gaussian correlation matrix between the enhanced features; S24: Use the overlap fraction matrix to mask the entries in the Gaussian correlation matrix that correspond to non-overlapping superpoints; S25: Based on the double soft maximum operation to suppress fuzzy matching, select Top-K entries to form the reliable superpoint correspondence set.

5. The method for rapid registration of multi-view leaf point clouds based on node-guided virtual correspondence according to claim 4, characterized in that, In S3, based on the superpoint correspondence set and the point-level features, node-guided, confidence-weighted virtual correspondence points are generated in the target point cloud for points in the source point cloud, forming a set of virtual correspondence point pairs, specifically including: S31: Calculate the point-by-point cost matrix between the source point cloud and the target point cloud based on the point-level features, and apply the Sinkhorn algorithm to solve for the soft correspondence confidence matrix. S32: Guided by the superpoint correspondence set, select the top-ranked bidirectional point-by-point matches from the soft correspondence confidence matrix as potential correspondence points; S33: For a point in the source point cloud, calculate its virtual corresponding point in the target point cloud as the confidence-weighted average position of the potential corresponding point, thereby forming the set of virtual corresponding point pairs.

6. The method for rapid registration of multi-view leaf point clouds based on node-guided virtual correspondence according to claim 5, characterized in that, In S4, based on the aforementioned set of virtual corresponding point pairs, the optimal rigid body transformation parameters are estimated using a bidirectional local-to-global hypothesis verification strategy, specifically including: S41: In the local phase, weights are assigned to each pair of virtual correspondences based on structural invariance information, and the weighted rotation center is calculated. S42: In the global phase, generate multiple candidate rotation centers and select the candidate rotation center that enables the maximum number of virtual corresponding point pairs to satisfy the interior point condition as the optimal posterior rotation center. S43: Based on the optimal posterior rotation center, solve for the final rigid body transformation.

Images