Material sample automatic processing method and system based on vision

Abstract

The invention discloses a material sample automatic processing method and system based on vision, and relates to the technical field of insulator detection, and the method comprises the steps: collecting a point cloud image of a to-be-detected insulator, carrying out the high-precision registration processing of the point cloud image, carrying out the segmentation and classification, and extracting a point cloud data set; performing principal component analysis on the point cloud data set to generate an initial reference framework; constructing a control grid based on the initial reference frame, performing B-spline smooth curved surface fitting on the point cloud by using the control grid, adjusting the position of a B-spline curved surface control point according to fitting error feedback, and generating a fitting curved surface after iterative optimization; according to the optimized fitting curved surface, key feature lines and a curved surface area of the insulator are extracted; based on the key characteristic line and the curved surface area of the insulator, key geometric parameters, plating thickness and electrical performance parameters of the insulator are calculated, and comprehensive analysis is carried out to generate a detection conclusion; according to the method, the problems of non-uniform point cloud attitudes, low curved surface fitting precision and incomplete key feature extraction are solved.

Description

Technical Field

[0001] This invention relates to the field of insulator testing technology, and more specifically, to a vision-based automatic material sample processing method and system. Background Technology

[0002] Material samples, especially complex structural material components such as insulators, are widely used in the fields of power, electronics and high-precision equipment. Their geometric dimensions, surface coating quality and electrical properties directly affect the safety and performance stability of the materials.

[0003] With the development of 3D vision technology and deep learning algorithms, automated detection methods based on point clouds have been gradually proposed. High-precision point cloud information of material surfaces can be acquired using 3D vision sensors such as structured light or laser scanning, and algorithms can be used to process and analyze the point clouds. However, existing technologies still have the following main drawbacks: Inconsistent point cloud poses make standardization difficult: Due to pose differences and spatial redundancy, direct geometric analysis of point clouds acquired from multiple perspectives can lead to error accumulation and make it difficult to obtain a consistent reference frame.

[0004] Insufficient surface fitting accuracy makes it difficult to reflect complex structural features: Traditional surface fitting methods often use fixed control grids or global smooth constraints, which are difficult to take into account both complex high curvature regions and flat regions. This results in large fitting errors in key structural parts, affecting the accuracy of geometric parameters and coating thickness measurements.

[0005] Incomplete extraction of key features and lack of structured analysis capabilities: Existing methods are unable to automatically and accurately identify key feature lines and curved areas of various functional parts of insulators, resulting in a lack of spatial structural correlation in the analysis of geometric parameters, coating thickness and electrical performance, and making it impossible to form a structured data set for comprehensive evaluation. Summary of the Invention

[0006] To overcome the above-mentioned defects of the prior art, embodiments of the present invention provide a vision-based automatic material sample processing method and system, which realizes high-precision registration of multi-view point clouds, structured surface fitting and key feature line extraction, thereby solving the problems of inconsistent point cloud poses, low surface fitting accuracy and incomplete key feature extraction.

[0007] To achieve the above objectives, the present invention provides the following technical solution: In a first aspect, this application provides a vision-based automatic material sample processing method, which includes: acquiring point cloud images of insulators to be tested; performing high-precision registration processing on the point cloud images; segmenting and classifying the processed point cloud images to extract point cloud data sets; performing principal component analysis on the point cloud data sets to determine the main directions and local coordinate references of the point clouds and generating an initial reference frame; constructing a control grid based on the initial reference frame; using the control grid to perform B-spline smoothing surface fitting on the point clouds; adjusting the positions of the B-spline surface control points according to the fitting error feedback to generate an iteratively optimized fitted surface; extracting key feature lines and surface regions of the insulator based on the optimized fitted surface; and calculating key geometric parameters, coating thickness, and electrical performance parameters of the insulator based on the key feature lines and surface regions of the insulator, and performing comprehensive analysis to generate test conclusions.

[0008] In one embodiment, acquiring a point cloud image of the insulator to be tested and performing high-precision registration processing on the point cloud image includes: preprocessing the point cloud image to generate a pre-processed point cloud image, and denoising it to obtain a denoised point cloud image; calculating a fast point feature histogram on the denoised point cloud image and generating a feature description set; performing coarse registration on the feature description set by matching the fast point feature histograms in the point cloud from different viewpoints to determine candidate matching point pairs; performing consistency screening on the candidate matching point pairs to obtain valid matching point pairs; calculating an initial spatial transformation relationship based on the valid matching point pairs, and transforming the point cloud according to the initial spatial transformation relationship to obtain a coarse registration result; and performing fine registration using an improved iterative nearest point method based on the coarse registration result to obtain a high-precision registered insulator point cloud image.

[0009] In one embodiment, based on the coarse registration result, an improved iterative nearest-point method is used for fine registration to obtain a high-precision registered insulator point cloud image. This includes: using the point cloud obtained from the coarse registration result as the point cloud to be registered, and using the target point cloud as the reference point cloud; in the point cloud to be registered, searching for the nearest point only within the local neighborhood of its corresponding target point cloud for each point, forming a point-to-point correspondence; for the initial point-to-point correspondence, obtaining the change in spatial distance between corresponding point pairs and the change in the angle between the normal vectors of corresponding points, and obtaining valid point pairs through consistency judgment; based on the valid point pairs, calculating the incremental rotation matrix and incremental translation vector in the current iteration; superimposing the incremental rotation matrix and incremental translation vector onto the current spatial pose of the point cloud to be registered for updating; if a preset condition is met, stopping the iteration, and outputting the high-precision registered insulator point cloud image.

[0010] In one embodiment, the processed point cloud image is segmented and classified to extract a point cloud dataset, including: determining the distance threshold for Euclidean clustering based on the average distance between adjacent points in the precisely registered insulator point cloud image, and performing clustering segmentation on the point cloud image based on the Euclidean distance between two points to obtain several spatially independent point cloud subsets; obtaining the number of points and spatial bounding size of each point cloud subset, and determining candidate point cloud subsets; performing point cloud subset preprocessing based on centroid alignment and scale normalization on each candidate point cloud subset; constructing input point cloud data for the processed candidate point cloud subsets; inputting the input point cloud data into a pre-trained PointNet neural network to output the structure category of each insulator subset; and summarizing point cloud subsets belonging to the same structure category according to the structure category determination result to generate a structured point cloud dataset.

[0011] In one embodiment, principal component analysis is performed on the point cloud dataset to determine the main directions and local coordinate references of the point cloud and generate an initial reference frame. This includes: calculating the covariance matrix of the three-dimensional coordinates based on the point cloud dataset; performing eigenvalue decomposition on the covariance matrix to obtain direction feature vectors, wherein the direction feature vectors include principal direction feature vectors, secondary directions, and plane normal direction feature vectors; and establishing a local coordinate reference frame for the point cloud using the direction feature vectors as coordinate axis directions to form the initial reference frame.

[0012] In one embodiment, a control grid is constructed based on an initial reference frame, and the point cloud is fitted with a B-spline smooth surface using the control grid. This includes: obtaining the number of neighboring points within a preset neighborhood radius for each point in the point cloud dataset initialized by principal component analysis, forming a local density distribution; setting thresholds for high-density and low-density regions based on the overall average density and standard deviation of the point cloud, and dividing the point cloud into high-density and low-density regions; generating an initial regular control grid within the spatial enclosure of the point cloud, and mapping the grid to the initial reference frame of the point cloud; adaptively adjusting the initial control grid according to the local density distribution, increasing the preset proportion of control points in high-density regions and decreasing the preset proportion of control points in low-density regions; and using the adjusted control grid to fit a B-spline smooth surface to the point cloud, generating an initial fitted surface.

[0013] In one embodiment, based on the fitting error feedback, the positions of the B-spline surface control points are adjusted to generate an iteratively optimized fitted surface. This includes: calculating the minimum Euclidean distance from each point in the point cloud dataset to the initial fitted surface, using this distance as the point-to-surface fitting error value, and forming an error distribution set; statistically analyzing the error distribution set to calculate the overall average error and identify high-error regions; mapping the high-error regions to the corresponding B-spline surface control point grid in the parameter space and determining the set of control points associated with the high-error regions; assigning error weight values ​​to the control points based on the magnitude of the error within the corresponding high-error regions; updating the control point positions along the point cloud normal direction based on the error weight values; and performing a second B-spline smoothing surface fitting based on the updated control point positions, followed by iterative optimization to output the iteratively optimized fitted surface that meets preset conditions.

[0014] In one embodiment, the key feature lines and surface regions of the insulator are extracted based on the optimized fitted surface, including: obtaining its parametric expression form based on the optimized fitted surface and mapping it to a two-dimensional parameter space; obtaining local geometric features corresponding to each position on the surface, including normal vectors, principal curvature, and rate of curvature change, within the parameter space; identifying regions with significant local extrema based on the statistical distribution of the principal curvature and rate of curvature change, and determining these regions as candidate feature regions; generating several feature line seed points based on the curvature extrema points within the candidate feature regions; continuously tracing along the principal curvature direction of the surface from the feature line seed points to generate key feature lines of the insulator; using the key feature lines as surface segmentation boundaries to divide the fitted surface into multiple independent surface regions; extracting the structural parameters of each surface region and labeling its structural attributes to form a set of structured surface regions; and outputting the set of key feature lines and surface regions of the insulator.

[0015] In one embodiment, based on the key feature lines and curved surface regions of the insulator, key geometric parameters, coating thickness, and electrical performance parameters of the insulator are calculated, and a comprehensive analysis is performed to generate a test conclusion. This includes: calculating the geometric parameters of each key structural part of the insulator based on the extracted key feature lines and curved surface regions; calculating the average thickness and standard deviation within the determined curved surface regions to form coating thickness data; collecting electrical performance data based on the key structural parts of the insulator to form a set of electrical performance parameters; comprehensively analyzing the geometric parameters, coating thickness data, and electrical performance parameters to determine the passability of each key structural part and generate passability marks for each parameter; and generating a test conclusion report based on the comprehensive analysis results.

[0016] Secondly, this application provides a vision-based automatic material sample processing system, which includes: an image processing module for acquiring point cloud images of the insulator to be tested, performing high-precision registration processing on the point cloud images, segmenting and classifying the processed point cloud images, and extracting a point cloud data set; a PCA initialization module for performing principal component analysis on the point cloud data set, determining the main directions and local coordinate references of the point cloud, and generating an initial reference frame; a fitted surface generation module for constructing a control grid based on the initial reference frame, using the control grid to perform B-spline smoothing surface fitting on the point cloud, and adjusting the positions of the B-spline surface control points according to the fitting error feedback to generate an iteratively optimized fitted surface; a feature extraction module for extracting key feature lines and surface regions of the insulator based on the optimized fitted surface; and a parameter detection module for calculating key geometric parameters, coating thickness, and electrical performance parameters of the insulator based on the key feature lines and surface regions of the insulator, and performing comprehensive analysis to generate detection conclusions.

[0017] As can be seen from the above technical solutions, the embodiments of this application have the following advantages: Starting with high-precision point cloud acquisition from multiple perspectives, denoising, and registration, and combining an initial reference frame based on PCA, B-spline smooth surface fitting, adaptive control mesh, and error feedback-driven iterative optimization, this approach can accurately extract key feature lines and structured surface regions of insulators. Based on this, it enables joint measurement and comprehensive analysis of creepage distance, geometric dimensions, coating thickness, and electrical performance. This scheme not only improves the accuracy and reliability of insulator testing for complex structures, ensuring the structured, consistent, and repeatable nature of measurement results, but also reduces manual intervention, significantly improving testing efficiency and providing technical support for the rapid, automated, and scientific generation of reliable testing conclusions. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of a vision-based automatic material sample processing method provided in an embodiment of this application.

[0019] Figure 2 This is a schematic diagram of a vision-based automatic material sample processing system provided in an embodiment of this application. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0021] Reference Figure 1 As shown, the present invention provides a schematic flowchart of a vision-based automatic material sample processing method, which includes the following steps: S1. Acquire point cloud images of the insulator to be tested, perform high-precision registration processing on the point cloud images, and segment and classify the processed point cloud images using Euclidean clustering and PointNet algorithms to extract point cloud data sets.

[0022] Among them, a three-dimensional vision sensor is used to scan the insulator under test from multiple perspectives. The sensor is a structured light or laser scanner with a scanning accuracy of 0.1~0.5 mm and a resolution of ≥640×480, and acquires point cloud images of the insulator under test.

[0023] In this embodiment, a point cloud image of the insulator to be tested is acquired, and the point cloud image is subjected to high-precision registration processing, including: The point cloud image is preprocessed to generate a preliminary processed point cloud image; The preprocessing includes isolated point removal and coordinate normalization. Isolated point removal is performed by calculating the number of neighboring points within a radius of 1 to 3 mm for each point. Points below a set threshold are considered isolated points and are removed. At the same time, the point cloud coordinates are normalized to a unified coordinate system.

[0024] The processed point cloud image is processed by a bilateral filtering algorithm, where the spatial weight controls the radius of 1~2 mm, the intensity weight controls the difference in normal vectors of 0.05~0.1, and the number of iterations is 2~5, in order to preserve edge and detail structure while denoising, and generate a denoised point cloud image. Calculate a fast point feature histogram for the denoised point cloud image with a neighborhood radius of 3-5 mm, and generate a feature description set for the point cloud. Specifically, the calculation of the fast point feature histogram involves: taking each point in the denoised point cloud image as the center, searching for neighboring points within a preset neighborhood radius, and obtaining the normal vector of the point based on the spatial distribution of the neighboring points; based on the relationship between the normal vectors of the point and its neighboring points, calculating the angular relationship (normal and direction relationship, the angle between normals, and the normal torsion relationship) and distance relationship (i.e., the Euclidean distance between the current point and its neighboring points) between the points, forming a set of local geometric relationships; statistically analyzing the set of local geometric relationships to generate a fast point feature histogram describing the local spatial structure of the point; and combining the fast point feature histogram features corresponding to each point in the denoised point cloud image to form a feature description set of the point cloud.

[0025] The specific calculation formula for the aforementioned angular relationship is as follows:

[0026]

[0027]

[0028] In the formula, The relationship between normal and direction. This is the vertical direction vector in the local coordinate system. Let be the normal vector of the neighborhood points. The angle between the normal directions. Let be the normal vector of the current point. The unit direction vector from the current point to its neighboring points. Let be the torsion angle of the neighboring point's normal vector relative to the current point's normal vector. This is the third orthogonal direction in the local coordinate system.

[0029] Coarse registration is performed based on the feature description set, and the fast point feature histograms in point clouds from different viewpoints are matched with each other. Candidate matching point pairs are determined by feature similarity search method. Consistency screening is performed on candidate matching point pairs to remove erroneous matching point pairs whose feature differences exceed a preset threshold, thus obtaining valid matching point pairs. Based on valid matching point pairs, the initial spatial transformation relationship is calculated, which includes the rotation matrix and translation parameters. The rotation matrix is ​​calculated using the following formula:

[0030] In the formula, To rotate and align the source point cloud to the target point cloud using a rotation matrix, Let be the principal direction basis matrix in the target point cloud space, and let its column vectors reflect the main spatial distribution directions of the target point cloud. The source point cloud principal direction basis matrix The transpose of is used to transform the source point cloud from its principal direction coordinate system to the standard coordinate system.

[0031] The specific calculation formula for the translation parameter is as follows:

[0032] In the formula, The translation vector required to move the rotated source point cloud to the target point cloud position. The geometric center of the point set is matched to the target point cloud to eliminate the overall translation effect of the target point cloud. It serves as the geometric center of the matching point set for the source point cloud, used to eliminate the overall translational effects of the source point cloud.

[0033] The point cloud is transformed according to the initial spatial transformation relationship to obtain a coarse registration result, thereby achieving the initial alignment of the point cloud from multiple perspectives. The specific calculation formula for the transformation is as follows:

[0034] In the formula, These are the points aligned with the target point cloud after transformation. For any point in the source point cloud, Let be the rotation matrix, and t be the translation parameter.

[0035] Based on the coarse registration results, an improved iterative nearest point method is used for fine registration to obtain a high-precision registered insulator point cloud image.

[0036] Furthermore, based on the coarse registration results, an improved iterative nearest-point method is used for fine registration to obtain a high-precision registered insulator point cloud image, including: The point cloud after coarse registration is used as the point cloud to be registered, and the target point cloud is used as the reference point cloud. Both are used as the initial input for fine registration in a unified spatial coordinate system. In the point cloud to be registered, for each point, the nearest point (i.e. the point with the smallest Euclidean distance) is searched only in the local neighborhood of its corresponding target point cloud to form a point pair correspondence; The local neighborhood range of the target point cloud is determined by the average registration error between the source point cloud and the target point cloud after coarse registration.

[0037] For the initial point pair correspondence, the consistency of the change in spatial distance between corresponding point pairs and the change in the angle between the normal vectors of corresponding points is judged, and point pairs that simultaneously satisfy the condition that both the change in distance and the change in the angle between normal vectors are less than a preset threshold are retained as valid point pairs. Based on the effective point pairs, the incremental rotation matrix and incremental translation vector under the current iteration are calculated using the least squares criterion, which are used to describe the incremental spatial transformation relationship between the point cloud to be registered and the target point cloud. The incremental rotation matrix and incremental translation vector are superimposed on the current spatial pose of the point cloud to be registered, and the position and orientation of the point cloud to be registered are updated. If the incremental spatial change obtained from two consecutive iterations is less than a preset threshold, the iteration stops and a high-precision registered insulator point cloud image is output.

[0038] It should be noted that the improvements of the improved iterative nearest-point method are as follows: First, it uses the coarse registration result as the initial pose input for fine registration, avoiding the problem of easily getting trapped in local optima caused by starting from random or zero pose in conventional iterative nearest-point methods; second, it adaptively determines the local neighborhood range of the nearest-point search based on the average registration error after coarse registration, rather than using a fixed search radius, thereby improving adaptability under different initial accuracy conditions; third, after establishing the point pair correspondence, it introduces a geometric consistency screening mechanism based on the change in spatial distance and the change in the angle between the normal vectors to eliminate mismatched point pairs, and then calculates the incremental spatial transformation based on the filtered valid point pairs. Through the above improvements, this fine registration method can significantly improve registration stability and accuracy while ensuring computational efficiency, reducing the impact of noise, occlusion, and complex curved surfaces on the registration results, and providing a reliable point cloud basis for subsequent high-precision measurement of insulator creepage distance, size, and electrical performance parameters.

[0039] Furthermore, the processed point cloud image is segmented and classified using Euclidean clustering and PointNet algorithms to extract the point cloud dataset, including: Based on the average distance between adjacent points in the insulator point cloud image after fine registration, the distance threshold for Euclidean clustering is determined, and the point cloud image is segmented into Euclidean clusters based on the Euclidean distance relationship between points to obtain several spatially independent subsets of the point cloud. Based on the point cloud subsets, obtain their point count and spatial bounding size respectively, and determine the point cloud subsets whose point count and spatial bounding size simultaneously meet the preset range conditions as candidate point cloud subsets; For each candidate point cloud subset, a point cloud subset preprocessing based on centroid alignment and scale normalization is performed, including: calculating its geometric centroid and performing translation alignment processing, while normalizing the point cloud coordinates according to its maximum spatial bounding size, so that each candidate point cloud subset is uniformly mapped to a preset scale range. The specific formula for calculating the geometric centroid is as follows:

[0040] In the formula, Let the geometric centroid coordinates of the candidate point cloud subset be... This represents the total number of points contained in the candidate point cloud subset. Let be the three-dimensional coordinate vector of the i-th point in the candidate point cloud subset.

[0041] For the processed subset of candidate point clouds, a point cloud representation containing a preset number of points is constructed by random sampling, which serves as the input point cloud data for the PointNet network; The input point cloud data is fed into a pre-trained PointNet neural network. Point-level features are encoded using a shared multilayer perceptron, which includes a multilayer node configuration, such as 64 nodes, 128 nodes, or 1024 nodes, with ReLU activation function to extract point-level features. Features are then aggregated using symmetric functions (such as max pooling or average pooling) to obtain a global feature vector representing the overall structural features of the candidate point cloud subset. This global feature vector is then input into a classifier (such as a two- or three-layer fully connected network) to determine the structural category and output the structural category of each subset of the insulator, such as cap, thread, or support rod. Based on the structural category determination results, the point cloud subsets belonging to the same structural category are aggregated to generate structured point cloud data sets corresponding to different structural parts of the insulator.

[0042] It should be noted that the point cloud data set extracted by the above technical solution can achieve automatic and accurate segmentation and classification of various structural parts of the insulator, ensuring the structured and consistent nature of the point cloud data and its high reliability for direct use in subsequent dimensional measurement and multi-parameter analysis.

[0043] S2 performs principal component analysis on the point cloud dataset to determine the main orientation and local coordinate reference of the point cloud and generate an initial reference frame.

[0044] In this embodiment, principal component analysis is performed on the point cloud dataset to determine the main orientation and local coordinate reference of the point cloud, generating an initial reference frame, including: Calculate the covariance matrix of the three-dimensional coordinates based on the point cloud dataset; The specific formula for calculating the covariance matrix is ​​as follows:

[0045] In the formula, Let covariance matrix be the variance matrix. The total number of points in the point cloud dataset. The coordinates of the geometric centroid of the point cloud in the point cloud dataset. Let be the three-dimensional coordinate vector of the i-th point.

[0046] Eigenvalue decomposition of the covariance matrix yields directional eigenvectors. , , ,in, The directional feature vector includes Main direction eigenvector, Secondary directional eigenvectors and Eigenvector of plane normal direction; Using directional feature vectors as coordinate axis directions, a local coordinate reference frame for the point cloud is established, where, As the main direction, and The secondary direction and the plane normal direction are respectively used to form the initial reference frame.

[0047] Specifically, the translated point cloud data set is aligned to the initial reference frame using a rotation matrix to ensure that the main direction of the rotated point cloud is consistent with the local coordinate system, and the rotation accuracy is controlled within 0.1°, providing a standardized reference for subsequent surface fitting or geometric feature extraction.

[0048] S3. Based on the initial reference frame, construct the control mesh required for the B-spline surface to ensure that the mesh covers the key areas of the point cloud. Use the control mesh to perform B-spline smoothing surface fitting on the point cloud, and adjust the position of the B-spline surface control points according to the fitting error feedback to generate the iteratively optimized fitted surface.

[0049] In this embodiment, based on the initial reference frame, a control mesh is constructed to meet the requirements of the B-spline surface, ensuring that the mesh covers the key areas of the point cloud. The control mesh is then used to perform B-spline smoothing surface fitting on the point cloud, including: For each point in the PCA-initialized point cloud dataset, obtain the number of its neighboring points within a preset neighborhood radius to form a local density distribution. This reflects the density of point clouds in space; PCA initialization refers to determining the main orientation and local coordinate reference of the point cloud by performing principal component analysis on the point cloud dataset, and aligning the point cloud translation and rotation to this initial reference frame to provide a unified coordinate system for subsequent processing.

[0050] Based on the overall average density of the point cloud and standard deviation Set a threshold for high-density areas threshold of low-density areas (k takes 1~2), the point cloud is divided into high-density and low-density regions to identify complex structures and flat areas; An initial rule control mesh is generated within the spatial bounding area of ​​the point cloud, and the mesh is mapped to the initial reference frame of the point cloud; Based on the local density distribution, the initial control grid is adaptively adjusted. If a preset proportion of control points is added in a high-density area, the distance between adjacent control points is reduced; if a preset proportion of control points is reduced in a low-density area, the distance between adjacent control points is increased. The adjustment process employs linear interpolation or quadratic smoothing transition methods to ensure a continuous and smooth distribution of control points.

[0051] The point cloud is fitted with a B-spline smooth surface using the adjusted control grid to generate an initial fitted surface.

[0052] Furthermore, based on the fitting error feedback, the positions of the B-spline surface control points are adjusted to generate an iteratively optimized fitted surface, including: Based on the initial fitted surface, the minimum Euclidean distance from each point in the point cloud dataset to the initial fitted surface is calculated as the corresponding point-to-surface fitting error value, and an error distribution set is formed. Statistical analysis is performed on the error distribution set to calculate the overall average error (mean of fitting error value), and the point cloud region with the average error exceeding the preset error threshold is identified as a high error region, while the remaining regions are identified as low error regions. The high-error region is mapped to the corresponding B-spline surface control point grid in the parameter space to determine the set of control points that have an influence relationship with the high-error region; The determination of the set of control points that have an influence relationship with the high error region can be understood as follows: based on the local support characteristics of the B-spline surface basis function (which is a function used to describe the influence relationship between control points in the surface parameter space and the shape of the local region of the surface, and has local support characteristics, so that each control point only affects the local region within the corresponding parameter interval of the surface), control points whose basis functions take non-zero values ​​within the corresponding parameter range of the high error region are selected as the set of control points that have an actual influence on the surface shape of the high error region.

[0053] For the set of control points, an error weight value is assigned to the control points according to the magnitude of the error in the corresponding high error region, so that the control points corresponding to the region with larger error have higher adjustment weight; First, the fitting error values ​​from each point cloud point in the high-error region to the current B-spline surface are statistically analyzed, and these error values ​​are used as error metrics. By normalizing the errors, error weight values ​​that are positively correlated with the error magnitude are generated. Then, based on the local support relationship of the B-spline surface basis functions, these error weights are mapped to control points that have an impact on the high-error region. This gives higher adjustment weights to control points that are closer to the high-error region and have a greater impact on the surface shape. As a result, the surface shape corresponding to the high-error region is corrected more preferentially and with greater magnitude during subsequent control point position updates, thus achieving adaptive optimization of fitting accuracy.

[0054] Based on the error weight values ​​of the control points, the positions of the control points are updated along the normal direction of the point cloud to reduce the fitting error of the point cloud to the B-spline surface and maintain the continuity and smoothness between adjacent control points. First, based on the error weight value corresponding to each control point, the adjustment range of the control point in one iteration is determined, and the normal directions of points in the point cloud that have an influence relationship with the control point are weighted and averaged as the main update direction of the control point. Then, the control point is updated by displacement along the main update direction, and the displacement is positively correlated with its error weight value. This makes the control points corresponding to high error regions move preferentially in the direction that reduces the point cloud-surface distance, thereby reducing the fitting error. At the same time, a smoothing constraint on the position of adjacent control points is introduced during the control point update process. For example, by limiting the displacement difference between adjacent control points, the control point adjustment maintains the continuity and smoothness of the overall surface while making local adaptive corrections, avoiding abrupt changes or oscillations, thereby achieving synergistic optimization of fitting accuracy and surface smoothness.

[0055] Based on the updated control point positions, a second B-spline smooth surface fitting is performed, and iterative optimization is carried out until the overall average error is less than the preset error threshold, and the iteratively optimized fitted surface is output.

[0056] It should be noted that PCA initialization first aligns the main direction and normalizes the coordinates of the point cloud, effectively eliminating attitude differences and spatial redundancy, so that the subsequent control mesh construction and surface fitting are based on a stable and unified reference frame. On this basis, the adaptive control mesh based on the local density of the point cloud and the error feedback-driven B-spline smooth surface fitting enable the surface to have higher fitting accuracy in structurally complex regions and maintain good smoothness in flat regions. This not only significantly reduces the overall fitting error, but also improves the continuity, stability and robustness of the surface, providing a high-quality surface foundation for subsequent geometric measurement and feature calculation.

[0057] S4. Based on the optimized fitted surface, extract the key feature lines and surface regions of the insulator.

[0058] In this embodiment, based on the optimized fitted surface, key feature lines and surface regions of the insulator are extracted, including: Based on the optimized fitted surface, its parametric expression is obtained, and the insulator surface is mapped to a unified two-dimensional parameter space to achieve a continuous and traceable representation of any point, curve, and region on the surface. Within the parameter space, first-order and second-order differential calculations are performed on the fitted surface to obtain the local geometric features corresponding to each position on the surface. The local geometric features include normal vectors, principal curvatures, and curvature change rates, which are used to characterize the local geometric morphology of the insulator surface. Based on the statistical distribution of the principal curvature and the rate of change of curvature of the surface, regions with significant local extrema are identified, and these regions are determined as candidate feature regions. Among them, the region with significant local extrema refers to the region with significant local extrema based on the statistical distribution of the principal curvature and the rate of change of curvature of the surface, such as the mean, variance or gradient distribution of curvature; then, the points with curvature higher than the surrounding neighborhood are identified, and the local regions where these points are located are the regions with significant local extrema.

[0059] Within the candidate feature region, several feature line seed points are generated based on the curvature extrema points, and these seed points serve as the starting constraints for subsequent feature line tracing. Starting from the seed point of the feature line, continuous tracing is performed along the principal curvature direction of the surface to generate a continuous spatial curve located on the fitted surface, which serves as the key feature line of the insulator. Using key feature lines as the surface segmentation boundary, the fitted surface is divided into multiple independent surface regions. Each surface region corresponds to a different structural part of the insulator, including the cap region, the skirt region, or the support rod region. For each surface region, its structural parameters are calculated, including average curvature, normal uniformity, and spatial dimension parameters. The surface regions are then labeled with structural attributes according to preset structural feature rules to form a set of structured surface regions with clear physical meaning. In this process, the average curvature (mean) of each region is calculated to reflect the overall curvature of the region, and the normal consistency (mean cosine similarity) is calculated to assess the uniformity of the point normal direction, thereby determining whether the region surface is smooth or has local unevenness. At the same time, the spatial dimensional parameters of the region (such as length, width, height or bounding box dimensions) are extracted to describe its geometric scale. Then, these parameters are matched with pre-defined structural feature rules (e.g., the cap region has high curvature and low normal consistency, the thread region has periodic protrusions, and the support rod region is slender and columnar with high normal consistency) to annotate the structural attributes of each surface region. Finally, all annotated surface regions are combined to form a set of structured surface regions. Each region not only has geometric coordinates and dimensional information, but also has a clear physical meaning, which can be used for further dimensional measurement, creepage distance analysis, or coating thickness detection.

[0060] Key feature lines and curved areas of output insulators.

[0061] It should be noted that the key feature lines of an insulator refer to the continuous spatial curves formed by tracing along the principal curvature direction of the surface from the curvature extremum point. These lines are used to mark the surface structural boundaries of the insulator and to divide different functional parts into geometric feature lines.

[0062] S5, based on the key feature lines and curved areas of the insulator, calculates the key geometric parameters, coating thickness and electrical performance parameters of the insulator, and performs comprehensive analysis to generate test conclusions.

[0063] In this embodiment, based on the key feature lines and curved surface regions of the insulator, the key geometric parameters, coating thickness, and electrical performance parameters of the insulator are calculated, and a comprehensive analysis is performed to generate detection conclusions, including: Based on the extracted key feature lines and curved surface regions, the geometric parameters of each key structural part of the insulator are calculated. The geometric parameters include creepage distance, overall length, cap diameter, thread height, and support rod spacing, specifically: Three-dimensional spatial measurements are performed on key feature lines, and the Euclidean distance between the start and end points is calculated as the creepage distance. For curved areas such as the cap body, threads, and support rods, boundary points are first extracted on the surface. The boundary can be identified using a surface convex hull algorithm. Then, the spatial bounding box dimensions of the region, including length, width, and height, are extracted based on the boundary points. At the same time, the curvature values ​​of all points in the region are statistically analyzed to calculate the average curvature, reflecting the morphological characteristics of the region. In addition, to ensure data consistency and reproducibility, all geometric parameters are calculated within a local coordinate reference frame obtained through PCA initialization. This allows for a unified comparison of measurement results from different insulators and provides an accurate spatial reference for subsequent coating thickness and electrical performance analysis.

[0064] Based on a defined curved surface area, thickness data is collected using a non-contact thickness measurement device (such as eddy current thickness measurement, laser interferometry, or optical reflection method), and the average thickness and standard deviation are calculated in each curved surface area to form coating thickness data, wherein the measurement points are evenly distributed on the curved surface and cover the high curvature area. Based on the key structural parts of the insulator, electrical performance data, including resistance value, leakage current and conductivity, are collected using an insulation resistance measuring device. The sampling frequency of the measurement voltage and current and the environmental conditions are all controlled according to preset standards to form a set of electrical performance parameters. By comprehensively analyzing geometric parameters, coating thickness data, and electrical performance parameters, and based on preset qualification standards and tolerance ranges (including upper and lower limits and deviation thresholds for each key parameter), the qualification of each key structural part is judged, and qualification marks for each parameter are generated. Based on the comprehensive analysis results, a test conclusion report is generated. The test conclusion report includes the geometric dimensions and deviations of each key structural part, the analysis results of the coating thickness, the electrical performance parameters, and whether they meet the standards.

[0065] It should be noted that by combining B-spline surface fitting with structured surface region extraction, high-precision joint measurement of key geometric parameters, coating thickness, and electrical performance of insulators is achieved. Based on multi-parameter comprehensive analysis, detection conclusions are automatically generated. This not only replaces traditional manual measurement and single two-dimensional image measurement methods, improving the accuracy and reliability of insulator detection for complex structures, but also reduces human interference and significantly improves detection efficiency. It provides a scientific and feasible technical means to prevent substandard or refurbished insulators from entering the power grid.

[0066] Reference Figure 2 As shown in the diagram, the present invention provides a vision-based automatic material sample processing system, which includes an image processing module, a PCA initialization module, a fitted surface generation module, a feature extraction module, and a parameter detection module. The modules are interconnected. The image processing module is used to acquire point cloud images of the insulator under test, perform high-precision registration processing on the point cloud images, segment and classify the processed point cloud images, and extract point cloud data sets. The PCA initialization module is used to perform principal component analysis on point cloud datasets, determine the main orientation and local coordinate reference of the point cloud, and generate an initial reference frame. The fitted surface generation module is used to construct a control mesh based on an initial reference frame, use the control mesh to perform B-spline smooth surface fitting on the point cloud, and adjust the position of the B-spline surface control points according to the fitting error feedback to generate an iteratively optimized fitted surface. The feature extraction module is used to extract key feature lines and surface regions of the insulator based on the optimized fitted surface. The parameter detection module is used to calculate the key geometric parameters, coating thickness, and electrical performance parameters of the insulator based on the key feature lines and curved areas of the insulator, and to perform comprehensive analysis to generate detection conclusions.

[0067] 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.

[0068] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product.

[0069] Those skilled in the art will recognize that the modules 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.

[0070] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0071] 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.

[0072] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A vision-based method for automated processing of material samples, characterized by, Includes the following steps: The point cloud image of the insulator to be tested is acquired, the point cloud image is registered with high precision, and the processed point cloud image is segmented and classified to extract the point cloud data set. Principal component analysis is performed on the point cloud dataset to determine the main orientation and local coordinate reference of the point cloud and generate an initial reference frame; Based on the initial reference frame, a control mesh is constructed, and the point cloud is fitted with a B-spline smooth surface using the control mesh. Based on the fitting error feedback, the positions of the control points of the B-spline surface are adjusted to generate the fitted surface after iterative optimization. Based on the optimized fitted surface, extract the key feature lines and surface regions of the insulator; Based on the key feature lines and curved areas of the insulator, the key geometric parameters, coating thickness and electrical performance parameters of the insulator are calculated, and a comprehensive analysis is performed to generate test conclusions.

2. The vision-based material coupon automated processing method of claim 1, wherein, The process of acquiring point cloud images of the insulator under test and performing high-precision registration processing on the point cloud images includes: The point cloud image is preprocessed to generate a preliminary processed point cloud image, and then denoised to obtain a denoised point cloud image. Calculate a fast point feature histogram for the denoised point cloud image and generate a feature description set; Perform coarse registration on the feature description set, and match the fast point feature histograms in the point cloud from different viewpoints to determine candidate matching point pairs; Consistency screening is performed on candidate matching point pairs to obtain valid matching point pairs; Based on the effective matching point pairs, the initial spatial transformation relationship is calculated, and the point cloud is transformed according to the initial spatial transformation relationship to obtain the coarse registration result; Based on the coarse registration results, an improved iterative nearest point method is used for fine registration to obtain a high-precision registered insulator point cloud image.

3. The vision-based material coupon automated processing method of claim 2, wherein, The process involves fine registration based on the coarse registration results, using an improved iterative nearest-point method to obtain a high-precision registered insulator point cloud image, including: The point cloud after coarse registration is used as the point cloud to be registered, and the target point cloud is used as the reference point cloud. In the point cloud to be registered, for each point, the nearest point is searched only within the local neighborhood of its corresponding target point cloud to form a point pair correspondence; For the initial point pair correspondence, obtain the change in spatial distance between corresponding point pairs and the change in the angle between the normal vectors of corresponding points, and obtain the valid point pairs through consistency judgment; Based on valid point pairs, calculate the incremental rotation matrix and incremental translation vector for the current iteration round; The incremental rotation matrix and incremental translation vector are superimposed on the current spatial pose of the point cloud to be registered for updating; If the preset conditions are met, the iteration stops and a high-precision registered insulator point cloud image is output.

4. The vision-based automatic material sample processing method according to claim 1, characterized in that, The step of segmenting and classifying the processed point cloud image and extracting the point cloud data set includes: Based on the average distance between adjacent points in the insulator point cloud image after fine registration, the distance threshold for Euclidean clustering is determined, and the point cloud image is clustered and segmented based on the Euclidean distance between two points to obtain several spatially independent point cloud subsets. Based on the point cloud subsets, obtain the number of points and the spatial bounding size of each subset, and determine the candidate point cloud subsets; For each candidate point cloud subset, perform point cloud subset preprocessing based on centroid alignment and scale normalization; For the processed subset of candidate point clouds, construct the input point cloud data; The input point cloud data is fed into a pre-trained PointNet neural network, which outputs the structure category of each subset of the insulator. Based on the structure category determination results, the point cloud subsets belonging to the same structure category are aggregated to generate a structured point cloud data set.

5. The vision-based automatic material sample processing method according to claim 1, characterized in that, The process of performing principal component analysis on the point cloud dataset to determine the main orientation and local coordinate reference of the point cloud and generate an initial reference frame includes: Calculate the covariance matrix of the three-dimensional coordinates based on the point cloud dataset; Eigenvalue decomposition is performed on the covariance matrix to obtain directional eigenvectors, which include principal directional eigenvectors, secondary directional eigenvectors, and plane normal directional eigenvectors. Using the directional feature vector as the coordinate axis direction, a local coordinate reference frame for the point cloud is established, forming an initial reference frame.

6. The vision-based automatic material sample processing method according to claim 1, characterized in that, The process of constructing a control mesh based on an initial reference frame and then using the control mesh to perform B-spline smoothing surface fitting on the point cloud includes: For each point in the point cloud dataset initialized by principal component analysis, obtain the number of its neighboring points within a preset neighborhood radius to form a local density distribution; Based on the overall average density and standard deviation of the point cloud, thresholds for high-density and low-density regions are set, and the point cloud is divided into high-density and low-density regions. An initial rule control mesh is generated within the spatial bounding area of ​​the point cloud, and the mesh is mapped to the initial reference frame of the point cloud; Based on the local density distribution, the initial control grid is adaptively adjusted, with preset proportional control points added to high-density areas and preset proportional control points reduced to low-density areas. The point cloud is fitted with a B-spline smooth surface using the adjusted control grid to generate an initial fitted surface.

7. The vision-based automatic material sample processing method according to claim 1, characterized in that, The step of adjusting the positions of the B-spline surface control points based on the fitting error feedback to generate an iteratively optimized fitting surface includes: Based on the initial fitted surface, the minimum Euclidean distance from each point in the point cloud dataset to the initial fitted surface is calculated as the point-to-surface fitting error value, and an error distribution set is formed. Statistical analysis is performed based on the error distribution set to calculate the overall average error and identify high error regions; The high-error region is mapped to the corresponding B-spline surface control point mesh in the parameter space, and the set of control points associated with the high-error region is determined. For a set of control points, an error weight value is assigned to each control point based on the magnitude of the error within the corresponding high-error region. The position of the control points is updated along the normal direction of the point cloud based on the error weight value of the control points; Based on the updated control point positions, a second B-spline smooth surface fitting is performed, followed by iterative optimization to output the iteratively optimized fitted surface that meets the preset conditions.

8. The vision-based automatic material sample processing method according to claim 1, characterized in that, The step of extracting key feature lines and surface regions of the insulator based on the optimized fitted surface includes: Based on the optimized fitted surface, its parametric expression is obtained and mapped to a two-dimensional parameter space; Within the parameter space, the local geometric features corresponding to each position on the surface, including the normal vector, principal curvature, and rate of change of curvature, are obtained. Based on the statistical distribution of the principal curvature and the rate of change of curvature of the surface, regions with significant local extrema are identified, and these regions are determined as candidate feature regions. Within the candidate feature region, several feature line seed points are generated based on the curvature extrema points; Starting from the seed point of the feature line, the key feature lines of the insulator are generated by continuously tracing along the principal curvature direction of the surface. Using key feature lines as the surface segmentation boundary, the fitted surface is divided into multiple independent surface regions; Extract the structural parameters of each surface region and annotate its structural attributes to form a set of structured surface regions; Set of key feature lines and curved surface regions of output insulators.

9. The vision-based automatic material sample processing method according to claim 1, characterized in that, Based on the key feature lines and curved surface regions of the insulator, the key geometric parameters, coating thickness, and electrical performance parameters of the insulator are calculated, and a comprehensive analysis is performed to generate test conclusions, including: Based on the extracted key feature lines and curved surface regions, the geometric parameters of each key structural part of the insulator are calculated; Within a defined curved surface region, the average thickness and standard deviation are calculated to generate coating thickness data. Based on the key structural parts of the insulator, electrical performance data are collected to form a set of electrical performance parameters. By comprehensively analyzing geometric parameters, coating thickness data, and electrical performance parameters, the passability of each key structural component is judged, and passability marks for each parameter are generated. Based on the comprehensive analysis results, a test conclusion report is generated.

10. A system using the vision-based automated material sample processing method as described in any one of claims 1-9, characterized in that, include: The image processing module is used to acquire point cloud images of the insulator under test, perform high-precision registration processing on the point cloud images, segment and classify the processed point cloud images, and extract point cloud data sets. The PCA initialization module is used to perform principal component analysis on point cloud datasets, determine the main orientation and local coordinate reference of the point cloud, and generate an initial reference frame. The fitted surface generation module is used to construct a control mesh based on an initial reference frame, use the control mesh to perform B-spline smooth surface fitting on the point cloud, and adjust the position of the B-spline surface control points according to the fitting error feedback to generate an iteratively optimized fitted surface. The feature extraction module is used to extract key feature lines and surface regions of the insulator based on the optimized fitted surface. The parameter detection module is used to calculate the key geometric parameters, coating thickness, and electrical performance parameters of the insulator based on the key feature lines and curved areas of the insulator, and to perform comprehensive analysis to generate detection conclusions.

Images