A three-dimensional imaging detection method for surface defects in connectors

Abstract

This application relates to the field of three-dimensional imaging technology, and in particular to a three-dimensional imaging detection method for surface defects in connectors. The proposed method involves: acquiring point cloud images of the connector; extracting suspected perturbation regions using morphological operators to generate target points; constructing an axisymmetric mapping relationship for the target points to generate a symmetrical residual distribution map; based on the perturbation spatial gradient field, eliminating perturbation convergence points and retaining perturbation divergence points; and further extracting normal abrupt changes and curvature minima using a defect identification network to output a set of three-dimensional coordinates of the defect region. This method effectively improves the accuracy of artifact removal and defect identification, and is suitable for robust three-dimensional defect detection tasks under complex internal wall structures.

Description

Technical Field

[0001] This application relates to the field of three-dimensional imaging technology, and in particular to a three-dimensional imaging detection method for surface defects of connectors. Background Technology

[0002] With the widespread application of miniature connectors in consumer electronics, automotive electronics, and medical devices, the manufacturing quality of their surface structure has become an increasingly critical factor affecting the reliability and safety of the entire device. Minor defects on the connector surface, such as scratches, dents, plating peeling, and bubbles, can lead to poor contact, electrical faults, or even mechanical failures during subsequent mating, connection, or sealing processes. Therefore, how to perform high-precision and robust detection of connector surface defects has become an urgent technical problem to be solved.

[0003] For example, Chinese patent application CN115953356A discloses a defect identification method for battery swapping connectors, including the following steps: Step 1: Defect localization, using an image acquisition system to collect and analyze the connector and locate the defect; Step 2: Defect classification, using DCNN to classify defects through convolutional layers; Step 3: Defect analysis, using CNN to extract defect features and analyze sample defects. This application method is used to detect whether there are abnormalities and defects in vehicle sockets and battery pack plugs before and after connection. It performs image-based automatic detection on specific connector sockets and plugs. The detection objects support user template addition and learning capabilities to continuously improve the system's adaptability, promptly detect abnormalities and defects, greatly improve the timeliness and efficiency of maintenance, and reduce users' economic losses.

[0004] For example, Chinese Patent CN113808131B discloses a connector defect identification method, system, device, and medium, relating to the field of smart industry. The method includes: obtaining a first image of a connector to be inspected; extracting a first region of interest (ROI) corresponding to the shell of the connector to be inspected from the first image; performing threshold segmentation on the first ROI to obtain a first binarized image; performing contour detection on the first binarized image to obtain a first contour detection result; calculating the contour perimeter and contour area based on the first contour detection result; and determining that the shell of the connector to be inspected has a detachment defect if the contour perimeter is greater than a perimeter threshold and the contour area is greater than an area threshold, thereby achieving automatic detection of connector detachment defects.

[0005] The above-mentioned existing technologies all have the problems mentioned in the background: since the cavity structure of the connector is usually mirror symmetric, it is easily affected by multipath reflection when collecting point clouds inside the structure, resulting in a large number of mirror artifacts. These artifacts are very similar in shape to real defects, which can easily lead to misjudgment. In order to solve the above problems, this application designs a three-dimensional imaging detection method for surface defects of connectors. Summary of the Invention

[0006] The technical problem to be solved by this application is to address the shortcomings of the existing technology by providing a three-dimensional imaging detection method for surface defects of connectors. The method involves acquiring point cloud images of the connector; extracting suspected disturbance regions using morphological operators to generate target points; constructing an axisymmetric mapping relationship of the target points to generate a symmetrical residual distribution map; eliminating disturbance convergence points and retaining disturbance divergence points based on the disturbance spatial gradient field; and further extracting normal abrupt change relationships and curvature minimum regions through a defect identification network to output a set of three-dimensional coordinates of the defect region.

[0007] To achieve the above objectives, this application provides the following technical solution:

[0008] A three-dimensional imaging detection method for surface defects in connectors, the method comprising:

[0009] Based on the collected point cloud data, a surface map in the depth direction is obtained;

[0010] Using the central axis of the surface plot as the axis of symmetry, target points are identified by morphological operators, and the spatial residual between each target point and its axisymmetric position point is calculated to obtain a symmetric residual distribution map.

[0011] In the symmetrical residual distribution map, an artifact point removal strategy is executed to obtain a connector point cloud image. The artifact point removal strategy determines whether there are artifact points based on the perturbation trend of the target points and removes the target points whose perturbation trend is convergent.

[0012] The connector point cloud image is input into a preset defect recognition network. The defect recognition network processes the connector point cloud image and outputs the defect region. The defect region is then used to create a three-dimensional image. The defect recognition network is trained using historical defect data of the connector.

[0013] Using the central axis of the surface plot as the axis of symmetry, the target point is identified by morphological operators, including:

[0014] Based on the distribution of point cloud in the axis of symmetry and surface diagram, determine the normal projection direction of point cloud, and calculate the projection signal curve based on the normal projection direction and point cloud spatial density.

[0015] Perform morphological opening and closing operations on the projection signal curve to obtain a morphological reference curve;

[0016] The residuals of the projected signal curve and the morphological reference curve are compared, and point clouds with residual amplitudes greater than the disturbance judgment threshold are selected to generate target points.

[0017] Perform morphological opening and closing operations on the projected signal curve to obtain a morphological reference curve, including:

[0018] Based on the shape of the surface plot and the density change of the point cloud in the normal projection direction, a nonlinear structural element is generated, wherein the nonlinear structural element is calculated based on the periodic characteristics of the density change.

[0019] Using the nonlinear structural element as a sliding window, erosion and dilation operations are performed on the projected signal curve, wherein the erosion operation is processed based on the local minimum value within the sliding window, and the dilation operation is processed based on the local maximum value within the sliding window.

[0020] The results of the erosion and dilation operations are fused into a smooth envelope curve, which serves as the morphological reference curve for the projected signal curve.

[0021] Calculate the spatial residual between each target point and its axisymmetric location, including:

[0022] Obtain the axisymmetric mapping point for each target point;

[0023] Calculate the spatial residual, normal residual, and curvature residual between the target point and the axisymmetric mapping point.

[0024] In the symmetrical residual distribution map, an artifact point removal strategy is performed to obtain a connector point cloud image, including:

[0025] Based on the symmetric residual distribution map, a perturbation spatial gradient field is constructed, and the perturbation intensity spatial gradient of each target point in the map is calculated.

[0026] The target point is divided into a disturbance convergence point and a disturbance divergence point according to the gradient direction of the spatial gradient of the disturbance intensity.

[0027] The disturbance convergence points are removed, and a connector point cloud image is generated based on the point cloud corresponding to the remaining disturbance divergence points.

[0028] The construction of the surface plot includes:

[0029] The point cloud data is projected into a polar coordinate space with the center of the connector as the reference axis;

[0030] Based on the projected point cloud data, multiple equidistant sub-regions are divided along the polar radius direction, and the point set that forms local extreme point density in the axial direction of each sub-region is extracted.

[0031] Using the density peak of the point set as the reference layer boundary, a fitted surface with normal continuity constraints is constructed based on the normal variation rate of the point set's neighborhood to generate a surface plot.

[0032] The defect identification network includes a point set geometric coding layer, a spatial perturbation extraction layer, and a defect determination layer, wherein:

[0033] The point set geometric coding layer is used to receive the connector point cloud image and construct a geometric feature vector based on the spatial coordinates, normal vector, and local curvature of each point cloud.

[0034] The spatial perturbation extraction layer is equipped with a graph convolutional network, which is used to calculate the normal abrupt change relationship and the local curvature minimum region in the spatial adjacency graph constructed by the geometric feature vectors;

[0035] The defect determination layer is used to output the location label of the defect region and its corresponding three-dimensional coordinate set based on the normal abrupt change relationship and the local curvature minimum region.

[0036] The spatial perturbation extraction layer includes an adjacency graph construction module, a graph convolution calculation module, and a differentiation module, wherein:

[0037] The adjacency graph construction module is used to generate a spatial adjacency graph with Euclidean distance as the edge weight based on the spatial coordinates and geometric feature vector of each point.

[0038] The graph convolution calculation module uses a shared-weight graph convolutional network to propagate and aggregate information on nodes in the spatial adjacency graph, and extracts the normal rate of change and curvature gradient of each point in its neighborhood.

[0039] The differentiation module determines whether the normal change rate of each point is greater than the weighted standard deviation of the mean normal change rate of the neighborhood based on the normal change rate and curvature gradient of each point in its neighborhood, so as to identify the abrupt change relationship of the normal. Under the condition that the curvature gradient is less than a preset first threshold and the curvature value is a local curvature minimum value in its neighborhood, the module identifies the local curvature minimum value region.

[0040] The defect determination layer includes:

[0041] Perform local connectivity analysis on the aforementioned abrupt change in normal direction and curvature minimum region to generate a candidate defect point set;

[0042] The candidate defect point set is clustered and grouped according to a clustering algorithm to obtain the clustering results;

[0043] Based on the spatial distribution density, curvature gradient variation range, and normal consistency of the point cloud in the clustering results, calculate the defect confidence of the point set region corresponding to the clustering results.

[0044] The point set region with a defect confidence level greater than the defect threshold is output as the defect region.

[0045] Performing three-dimensional imaging of the defective region includes:

[0046] Extract the set of three-dimensional point cloud coordinates corresponding to the defect area, and construct a spatial voxel mesh based on the set of three-dimensional point cloud coordinates;

[0047] Interpolation fitting is performed on the spatial voxel mesh to generate a three-dimensional image of the defect.

[0048] Compared with the prior art, the beneficial effects of this application are:

[0049] This application achieves efficient pre-screening of suspected defect regions by constructing a projection signal curve dominated by normal perturbation and introducing morphological operators to extract abnormal perturbation points; further, it constructs a symmetric residual distribution map through axisymmetric mapping and combines it with perturbation gradient field directionality analysis to accurately remove mirror artifact points; on this basis, it combines spatial adjacency graph and graph structure convolutional units to perform multidimensional geometric analysis on candidate regions, and finally achieves accurate identification of real defect regions. Attached Figure Description

[0050] Other features, objects, and advantages of this application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0051] Figure 1 This is a schematic diagram illustrating an exemplary application scenario of an embodiment of this application;

[0052] Figure 2 This is a schematic diagram of artifact imaging in an embodiment of this application;

[0053] Figure 3 This is a flowchart illustrating a three-dimensional imaging detection method for surface defects in a connector, according to an embodiment of this application.

[0054] Figure 4 This is a schematic diagram of the target point identification process in an embodiment of this application;

[0055] Figure 5 This is a schematic diagram illustrating the principle of generating morphological reference curves in an embodiment of this application.

[0056] Figure 6 This is a schematic diagram illustrating the principle of generating axisymmetric mapping points in an embodiment of this application. Detailed Implementation

[0057] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments.

[0058] The term "embodiment" as used herein means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0059] This application applies to connector products with cavity-type mirror symmetry structure and strong surface reflectivity. The three-dimensional defect detection process is limited by the external viewing angle acquisition method of the imaging equipment. Artifacts caused by structural reflections interfere near the defect area. Application scenarios include, but are not limited to:

[0060] High-speed signal connectors featuring multiple rows of metal pins and miniature deep cavity ports;

[0061] The pins inside the circular connector are arranged symmetrically around the circular axis, which can easily cause secondary or multiple reflection interference in areas with high reflection angles during imaging.

[0062] The selection of application scenarios is based on the common characteristics of connectors. The selected application scenarios should have at least one of the following characteristics:

[0063] The defect is located in imaging blind zones such as deep cavities, slits, or inverted cones.

[0064] The surface has specular reflection properties, and multipath reflection easily forms symmetrical artifacts;

[0065] Due to the limited external acquisition angle, the point cloud distribution exhibits a unilateral tilt or sparse state.

[0066] The defect is a non-structural geometrical abrupt change, manifesting only as a very small perturbation.

[0067] It should be noted that the three-dimensional imaging detection method for connector surface defects proposed in this application does not rely on the explicit structural partitioning or fixed cavity division of the target connector as a prerequisite, but is oriented towards imaging scenarios with typical specular reflection path constraints and viewing angle occlusion limitations.

[0068] It is worth noting that the artifact removal strategy described in this application is not designed for a specific structural module or connector series, but can also be applied to:

[0069] In engineering imaging cabling scenarios where there is a possibility of multiple optical path reversals;

[0070] Scenarios where default projection or receive path dependencies exist within the structural space;

[0071] The same defect behavior can be mirrored or perturbed in the model space of the point cloud.

[0072] Please see Figure 1 This figure is a schematic diagram of an exemplary application scenario provided by an embodiment of this application.

[0073] like Figure 1 As shown, the method of this application is applicable to defect detection of pin holes in connectors. Connectors typically include a housing body structure and multiple pin hole array units. The pin hole array units are evenly arranged around the central axis of symmetry and present a deep cavity columnar shape, which is a typical axisymmetric mirror structure.

[0074] Figure 1 The imaging system shown includes three parts: a signal source, a connector, and an imaging device. In this application, the signal source can be a structured light projection device, a laser scanner, or a controllable light source with modulation and coding capabilities, which projects onto the connector structure to be inspected.

[0075] Figure 1 The diagram illustrates how, during the detection process, a signal emitted by a signal source illuminates the inner wall of the connector. After surface reflection, the reflected signal image is acquired by an imaging device, which then generates a three-dimensional point cloud image of the connector.

[0076] It is understood that the imaging device mentioned in this application refers to a depth imaging system with the ability to acquire three-dimensional structural information, used to extract the spatial location point set of the connector surface from structured light or laser reflection maps.

[0077] Furthermore, the imaging device in this application also integrates or connects a processor for performing logic for point cloud data preprocessing, symmetric residual distribution map construction, perturbation behavior analysis, and defect determination. The processor can be a computing unit with point cloud processing capabilities, such as a central processing unit, graphics processing unit, edge AI chip, FPGA, or DSP.

[0078] Please see Figure 2 This figure is a schematic diagram of artifact imaging provided in an embodiment of this application.

[0079] Figure 2 The diagram illustrates a signal source projecting a signal into the pinhole area inside a connector. After specular reflection on the inner wall of the pinhole, the reflected signal is received by an imaging device and used to generate 3D point cloud data. Ideally, the imaging device should acquire real points representing the actual surface of the structure to accurately reconstruct the geometry of the connector's inner wall.

[0080] It is easy to understand that, since the connector is usually axially symmetrical and the inner wall is mostly made of metal or electroplated material, it has a strong specular reflection capability. The signal light may be reflected once or multiple times between the two sides of the pinhole and is eventually received by the imaging device. This results in some reflected signals not coming directly from the actual surface being measured, but from non-target positions on the optical path reflection path inside the structure.

[0081] Figure 2 The image shows that these indirectly reflected imaging signals are misinterpreted by the system as point cloud information of the actual surface. In the point cloud image, they appear as a set of artifact points. Especially in symmetrical structures, these artifact points often appear in a mirror image around the central axis of the connector and have a symmetrical mapping relationship with the real structural points in space, forming artifact interference.

[0082] Those skilled in the art will know that artifacts are highly similar to real points in spatial distribution, but their positions have no actual physical meaning, severely interfering with the stability and accuracy of defect identification. This application addresses this technical problem by proposing related logic for perturbation residual analysis and spatial gradient discrimination to remove artifacts caused by multipath reflections from the point cloud, thereby improving the defect detection quality of 3D imaging in the inner wall structure of connectors.

[0083] Next, with reference to the accompanying drawings, a three-dimensional imaging detection method for surface defects of a connector provided in an embodiment of this application will be described. Figure 3 The method shown includes the following steps S1-S4:

[0084] S1: Based on the collected point cloud data, obtain the surface map in the depth direction;

[0085] In this application, point cloud data can originate from 3D imaging devices based on imaging mechanisms such as structured light projection, laser triangulation, time-of-flight depth sensing, and multi-view stereo matching. Specific devices may include industrial-grade structured light 3D scanners, laser profilometers, ToF depth cameras, or imaging devices with active illumination and spatial reconstruction capabilities. Point cloud data is typically constructed based on the 3D coordinates of each spatial point, and some detection methods can further provide additional attributes such as normal, reflectivity, or grayscale intensity. The point cloud acquisition process can combine different viewpoints, different exposure strategies, or active focusing to optimize data density and accuracy. Those skilled in the art will understand that the point cloud data acquisition method can be flexibly selected based on the specific detection environment, connector material, and hole structure characteristics, and is not limited herein.

[0086] Furthermore, to enhance the directionality and distribution regularity of structural analysis, the point cloud data is projected onto a polar coordinate space with the connector center as the reference, and a surface plot representing the depth perturbation trend is generated based on the point cloud normal direction.

[0087] S2: Using the central axis of the surface plot as the axis of symmetry, the target point is identified by morphological operators, and the spatial residual between each target point and its axisymmetric position point is calculated to obtain the symmetric residual distribution map.

[0088] In this embodiment, based on the surface map generated by S1, the connector's central axis is set as the symmetry reference axis. Projection density analysis is performed along the point cloud normal direction to extract the projection signal curve, and opening and closing operations are performed to obtain the morphological reference curve. Further, the residual segment between the original projection signal curve and the morphological reference curve is calculated to identify local target points with significant curvature perturbation amplitudes. Subsequently, for each target point, its mirror mapping position about the symmetry axis is constructed, and the spatial coordinate residual is calculated, thereby forming a symmetric residual distribution map containing all target point pairs.

[0089] S3: In the symmetrical residual distribution map, perform the artifact point removal strategy to obtain the connector point cloud image;

[0090] In this embodiment, for the symmetric residual distribution map generated by S2, the perturbation intensity gradient direction of each target point in the local space is further calculated. Based on the local distribution trend of the perturbation value, it is determined whether the point is a perturbation convergence point or a perturbation divergence point; and further combined with the spatial distribution relationship between symmetric point pairs, the angle between the gradient directions and the consistency of the perturbation direction for joint analysis. When a symmetric point pair is simultaneously a perturbation convergence point and has gradient convergence behavior with consistent symmetric directions, it is marked as a high-confidence artifact point pair and overall removal is performed.

[0091] It is understandable that the processing logic corresponding to this step is based on the multipath reflection physical mechanism disclosed in this application. By structurally modeling the perturbation behavior characteristics, it avoids the problem of false deletion of real defects caused by the coarse threshold judgment in traditional methods, and is adapted to typical connector configurations with central axis structure and complex reflection interference paths.

[0092] S4: Input the connector point cloud image into a preset defect recognition network, process the connector point cloud image through the defect recognition network, output the defect area, and perform three-dimensional imaging of the defect area;

[0093] In this embodiment, a point cloud image with artifacts removed is used as network input, and a pre-trained defect recognition neural network model is invoked. The model includes a point set geometric encoding layer, a spatial perturbation extraction layer, and a defect determination layer. Specifically, a graph-structured convolution mechanism is used to extract the normal abrupt change relationship and curvature perturbation trend between point clouds, identifying the true defect region and further outputting the precise location result of the defect region in three-dimensional space.

[0094] Before delving into the specific technical details of the steps, this application's embodiments need to reiterate that in typical connector pinhole structures, due to the limited space and regular structure, the inner walls are often made of metal or electroplated materials, possessing strong specular reflection characteristics. When using structured light or other methods for 3D imaging acquisition, the light signal may undergo one or more reflections within the cavity, causing non-real surface locations to be identified by the receiver as valid signal points on the reflection path, ultimately forming regularly distributed false structural information in the point cloud image. Unlike actual defects, these points often exhibit a symmetrical mapping relationship around the central axis with a real point or secondary refraction point, and their geometric perturbation trend is similar to that of real defects. If only conventional symmetry identification or error filtering is relied upon, artifact points are often mistakenly retained as defect points due to overlapping perturbation amplitudes or similar textures, or real defects are also rejected.

[0095] The processing logic employed in this embodiment does not focus on improving defect identification itself, but rather on calculating the changing trend of spatial perturbation behavior. Specifically, it uses the gradient directionality of perturbation values ​​as an additional identification criterion, thus adding a discriminative clue beyond the defect morphology. Unlike traditional judgment criteria, this application determines whether the perturbation aggregates or disperses in space. In engineering practice, this can be viewed as whether the interference behavior undergoes path folding or projection overlap due to the mirror structure. Furthermore, when the gradient directions of multiple target points are found to be symmetrical and pointing towards each other, it can be logically deduced that this is the result of mirror interference imaging, and high-confidence rejection can be performed accordingly. In comparison, the method of this application has higher stability and adaptability for imaging artifacts in pinhole and blind cavity connector structures.

[0096] Next, we will further elaborate on the part of the method of this application regarding curved surface images.

[0097] In deep cavity structures such as connector pinholes, the point cloud data acquired during acquisition exhibits spatial non-uniformity due to strong reflections from the inner walls, limited incident angles, and the fact that imaging equipment is typically located externally. Particularly in the depth direction, the number of point clouds often decreases exponentially, with only a small number of points concentrated within a narrow effective layer, resulting in a discontinuous and irregular sheet-like structure.

[0098] Understandably, under the influence of this sheet-like structure, if patch fitting is performed using a conventional Cartesian coordinate system, geometric distortion may occur due to insufficient support points in the fitting region or strong edge jumps. In this embodiment, the original point cloud data is projected onto a two-dimensional plane corresponding to the normal direction, and a fitting method in a polar coordinate system is used to interpolate and smooth the local high-density point areas, thereby more stably restoring the true curved surface structure of the pinhole inner wall and the specific location of the point cloud.

[0099] In one example, the specific steps for constructing the surface plot of S1 are as follows:

[0100] S1.1: Project the point cloud data into a polar coordinate space with the connector center as the reference axis;

[0101] In this embodiment, the connector center axis can be obtained by extracting the overall point cloud axially using principal component analysis. Specifically, the covariance matrix is ​​calculated in the original three-dimensional point set, the first principal direction is extracted as the initial axis direction, and then outlier interference points are removed by the RANSAC algorithm to further accurately fit the structural axis of the connector pinhole.

[0102] Furthermore, the polar radius, axial direction, and polar angle of all points are calculated using the structural axis as the polar axis, thereby completing the transformation to polar coordinate space. This will not be elaborated upon here.

[0103] S1.2: Based on the projected point cloud data, divide the area into multiple equidistant sub-regions along the polar radius direction, and extract the point set that forms a local extreme point density in the axial direction in each sub-region;

[0104] Specifically, after projection onto the polar coordinate system, the distribution of the point cloud in each polar radius direction reflects the structural morphology of the connector's inner wall and deep cavity. Effective imaging points are typically concentrated within a certain axial height range. In this step, the polar radius direction is equidistantly divided into several sub-regions. Each sub-region is a fan-shaped annular flat strip region, and its radial width can be adaptively set according to the total polar radius range to ensure resolution capability in connectors of different sizes.

[0105] In this embodiment, to eliminate isolated high-density errors caused by a single strong reflection, the point cloud in each sub-region is statistically analyzed along the axial direction, and a sliding window extreme value extraction method is used to identify the axial location point set where the main density peak is located. During the density statistics process, the minimum distance between local point clouds is considered to avoid high-frequency fluctuations caused by angular occlusion or projection errors. Simultaneously, a kernel-based smoothing filter can be superimposed, ensuring that the extracted extreme value point set not only represents the real physical structure surface but also possesses a certain degree of spatial continuity. These extracted point sets constitute the candidate data for surface plot fitting.

[0106] S1.3: Using the density peak of the point set as the reference layer boundary, construct a fitting surface with normal continuity constraints based on the normal variation rate of the neighborhood of the point set to generate a surface plot;

[0107] Specifically, in pinhole-like structures, the effective surface typically exhibits a dense and reliable data distribution only within certain spatial regions. Therefore, this application selects only the identified set of extreme points as the core region for geometric reconstruction. Based on the core region, the normal estimation information of each point is further utilized to analyze the rate of change of the normal in its local neighborhood, thereby determining whether the region possesses continuous and smooth morphological characteristics.

[0108] In this embodiment, the normal vector is obtained using a PCA-based local neighborhood fitting method. The neighborhood radius or the number of K neighboring points can be dynamically adjusted to adapt to the point density changes in different polar diameter regions.

[0109] Furthermore, this application introduces a normal continuity constraint during the process of generating the fitted surface. Specifically, this can be achieved by jointly evaluating the variation range of the fitting residual and the normal angle at each point, thereby optimizing the local interpolation model of the surface.

[0110] Please see Figure 4 , Figure 4 This is a schematic diagram of the target point identification process in an embodiment of this application. It can be understood that this application is aimed at the three-dimensional imaging of defective parts rather than the overall three-dimensional imaging of the connector. There are two solutions to achieve this goal. The first solution is to image the collected point cloud data as a whole and then locate the defective area from it. The second solution is to preprocess the point cloud and filter out the point clouds with abnormal shapes separately as point clouds that may have defective areas.

[0111] Understandably, in typical micro-connectors, pinhole interfaces, and housing internal structures, the compact design and complex reflection paths make overall point cloud imaging highly susceptible to noise interference, redundancy artifacts, and optical distortion. Performing symmetry analysis and 3D reconstruction on the entire point cloud indiscriminately would significantly increase computational costs and lead to a higher misclassification rate.

[0112] Specifically, Figure 4 The method shown is based on the normal perturbation and spatial density change of the point cloud in the surface image. It pre-screens the regions with abnormal perturbations in the local structure, focuses on the selective modeling of the real defect region from the overall point cloud, reduces the invalid computational overhead of the complete data, and improves the reliability of identifying local defects in complex structures and the imaging resolution.

[0113] Next, we will further elaborate on the target point identification aspect of the method in this application.

[0114] Figure 4 The method shown can be applied to S2 above, and the specific steps are as follows:

[0115] S2.1: Based on the distribution of the point cloud in the symmetry axis and surface diagram, determine the normal projection direction of the point cloud, and calculate the projection signal curve based on the normal projection direction and the spatial density of the point cloud;

[0116] Specifically, in the surface map constructed from point clouds, the point clouds are distributed on the three-dimensional surface of the inner wall of the connector cavity. The entire surface is constructed with the central axis of the socket as the axis of symmetry. Therefore, the projection direction can be established through the radial or normal direction defined by this axis to capture the shape perturbation distribution of the surface in this direction.

[0117] In this embodiment, the structural center axis of the connector is first obtained by extracting dense point cloud regions from the surface image and using least-squares fitting. Then, the normal direction with the structural center axis as a reference is defined as the principal projection direction, and all point clouds are mapped one-dimensionally along this direction. During the one-dimensional mapping process, the local spatial density of the point cloud is introduced as a weighting parameter to enhance the resolution of sparse regions, resulting in a representative projection signal curve.

[0118] S2.2: Perform morphological opening and closing operations on the projection signal curve to obtain a morphological reference curve;

[0119] Specifically, morphological operations, as a spatial structure recognition tool, are used in this step to construct a reference signal that reflects the ideal state of the surface.

[0120] Understandably, the signals from defective regions in the projected signal curve of a surface plot are often weak and have small fluctuations, making it difficult to directly identify the perturbation location corresponding to the defect through the projected signal curve. Therefore, morphological operators are used to eliminate unstructured perturbations, thereby obtaining a morphological reference curve that can be used as a comparison benchmark.

[0121] For example, you can refer to the curve. Figure 5 To understand, Figure 5 This is a schematic diagram illustrating the principle of generating morphological reference curves in an embodiment of this application. Figure 5 The projection signal curve of a certain segment of a point cloud is shown.

[0122] Figure 5 The diagram shows a projection signal curve based on point cloud data in the normal direction, reflecting the geometric perturbation of the connector surface in a certain region. Several tiny local fluctuations can be observed in the curve, which may correspond to minor defects, material undulations, or interference points caused by specular reflection in the actual structure.

[0123] Figure 5A set of nonlinear structural elements is shown to adaptively match the periodic and radial geometric variations exhibited in the connector cavity structure. Based on these nonlinear structural elements, erosion and dilation operations are performed on the projected signal curve.

[0124] Figure 5 The location where the corrosion operation is applied is shown, i.e., the minimum value is selected within the sliding window of the structural element, so that the local high point at that location is flattened.

[0125] Figure 5 It also shows the location of the expansion operation, which is to take the maximum value within the window to fill the concave segment.

[0126] Figure 5 The diagram further illustrates that after the above-described erosion and expansion operations, a morphological reference curve, as shown by the dashed line, can be formed to reflect the ideal structural trend of the point cloud.

[0127] Understandable Figure 5 The image shown is merely an exemplary fragment used to illustrate the application process of morphological processing methods in projected signal curves, and does not limit the specific shape and dimension of the surface map or point cloud data in this application.

[0128] In one example, the steps for generating the morphological reference curve are as follows:

[0129] S2.2.1: Based on the shape of the surface plot and the density change of the point cloud in the normal projection direction, generate nonlinear structural elements, wherein the nonlinear structural elements are calculated according to the periodic characteristics of the density change;

[0130] Specifically, point clouds often exhibit repetitive peak and trough characteristics within the projection domain, but different regions may show periodic fine-tuning or morphological anomalies. Using linear structure elements would not be able to accommodate these changing characteristics. Therefore, in this embodiment, a periodic template signal is first extracted based on the axial projection density function of the connector. The template signal is based on the peak position of the point cloud distribution in the projection direction, combined with the calculation result of the local average period width, to construct a set of structure elements with a non-uniform width distribution.

[0131] Preferably, the shape of the structural element during the sliding process can be automatically adjusted according to the changes in the shape of the local curved surface. Its function is to perform adaptive feature extraction for the changes in the radius of curvature and arrangement period of the connector surface at different positions, so as to improve the adaptability of corrosion and expansion operations to morphological changes.

[0132] S2.2.2: Using the nonlinear structural element as a sliding window, perform erosion and dilation operations on the projected signal curve, wherein the erosion operation is processed based on the local minimum value within the sliding window, and the dilation operation is processed based on the local maximum value within the sliding window;

[0133] Specifically, the nonlinear structuring element slides along the projected signal curve with a fixed sliding step size, performing morphological operations on the signal data within each window region. The erosion operation searches for the local minimum value of the current segment within each sliding window and replaces the window center value with this minimum value, thereby weakening the micro-protrusion structure globally. Correspondingly, the dilation operation selects the local maximum value within each window for replacement at the center position, thereby effectively filling the micro-depression region.

[0134] S2.2.3: The results of the erosion and dilation operations are fused into a smooth envelope curve, which serves as the morphological reference curve for the projection signal curve;

[0135] Specifically, after the morphological operation is completed, the erosion curve and the expansion curve are weighted and fused in the embodiment. The weight can be adjusted based on the directionality and rate of change of the curve fluctuation, and finally an envelope curve that is highly consistent with the original projection signal in terms of trend but whose local perturbation components are smoothed is constructed.

[0136] S2.3: Compare the residuals of the projection signal curve and the shape reference curve, and filter out point clouds with residual amplitudes greater than the disturbance judgment threshold to generate target points;

[0137] Next, we will further elaborate on the part of the method for target point selection in this application.

[0138] S2.4: Obtain the axisymmetric mapping point of each target point, calculate the normal residual and curvature residual between the target point and the axisymmetric mapping point, and generate a symmetric residual distribution map based on the normal residual and curvature residual;

[0139] Understandably, the target point set is a broad candidate set that includes both real defect points and artifact points.

[0140] In this application, the inner wall of the connector has a cavity-type mirror-symmetric structure. It can be understood that the artifact points must originate from the reflection effect of real points or other artifact points, that is to say:

[0141] In one scenario, if a target point does not have a corresponding axisymmetric mapping point, then theoretically, the target point is a real defect point that does not have a corresponding reflection during the imaging process, and there is no need to further screen for artifact points.

[0142] In another case, if a target point has one or more corresponding axisymmetric mapping points, these points form a set of reflection points, which contains only one real defect point. This application needs to further process the set of reflection points to remove artifact points and obtain the real defect point.

[0143] In this embodiment, the central axis of the surface map is first used as the reference axis to perform symmetrical mapping on each identified target point. This symmetrical mapping process is carried out in a preset polar coordinate space. By performing an axisymmetric transformation on the polar angular components of the target point, the theoretical position of the symmetrical mapping point is obtained. The actual point correspondence is then searched in the original point cloud. If there is no accurate mapping point, there is no need to process the target point further. The target point is selected and aggregated with the target points obtained from further processing of the candidate set to generate a connector point cloud image.

[0144] Furthermore, target points with axisymmetric mapping points are aggregated. For each pair of symmetrical points, using the radial position and polar angle information in polar coordinate space as references, the difference in the angle between the normal vectors and the difference in curvature values ​​between the point pairs are calculated. This yields a set of corresponding structural perturbation consistency indices. Subsequently, all point pairs with symmetrical relationships are used as sample points to construct a symmetrical residual distribution map.

[0145] It is understandable that the symmetric residual distribution map uses spatial location as the basic coordinate and normal residual and curvature residual as multidimensional perturbation amplitude indicators to describe the structural consistency pattern that can be derived from symmetric mapping in the entire surface map.

[0146] For example, regarding axisymmetric mapping points, please refer to... Figure 6 , Figure 6 This is a schematic diagram illustrating the principle of axisymmetric mapping point generation in an embodiment of this application. Figure 6 Three cases are shown: no symmetric relationship, one-to-one mapping symmetric relationship, and multiple pairs of symmetric mapping relationships.

[0147] Figure 6 The diagram shows that target point 1, target point 2, and target point 3 have multiple pairs of symmetrical mapping relationships. Therefore, these three target points form a group, and in subsequent processing, only the corresponding three target points need to be analyzed to obtain the final true defect points.

[0148] Figure 6 The results further demonstrate that target point four and target point six have a one-to-one mapping symmetric relationship. Therefore, these two target points are a group, and in subsequent processing, only the corresponding two target points need to be analyzed to obtain the final true defect points.

[0149] Figure 6 It is further shown that the axisymmetric mapping point of target point five does not have a corresponding target point, that is, target point five does not have a symmetry relationship. In this application, there is no need to process target point five further. We can wait for the other target points to be processed to generate the connector point cloud image.

[0150] Next, we will further elaborate on the part of the method for artifact point removal in this application.

[0151] Understandably, during structured light or laser imaging, defects or surface disturbances can generate multipath reflections through the inner walls, forming symmetrically distributed virtual image points. These artifact points are extremely similar to real defect points in terms of surface morphology, exhibiting highly overlapping local normal features and curvature perturbation characteristics. However, they typically show a certain axisymmetric offset from the real points in spatial location. The main characteristics of these artifact points are: a single source, multiple results, and a regular, symmetrical clustered distribution. Conversely, real defect points, because they actually exist on the structural surface, are not generated by reflections from other points and are therefore more likely to exist independently in a divergent manner.

[0152] Those skilled in the art will now understand that this application believes that if a certain perturbation point is located at the starting point of a relationship between multiple low residual symmetric points in the symmetric residual distribution map, and its perturbation direction shows a significant gradient divergence trend compared with the corresponding axisymmetric mapping point, then the point is more likely to be the source point of the actual defect; while those points that show convergence characteristics in the perturbation direction are more likely to be artifact points generated by reflection paths or structural coupling.

[0153] In one example, the specific steps of S3 are as follows:

[0154] S3.1: Construct a perturbation spatial gradient field based on the symmetric residual distribution map, and calculate the perturbation intensity spatial gradient at each target point in the map;

[0155] Specifically, for each target point, its corresponding normal residual and curvature residual values ​​are obtained, and its disturbance intensity spatial gradient is calculated based on the spatial distribution of its neighboring point set in the three-dimensional polar coordinate system. The calculation process involves traversing the residual variation trends of each point's neighboring points in the polar radius and polar angle directions to form a disturbance vector centered on the target point in the local disturbance field. The direction and amplitude of the disturbance vector reflect the dominant flow direction of that point in the surrounding disturbance field.

[0156] S3.2: Divide the target point into a disturbance convergence point and a disturbance divergence point according to the gradient direction of the spatial gradient of the disturbance intensity;

[0157] Specifically, in the aforementioned steps, target points with axisymmetric relationships have been aggregated. That is, the gradient direction of each target point is analyzed using the perturbation spatial gradient field constructed from the point cloud in polar coordinate space. Specifically, for a given target point, the perturbation gradient vector is formed by statistically analyzing the changing trends of the perturbation intensity of its adjacent points in the radial and axial directions. This vector reflects whether the point is in the direction of initiation or reception of the perturbation. If multiple perturbation gradient vectors pointing to a point exist among its adjacent points, it indicates that the point has gathered perturbation energy from multiple directions and is judged as a perturbation convergence point; conversely, if most of the perturbation vectors of a point point point to its adjacent points, that is, the perturbation diverges outward from that point, then the point is considered a perturbation divergence point.

[0158] S3.3: Remove the disturbance convergence points and generate a connector point cloud image based on the point cloud corresponding to the remaining disturbance divergence points;

[0159] It is understandable that, in addition to the point cloud corresponding to the remaining perturbation divergence points, this application also generates a connector point cloud image based on the point cloud corresponding to the target point that does not have an axisymmetric mapping point.

[0160] Those skilled in the art will understand that how to generate point cloud images from point clouds is common knowledge and will not be elaborated here. However, it should be noted that since the aforementioned steps involve multiple screenings of the point cloud, the final point cloud data may be sparse and fragmented. Therefore, it is necessary to aggregate the point clouds that were previously screened out in the aforementioned steps with the final point cloud to generate a complete point cloud image.

[0161] As a preferred embodiment, this application further analyzes the elimination of disturbance convergence points.

[0162] Specifically, a perturbation convergence point refers to a point in a symmetrical residual distribution map where the spatial gradient direction of the perturbation intensity tends to point towards itself. Spatially, this gradient convergence phenomenon may reflect a typical characteristic of artifact interference. However, in actual point cloud data, some real defect regions may also exhibit a geometric pattern of local perturbation convergence. For example, the closed endpoints of crack-like defects often appear as high-density edge structures in imaging due to their microscopic morphology, forming a perturbation-centric aggregation effect. Similarly, concave surface regions formed by pits, ablation, or erosion may also generate convergence gradient patterns similar to artifact points in the perturbation field due to their abrupt changes in normal and curvature concentration.

[0163] It is easy to understand that if all gradient-converging target points are directly excluded when removing artifact points, it is highly likely that point clouds corresponding to the aforementioned real defect features will be mistakenly deleted, thus affecting the completeness and confidence of the final detection. Furthermore, this embodiment also considers that in actual imaging, the cavity structure is limited by factors such as the incident angle, non-ideal reflection path, and inconsistent surface coatings. Some artifact points will not strictly exhibit typical gradient convergence behavior, but rather exist in a weaker perturbation gradient or a local asymmetric form. If only the gradient directionality is used as the exclusion criterion, these artifact regions may also be missed, leading to misidentification.

[0164] In this preferred embodiment, it is further determined whether these convergence points constitute a mirror point pair relationship under an axisymmetric structure. Based on the connector's central axis as the geometric reference, a symmetry mapping operation is performed on each disturbed convergence point, and a search is conducted in the actual point cloud to determine if a corresponding point exists at its axisymmetric position. If a corresponding point exists, the following conditions are further determined:

[0165] Firstly, the symmetrical point itself is also a point of convergence of disturbances, and it has similar local disturbance intensity gradient convergence characteristics to the original point.

[0166] Secondly, the gradient directions between the original point and its symmetrical point are mirror images in space, meaning that the perturbation gradient directions of the two are basically symmetrical about the central axis.

[0167] Thirdly, the fact that the disturbance intensity values ​​at the two points are close within a certain tolerance range suggests that they may originate from the same imaging interference or reflection path response.

[0168] When all three conditions are met, it can be reasonably assumed that the point pair is a typical mirror artifact pair, with symmetrical spatial distribution, similar perturbation behavior and consistent interference source, and therefore it should be eliminated as a whole.

[0169] Furthermore, for convergence points that fail to form a mirror pair relationship, i.e., where there is no corresponding perturbation convergence point at their symmetrical position, or where the perturbation direction / intensity does not meet the consistency requirements, this embodiment does not directly remove them. These asymmetric convergence points are temporarily retained as part of the perturbation region to be determined and enter the subsequent defect identification network. The network then determines whether they constitute a real defect point based on deeper geometric features and contextual information, thereby achieving a high-precision, high-fidelity defect identification process.

[0170] As a preferred option, this application further analyzes the retention of perturbation divergence points.

[0171] Specifically, real defects in a perturbation gradient field often exhibit an outward divergence of the spatial gradient of the perturbation intensity. Therefore, preserving all divergence points helps ensure the complete capture of the defect region. However, under specific structural conditions, some artifacts may also form visual features of perturbation gradient divergence under the interference of three or more optical reflection paths, thus causing interference and misjudgment of the recognition results.

[0172] It is easy to understand that, due to the typical mirror-symmetric cavity structure inside the connector, after multiple reflections of the laser or imaging beam on the inner wall, secondary artifacts may form under a non-principal axisymmetric structure. This is especially true when the reflection path involves multiple surface intersections or multi-cavity transitions. The spatial location of these artifacts deviates from the initial reflection symmetry axis, and the gradient direction of their perturbation no longer exhibits strong symmetry, but rather tends to be discrete or locally perturbed divergent. These higher-order reflection artifacts are difficult to capture using traditional residual analysis and axisymmetric sieving mechanisms, and may appear as pseudo-divergent points in the perturbed gradient field, thus being misclassified as real defect points.

[0173] In this preferred embodiment, among the target points identified as disturbance divergence points, the angle between the disturbance direction and the normal of the connector inner wall is further calculated to determine whether the source of the disturbance conforms to the single reflection propagation law; if the disturbance direction deviates from the normal of the inner wall and the disturbance value is less than the preset lower limit, it is marked as an artifact inheritance point formed by multiple reflections and is removed.

[0174] The defect identification network includes a point set geometric coding layer, a spatial perturbation extraction layer, and a defect determination layer, wherein:

[0175] The point set geometric coding layer is used to receive the connector point cloud image and construct a geometric feature vector based on the spatial coordinates, normal vector, and local curvature of each point cloud.

[0176] The spatial perturbation extraction layer is equipped with a graph convolutional network for calculating normal abrupt changes and local curvature minima in the spatial adjacency graph constructed from geometric feature vectors. The spatial perturbation extraction layer includes an adjacency graph construction module, a graph convolution calculation module, and a differentiation module.

[0177] The adjacency graph construction module is used to generate a spatial adjacency graph with Euclidean distance as the edge weight based on the spatial coordinates and geometric feature vector of each point.

[0178] The graph convolution calculation module uses a shared-weight graph convolutional network to propagate and aggregate information on nodes in the spatial adjacency graph, and extracts the normal rate of change and curvature gradient of each point in its neighborhood.

[0179] The differentiation module determines whether the normal change rate of each point is greater than the weighted standard deviation of the mean normal change rate of the neighborhood based on the normal change rate and curvature gradient of each point in its neighborhood, so as to identify the abrupt relationship of normal change. Under the condition that the curvature gradient is less than a preset first threshold and the curvature value is a local curvature minimum value in its neighborhood, the module identifies the local curvature minimum value region.

[0180] The defect determination layer is used to output the location label of the defect region and its corresponding three-dimensional coordinate set based on the normal abrupt change relationship and the local curvature minimum region. The defect determination layer includes:

[0181] Perform local connectivity analysis on the aforementioned abrupt change in normal direction and curvature minimum region to generate a candidate defect point set;

[0182] The candidate defect point set is clustered and grouped according to a clustering algorithm to obtain the clustering results;

[0183] Based on the spatial distribution density, curvature gradient variation range, and normal consistency of the point cloud in the clustering results, calculate the defect confidence of the point set region corresponding to the clustering results.

[0184] The point set region with a defect confidence level greater than the defect threshold is output as the defect region.

[0185] In the embodiments of this application, a defect identification network is used to identify real defect regions from connector point cloud images after artifact removal processing. Its core objective is to achieve accurate identification and localization of defect points with subtle geometric perturbation features. The network structure is specially designed to fully combine the spatial geometric characteristics of the point cloud and the structural regularity of the connector's inner wall to support a three-stage identification process for fine-grained perturbation detection, including geometric feature encoding, spatial perturbation extraction, and defect region determination.

[0186] In this embodiment, the point set geometric coding layer first encodes the input point cloud image. By combining geometric descriptors such as the three-dimensional spatial coordinates, normal vector direction, and local curvature of each point, a geometric feature vector describing the local structural state of the point is constructed. This feature representation not only preserves the positional distribution of the points but also enhances their distinguishability under minute geometric changes.

[0187] The spatial perturbation extraction layer further constructs a spatial adjacency graph based on the aforementioned geometric features and performs feature propagation and aggregation through a graph convolutional network, thereby encoding the geometric perturbation relationship between each point and its neighborhood. This process effectively captures spatial perturbation features such as abrupt changes in normal direction and curvature gradient changes through the graph convolutional computation module, and the differentiation module extracts regional points that meet certain abrupt changes and minimal curvature conditions in the local neighborhood of each point, laying the foundation for subsequent identification of typical morphological changes such as defect edges, pores, or protrusions.

[0188] The defect determination layer then fuses and judges the extracted spatial perturbation information. Through local connectivity analysis, the network first filters out candidate point sets that may constitute defect boundaries or internal perturbations, and then aggregates these points structurally using a clustering algorithm. Finally, based on multi-dimensional features such as spatial density, curvature variation amplitude, and normal vector consistency of points in the clustered region, a confidence score for the region as a defect is comprehensively calculated. Only when this score exceeds a certain threshold is it identified as a valid defect region and its three-dimensional coordinate set is output. This processing method enables this application not only to identify visible surface structural perturbations, but also to effectively remove artifact points and retain true defects in point cloud data with complex structures and strong reflection interference.

[0189] Those skilled in the art will understand that although this embodiment uses a graph convolutional network structure and specific geometric rules as the implementation form of the defect recognition network, it is not limited to this. Other neural network architectures, such as those based on PointNet, DGCNN, or Transformer, can also be used to achieve the defect recognition objective of this application, provided that the requirement for local perturbation perception is met. The construction principle of the defect recognition network is based on the structural consistency of point clouds in space, local perturbation patterns, and geometric abrupt change characteristics. Through high-dimensional feature aggregation and confidence modeling strategies, it achieves stable identification and accurate calibration of defects on the inner wall of connectors. This application does not impose specific limitations on this.

[0190] In one example, performing three-dimensional imaging of the defective region includes:

[0191] Extract the set of three-dimensional point cloud coordinates corresponding to the defect area, and construct a spatial voxel mesh based on the set of three-dimensional point cloud coordinates;

[0192] Interpolation fitting is performed on the spatial voxel mesh to generate a three-dimensional image of the defect.

[0193] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. A three-dimensional imaging detection method for surface defects in connectors, characterized in that, The method includes: Based on the collected point cloud data, a surface map in the depth direction is obtained; Using the central axis of the surface plot as the axis of symmetry, target points are identified by morphological operators, and the spatial residual between each target point and its axisymmetric position point is calculated to obtain a symmetric residual distribution map. In the symmetrical residual distribution map, an artifact point removal strategy is executed to obtain a connector point cloud image. The artifact point removal strategy determines whether there are artifact points based on the perturbation trend of the target points and removes the target points whose perturbation trend is convergent. The connector point cloud image is input into a preset defect recognition network. The defect recognition network processes the connector point cloud image and outputs the defect region. The defect region is then used to create a three-dimensional image. The defect recognition network is trained using historical defect data of the connector. Using the central axis of the surface plot as the axis of symmetry, the target point is identified by morphological operators, including: Based on the distribution of point cloud in the axis of symmetry and surface diagram, determine the normal projection direction of point cloud, and calculate the projection signal curve based on the normal projection direction and point cloud spatial density. Perform morphological opening and closing operations on the projection signal curve to obtain a morphological reference curve; The residuals of the projection signal curve and the shape reference curve are compared, and point clouds with residual amplitudes greater than the disturbance judgment threshold are selected to generate target points. Perform morphological opening and closing operations on the projected signal curve to obtain a morphological reference curve, including: Based on the shape of the surface plot and the density change of the point cloud in the normal projection direction, a nonlinear structural element is generated, wherein the nonlinear structural element is calculated based on the periodic characteristics of the density change. Using the nonlinear structural element as a sliding window, erosion and dilation operations are performed on the projected signal curve, wherein the erosion operation is processed based on the local minimum value within the sliding window, and the dilation operation is processed based on the local maximum value within the sliding window. The results of the erosion and dilation operations are fused into a smooth envelope curve, which serves as the morphological reference curve for the projected signal curve. In the symmetrical residual distribution map, an artifact point removal strategy is performed to obtain a connector point cloud image, including: Based on the symmetric residual distribution map, a perturbation spatial gradient field is constructed, and the perturbation intensity spatial gradient of each target point in the map is calculated. The target point is divided into a disturbance convergence point and a disturbance divergence point according to the gradient direction of the spatial gradient of the disturbance intensity. The disturbance convergence points are removed, and a connector point cloud image is generated based on the point cloud corresponding to the remaining disturbance divergence points.

2. The three-dimensional imaging detection method for surface defects of a connector according to claim 1, characterized in that, Calculate the spatial residual between each target point and its axisymmetric location, including: Obtain the axisymmetric mapping point for each target point; Calculate the normal residual and curvature residual between the target point and the axisymmetric mapping point.

3. The three-dimensional imaging detection method for surface defects of a connector according to claim 1, characterized in that, The construction of the surface plot includes: The point cloud data is projected into a polar coordinate space with the center of the connector as the reference axis; Based on the projected point cloud data, multiple equidistant sub-regions are divided along the polar radius direction, and the point set that forms local extreme point density in the axial direction of each sub-region is extracted. Using the density peak of the point set as the reference layer boundary, a fitted surface with normal continuity constraints is constructed based on the normal variation rate of the point set's neighborhood to generate a surface plot.

4. The three-dimensional imaging detection method for surface defects of a connector according to claim 1, characterized in that, The defect identification network includes a point set geometric coding layer, a spatial perturbation extraction layer, and a defect determination layer, wherein: The point set geometric coding layer is used to receive the connector point cloud image and construct a geometric feature vector based on the spatial coordinates, normal vector, and local curvature of each point cloud. The spatial perturbation extraction layer is equipped with a graph convolutional network, which is used to calculate the normal abrupt change relationship and the local curvature minimum region in the spatial adjacency graph constructed by the geometric feature vectors; The defect determination layer is used to output the location label of the defect region and its corresponding three-dimensional coordinate set based on the normal abrupt change relationship and the local curvature minimum region.

5. The three-dimensional imaging detection method for surface defects of a connector according to claim 4, characterized in that, The spatial perturbation extraction layer includes an adjacency graph construction module, a graph convolution calculation module, and a differentiation module, wherein: The adjacency graph construction module is used to generate a spatial adjacency graph with Euclidean distance as the edge weight based on the spatial coordinates and geometric feature vector of each point. The graph convolution calculation module uses a shared-weight graph convolutional network to propagate and aggregate information on nodes in the spatial adjacency graph, and extracts the normal rate of change and curvature gradient of each point in its neighborhood. The differentiation module determines whether the normal change rate of each point is greater than the weighted standard deviation of the mean normal change rate of the neighborhood based on the normal change rate and curvature gradient of each point in its neighborhood, so as to identify the abrupt change relationship of the normal. Under the condition that the curvature gradient is less than a preset first threshold and the curvature value is a local curvature minimum value in its neighborhood, the module identifies the local curvature minimum value region.

6. The three-dimensional imaging detection method for surface defects of a connector according to claim 4, characterized in that, The defect determination layer includes: Perform local connectivity analysis on the aforementioned abrupt change in normal direction and curvature minimum region to generate a candidate defect point set; The candidate defect point set is clustered and grouped according to a clustering algorithm to obtain the clustering results; Based on the spatial distribution density, curvature gradient variation range, and normal consistency of the point cloud in the clustering results, calculate the defect confidence of the point set region corresponding to the clustering results. The point set region with a defect confidence level greater than the defect threshold is output as the defect region.

7. The three-dimensional imaging detection method for surface defects of a connector according to claim 1, characterized in that, Performing three-dimensional imaging of the defective region includes: Extract the set of three-dimensional point cloud coordinates corresponding to the defect area, and construct a spatial voxel mesh based on the set of three-dimensional point cloud coordinates; Interpolation fitting is performed on the spatial voxel mesh to generate a three-dimensional image of the defect.

Images