A kind of automobile steel wire surface defect detection system based on structured light three-dimensional measurement

Abstract

This application relates to the field of defect detection technology, and in particular to a defect detection system for automotive steel wire surfaces based on structured light 3D measurement. The system includes an acquisition module configured to acquire 3D point cloud data of the automotive steel wire and determine the local geometric variation index of each data point in the 3D point cloud data; a construction module configured to determine geometric confidence weights using the local geometric variation indexes corresponding to the data points, and construct an anisotropic weighting function using the geometric confidence weights; a fitting module configured to perform moving least squares fitting on other points within the local geometric neighborhood using the anisotropic weighting function to obtain an ideal reference surface passing through the data points; and an output module configured to output the defect detection results for the automotive steel wire surface in conjunction with the ideal reference surface. Through the above technical solution, defect detection of the automotive steel wire surface can be achieved.

Description

Technical Field

[0001] This application relates to the field of defect detection technology, and in particular to a defect detection system for automotive steel wire surfaces based on structured light three-dimensional measurement. Background Technology

[0002] As a key raw material for core transmission and load-bearing components of automobiles, automotive steel wire is typically made from high-carbon cold-rolled steel wire through multiple processes such as drawing, heat treatment, and coating. Automotive steel wire exhibits high strength and high toughness. Its cross-section is mostly circular, and its surface undergoes anti-corrosion treatments such as galvanizing or copper plating to improve weather resistance and wear resistance. By controlling the carbon content and heat treatment process, automotive steel wire is ensured to possess excellent tensile strength, elastic limit, and fatigue life, making it suitable for the mechanical performance requirements of different automotive components.

[0003] In the manufacturing and operation of automobiles, automotive steel wires can be used in core components such as braking systems, suspension systems, transmission systems, and tire reinforcement. In the braking system, brake cable steel wires need to transmit braking force to ensure timely and reliable braking response. In the suspension system, shock absorber spring steel wires absorb road impacts through elastic deformation, improving driving stability and comfort. Tire reinforcement steel wires significantly improve the tire's load-bearing capacity and tear resistance, extending tire life. Automotive steel wires are fundamental components that ensure the normal operation of the overall function of a vehicle.

[0004] During the production, processing, transportation, and assembly of automotive steel wires, various surface defects such as scratches, dents, protrusions, or cracks are easily generated. These defects can affect the performance and safety of automotive steel wires. Defects in automotive steel wires can reduce their tensile strength and elastic limit, potentially leading to serious safety accidents such as brake failure or suspension system malfunction, threatening the lives of drivers and passengers. Therefore, it is necessary to conduct defect detection on automotive steel wires. Summary of the Invention

[0005] To detect defects in automotive steel wires, this application provides a surface defect detection system for automotive steel wires based on structured light 3D measurement, comprising: an acquisition module configured to acquire 3D point cloud data of the automotive steel wires, construct a local geometric neighborhood for each data point in the 3D point cloud data, perform differential geometric analysis on the distribution of the point set within the local geometric neighborhood, and determine the local geometric variation index corresponding to each data point; the local geometric variation index is used to characterize the degree of microscopic geometric abrupt change of the data point on the surface of the automotive steel wires; and a construction module configured to determine the geometric position using the local geometric variation index corresponding to the data point. The system employs a confidence weighting mechanism and utilizes geometric confidence weights to construct an anisotropic weighting function. A fitting module is configured to use the anisotropic weighting function to perform moving least squares fitting on other points within the local geometric neighborhood to obtain an ideal reference surface passing through the data points. This ideal reference surface is used to fit the macroscopic surface morphology of the automotive steel wire after defect removal. An output module is configured to calculate the directed normal distance between the data points and their projection points on the ideal reference surface, determine the adaptive threshold corresponding to each data point, and output the defect detection results for the automotive steel wire surface based on the directed normal distance and the adaptive threshold.

[0006] In this way, by introducing the local geometric variation index, the system can keenly perceive the micro-geometric abrupt changes on the surface of the steel wire and construct geometric confidence weights accordingly. When performing moving least squares fitting, the anisotropic weighting function constructed using the weights can automatically reduce the influence of potential defect points on the fitting process, thereby reconstructing an ideal reference surface that avoids defect interference and obtaining more accurate defect detection results for the surface of automotive steel wire.

[0007] Optionally, the local geometric variability index corresponding to the data points is constructed as follows: Eigenvalue decomposition of the covariance matrix is ​​performed on the point set within the local geometric neighborhood to obtain a set of eigenvalues ​​and a set of normal vectors. The surface curvature component is calculated using the eigenvalue set, and the normal vector consistency component is calculated using the set of normal vectors. ,in, , and Let be three non-negative eigenvalues ​​in the eigenvalue set, and , This represents the number of points within the local geometric neighborhood. The unit normal vector of the data point and the first The dot product of the unit normal vectors of the neighboring points, It is a natural exponential function.

[0008] In this way, by fusing the surface curvature component and the normal vector consistency component, it is possible to identify drastic changes in depth as well as minute perturbations in the surface normal.

[0009] Optionally, the geometric confidence weights are determined in the following way: ,in, Geometric confidence weights; This is the sensitivity adjustment coefficient, used to control the decay rate of the weight as the difference changes; The local geometric variation index of the data points; The median of the local geometric variation index of all points within the local geometric neighborhood; To take the absolute value.

[0010] Optionally, the anisotropy weighting function of the target data point is obtained by multiplying the spatial distance weight and the geometric confidence weight; the spatial distance weight is determined by using a Gaussian function based on the Euclidean distance from the neighboring data points in the local geometric neighborhood to the target data point, and it decays exponentially with the increase of the Euclidean distance.

[0011] In this way, by considering the proximity of neighboring data points to the target data point, as well as the local geometric variation index of the neighboring data points, we can determine higher geometric confidence weights for neighboring data points that are closer and have higher geometric confidence weights.

[0012] Optionally, the adaptive threshold corresponding to the data point is determined in the following way: ,in, An adaptive threshold for the target data points; Based on the basic noise threshold, The preset confidence coefficient, The number of data points within the local geometric neighborhood of the target data point; For the local geometric neighborhood, the first The depth of each data point; This is the average depth of all data points within the local geometric neighborhood.

[0013] In this way, the threshold can be dynamically adjusted based on local depth fluctuations.

[0014] The technical solutions provided by the embodiments of this application may include the following beneficial effects: acquiring three-dimensional point cloud data of automotive steel wires and determining the local geometric variation index of each data point in the three-dimensional point cloud data; determining the geometric confidence weight using the local geometric variation index corresponding to the data point, which can assign a lower geometric confidence weight to pixels with a higher probability of anomalies, and constructing an anisotropic weighting function using the geometric confidence weight; performing moving least squares fitting on other points in the local geometric neighborhood using the anisotropic weighting function to obtain an ideal reference surface passing through the data point; and outputting more accurate defect detection results for the surface of automotive steel wires by using the directed normal distance between the data point and the projection point on the ideal reference surface.

[0015] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of a structural system for detecting surface defects in automotive steel wires based on structured light three-dimensional measurement, according to an exemplary embodiment. Figure 2 This is a schematic diagram of a two-dimensional parametric space image obtained by unfolding and mapping three-dimensional point cloud data; Figure 3 This application provides a schematic diagram illustrating the defect detection results of automotive steel wires. Figure 4 A schematic diagram showing the results of defect detection of automotive steel wire using existing technology. Detailed Implementation

[0017] To detect defects in automotive steel wires, this application provides a 1000 automotive steel wire surface defect detection system 1000 based on structured light three-dimensional measurement. Figure 1 This is a schematic diagram illustrating the structure of an automotive steel wire surface defect detection system 1000 based on structured light three-dimensional measurement, according to an exemplary embodiment. Figure 1 As shown, the automotive steel wire surface defect detection system 1000 based on structured light three-dimensional measurement includes: an acquisition module 1100, a construction module 1200, a fitting module 1300, and an output module 1400.

[0018] The acquisition module 1100 is configured to acquire three-dimensional point cloud data of automotive steel wire, construct a local geometric neighborhood for each data point in the three-dimensional point cloud data, perform differential geometric analysis on the distribution of point sets within the local geometric neighborhood, and determine the local geometric variation index corresponding to each data point; the local geometric variation index is used to characterize the degree of microscopic geometric mutation of the data point on the surface of the automotive steel wire.

[0019] In one embodiment, the acquisition module 1100 is further configured to perform the following processing steps: control the structured light scanning system to acquire stripe images using multiple sets of different exposure times, and calculate the intensity modulation schedule for each pixel in the stripe image; the intensity modulation schedule is used to characterize the amplitude intensity of the periodic change in light intensity of the pixel in multiple sets of stripe images; determine whether the pixel is in a saturated state at the exposure time based on the light intensity gray value of the pixel, and perform interpolation repair on the pixels in a saturated state in all exposure channels to obtain repair data, so as to convert the pixels in a non-saturated state and the repair data to three-dimensional spatial coordinates to obtain three-dimensional point cloud data.

[0020] The surface of automotive steel wire is usually a cylindrical metal curved surface. The light conditions are unevenly distributed on the curved surface. The top area of ​​the automotive steel wire surface is prone to saturation of the camera sensor due to specular reflection, while the side edge areas have weak reflected light due to the large incident angle.

[0021] The detection system can preset three or more different exposure time parameters, such as short exposure of 50 microseconds, medium exposure of 200 microseconds and long exposure of 800 microseconds.

[0022] The short exposure time of 50 microseconds is mainly used to capture the stripe information of the highly reflective area at the top of the steel wire. At this time, the light flux entering the sensor is relatively small, which can effectively suppress the overflow of highlights. The medium exposure time of 200 microseconds is used to collect information of the main area of ​​the steel wire. The long exposure time of 800 microseconds is used to improve the signal-to-noise ratio of the dark areas on both sides of the steel wire.

[0023] During the acquisition process, the detection system can calculate the intensity modulation schedule for each pixel across the three exposure channels. Intensity modulation schedule is a physical quantity that measures the contrast of sinusoidal fringes. The calculation method for intensity modulation schedule is based on the intensity difference value of the four-step phase-shift diagram in the phase-shifting method. ,in For the first The grayscale value of the phase shift is the value of the stripe. The larger the value, the higher the signal-to-noise ratio of the stripe. n is equal to 0, 1, 2, 3 or 4.

[0024] The detection system can set a grayscale saturation threshold. For any pixel, it can be detected in order from long exposure to short exposure. If the grayscale value under the current exposure time exceeds the saturation threshold, it is determined to be saturated and automatically switches to the next level of shorter exposure time data.

[0025] If a pixel is not saturated under all exposure times, the set of data with the largest modulation and demodulation schedule is selected as the effective input. For a few pixels that are still saturated under the shortest exposure, or blind spots that still have a low signal-to-noise ratio under the longest exposure, the detection system can use a neighborhood-based Laplacian interpolation algorithm to generate repair data to ensure that the final generated phase map is continuous and complete.

[0026] Using pre-calibrated structured light system parameters, including camera intrinsic matrix, projector parameters, and the rotation and translation matrix between them, the optimized phase data is mapped into three-dimensional spatial coordinates through the principle of triangulation.

[0027] The acquisition strategy based on multi-exposure fusion in this embodiment avoids the data loss that may be caused by reflections from the curved metal surface, and provides an accurate point cloud basis for subsequent defect detection.

[0028] In one embodiment, the acquisition module 1100 is further configured to perform the following processing steps: acquire initial point cloud data of the automobile wire mesh from a multi-view scanning perspective; determine the target point cloud as the registration reference and the source point cloud to be transformed into coordinates from the initial point cloud data; transform the source point cloud into the coordinate system of the target point cloud by minimizing the objective function to solve the rotation matrix and translation vector, so as to achieve the stitching of the panoramic view point cloud of the automobile wire mesh to obtain three-dimensional point cloud data; wherein, the objective function is used to constrain the geometric fit between the source point cloud and the target point cloud on the surface, and to restrict the displacement degree of freedom of the source point cloud relative to the target point cloud along the central axis of the automobile wire mesh.

[0029] Since the cross-section of the steel wire is usually circular, a single structured light camera can only cover a field of view of about 120 degrees. In order to achieve 360-degree full-circumference surface defect detection, the detection system can arrange three sets of structured light sensors evenly along the circumference of the steel wire, with each set of sensors at a 120-degree angle to the others.

[0030] Mechanical installation may have tolerances and the steel wire may vibrate during movement. The initial point cloud data collected by these three sets of sensors are in three independent coordinate systems, and spatial registration of these three sets of sensors is required. However, the automotive steel wire has axial translation invariance.

[0031] Traditional iterative nearest-point algorithms may slip along the axial direction, causing misalignment in the length direction of the stitched model. In the embodiments of this application, the objective function not only includes the Euclidean distance error term from the point to the tangent plane, but also includes an axial displacement penalty term, which enables the detection system to identify the inherent drawing texture feature points on the surface of the steel wire, or to use the micro-geometric features of the overlapping area of ​​the field of view as anchor points to construct point-pair constraints.

[0032] The target point cloud and the source point cloud are different subsets of the point cloud obtained from scanning the initial point cloud data under different multi-view conditions; the rotation matrix is ​​solved during the optimization process (e.g., Matrix) and translation vector (e.g.) The vector minimizes the geometric error between the transformed source point cloud and the target point cloud in the overlapping region, while restricting the relative displacement component along the central axis of the wire to be closer to zero.

[0033] In this way, by aligning the point clouds from multiple perspectives along the axial direction, a more complete three-dimensional point cloud information of the automotive steel wire surface is obtained, avoiding the interference of artifacts caused by splicing misalignment with subsequent defect identification.

[0034] In one embodiment, the geometric neighborhood of the data points is constructed as follows: the central axis of the 3D point cloud data is extracted, and an orthogonal curve coordinate system is constructed based on the central axis to expand and map the 3D point cloud data into a 2D parametric space image using a manifold mapping algorithm; the horizontal coordinate of the 2D parametric space image is the arc length along the central axis, and the vertical coordinate is the circumferential angle; the 3D point cloud data corresponding to the adjacent pixels of the target pixel in the 2D parametric space image are used as data points in the local geometric neighborhood of the target pixel.

[0035] When dealing with complex curved steel wires, the traditional K-nearest neighbor search algorithm based on three-dimensional Euclidean distance may be subject to topological errors. For example, when the steel wire is spirally curved, two points that are close in space may actually be located on different turns of the steel wire, that is, they are not adjacent on the surface topology. If non-topologically adjacent points are included in the neighborhood calculation, it will lead to distortion in the estimation of surface normal vector and curvature.

[0036] This embodiment reduces the computation in the three-dimensional bending space to the two-dimensional parameter domain, which can establish a Frenet frame or a rotation minimization frame along the central axis of the wire, and define the axial arc length parameter and the circumferential angle parameter.

[0037] For each data point in the point cloud, the position of the projection point of the data point on the central axis can be calculated to determine the value of the axial arc length parameter, and the angle of the data point in the polar coordinate system of the cross section can be calculated to determine the value of the circumferential angle parameter. Through this mapping, the twisted steel wire surface in three-dimensional space is unfolded into a two-dimensional depth image. The adjacency relationship between pixels in the two-dimensional image corresponds to the topological adjacency relationship of the steel wire surface.

[0038] Determining the local geometric neighborhood simplifies to selecting a region centered on the target pixel in the two-dimensional image. (For example or The pixel window reduces the time complexity of neighborhood search from logarithmic to constant, ensuring that the input data for subsequent differential geometric analysis is continuous and compact on the physical surface.

[0039] In one embodiment, the central axis of the 3D point cloud data is obtained by: performing iterative shrinkage calculations on the 3D point cloud data using a skeleton extraction algorithm; finding the target spatial location of the skeleton point to be determined during the iterative shrinkage calculation process; minimizing the sum of the Euclidean distances of all data points in the local neighborhood of the target spatial location to the skeleton point to be determined; determining the target spatial location as the skeleton point; and fitting and connecting all the skeleton points determined by the skeleton extraction algorithm in sequence to obtain the central axis of the 3D point cloud data.

[0040] To obtain the central axis required for manifold mapping, the L1-median skeleton extraction algorithm can be used. Compared with the traditional least squares centroid method, the L1-median method is robust to outliers.

[0041] The core idea of ​​the L1-median skeleton extraction algorithm is to simulate the physical process of point cloud data shrinking towards the geometric center. For each skeleton node to be determined, the algorithm iteratively searches for an optimal position in three-dimensional space, such that the sum of the Euclidean distances from the optimal position to all data points in the surrounding neighborhood reaches the global minimum.

[0042] When determining the target spatial location, the sum of the Euclidean distances of all data points in the local neighborhood of the target spatial location to the skeleton point to be determined is minimized, which conforms to the sparsity property of the L1 norm. The calculated center point can be more stably maintained on the geometric center line of the wire.

[0043] After completing the shrinkage calculations for all sections, spline curves can be used to smoothly fit these discrete skeleton points, generating a continuous and differentiable three-dimensional space curve as the central axis. This central axis forms the reference for subsequent coordinate system transformations and unfolding operations.

[0044] In one embodiment, the local geometric variability index corresponding to a data point is constructed as follows: the covariance matrix eigenvalue decomposition is performed on the point set within the local geometric neighborhood to obtain an eigenvalue set and a normal vector set; the surface curvature component is calculated using the eigenvalue set; and the normal vector consistency component is calculated using the normal vector set. ,in, , and Let be three non-negative eigenvalues ​​in the eigenvalue set, and , This represents the number of points within the local geometric neighborhood. The unit normal vector of the data point and the first The dot product of the unit normal vectors of the neighboring points, It is a natural exponential function.

[0045] In the formula for calculating the local geometric variation index, the local geometric variation index is used to capture microscopic defect features from two orthogonal dimensions; one of these dimensions is the surface curvature component based on eigenvalues. By constructing a set of points within a local neighborhood By analyzing the covariance matrix and performing principal component analysis, three eigenvalues ​​can be obtained.

[0046] and This represents the distribution range of the point set within the tangent plane, while This represents the degree of dispersion of the point set along the normal direction. For an ideal smooth surface, the data points can more comprehensively fall on the tangent plane. The variance approaches zero and takes smaller values; however, when there are pits or protrusions on the surface, the variance of the data points in the normal direction increases dramatically, leading to... The value of the component increases significantly.

[0047] In addition to the surface curvature component based on eigenvalues, another dimension of the local geometric variation index is the normal vector consistency component based on the dot product of normal vectors. , The normal vector at the center point, Let be the normal vector of a neighboring point, and let be the absolute value of their dot product. It represents the cosine value of the angle between the normal vectors.

[0048] In flat regions, the normal vectors are more parallel, making the dot product closer to 1. Closer to 0, the exponent term is closer to However, at the edge of a scratch or at a crack, the surface normal will be flipped more drastically, resulting in a smaller value for the dot product and thus causing the exponential term to grow non-linearly.

[0049] By multiplying the two components of the local geometric variation index, the variation index will reach a higher value when the depth changes abruptly and the normal becomes disordered. This suppresses the misjudgment that may be introduced by single-dimensional features. For example, a gentle slope area with only depth changes but consistent normal will not be identified as a defect, thus ensuring the capture of real defect features.

[0050] The construction module 1200 is configured to determine the geometric confidence weights using the local geometric variation index corresponding to the data points, and to construct an anisotropic weighting function using the geometric confidence weights.

[0051] In one embodiment, the geometric confidence weights are determined in the following manner: ,in, Geometric confidence weights; is the sensitivity adjustment coefficient, used to control the decay rate of the weight as the difference changes; Q is the local geometric variation index of the data point; The median of the local geometric variation index of all points within the local geometric neighborhood; To take the absolute value.

[0052] In the process of constructing an ideal reference surface, if all points have the same weight, the defect points themselves will also participate in the fitting of the reference surface, causing the reference surface to be concave in the direction of the defect, thereby masking the true depth of the defect.

[0053] Under normal circumstances, the proportion of defect points in a local neighborhood is smaller. Therefore, the median of the local geometric variation index of all points in the local geometric neighborhood can accurately represent the roughness level of the normal surface of the region.

[0054] The geometric confidence weight is calculated using the denominator term of the formula. When the variance index of a certain data point is close to the background median, the denominator is close to 1, making the value of the geometric confidence weight closer to 1, indicating that the point where the variance index is close to the median is a credible normal point; the sensitivity adjustment coefficient can be set according to the surface finish of the steel wire, for example, a value of 100.

[0055] When the local geometric variability index at a certain point deviates from the median of the local geometric variability index due to the presence of defects, the squared term of the difference will be further amplified by the local geometric variability index, which increases the denominator of the formula for calculating the weight geometric confidence weight, and the weight geometric confidence weight will be closer to 0.

[0056] In one embodiment, the anisotropy weighting function of the target data point is obtained by multiplying the spatial distance weight and the geometric confidence weight; the spatial distance weight can be determined by using a Gaussian function based on the Euclidean distance from the neighboring data points in the local geometric neighborhood to the target data point, and it decays exponentially as the Euclidean distance increases.

[0057] Anisotropic weighted functions can be constructed, which combine spatial proximity and geometric similarity, with spatial distance weights following a Gaussian distribution model. ,in Let be the Euclidean distance from the neighboring points to the target point. Here, is the preset bandwidth parameter, and exp is the natural exponential function.

[0058] By performing a point-to-point product operation between the spatial distance weight and the aforementioned geometric confidence weight, in a flat and continuous surface region, the geometric weight is uniform and closer to 1. The weighting function can act as a standard isotropic Gaussian filter, thereby achieving a smoothing and noise reduction effect.

[0059] When the neighboring data points of the target data point are located at the edge of defects such as scratches or dents, the geometric weight of the neighboring data points belonging to the defect is closer to 0, which reduces the weight contribution of the defect direction. This makes the filter kernel exhibit anisotropic shape in space, and the normal points on the non-defect side are used for fitting, avoiding blurring the edge features while removing noise.

[0060] The fitting module 1300 is configured to use an anisotropic weighted function to perform moving least squares fitting on other points in the local geometric neighborhood to obtain an ideal reference surface passing through the data points; the ideal reference surface is used to fit the macroscopic surface morphology of the automotive steel wire after defects are removed.

[0061] In one embodiment, the ideal reference surface passing through the data points is obtained as follows: a second-order polynomial model is established in the local tangent plane coordinate system with the data points as the origin, and a weighted residual energy function is constructed; the weighted residual energy function is equal to the sum of the products of the squares of the fitting residuals of each point in the local geometric neighborhood and the corresponding anisotropic weighting functions; the weighted residual energy function is minimized to obtain the polynomial coefficient vector that minimizes the weighted residual energy function; the fitting coordinates of the data points in the local tangent plane coordinate system are calculated using the polynomial coefficient vector, and the fitting coordinates are determined as the projection points of the data points on the ideal reference surface, thereby obtaining the ideal reference surface passing through the data points through the projection points.

[0062] The moving least squares algorithm is used to accurately reconstruct the macroscopic geometry of the steel wire. For each data point to be processed, it can be transformed into a local coordinate system with the data point as the origin. In this local coordinate system, a second-order polynomial surface function can be defined. .

[0063] The local geometric features of the steel wire surface are mainly determined by curvature. A second-order model is sufficient to accurately describe the local bending characteristics of the cylindrical surface and can avoid overfitting oscillations that may be caused by higher-order models.

[0064] To solve for the six unknown coefficients of the polynomial of the second-order polynomial surface function to We can construct a weighted residual energy function. , The anisotropic weighting function of the target pixel is equivalent to the i-th neighboring pixel. , as well as Let be the coordinates of the i-th neighboring pixel in the local coordinate system.

[0065] The purpose of the energy function is to find a surface such that the sum of the squared weighted distances between the surface and all its neighboring points is minimized by the polynomial coefficient vector. By taking the partial derivative of the polynomial coefficient vector and setting the derivative to zero, the ideal reference surface passing through the data points can be obtained.

[0066] The output module 1400 is configured to calculate the directed normal distance between the data point and the projection point on the ideal reference plane, and determine the adaptive threshold corresponding to the data point. Based on the directed normal distance corresponding to the data point and the adaptive threshold, it outputs the defect detection results of the automotive steel wire surface.

[0067] In one embodiment, the adaptive threshold corresponding to the data point is determined in the following way: ,in, An adaptive threshold for the target data points; Based on the basic noise threshold, The preset confidence coefficient, The number of data points within the local geometric neighborhood of the target data point; For the local geometric neighborhood, the first The depth of each data point; This is the average depth of all data points within the local geometric neighborhood.

[0068] In industrial settings, a single fixed threshold often cannot meet the detection needs of different areas. For example, the surface of a steel wire may have some areas with high roughness due to the drawing process, while other areas may be very smooth. If a uniform low threshold is used, the rough areas will generate a large number of false alarms; if a uniform high threshold is used, the small defects in the smooth areas will be missed.

[0069] The adaptive threshold calculation formula proposed in this embodiment is based on the principle of statistical process control, realizing the dynamic floating of the detection standard; the square root term in the adaptive threshold calculation formula The background noise level or surface roughness of the current area is quantified, and the basic noise threshold is set. It is the noise threshold value determined by the physical resolution of the imaging system.

[0070] The confidence coefficient can be set to 3 to 4, corresponding to the three sigma or four sigma criteria in statistics. A data point is identified as an outlier only when its depth deviation exceeds the statistical distribution range of the background noise in the region.

[0071] When the detection scan encounters a section of smooth steel wire, the local standard deviation is even smaller, for example, equal to 0.005 mm. At this point, the adaptive threshold corresponding to the data point decreases to... The detection system operates at a millimeter level and is in high-sensitivity mode, enabling it to detect even finer scratches.

[0072] When a region with obvious processing texture is scanned, the local standard deviation increases to 0.02 mm, at which point the adaptive threshold corresponding to the data point increases to [a higher value]. The detection system automatically relaxes the standard to millimeters, thus effectively shielding the interference of normal texture signals.

[0073] Figure 2 This is a schematic diagram of a two-dimensional parametric space image obtained by unfolding and mapping three-dimensional point cloud data. Figure 2 The local geometric variation index was plotted at different unfolded locations. The local geometric variation index varies at different locations, and the local geometric variation index of some abnormal regions is higher than that of other regions.

[0074] Figure 3 This is a schematic diagram of the defect detection results of automotive steel wire using existing technology. Specifically, it is the detection result obtained by directly determining the Euclidean distance from the neighboring data points in the local geometric neighborhood to the target data point, without considering the geometric confidence weight of the data points.

[0075] Figure 4 This is a schematic diagram illustrating the defect detection results of automotive steel wires implemented in this application. In this embodiment, by considering the geometric confidence weight of data points, a higher value of the directed normal distance can be determined for data points that may be abnormal. This can prevent minor defects that may exist on the surface of automotive steel wires from being missed during the detection process, thereby improving the accuracy of the detection process for automotive steel wires.

[0076] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and embodiments are to be considered exemplary only.

[0077] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope.

Claims

1. A system for detecting surface defects in automotive steel wires based on structured light three-dimensional measurement, characterized in that, include: The acquisition module is configured to acquire three-dimensional point cloud data of automotive steel wire, construct a local geometric neighborhood for each data point in the three-dimensional point cloud data, perform differential geometric analysis on the distribution of point sets within the local geometric neighborhood, and determine the local geometric variation index corresponding to each data point; the local geometric variation index is used to characterize the degree of microscopic geometric abrupt change of the data point on the surface of the automotive steel wire. The construction module is configured to determine the geometric confidence weights using the local geometric variability index corresponding to the data points, and to construct an anisotropic weighting function using the geometric confidence weights; The fitting module is configured to use an anisotropic weighted function to perform moving least squares fitting on other points in the local geometric neighborhood to obtain an ideal reference surface passing through the data points; the ideal reference surface is used to fit the macroscopic surface morphology of the automotive steel wire after defects are removed. The output module is configured to calculate the directed normal distance between the data point and the projection point on the ideal reference plane, determine the adaptive threshold corresponding to the data point, and output the defect detection results of the automotive steel wire surface based on the directed normal distance corresponding to the data point and the adaptive threshold.

2. The automotive steel wire surface defect detection system according to claim 1, characterized by, The local geometric variation index corresponding to the data point is constructed in the following way: The covariance matrix eigenvalues ​​are decomposed into eigenvalues ​​and normal vectors for the point set in the local geometric neighborhood. The surface curvature components are calculated using the eigenvalue set and the normal vector consistency components are calculated using the normal vector set. Local geometric variation index ,in, , and Let be three non-negative eigenvalues ​​in the eigenvalue set, and , This represents the number of points within the local geometric neighborhood. The unit normal vector of the data point and the first The dot product of the unit normal vectors of the neighboring points, where exp is the natural exponential function.

3. The automotive steel wire surface defect detection system according to claim 1, wherein The geometric confidence weights are determined in the following way: ,in, Geometric confidence weights; is the sensitivity adjustment coefficient, used to control the decay rate of the weight as the difference changes; Q is the local geometric variation index of the data point; The median of the local geometric variation index of all points within the local geometric neighborhood; To take the absolute value.

4. The automotive steel wire surface defect detection system according to claim 1, characterized by, The anisotropy weighting function of the target data point is obtained by multiplying the spatial distance weight and the geometric confidence weight. The spatial distance weight is determined by using a Gaussian function based on the Euclidean distance from the neighboring data points in the local geometric neighborhood to the target data point, and it decays exponentially with the increase of the Euclidean distance.

5. The automotive steel wire surface defect detection system according to claim 1, wherein The adaptive threshold corresponding to the data point is determined in the following way: ,in, An adaptive threshold for the target data points; Based on the basic noise threshold, The preset confidence coefficient, The number of data points within the local geometric neighborhood of the target data point; For the local geometric neighborhood, the first The depth of each data point; This is the average depth of all data points within the local geometric neighborhood.

6. The automotive steel wire surface defect detection system according to claim 1, characterized in that, The local geometric neighborhood of a data point is constructed in the following way: The central axis of the 3D point cloud data is extracted, and an orthogonal curve coordinate system is constructed based on the central axis. The 3D point cloud data is then expanded and mapped to a 2D parametric space image using a manifold mapping algorithm. The horizontal axis of the two-dimensional parametric space image is the arc length along the central axis, and the vertical axis is the circumferential angle. The 3D point cloud data corresponding to the neighboring pixels of the target pixel in the 2D parameter space image are used as data points in the local geometric neighborhood of the target pixel.

7. The automotive steel wire surface defect detection system according to claim 6, wherein The central axis of the 3D point cloud data is obtained in the following way: The skeleton extraction algorithm is used to perform iterative shrinkage calculation on 3D point cloud data. During the iterative shrinkage calculation, the target spatial location of the skeleton point to be determined is found. The sum of the Euclidean distances of all data points in the local neighborhood of the target spatial location to the skeleton point to be determined is minimized. The target spatial location is determined as the skeleton point, and all skeleton points determined by the skeleton extraction algorithm are fitted and connected in sequence to obtain the central axis of the 3D point cloud data.

8. The automotive steel wire surface defect detection system of claim 1, wherein The ideal reference plane passing through the data points is obtained in the following way: A second-order polynomial model is established in the local tangent plane coordinate system with the data points as the origin, and a weighted residual energy function is constructed. The weighted residual energy function is equal to the sum of the products of the squares of the fitting residuals of each point in the local geometric neighborhood and the corresponding anisotropic weighted functions. Minimize the weighted residual energy function to obtain the polynomial coefficient vector that minimizes the weighted residual energy function; use the polynomial coefficient vector to calculate the fitted coordinates of the data points in the local tangent plane coordinate system, and determine the fitted coordinates as the projection points of the data points on the ideal reference plane, thereby obtaining the ideal reference plane passing through the data points through the projection points.

9. The automotive steel wire surface defect detection system of claim 1, wherein, The acquisition module is also configured to perform the following processing steps: The structured light scanning system is controlled to acquire stripe images using multiple sets of different exposure times and to calculate the intensity modulation and scheduling of each pixel in the stripe image. Light intensity modulation and scheduling are used to characterize the amplitude of periodic changes in light intensity of pixels in multiple stripe images; The light intensity grayscale value of the pixel determines whether the pixel is in a saturated state under the exposure time. For pixels in a saturated state in all exposure channels, interpolation is performed to obtain repair data. The pixels in a non-saturated state and the repair data are then converted to three-dimensional spatial coordinates to obtain three-dimensional point cloud data.

10. The automotive steel wire surface defect detection system according to claim 1, characterized in that, The acquisition module is also configured to perform the following processing steps: Acquire initial point cloud data of the automotive steel wire from multiple perspectives, and determine the target point cloud as the registration reference and the source point cloud to be transformed from the initial point cloud data. By minimizing the objective function to solve for the rotation matrix and translation vector, the source point cloud is transformed into the coordinate system of the target point cloud, so as to obtain three-dimensional point cloud data by stitching together the panoramic point cloud of the car steel wire. The objective function is used to constrain the geometric fit between the source point cloud and the target point cloud on the surface, and to restrict the displacement degree of freedom of the source point cloud relative to the target point cloud along the central axis of the car wire.

Images