Tensor voting principal component analysis based method for defect detection of plate point cloud

Abstract

The present application relates to a kind of plate point cloud defect detection methods based on tensor voting principal component analysis, comprising the following steps, the three-dimensional point cloud data of the plate to be detected is obtained, and the plate point cloud data obtained is preprocessed;Plate point cloud kdtree is built, all point neighborhood point set is found, and is saved;Normal voting tensor is constructed;The voting tensor of point is constructed;The spectral component of normal voting tensor is calculated;The spectral component of point voting tensor is calculated;Defect index is calculated;Defect index threshold is set, and defect region is obtained;Defect point cloud is clustered, and delete threshold is set, and interference defect region is deleted, and finally defect region point cloud minimum bounding box is obtained.The present application has obvious advantages compared with traditional two-dimensional image defect detection, and the detection result is more accurate;Detection speed is faster.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of industrial defect detection, and in particular to a plate point cloud defect detection method based on tensor voting principal component analysis. BACKGROUND

[0002] The difficulties of plate surface defect detection are as follows: 1) environmental light and system light source interference affect imaging effect; 2) strong reflection on the plate surface, the environmental reflection on the plate surface increases the difficulty of on-site installation personnel debugging; 3) and on-site dust, oil stains pollute the plate surface, seriously affecting the image processing process, reducing the yield of production products.

[0003] For strong reflection and interference of plate surface defect detection, traditional two-dimensional image processing algorithm detection is difficult, and practicality is insufficient. Three-dimensional vision information can obtain height information, use three-dimensional vision scanner to extract point cloud data of object surface, then extract features of defect points, calculate difference with standard model, judge whether there is defect through threshold value, and use surrounding calculation defect size, which is a good scheme.

[0004] In the prior art, CN 115100116A proposes a plate defect detection method based on three-dimensional point cloud, which measures related parameters by filtering, coordinate conversion, fitting plane and other operations on the plate three-dimensional point cloud data collected in the production site, and compares with the set standard value and threshold value to judge whether there is defect in size and flatness. However, the core detection method of this method is to fit the plane where the plate is located, set the threshold value, and if there are points higher or lower than the set threshold value, it is determined as a defect. When collecting point cloud, due to the rotation of the motor, the collected point cloud will show sinusoidal interference, and the plate plane will have a certain degree of fluctuation, which will affect the detection of defects and reduce the detection accuracy.

[0005] CN 111539949A proposes a lithium battery pole piece surface defect detection method based on point cloud data, which directly processes coordinate point data through end-to-end learning, and classifies and locates scratches, cracks, bubbles and particles. However, this invention uses three-dimensional deep learning method for defect detection, which needs to collect a large number of defect samples and normal samples, which is time-consuming and laborious, and the detection accuracy is reduced when dealing with new uncertain defects.

[0006] CN 112326671A proposes a kind of metal plate surface defect detection method based on machine vision Red stripe light is projected to plate surface, so that pit, protrusion, scratch and other defects are highlighted, and CCD camera is used to collect stripe projection image;Color image is decomposed, and the color information of light source is highlighted;The center of stripe is extracted, and the distortion of line is judged by algorithm, to reflect the size of defect.The lighting mode of stripe is used, which can eliminate the influence of mirror reflection effect, enhance the appearance of plate surface defect, and obtain defect image with high quality.However, the invention adopts traditional two-dimensional detection method, and there are the following three problems: environmental light and system light source form interference, affect imaging effect;Plate surface exists strong light reflection, and the environment of reflection on plate surface is buried in certain hidden danger after imaging post-processing, which increases the debugging difficulty of on-site installation personnel;And dust, oil stain on site pollutes plate surface and interferes, seriously affects image processing process, and reduces the yield of production product.

[0007] And the disclosure document: "Highly reflective metal surface defect detection method based on HDRI" according to the reflection of object surface, the measured object is divided into weak mirror reflection and strong mirror reflection two kinds, a solution is proposed to realize the measurement of object with high dynamic range or mirror surface by combining adaptive optimal exposure time, adding orthogonal polarizer in front of camera and projector, and appropriately changing the angle between the transmission axes of the used polarizer.However, this document realizes the measurement of object with high dynamic range or mirror surface by combining adaptive optimal exposure time, adding orthogonal polarizer in front of camera and projector, and appropriately changing the angle between the transmission axes of the used polarizer.The disadvantages and causes of this method are consistent with the prior art CN 112326671A, which belongs to two-dimensional imaging range. SUMMARY

[0008] The technical problem to be solved by the present application is to provide a kind of plate point cloud defect detection method based on tensor voting principal component analysis, by pre-processing the three-dimensional point cloud data of plate collected on site, using tensor voting principal component analysis method, calculating the defect index of each point, defect region minimum bounding box and other related parameters, comparing with the set standard value and threshold value, judging whether there is protrusion, depression, scratch and other defects, and giving defect size.

[0009] The technical solution adopted by the present application to solve its technical problems is: a kind of plate point cloud defect detection method based on tensor voting principal component analysis, comprising the following steps,

[0010] S1, obtain the three-dimensional point cloud data of the plate to be detected, and pre-process the obtained plate point cloud data;

[0011] S2, construct a plate material point cloud kdtree, find all point neighborhood point sets, and save;

[0012] S3, construct a normal voting tensor;

[0013] S4, construct a point voting tensor;

[0014] S5, calculate the spectral component of the normal voting tensor;

[0015] S6, calculate the spectral component of the point voting tensor;

[0016] S7, calculate a defect index;

[0017] S8, set a defect index threshold to obtain a defect area;

[0018] S9, cluster the defect point cloud, set a deletion threshold, delete interference defect areas, and finally obtain a defect area point cloud minimum bounding box.

[0019] Further, in step S3 of the application, the normal voting tensor of the i-th point is the sum of the weighted covariance matrices of the normal of the point set in the kdtree neighborhood, and is:

[0020]

[0021] wherein, is a Gaussian weighting function of the neighborhood point, pi and pj represent the i-th point and the j-th point of the point cloud, N(j) represents the set of all points in the neighborhood of the i-th point, and increases with the distance between pi and pj decreasing; the standard deviation σ of the point is usually set to the average distance between the kdtree neighborhood points according to experience; ni and nj represent the normal of the i-th point and the j-th point of the point cloud, the symbol represents the outer product of the normal of the point

[0022] Further, in step S4 of the application, the point voting tensor of the i-th point is the sum of the weighted covariance matrices of the normal of the point set in the kdtree neighborhood, and is:

[0023]

[0024] wherein, is a Gaussian weighting function of the neighborhood point, and increases with the distance between ni and nj decreasing; and the normal standard deviation σ is set according to different modes, pi and pj represent the i-th point and the j-th point of the point cloud, and ni and nj represent the normal of the i-th point and the j-th point of the point cloud.

[0025] Further, in step S5 of the application, the spectral component of the normal voting tensor is calculated:

[0026]

[0027] wherein, and is the eigenvalue and eigenvector corresponding to the i-th point voting tensor.

[0028] Further, in the step S6, the spectral components of the point voting tensor are calculated.

[0029]

[0030] wherein, and is the eigenvalue and eigenvector corresponding to the i-th point voting tensor.

[0031] Further, in the step S7, the defect index R f，i of the i-th point is calculated as follows:

[0032]

[0033] wherein, is the eigenvalue corresponding to the i-th point voting tensor, is the eigenvalue corresponding to the i-th point voting tensor.

[0034] Further, in the step S9, the clustering method includes kmean clustering, meanshif clustering, dichotomy or DBSCAN clustering.

[0035] The present application has the advantages of solving the defects in the prior art, having obvious advantages over traditional two-dimensional image defect detection in three-dimensional space defects such as scratches, grooves and breakage, containing more details, being more accurate in detecting spatial information such as defect volume and surface area, eliminating the influence of uncertain factors such as image surface texture, dirt and reflection, and being more accurate in detection results; eliminating the large amount of time consumed by point cloud registration, being more accurate in detection results, being faster in detection speed, and being less labor-intensive in automatic detection. BRIEF DESCRIPTION OF DRAWINGS

[0036] Figure 1 is a flowchart of the present application. DETAILED DESCRIPTION

[0037] The present application will now be further described in detail by reference to the drawings and preferred embodiments. These drawings are simplified schematic diagrams and only show the basic structure of the present application in a schematic manner, and thus only show the components related to the present application.

[0038] As Figure 1A plate point cloud defect detection method based on tensor voting principal component analysis is shown, comprising the following steps,

[0039] 1. Point cloud acquisition

[0040] Acquire the three-dimensional point cloud data of the plate to be detected by a 3D device, which includes a 3D laser contour scanner, a binocular structured light, etc.

[0041] 2. Point cloud preprocessing

[0042] The obtained plate point cloud is preprocessed, which includes point cloud simplification, outlier removal, point cloud smoothing, etc.

[0043] 3. Defect detection

[0044] 1) Construct the plate point cloud kdtree, find all point sets in the neighborhood, and save;

[0045] 2) Construct the normal voting tensor, the normal voting tensor of the ith point is the sum of the weighted covariance matrices of the normal of its kdtree neighborhood point set, which is:

[0046]

[0047] Wherein, is a Gaussian weighting function of the neighborhood point, pi and pj represent the ith point and the jth point of the point cloud, N(j) represents the set of all points in the neighborhood of the ith point, which increases with the distance between pi and pj decreasing; The standard deviation σ of the point is usually set according to experience as the average distance between the kdtree neighborhood points; ni and nj represent the normal of the ith point and the jth point of the point cloud, The symbol represents the outer product of the point normal

[0048] 3) Construct the point voting tensor, the point voting tensor of the ith point is the sum of the weighted covariance matrices of the normal of its kdtree neighborhood point set, which is:

[0049]

[0050] Wherein, is a Gaussian weighting function of the neighborhood point, which increases with the distance between ni and nj decreasing; and The normal standard deviation σ is set according to different modes, pi and pj represent the ith point and the jth point of the point cloud, and ni and nj represent the normal of the ith point and the jth point of the point cloud.

[0051] 4) Calculate the spectral components of the normal voting tensor according to step 2):

[0052]

[0053] where, and are the eigenvalue and eigenvector of the i-th point normal voting tensor.

[0054] 5) Calculate the spectral components of the point voting tensor from step 3):

[0055]

[0056] where, and are the eigenvalue and eigenvector of the i-th point voting tensor.

[0057] 6) Calculate the i-th point defect indicator R f，i from steps 4 and 5:

[0058]

[0059] where, is the eigenvalue of the i-th point normal voting tensor, is the eigenvalue of the i-th point voting tensor.

[0060] 7) Set the threshold values of the defect indicators and Loop through each point in the plate, calculate the values of and If both and are greater than the threshold values, it is determined as a defect point, otherwise it is a non-defect point.

[0061] 4. Cluster the extracted defect point cloud, the clustering methods include kmean clustering, meanshif clustering, dichotomy, DBSCAN clustering, etc.

[0062] 5. Set a deletion threshold, when the number of points in the clustered block is less than the value, exclude the interference defect area, and finally the remaining is the defect area point cloud.

[0063] 6. Find the minimum bounding box of the defect area point cloud to get the defect size and defect position.

[0064] In this method, "plate" refers to a series of planar materials such as wood, steel, glass, etc. According to other embodiments, the "plate" category can be different, the pre-processing method can be different, the point cloud acquisition method can be different, and the defect area selection method can also be different.

[0065] The application is a kind of new point cloud defect detection method; in the calculation, the interference of other points in the neighborhood is considered, the distance and normal relationship of the points in its neighborhood are used to exclude the interference of some wrong neighborhood points, the weight of the wrong neighborhood points is reduced, and the effect is improved. In engineering application, whether the setting of parameters is convenient or not is also an important index, which relates to whether the algorithm is convenient to land in industry. Obviously, the principal component analysis method of tensor voting has a larger parameter adjustment space, which makes parameter adjustment more convenient and facilitates correct use by personnel who are not familiar with the algorithm. And the principal component analysis method of tensor voting detection does not belong to the field of deep learning, and does not need to collect a large number of sample sets; at the same time, the 3D imaging method is adopted, which greatly reduces the interference of environmental light, material reflection and oil dirt.

[0066] The above description is only a specific embodiment of the application, and various examples do not limit the essential content of the application, and those skilled in the art can modify or deform the previously described specific embodiments without departing from the essence and scope of the application after reading the specification.

Claims

1. A method for defect detection of a plate point cloud based on tensor voting principal component analysis, characterized in that: The method comprises the following steps: S1, obtaining three-dimensional point cloud data of a to-be-detected board, and preprocessing the obtained board point cloud data; S2, constructing a board point cloud kdtree, finding a point set in a neighborhood of all points, and saving the point set; S3, constructing a normal voting tensor, and the normal voting tensor of an i-th point is a sum of weighted covariance matrices of normal points in a neighborhood of the i-th point; S4, constructing a point voting tensor, and the point voting tensor of the i-th point is a sum of weighted covariance matrices of points in a neighborhood of the i-th point; S5, calculating a spectral component of the normal voting tensor: wherein, and is the eigenvalue and eigenvector corresponding to the i-th point normal voting tensor; S6, calculating a spectral component of the point voting tensor: wherein, and are the eigenvalue and eigenvector corresponding to the ith point voting tensor; S7, calculating a defect index of the i-th point; S8, setting a defect index threshold value, and obtaining a defect region; S9, clustering defect point clouds, setting a deletion threshold value, deleting an interference defect region, and finally obtaining a minimum bounding box of the defect region point clouds.

2. The method of claim 1, wherein the method is based on tensor voting principal component analysis. In the step S3, the normal voting tensor of the i-th point is: wherein, is a Gaussian weighting function of the neighborhood points, pi and pj represent the i-th point and the j-th point of the point cloud, N(i) represents the set of all neighborhood points of the i-th point, and increases with the distance between pi and pj decreasing; the standard deviation σ of the point is usually set as kdtree the average distance between the neighborhood points; nj represents the normal of the j-th point of the point cloud, the symbol represents the outer product of the normal of the point 3. The method of claim 1, wherein the method is based on tensor voting principal component analysis. In the step S4, the point voting tensor of the i-th point is: wherein, is a Gaussian weighting function of the neighborhood points, which increases with the distance between ni and nj; and The normal standard deviation σ is set according to different modes, pj represents the jth point of the point cloud, N(i) represents the set of all points in the neighborhood of the ith point, and ni and nj represent the normal of the ith point and the jth point of the point cloud, respectively.

4. The method of claim 1, wherein the method is based on tensor voting principal component analysis for defect detection in a sheet metal point cloud. In step S7, the defect index R at the i-th point f，i The calculation method is as follows: wherein, is the eigenvalue corresponding to the i-th point normal voting tensor, is the eigenvalue corresponding to the i-th point voting tensor.

5. The method of claim 1, wherein the method is based on tensor voting principal component analysis for defect detection in a sheet metal point cloud. In the step S9, the clustering method comprises kmean clustering, meanshif clustering, dichotomy, or DBSCAN clustering.

Images