A bearing rolling body surface defect visual detection method

Abstract

The application relates to the technical field of defect identification, and discloses a bearing rolling element surface defect visual detection method, which comprises the following steps: constructing a mapping relationship from an image sequence to a three-dimensional spherical surface of a bearing rolling element; performing sub-pixel level registration on multiple surface images, and performing shape-preserving mapping on the multiple surface images to obtain distortion-free developed images; performing feature decoupling on the distortion-free developed images to obtain a multi-dimensional defect sensitive feature set; performing multi-channel defect feature classification on the distortion-free developed images to obtain a suspicious area list; performing confidence evaluation on defect probabilities of the suspicious area list, and performing super-resolution re-imaging on defect areas with a confidence lower than a preset threshold to obtain defect accurate positioning and classification results; and comprehensively evaluating the defect accurate positioning and classification results to obtain a surface quality evaluation report; and the application can improve the efficiency of bearing rolling element surface defect visual detection.

Description

Technical Field

[0001] This invention relates to the field of defect identification technology, and in particular to a visual inspection method for defects on the surface of bearing rolling elements. Background Technology

[0002] In the field of visual inspection of surface defects on bearing rolling elements, traditional inspection methods struggle to construct accurate mappings from multiple surface images and rotational pose information to the three-dimensional spherical surface of the rolling element. Trajectory tracking of key feature points lacks spatiotemporal consistency constraints, and the spatial curve fitting accuracy after backprojection is low. This results in a lack of reliable coordinate references for subsequent image registration, failing to provide a high-quality image foundation for defect detection. Traditional methods also struggle to achieve sub-pixel accuracy when registering multiple surface images. Furthermore, they lack effective conformal mapping techniques when unfolding spherical images, leading to severely distorted unfolded images that fail to accurately reproduce the texture and defect information of the rolling element surface, directly hindering the effective extraction of subsequent defect features.

[0003] Existing defect detection technologies lack sufficient feature decoupling capabilities for distortion-free unfolded images, failing to adequately separate key defect features such as orientation sensitivity, enhanced local contrast, and color space anomalies. The extracted features are often one-dimensional and weakly correlated, making it difficult to form a comprehensive multi-dimensional set of defect-sensitive features. Furthermore, existing technologies lack scientific confidence assessment criteria when evaluating the probability of defects in suspicious areas. They do not employ effective super-resolution re-imaging mechanisms for low-confidence regions, relying solely on raw image information for defect determination. This results in insufficient accuracy in defect localization, poor reliability of defect classification results, and ultimately, surface quality assessment reports that fail to meet the stringent requirements of high-precision inspection in industrial production. Summary of the Invention

[0004] This invention provides a visual inspection method for surface defects of bearing rolling elements to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides a visual inspection method for surface defects of bearing rolling elements, comprising:

[0006] S1. Based on the multi-frame surface images and rotational pose information of the bearing rolling elements, construct the mapping relationship between the bearing rolling elements from the image sequence to the three-dimensional spherical surface of the rolling elements;

[0007] S2. Based on the mapping relationship, subpixel-level registration is performed on the multi-frame surface images, and conformal mapping is performed on the aligned multi-frame surface images to obtain a distortion-free unfolded image of the bearing rolling element.

[0008] S3. Decouple the orientation sensitivity, local contrast enhancement and color space anomaly of the distortion-free unfolded image to obtain the multidimensional defect-sensitive feature set of the bearing rolling element;

[0009] S4. Based on the multidimensional defect sensitive feature set, perform multi-channel defect feature classification on the linear texture direction feature, morphological depression feature and color variation feature of the distortion-free unfolded image to obtain a list of suspicious areas of the bearing rolling element.

[0010] S5. Calculate the confidence level of the defect probability in the list of suspicious areas, and perform super-resolution re-imaging on defect areas with confidence levels below a preset threshold to obtain the accurate location and classification results of defects in the bearing rolling elements.

[0011] S6. The precise location of the defect and the classification results are comprehensively evaluated to obtain a surface quality assessment report of the bearing rolling element.

[0012] In a preferred embodiment, the step of constructing a mapping relationship between the bearing rolling elements and the three-dimensional spherical surface of the rolling elements from the image sequence, based on multi-frame surface images and rotational pose information of the bearing rolling elements, includes:

[0013] From the multi-frame surface images of the bearing rolling elements, key feature frames that meet the requirements of clarity and view coverage are selected to obtain the key feature point set of the bearing rolling elements.

[0014] Based on the rotational pose information of the bearing rolling elements, spatiotemporal consistency tracking is performed on a continuous set of key feature points to obtain the trajectory tracking chain of the key feature point set.

[0015] The trajectory tracking chain is back-projected onto the candidate space curve of the three-dimensional sphere of the rolling body, and the candidate space curve is optimized and fitted to obtain the unique spatial position of the key feature point set.

[0016] Using the unique spatial location as the control point, progressive surface parameterization is employed to construct the mapping relationship between the bearing rolling element and the three-dimensional spherical surface of the rolling element from the image sequence.

[0017] In a preferred embodiment, the step of performing subpixel-level registration on the multi-frame surface images based on the mapping relationship, and performing conformal mapping on the aligned multi-frame surface images to obtain a distortion-free unfolded image of the bearing rolling element includes:

[0018] Based on the mapping relationship, dense matching is performed on the multi-frame surface images to obtain the sub-pixel level displacement field of the bearing rolling element;

[0019] Local deformation correction is performed on the subpixel-level displacement field to obtain the global deformation field of the bearing rolling element;

[0020] Based on the global deformation field, the multi-frame surface images are interpolated and resampled to obtain a high-precision aligned image of the bearing rolling elements;

[0021] Based on the three-dimensional spherical surface of the rolling element, the high-precision aligned image is divided into spherical triangular meshes to obtain the spherical parametric mesh system of the bearing rolling element.

[0022] Based on the spherical parametric mesh system, conformal mapping is performed on the texture information in the high-precision aligned image to obtain a distortion-free unfolded image of the bearing rolling element.

[0023] In a preferred embodiment, the step of performing conformal mapping on the texture information in the high-precision aligned image according to the spherical parametric mesh system to obtain a distortion-free unfolded image of the bearing rolling element includes:

[0024] Texture extraction is performed on the spherical triangular facets in the high-precision aligned image to obtain local texture blocks of the bearing rolling elements;

[0025] Based on the adjacent topological relationships and geometric features of the spherical triangular facets, conformal mapping is performed on the local texture blocks to obtain the planar texture blocks of the bearing rolling elements;

[0026] Based on the mesh density distribution of the spherical parametric mesh system, the planar texture blocks are divided into fusion regions to obtain a multi-level fusion strategy for the bearing rolling elements;

[0027] The multi-level fusion strategy is executed to progressively stitch the planar texture blocks and perform gradient domain fusion during the stitching process to obtain a distortion-free unfolded image of the bearing rolling elements.

[0028] In a preferred embodiment, the decoupling of orientation sensitivity, local contrast enhancement, and color space anomalies in the distortion-free unfolded image to obtain a multidimensional defect-sensitive feature set of the bearing rolling element includes:

[0029] Based on the gray-level gradient distribution of the distortion-free unfolded image, the texture direction of the distortion-free unfolded image is evaluated, and directional filtering is performed based on the evaluated main texture direction to obtain the directional enhancement feature map of the bearing rolling element.

[0030] The distortion-free unfolded image is divided into multi-scale local regions. The gray-level distribution range and standard deviation within each region are statistically analyzed at each scale. Based on the statistical results, the original gray-level values ​​of the distortion-free unfolded image are subjected to background texture suppression to obtain the local contrast depth map of the bearing rolling element.

[0031] Based on the Mahalanobis distance between the pixel color components in the distortion-free unfolded image and the preset normal color template, the pixels in the distortion-free unfolded image are subjected to region clustering growth to obtain the color abnormality region map of the bearing rolling element.

[0032] The formula for calculating the Mahalanobis distance is as follows:

[0033] ;

[0034] In the formula, The Mahalanobis distance is... The observed sample feature value in the pixel color component, This is the mean vector of the pixel color components. Let covariance matrix be the variance matrix. It is the inverse of the covariance matrix. This is the transpose operation for a matrix.

[0035] The directional enhancement feature map, the local contrast depth map, and the color abnormality region map are fused to obtain a multidimensional defect-sensitive feature set for the bearing rolling element.

[0036] In a preferred embodiment, the step of evaluating the texture orientation of the distortion-free unfolded image based on its gray-level gradient distribution, and performing directional filtering based on the evaluated main texture orientation to obtain the orientation enhancement feature map of the bearing rolling element includes:

[0037] Orientation statistics are performed on the undistorted unfolded image to obtain the main texture orientation field of the bearing rolling element;

[0038] Based on the main texture direction field, the direction-selective filter kernel of the distortion-free unfolded image is adaptively generated to obtain the filter kernel of the bearing rolling element;

[0039] Based on the filter kernel, the distortion-free unfolded image is convolutionally filtered to obtain the initial response map of the bearing rolling element;

[0040] Non-maximum suppression is applied to the initial response map to obtain the directional enhancement feature map of the bearing rolling element.

[0041] In a preferred embodiment, based on the multidimensional defect-sensitive feature set, multi-channel defect feature classification is performed on the linear texture direction features, morphological depression features, and color variation features of the distortion-free unfolded image to obtain a list of suspicious regions of the bearing rolling element, including:

[0042] Based on the multidimensional defect-sensitive feature set, structural tensor analysis is performed on the linear texture direction features of the distortion-free unfolded image to obtain the linear structural skeleton diagram of the bearing rolling element.

[0043] Morphological filling is performed on the linear structural skeleton diagram to obtain the candidate region for crack defects in the bearing rolling element;

[0044] Based on the multidimensional defect-sensitive feature set, local extreme points are detected for the morphological concavity features of the distortion-free unfolded image, and region growth is performed with each extreme point as the center to obtain the candidate concavity region of the bearing rolling element.

[0045] The candidate depression region is reconstructed in three dimensions to obtain the candidate pit defect region of the bearing rolling element;

[0046] Based on the multidimensional defect-sensitive feature set, color variation features of the distortion-free unfolded image are clustered in color saturation space to obtain candidate areas for burn defects of the bearing rolling elements.

[0047] Spatial topological relationship analysis is performed on the candidate areas of crack defects, pit defects, and burn defects to obtain a list of suspicious areas of the bearing rolling elements.

[0048] In a preferred embodiment, the step of assessing the confidence level of the defect probability in the list of suspected areas and performing super-resolution re-imaging on defect areas with a confidence level below a preset threshold to obtain the precise location and classification results of defects in the bearing rolling elements includes:

[0049] Extract the response intensity and spatial consistency features of the list of suspicious regions on the multidimensional defect-sensitive feature set to obtain the confidence evaluation data of the bearing rolling elements;

[0050] Based on the confidence assessment data, all regions in the list of suspicious regions are divided into high-confidence regions and low-confidence regions, and the low-confidence regions are classified as the targets to be re-inspected for the bearing rolling elements;

[0051] Based on the position information of the target to be re-inspected in the distortion-free unfolded image, the original image patch set of the target to be re-inspected in the original multi-frame surface image sequence is back-located.

[0052] High-frequency information fusion is performed on the original image patch set to obtain a super-resolution local image of the bearing rolling element;

[0053] Feature extraction and analysis are performed on the orientation sensitivity, local contrast enhancement, and color space anomalies of the super-resolution local image to obtain the resampling feature description of the bearing rolling element;

[0054] The re-sampling feature description is compared and corrected with the original feature description of the corresponding region in the list of suspicious regions to obtain the accurate location and classification result of the defect of the bearing rolling element.

[0055] In a preferred embodiment, the step of performing high-frequency information fusion on the original image patch set to obtain a super-resolution local image of the bearing rolling element includes:

[0056] Based on the phase correlation method, the original image block set is finely aligned to obtain the sub-image block sequence of the bearing rolling element;

[0057] The grayscale observations of the sub-image block sequence are arranged and collected to obtain the temporal grayscale observation set of the bearing rolling element;

[0058] By performing iterative back projection on the set of grayscale observations, an initial high-resolution image of the bearing rolling element is obtained;

[0059] Noise suppression is applied to the initial high-resolution image to obtain a super-resolution local image of the bearing rolling element.

[0060] In a preferred embodiment, the step of comprehensively evaluating the precise location of the defect and the classification results to obtain a surface quality assessment report for the bearing rolling element includes:

[0061] From the precise location of the defects and the classification results, the defect type, precise boundary coordinates and coverage area of ​​the bearing rolling element are extracted to obtain the defect attribute set of the bearing rolling element.

[0062] Based on the preset allowable standards for the size and quantity of various defects, the defect attribute set is classified into levels to obtain the defect level list of the bearing rolling elements;

[0063] The defect attribute set and the defect level list are rearranged and reorganized to obtain the surface quality assessment report of the bearing rolling elements.

[0064] Compared with the prior art, the present invention has the following beneficial effects:

[0065] 1. This invention constructs a precise mapping relationship between image sequences and a three-dimensional sphere based on multi-frame surface images and rotational pose information of the bearing rolling elements. Through sub-pixel level registration and conformal mapping processing, distortion-free unfolded images are obtained, laying a high-quality image foundation for subsequent defect detection. The distortion-free unfolded images undergo feature decoupling based on orientation sensitivity, local contrast enhancement, and color space anomalies to form a multi-dimensional defect-sensitive feature set. This set can comprehensively capture various defect-related features of the rolling element surface, improving the completeness and accuracy of feature extraction.

[0066] 2. This invention utilizes a multi-dimensional defect-sensitive feature set to perform multi-channel defect feature classification, accurately identifying different types of defect candidate areas such as cracks, pits, and burns. Combined with confidence assessment and super-resolution re-imaging technology, low-confidence defect areas are re-examined and feature-corrected, achieving precise defect location and classification. The final surface quality assessment report includes detailed defect types, boundary coordinates, coverage area, and defect level, providing a comprehensive and reliable basis for judging the surface quality of bearing rolling elements, improving overall detection efficiency and judgment accuracy. Attached Figure Description

[0067] Figure 1 This is a flowchart illustrating a visual inspection method for surface defects of rolling elements in a bearing, according to an embodiment of the present invention.

[0068] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0069] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0070] This application provides a visual inspection method for surface defects of bearing rolling elements. The execution subject of this method includes, but is not limited to, at least one of the following electronic devices that can be configured to execute the method provided in this application: a server, a terminal, etc. In other words, the visual inspection method for surface defects of bearing rolling elements can be executed by software or hardware installed on a terminal device or a server device. The server includes, but is not limited to, a single server, a server cluster, a cloud server, or a cloud server cluster. The server can be an independent server or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDN), and big data and artificial intelligence platforms.

[0071] Reference Figure 1 The diagram shown is a flowchart illustrating a visual inspection method for surface defects of bearing rolling elements according to an embodiment of the present invention. In this embodiment, the visual inspection method for surface defects of bearing rolling elements includes:

[0072] S1. Based on the multi-frame surface images and rotational pose information of the bearing rolling elements, construct the mapping relationship between the bearing rolling elements from the image sequence to the three-dimensional spherical surface of the rolling elements;

[0073] In this embodiment of the invention, the step of constructing a mapping relationship between the bearing rolling element and the three-dimensional spherical surface of the rolling element based on multi-frame surface images and rotational pose information of the bearing rolling element includes:

[0074] From the multi-frame surface images of the bearing rolling elements, key feature frames that meet the requirements of clarity and view coverage are selected to obtain the key feature point set of the bearing rolling elements.

[0075] Based on the rotational pose information of the bearing rolling elements, spatiotemporal consistency tracking is performed on a continuous set of key feature points to obtain the trajectory tracking chain of the key feature point set.

[0076] The trajectory tracking chain is back-projected onto the candidate space curve of the three-dimensional sphere of the rolling body, and the candidate space curve is optimized and fitted to obtain the unique spatial position of the key feature point set.

[0077] Using the unique spatial location as the control point, progressive surface parameterization is employed to construct the mapping relationship between the bearing rolling element and the three-dimensional spherical surface of the rolling element from the image sequence.

[0078] From multiple frames of surface images of the bearing rolling element, the gray-level difference of each pixel neighborhood in each frame is calculated to obtain the gray-level gradient amplitude. Images with gray-level gradient amplitudes greater than a preset sharpness threshold and whose single-frame image view covers a proportion of the three-dimensional spherical area of ​​the rolling element reaching a preset coverage threshold are selected. These images that meet the conditions are determined as key feature frames. Then, from the key feature frames, pixels with edge gradient values ​​higher than a preset edge threshold, clear edge contours, and stable texture information with regular pixel gray-level distribution are selected. These pixels are integrated to obtain the key feature point set of the bearing rolling element.

[0079] Based on the rotation angle values ​​corresponding to each frame of the image contained in the rotation pose information of the bearing rolling element and the relative spatial position coordinates of the image acquisition device and the rolling element, the position distance and texture similarity of pixels in adjacent key feature frames are calculated. Key feature points that are close in position and have consistent texture features with a position distance lower than a preset distance threshold and a texture similarity higher than a preset similarity threshold are associated frame by frame, so that each key feature point forms a continuous and uninterrupted spatial movement path, and the trajectory tracking chain of the key feature point set is obtained.

[0080] Based on the standard geometric model of the three-dimensional sphere of the rolling element and the preset sphere radius, combined with the internal and external parameters of image acquisition, the two-dimensional image coordinates of each key feature point on the trajectory tracking chain are converted into corresponding three-dimensional spherical space coordinates. Then, the converted three-dimensional spherical space coordinates are projected onto multiple candidate three-dimensional spherical space curves with different curvatures pre-set according to the geometric characteristics of the rolling element's sphere. The vertical distance between the projection point and the curve is calculated point by point. By adjusting the curvature value and spatial position parameters of the candidate space curves, the positional deviation values ​​between the curves and all projection points are made lower than the preset deviation threshold, thus completing the optimization fitting operation of the candidate space curves and obtaining the unique spatial position of the key feature point set.

[0081] Using the unique spatial location as the basic control point, the entire area of ​​the three-dimensional sphere of the rolling element is gradually divided into grids according to a preset grid density. Three-dimensional coordinate information is assigned to each grid cell after division, establishing a one-to-one correspondence between each pixel in the image sequence and the spatial coordinates of the three-dimensional sphere of the rolling element. A progressive surface parameterization method is adopted, which gradually expands from the control point to the surrounding adjacent grid cells, to construct the mapping relationship between the bearing rolling element and the three-dimensional sphere of the rolling element from the image sequence.

[0082] The beneficial effects include screening key feature frames that meet the requirements of clarity and view coverage, and extracting key feature point sets, which can ensure the basic data quality of subsequent processing. Based on the rotational pose information of the bearing rolling elements, spatiotemporal consistency tracking of continuous key feature point sets is carried out to form a trajectory tracking chain, which can enhance the spatial correlation of feature points. The trajectory tracking chain is back-projected onto the candidate space curve of the three-dimensional spherical surface of the rolling element and optimized and fitted to obtain the unique spatial position of the key feature point set, which can improve the accuracy of feature point spatial positioning. Using this unique spatial position as the control point, the mapping relationship between the image sequence and the three-dimensional spherical surface of the rolling element constructed by progressive surface parameterization can realize the precise correspondence between image pixels and spherical spatial coordinates, providing a reliable coordinate reference for subsequent sub-pixel level registration and distortion-free unfolded image acquisition, thereby improving the accuracy and reliability of the overall visual inspection of bearing rolling element surface defects.

[0083] S2. Based on the mapping relationship, subpixel-level registration is performed on the multi-frame surface images, and conformal mapping is performed on the aligned multi-frame surface images to obtain a distortion-free unfolded image of the bearing rolling element.

[0084] In this embodiment of the invention, the step of performing sub-pixel level registration on the multi-frame surface images based on the mapping relationship, and performing conformal mapping on the aligned multi-frame surface images to obtain a distortion-free unfolded image of the bearing rolling element includes:

[0085] Based on the mapping relationship, dense matching is performed on the multi-frame surface images to obtain the sub-pixel level displacement field of the bearing rolling element;

[0086] Local deformation correction is performed on the subpixel-level displacement field to obtain the global deformation field of the bearing rolling element;

[0087] Based on the global deformation field, the multi-frame surface images are interpolated and resampled to obtain a high-precision aligned image of the bearing rolling elements;

[0088] Based on the three-dimensional spherical surface of the rolling element, the high-precision aligned image is divided into spherical triangular meshes to obtain the spherical parametric mesh system of the bearing rolling element.

[0089] Based on the spherical parametric mesh system, conformal mapping is performed on the texture information in the high-precision aligned image to obtain a distortion-free unfolded image of the bearing rolling element.

[0090] The step of performing conformal mapping on the texture information in the high-precision aligned image according to the spherical parametric mesh system to obtain a distortion-free unfolded image of the bearing rolling element includes:

[0091] Texture extraction is performed on the spherical triangular facets in the high-precision aligned image to obtain local texture blocks of the bearing rolling elements;

[0092] Based on the adjacent topological relationships and geometric features of the spherical triangular facets, conformal mapping is performed on the local texture blocks to obtain the planar texture blocks of the bearing rolling elements;

[0093] Based on the mesh density distribution of the spherical parametric mesh system, the planar texture blocks are divided into fusion regions to obtain a multi-level fusion strategy for the bearing rolling elements;

[0094] The multi-level fusion strategy is executed to progressively stitch the planar texture blocks and perform gradient domain fusion during the stitching process to obtain a distortion-free unfolded image of the bearing rolling elements.

[0095] Based on the mapping relationship, the frame with the highest imaging clarity and the largest viewing angle coverage area among multiple surface images is selected as the reference frame, and all other frames are used as the frames to be matched. The gray-level distribution of the corresponding pixels in the frames to be matched and the reference frame and their 3×3 neighborhoods are compared pixel by pixel. Pixels whose absolute values ​​of the gray-level difference between the corresponding pixels and the neighborhood are all lower than a preset difference threshold are determined as matching points with gray-level similarity higher than a preset similarity threshold. The sub-pixel level position deviation of each matching point in the frames to be matched and the reference frame is determined by gray-level interpolation. The sub-pixel level position deviation data of all matching points are integrated to obtain the sub-pixel level displacement field of the bearing rolling element.

[0096] The subpixel-level displacement field is divided into multiple independent local regions according to the preset region size set by the spherical curvature of the rolling element. The difference between the maximum and minimum values ​​of the displacement of all pixels in each region is calculated as the fluctuation value of the displacement in the region. For regions where the fluctuation value exceeds the preset fluctuation threshold, the abnormal displacement in the region is corrected point by point with reference to the average displacement of the three adjacent regions around the region, so that the displacement in the corrected region is consistent with the displacement in the surrounding regions without abrupt changes. The displacement data of all corrected local regions are integrated to obtain the global deformation field of the bearing rolling element.

[0097] Based on the pixel displacement data recorded in the global deformation field, the position of each pixel in the multi-frame surface image is precisely translated and adjusted pixel by pixel. For the pixel gaps that appear after adjustment, the gray value of all effective pixels in the 3×3 neighborhood around the gap is extracted and filled. For the pixel repetitions that appear after adjustment, the gray value variance between each repetition pixel and the surrounding 8 pixels is calculated. The pixel information with the smallest gray value variance, that is, the pixel with the highest gray value coordination with the surrounding pixels, is retained. The position adjustment and pixel filling operations of all frames of images are completed to obtain the high-precision aligned image of the bearing rolling element.

[0098] Based on the standard geometric dimensions such as the actual spherical diameter and radius of curvature of the rolling element's three-dimensional sphere, as well as the range of three-dimensional spatial coordinates of the sphere, the preset mesh density is adjusted according to the curvature of different regions of the sphere. Regions with larger curvature values ​​are assigned higher mesh densities. The entire sphere is divided into multiple continuous and non-overlapping triangular patches. Each triangular patch is uniquely encoded, including regional position information and vertex coordinate information, and its three-dimensional spherical spatial coordinates of the three vertices are recorded in detail. By integrating the unique encoding information and vertex coordinate information of all triangular patches, the spherical parametric mesh system of the bearing rolling element is obtained.

[0099] Based on the adjacent topological relationships and positional encoding order of each triangular facet in the spherical parametric mesh system, texture information of the pixel region that precisely corresponds to each triangular facet in the high-precision aligned image is extracted. During the conformal mapping process of converting spherical texture information into planar texture information, the interior angle and side length ratio of the texture remain unchanged. All converted planar texture information is sequentially stitched together according to the unique encoding order of the triangular facets. During the stitching process, grayscale gradient calculation is performed on the edge pixels of adjacent texture blocks. The grayscale values ​​of the edge pixels are adjusted point by point according to the gradient change trend to perform smooth transition processing, completely eliminating stitching gaps, and obtaining the distortion-free unfolded image of the bearing rolling element.

[0100] Based on the unique encoding information and precise three-dimensional spherical space coordinates of each triangular facet in the spherical parameterized mesh system, the pixel region that completely corresponds to each triangular facet is located in the high-precision aligned image by coordinate mapping. The boundary vertex coordinates of the pixel region correspond one-to-one with the three-dimensional spherical space coordinates of the triangular facet. All texture detail information and pixel-by-pixel grayscale distribution information within the pixel region are extracted. The matching degree between the boundary pixel coordinates of the texture region and the vertex coordinates of the triangular facet is checked point by point to ensure that the matching deviation value is lower than the preset coordinate deviation threshold, thereby obtaining the local texture block of the bearing rolling element.

[0101] The adjacent topological relationships of local texture blocks are determined by referring to the adjacent connection relationships of the spherical triangular facets. That is, the number and relative positions of adjacent texture blocks of each local texture block are determined by querying the adjacent association table of the triangular facet encoding. Based on the actual side length and interior angles of the spherical triangular facets, the interior angle values ​​of the texture blocks are measured and recorded in real time during the process of converting the spherical local texture blocks into planar texture blocks. This ensures that the interior angle values ​​of the converted planar texture blocks are completely consistent with the interior angle values ​​of the original spherical triangular facets. The side lengths of the spherical triangular facets are scaled according to a uniform ratio. The scaling ratio is determined by the actual radius of the three-dimensional sphere of the rolling element and the preset planar unfolding reference size, so that the topological structure of the converted planar texture blocks is completely consistent with that of the original spherical local texture blocks, thus obtaining the planar texture blocks of the bearing rolling element.

[0102] Based on the mesh density distribution of the spherical parametric mesh system, the mesh density is determined by the number of triangular facets contained in a unit area. The planar texture blocks corresponding to the areas with mesh density higher than a preset density threshold are divided into fine-grained fusion regions, which correspond to key parts of the rolling element surface that are prone to minor defects. The planar texture blocks corresponding to the areas with mesh density lower than the preset density threshold are divided into coarse-grained fusion regions, which correspond to the conventional parts of the rolling element surface with a low defect incidence. The priority of splicing fine-grained fusion regions over coarse-grained fusion regions is clarified. At the same time, the specific rules for using pixel-by-pixel grayscale transition for fine-grained regions and using regional average grayscale transition for coarse-grained regions are determined, thus obtaining the multi-level fusion strategy of the bearing rolling element.

[0103] According to the stitching priority set by the multi-level fusion strategy, the planar texture blocks are stitched step by step starting from the center of the fine-grained fusion region. During the stitching process, the gray-level gradient value of the edge pixels of adjacent planar texture blocks is calculated point by point. The gray-level gradient value is the ratio of the gray-level difference between adjacent pixels to the pixel spacing. The gray-level information of the edge pixels is adjusted point by point according to the gradient change trend. The adjustment range is based on the standard that the gray-level difference between adjacent pixels is lower than the preset gray-level difference threshold, so that the gray-level transition between adjacent texture blocks is continuous without abrupt changes. After the stitching is completed, the gray-level consistency of all stitched parts is checked. Parts that do not meet the gray-level transition requirements are readjusted. The stitching and fusion operation of all planar texture blocks is completed, and the distortion-free unfolded image of the bearing rolling element is obtained.

[0104] The beneficial effects are as follows: dense matching of multiple surface images based on mapping relationships and acquisition of sub-pixel displacement fields can accurately capture pixel-level positional deviations between images; local deformation correction of sub-pixel displacement fields yields global deformation fields, eliminating abrupt displacement changes in local areas and ensuring the spatial continuity of the displacement fields; interpolation and resampling of multiple surface images based on the global deformation fields yields high-precision aligned images, providing a unified and high-quality image foundation for subsequent processing; spherical triangular meshing of high-precision aligned images based on the three-dimensional sphere of the rolling body yields a spherical parametric mesh system, enabling the orderly division and management of spherical textures; local texture blocks are extracted based on the spherical parametric mesh system and conformal mapping is performed to obtain planar texture blocks, maintaining the topological structure and geometric features during texture conversion; a multi-level fusion strategy is formulated based on mesh density distribution, and progressive stitching and gradient domain fusion of planar texture blocks are completed, eliminating texture block stitching gaps and obtaining distortion-free unfolded images, laying a solid foundation for the accurate extraction and classification of subsequent defect features.

[0105] S3. Decouple the orientation sensitivity, local contrast enhancement and color space anomaly of the distortion-free unfolded image to obtain the multidimensional defect-sensitive feature set of the bearing rolling element;

[0106] In this embodiment of the invention, the decoupling of the orientation sensitivity, local contrast enhancement, and color space anomaly of the distortion-free unfolded image to obtain the multidimensional defect-sensitive feature set of the bearing rolling element includes:

[0107] Based on the gray-level gradient distribution of the distortion-free unfolded image, the texture direction of the distortion-free unfolded image is evaluated, and directional filtering is performed based on the evaluated main texture direction to obtain the directional enhancement feature map of the bearing rolling element.

[0108] The distortion-free unfolded image is divided into multi-scale local regions. The gray-level distribution range and standard deviation within each region are statistically analyzed at each scale. Based on the statistical results, the original gray-level values ​​of the distortion-free unfolded image are subjected to background texture suppression to obtain the local contrast depth map of the bearing rolling element.

[0109] Based on the Mahalanobis distance between the pixel color components in the distortion-free unfolded image and the preset normal color template, the pixels in the distortion-free unfolded image are subjected to region clustering growth to obtain the color abnormality region map of the bearing rolling element.

[0110] The formula for calculating the Mahalanobis distance is as follows:

[0111] ;

[0112] In the formula, The Mahalanobis distance is... The observed sample feature value in the pixel color component, This is the mean vector of the pixel color components. Let covariance matrix be the variance matrix. It is the inverse of the covariance matrix. This is the transpose operation for a matrix.

[0113] The directional enhancement feature map, the local contrast depth map, and the color abnormality region map are fused to obtain a multidimensional defect-sensitive feature set for the bearing rolling element.

[0114] The process involves evaluating the texture direction of the distortion-free unfolded image based on its grayscale gradient distribution, and performing directional filtering based on the evaluated main texture direction to obtain the directional enhancement feature map of the bearing rolling element, including:

[0115] Orientation statistics are performed on the undistorted unfolded image to obtain the main texture orientation field of the bearing rolling element;

[0116] Based on the main texture direction field, the direction-selective filter kernel of the distortion-free unfolded image is adaptively generated to obtain the filter kernel of the bearing rolling element;

[0117] Based on the filter kernel, the distortion-free unfolded image is convolutionally filtered to obtain the initial response map of the bearing rolling element;

[0118] Non-maximum suppression is applied to the initial response map to obtain the directional enhancement feature map of the bearing rolling element.

[0119] Based on the gray-level gradient distribution of the distortion-free unfolded image, the gray-level change of each pixel in the image in the horizontal and vertical directions is calculated. Pixels whose sum of gray-level changes is higher than a preset gradient magnitude threshold are marked as texture pixels. The gradient directions of all texture pixels are counted, and the direction with the highest frequency is determined as the main texture direction. A direction-selective filter kernel is generated based on the main texture direction. The distortion-free unfolded image is convolved using the filter kernel to retain the pixel information of the main texture direction and weaken the interference information of other directions to obtain an initial response map. Non-maximum suppression is performed on the initial response map to remove non-extreme pixels with response values ​​lower than those of neighboring pixels, thereby obtaining the direction enhancement feature map of the bearing rolling element.

[0120] The distortion-free unfolded image is divided into multi-scale local regions according to three preset region templates of different sizes. In each local region corresponding to each scale, the difference between the maximum and minimum gray values ​​of all pixels (i.e., the gray-level distribution range) and the deviation of all pixel gray values ​​from the mean gray value of the region (i.e., the standard deviation) are calculated. Pixels whose gray values ​​in the region are lower than the difference between the gray values ​​of the background texture and the gray values ​​of the background texture are identified as background texture pixels. The gray-level weight of these background texture pixels is reduced to obtain the local contrast depth map of the bearing rolling element.

[0121] A preset normal color template for bearing rolling elements is retrieved. This template contains the standard color component range of the bearing rolling elements under defect-free conditions. The Mahalanobis distance between the color component of each pixel in the undistorted unfolded image and the corresponding color component in the normal color template is calculated. Pixels whose Mahalanobis distance exceeds a preset distance threshold are marked as color aberration pixels. With each color aberration pixel as the center, pixels in the neighborhood that meet the color component similarity requirements are included in the same region. The region clustering and growth operation is completed to obtain the color aberration region map of the bearing rolling elements.

[0122] The orientation enhancement feature map, local contrast depth map, and color abnormality area map are precisely aligned according to pixel position. The feature information corresponding to each pixel in the three feature maps is extracted. These feature information are integrated and a one-to-one correspondence is established to form an information set containing three types of features: orientation, contrast, and color. This results in a multidimensional defect sensitive feature set for the bearing rolling element.

[0123] The distortion-free unfolded image is divided into multiple non-overlapping pixel blocks according to a preset fixed size. The grayscale change of all pixels in each pixel block in the horizontal and vertical directions is calculated. Pixels whose sum of grayscale change is higher than a preset gradient magnitude threshold are included in the direction statistics range. The gradient direction of all pixels in this range is counted. The gradient direction with the highest frequency in each pixel block is determined as the main texture direction of the pixel block. The main texture direction information of all pixel blocks is integrated to obtain the main texture direction field of the bearing rolling element.

[0124] Based on the main texture direction of each pixel block in the main texture direction field, the direction of the direction-selective filter kernel is determined to be consistent with the main texture direction. The window size of the filter kernel is determined according to the density of the texture within the pixel block. Pixel blocks with dense texture distribution correspond to smaller filter kernels, and pixel blocks with sparse texture distribution correspond to larger filter kernels. Weights are assigned to pixels at different positions within the filter kernel. The pixel weights in the main texture direction are higher than those in other directions. The parameter setting and structure generation of the filter kernel are completed, and the filter kernel of the bearing rolling element is obtained.

[0125] Based on the filter kernel, a pixel-by-pixel sliding operation is performed on the distortion-free unfolded image. Each pixel within the coverage area of ​​the filter kernel is weighted according to its corresponding weight, retaining the pixel grayscale information in the main texture direction and weakening the pixel grayscale interference in the non-main texture direction. The weighted calculation result of each pixel position is used as the response value of that pixel. By integrating the response value information of all pixels, the initial response map of the bearing rolling element is obtained.

[0126] The initial response map is traversed pixel by pixel. A neighborhood range of a preset size is selected with each pixel as the center. The response value of the center pixel is compared with the response values ​​of all pixels in the neighborhood one by one. Only the pixel information with the response value of the center pixel as the maximum value in the neighborhood is retained. The response values ​​of the remaining pixels are adjusted to the background grayscale value. The local maximum value filtering and retention operation is completed to obtain the direction enhancement feature map of the bearing rolling element.

[0127] The feature values ​​of the observation samples are derived from the color components of each pixel in the distortion-free unfolded image. The color component values ​​corresponding to each pixel in the distortion-free unfolded image are directly extracted as the feature values ​​of the observation samples.

[0128] The mean vector is derived from the color components of all pixels in the distortion-free unfolded image. The average value of all pixel color components in the distortion-free unfolded image is calculated, and all average values ​​are combined to form the mean vector.

[0129] The covariance matrix is ​​derived from the color components of all pixels in the distortion-free unfolded image. The covariance of the color components of all pixels in the distortion-free unfolded image is calculated, and the results of all covariance calculations are integrated to form the covariance matrix.

[0130] The inverse of the covariance matrix is ​​obtained by inverting the covariance matrix. The inverse of the covariance matrix is ​​obtained by performing the inversion operation on the generated covariance matrix.

[0131] The transpose operation of a matrix is ​​derived from the transpose of the difference between the eigenvalues ​​of the observed samples and the mean vector. After calculating the difference between the eigenvalues ​​of the observed samples and the mean vector, the transpose operation is performed on the difference to obtain the transpose result.

[0132] This formula is used to calculate the distance between the color component of each pixel in a distortion-free unfolded image and a preset normal color template.

[0133] The distance calculated by the formula is compared with the preset distance threshold. Pixels whose distance exceeds the preset distance threshold will be marked as pixels with abnormal color.

[0134] The marked pixels with abnormal color will serve as the basis for the region clustering and growth operation, which will include pixels in the neighborhood that meet the color component similarity requirements as the center of each pixel with abnormal color.

[0135] The region clustering growth operation ultimately yielded a map of the abnormal color regions of the bearing rolling elements.

[0136] The abnormal color region map will be fused with the directional enhancement feature map and the local contrast depth map to form a multi-dimensional defect-sensitive feature set of the bearing rolling elements.

[0137] The beneficial effects include: performing texture orientation evaluation and directional filtering operations on distortion-free unfolded images; accurately capturing and enhancing the texture orientation features of the rolling surface; suppressing interference information from non-primary texture orientations; improving the targeting and effectiveness of filtering operations by constructing the primary texture orientation field and generating adaptive filtering kernels; further filtering and retaining effective feature information by non-maximum suppression; and obtaining orientation enhancement feature maps that clearly present texture direction details. Multi-scale local region segmentation and statistical analysis of grayscale distribution data in distortion-free unfolded images, followed by background texture suppression, effectively enhances local image contrast, weakens the obscuring effect of background textures on defect features, and highlights potential subtle defects on the surface. Based on Mahalanobis distance and region clustering growth operations, abnormal color regions can be located, accurately identifying areas with deviations from normal colors and avoiding the omission of color-related defects. Fusion of orientation enhancement feature maps, local contrast depth maps, and abnormal color region maps integrates multi-dimensional defect-related features, forming a comprehensive and accurate multi-dimensional defect-sensitive feature set. This provides a solid and reliable feature basis for subsequent multi-channel defect feature classification, improving the accuracy and effectiveness of overall defect detection.

[0138] S4. Based on the multidimensional defect sensitive feature set, perform multi-channel defect feature classification on the linear texture direction feature, morphological depression feature and color variation feature of the distortion-free unfolded image to obtain a list of suspicious areas of the bearing rolling element.

[0139] In this embodiment of the invention, based on the multidimensional defect-sensitive feature set, multi-channel defect feature classification is performed on the linear texture direction features, morphological depression features, and color variation features of the distortion-free unfolded image to obtain a list of suspicious areas of the bearing rolling element, including:

[0140] Based on the multidimensional defect-sensitive feature set, structural tensor analysis is performed on the linear texture direction features of the distortion-free unfolded image to obtain the linear structural skeleton diagram of the bearing rolling element.

[0141] Morphological filling is performed on the linear structural skeleton diagram to obtain the candidate region for crack defects in the bearing rolling element;

[0142] Based on the multidimensional defect-sensitive feature set, local extreme points are detected for the morphological concavity features of the distortion-free unfolded image, and region growth is performed with each extreme point as the center to obtain the candidate concavity region of the bearing rolling element.

[0143] The candidate depression region is reconstructed in three dimensions to obtain the candidate pit defect region of the bearing rolling element;

[0144] Based on the multidimensional defect-sensitive feature set, color variation features of the distortion-free unfolded image are clustered in color saturation space to obtain candidate areas for burn defects of the bearing rolling elements.

[0145] Spatial topological relationship analysis is performed on the candidate areas of crack defects, pit defects, and burn defects to obtain a list of suspicious areas of the bearing rolling elements.

[0146] Based on the multidimensional defect-sensitive feature set, pixel information corresponding to the linear texture direction features in the distortion-free unfolded image is extracted. The extension direction of the texture and the number of consecutive pixels are analyzed pixel by pixel. Pixels with texture direction deviation below a preset direction deviation threshold and the number of consecutive pixels above a preset length threshold are completely retained. Isolated pixels without texture pixels in the same direction within a 3×3 neighborhood and pixels with the number of consecutive pixels below the threshold are all removed. The relevant operations of structural tensor analysis are completed to obtain the linear structural skeleton diagram of the bearing rolling element.

[0147] A morphological dilation operation is performed on the linear structure skeleton diagram. The number of dilation operations is determined based on the width of the fine fracture gaps in the linear texture to ensure that all fracture parts can be connected after dilation. Then, a morphological erosion operation corresponding to the number of dilation operations is performed to restore the original width of the linear texture. Slender linear regions with a length-to-width ratio higher than a preset ratio threshold are selected to obtain the candidate areas for crack defects in the bearing rolling element.

[0148] Based on the multidimensional defect-sensitive feature set, local contrast depth map information corresponding to the morphological depression features in the distortion-free unfolded image is extracted. A 3×3 neighborhood range is selected with each pixel as the center, and the gray values ​​of the center pixel are compared with those of all pixels in the neighborhood pixel. Pixels with gray values ​​lower than those of all pixels in the neighborhood pixel and gray value differences higher than a preset difference threshold are marked as local extrema. With each local extrema point as the center, pixels with gray value differences lower than a preset similarity threshold (i.e., gray value similarity higher than a preset similarity threshold) in the neighborhood pixel are included in the same region to complete the region growing operation and obtain the candidate depression region of the bearing rolling element.

[0149] The three-dimensional spherical mapping relationship of the bearing rolling element is retrieved, and the two-dimensional planar coordinates of the candidate recessed area are converted into the corresponding three-dimensional spherical spatial coordinates. The three-dimensional morphology of the area on the rolling element surface is restored based on the height difference of the spherical spatial coordinates. The average depth value in the area is calculated, and it is determined whether the average depth value is higher than the preset depth threshold. Areas whose average depth value and area both meet the characteristics of pit defects are selected to obtain the candidate pit defect area of ​​the bearing rolling element.

[0150] Based on the multidimensional defect sensitive feature set, the color abnormality region map information corresponding to the color variation features in the distortion-free unfolded image is extracted. The chroma and saturation components of the image are separated according to the color channel. Pixels with chroma difference and saturation difference both lower than the preset clustering threshold are divided into the same category. The chroma and saturation features of each category are compared with the preset burn color features one by one. Regions with feature matching degree higher than the preset matching degree threshold are selected to obtain the burn defect candidate area of ​​the bearing rolling element.

[0151] The spatial location and range of the candidate areas for crack defects, pit defects, and burn defects are accurately compared. The overlap and adjacency relationships between different types of candidate areas are analyzed. The spatial range union of overlapping areas is taken. Adjacent areas with similar features are merged. Texture continuity is supplemented for areas where edge pixels are not completely covered. Duplicate marked areas are removed and missing edge areas are supplemented. All area information that meets the defect characteristics is classified and integrated according to the defect type to obtain a list of suspicious areas of the bearing rolling element.

[0152] The beneficial effects include multi-channel defect feature classification based on a multi-dimensional defect-sensitive feature set for distortion-free unfolded images; extraction of linear texture direction features through structural tensor analysis to generate a linear structural skeleton map; accurate location of crack defect candidate areas through morphological filling; local extremum point detection and region growing operations for morphological depression features; accurate identification of pit defect candidate areas by combining 3D shape reconstruction; effective locking of burn defect candidate areas by relying on color variation features through color saturation spatial clustering analysis; and spatial topological relationship analysis of various defect candidate areas to form a list of suspicious areas. This enables accurate differentiation and comprehensive collection of different types of defects, avoiding omissions and duplicate annotations, and providing reliable and comprehensive basic data for subsequent defect confidence assessment and super-resolution re-imaging, thereby improving the accuracy and effectiveness of the overall defect detection process.

[0153] S5. Calculate the confidence level of the defect probability in the list of suspicious areas, and perform super-resolution re-imaging on defect areas with confidence levels below a preset threshold to obtain the accurate location and classification results of defects in the bearing rolling elements.

[0154] In this embodiment of the invention, the step of assessing the confidence level of the defect probability in the list of suspicious areas and performing super-resolution re-imaging on defect areas with a confidence level lower than a preset threshold to obtain the precise location and classification results of defects in the bearing rolling elements includes:

[0155] Extract the response intensity and spatial consistency features of the list of suspicious regions on the multidimensional defect-sensitive feature set to obtain the confidence evaluation data of the bearing rolling elements;

[0156] Based on the confidence assessment data, all regions in the list of suspicious regions are divided into high-confidence regions and low-confidence regions, and the low-confidence regions are classified as the targets to be re-inspected for the bearing rolling elements;

[0157] Based on the position information of the target to be re-inspected in the distortion-free unfolded image, the original image patch set of the target to be re-inspected in the original multi-frame surface image sequence is back-located.

[0158] High-frequency information fusion is performed on the original image patch set to obtain a super-resolution local image of the bearing rolling element;

[0159] Feature extraction and analysis are performed on the orientation sensitivity, local contrast enhancement, and color space anomalies of the super-resolution local image to obtain the resampling feature description of the bearing rolling element;

[0160] The re-sampling feature description is compared and corrected with the original feature description of the corresponding region in the list of suspicious regions to obtain the accurate location and classification result of the defect of the bearing rolling element.

[0161] The step of fusing high-frequency information from the original image patch set to obtain a super-resolution local image of the bearing rolling element includes:

[0162] Based on the phase correlation method, the original image block set is finely aligned to obtain the sub-image block sequence of the bearing rolling element;

[0163] The grayscale observations of the sub-image block sequence are arranged and collected to obtain the temporal grayscale observation set of the bearing rolling element;

[0164] By performing iterative back projection on the set of grayscale observations, an initial high-resolution image of the bearing rolling element is obtained;

[0165] Noise suppression is applied to the initial high-resolution image to obtain a super-resolution local image of the bearing rolling element.

[0166] The feature information of each region in the list of suspicious regions is extracted from the direction enhancement feature map, local contrast depth map and color abnormality region map corresponding to the multidimensional defect sensitive feature set. The feature response value of each region is compared with the preset defect feature response standard value, and the feature matching degree is calculated to obtain the response intensity data. At the same time, the continuity of feature distribution in each region is analyzed, and regions with a continuous pixel ratio higher than a preset ratio threshold are judged as meeting the spatial consistency standard. The response intensity data of all regions and the spatial consistency judgment results are integrated to obtain the confidence evaluation data of the bearing rolling element.

[0167] Based on the confidence assessment data, areas with response intensity higher than the preset confidence threshold and spatial consistency meeting the standard are classified as high-confidence areas, while areas with response intensity lower than the preset confidence threshold or spatial consistency not meeting the standard are classified as low-confidence areas. All low-confidence areas are uniformly classified as the bearing rolling elements to be re-inspected targets.

[0168] Based on the two-dimensional plane coordinates of the target to be re-inspected in the distortion-free unfolded image, the mapping relationship between the previously constructed image sequence and the three-dimensional sphere of the rolling element is retrieved, and the two-dimensional plane coordinates are converted in reverse to the corresponding three-dimensional sphere space coordinates. Then, the corresponding pixel region in the original multi-frame surface image sequence is located according to the three-dimensional sphere space coordinates, and the image information of all corresponding pixel regions is extracted to obtain the original image block set of the bearing rolling element.

[0169] The original image block set is subjected to block-by-block grayscale information analysis to extract high-frequency texture detail information in each image block. The high-frequency texture detail information at the same position in different image blocks is superimposed and fused to supplement the missing details in the original image block set. At the same time, redundant information that occurs during the fusion process is removed to enhance the clarity and detail of the image blocks, thereby obtaining a super-resolution local image of the bearing rolling element.

[0170] Direction-sensitive feature analysis is performed on the super-resolution local image to determine the main texture direction of the image and generate direction enhancement feature information. The super-resolution local image is divided into multi-scale local regions. The gray-level distribution range and standard deviation within the region are statistically analyzed to obtain local contrast enhancement information. The pixel color components of the super-resolution local image are compared with the preset normal color template to obtain color space anomaly information. The three types of feature information are integrated to obtain the resampling feature description of the bearing rolling element.

[0171] The re-sampled feature description is compared dimension by dimension with the original feature description of the corresponding region in the list of suspicious regions. For parts with inconsistent directional features, the re-sampled feature description is used as the standard for correction. For parts with deviations in local contrast and color space abnormal features, the defect is re-judged. The corrected feature information and the judgment result are integrated to determine the accurate location boundary and specific type of the defect, and the precise location and classification result of the defect of the bearing rolling element is obtained.

[0172] Based on the phase correlation method, the image block with the highest clarity in the original image block set is selected as the reference block. The phase distribution difference between each of the remaining image blocks and the reference block is calculated. The spatial position of the corresponding image blocks is adjusted according to the phase difference until the phase difference value between all image blocks and the reference block is lower than the preset phase difference threshold. The fine alignment operation of the original image block set is completed, and the sub-image block sequence of the bearing rolling element is obtained.

[0173] According to the order of the sub-image block sequence, the gray values ​​of all pixels in each image block are extracted one by one. The gray value of each pixel is associated with its coordinate position in the image block. Then, all gray values ​​are arranged in order according to the order of the image blocks and the order of the pixel coordinates to obtain the time-domain gray-scale observation set of the bearing rolling element.

[0174] An initial image template of the same size as the sub-image block sequence is constructed. Gray values ​​from the temporal gray-scale observation set are projected onto the corresponding pixel positions of the initial image template. The deviation between the projected pixel gray values ​​and the corresponding gray values ​​in the observation set is calculated. The pixel gray-scale distribution of the initial image template is adjusted according to the deviation. The projection and adjustment operations are repeated until the deviation is lower than a preset deviation threshold, thereby obtaining the initial high-resolution image of the bearing rolling element.

[0175] The initial high-resolution image is traversed pixel by pixel. A neighborhood range of a preset size is selected with each pixel as the center. The gray value of all pixels in the neighborhood is calculated. It is determined whether the difference between the gray value of the center pixel and the gray value of the neighborhood is higher than a preset noise judgment threshold. Pixels with a difference higher than the threshold are judged as noise pixels. The gray value of the noise pixel is replaced with the gray value of the neighborhood. The correction process of all noise pixels is completed, and the super-resolution local image of the bearing rolling element is obtained.

[0176] The beneficial effects include extracting the response intensity and spatial consistency features of suspicious areas on a multidimensional defect-sensitive feature set for confidence assessment, which can scientifically determine the credibility of defects in each area. By dividing high and low confidence areas and listing low confidence areas as targets to be re-inspected, areas requiring further verification can be targeted, reducing invalid detection work. Based on the location of the targets to be re-inspected, the original image block set can be traced back to obtain more original image information for subsequent processing. High-frequency information fusion related operations are performed on the original image block set, and the sub-image block sequence is obtained by fine alignment using the phase correlation method, which can ensure the positional consistency between image blocks. Collecting grayscale observations to form a temporal grayscale observation set can integrate the grayscale information of multiple frames of images. Iterative back-projection of the grayscale observation set can improve the image resolution to obtain an initial high-resolution image. After noise suppression processing, image interference can be reduced and clarity improved, and finally, a super-resolution local image is generated. Feature extraction is performed on the super-resolution local image to obtain a re-sampled feature description, which is compared and corrected with the original feature description. This can accurately calibrate the location and type information of defects, realize the precise location and classification of defects, and provide an accurate and reliable basis for subsequent surface quality assessment.

[0177] S6. The precise location of the defect and the classification results are comprehensively evaluated to obtain a surface quality assessment report of the bearing rolling element.

[0178] In this embodiment of the invention, the step of comprehensively evaluating the precise location of the defect and the classification results to obtain a surface quality assessment report for the bearing rolling element includes:

[0179] From the precise location of the defects and the classification results, the defect type, precise boundary coordinates and coverage area of ​​the bearing rolling element are extracted to obtain the defect attribute set of the bearing rolling element.

[0180] Based on the preset allowable standards for the size and quantity of various defects, the defect attribute set is classified into levels to obtain the defect level list of the bearing rolling elements;

[0181] The defect attribute set and the defect level list are rearranged and reorganized to obtain the surface quality assessment report of the bearing rolling elements.

[0182] From the precise location and classification results of the defects, the specific type of each defect is identified one by one. At the same time, the two-dimensional boundary coordinates of each defect in the undistorted unfolded image and the three-dimensional spatial boundary coordinates after transformation by the three-dimensional spherical mapping relationship are retrieved. The actual coverage area of ​​each defect on the rolling element surface is calculated based on the boundary coordinates. All defect type information, precise boundary coordinate information and coverage area information are classified and integrated to obtain the defect attribute set of the bearing rolling element.

[0183] The preset allowable standards for the size and quantity of various defects are retrieved. These standards include the allowable length value for crack defects, the allowable depth and area value for pit defects, the allowable area value for burn defects, and the allowable quantity value for various defects on the surface of the rolling element. The size and quantity parameters of each defect in the defect attribute set are compared one by one with the preset allowable standards. According to the comparison results, each defect is classified into a corresponding level. Then, the level information of all defects is integrated to obtain the defect level list of the bearing rolling element.

[0184] The defect attribute set and defect level list are arranged and reorganized according to the preset report format. First, the overall quality judgment conclusion of the bearing rolling elements is clarified. Then, the specific attribute information and corresponding level of each defect are listed in turn according to the defect type. Finally, the judgment basis and quality improvement suggestions are attached to complete the orderly organization and standardized presentation of all information, and the surface quality assessment report of the bearing rolling elements is obtained.

[0185] The beneficial effects include extracting defect types, precise boundary coordinates, and coverage areas from the accurate defect location and classification results to form a defect attribute set. This system can integrate all defect-related information on the surface of bearing rolling elements, ensuring the integrity and accuracy of defect data. Based on preset allowable standards for various defect sizes and quantities, the defect attribute set is graded and a defect level list is generated, enabling standardized judgment of defect severity and avoiding bias caused by subjective judgment. The defect attribute set and defect level list are then arranged and reorganized to obtain a surface quality assessment report, which clearly and systematically presents the surface quality status of the bearing rolling elements, providing a comprehensive and reliable reference for subsequent quality improvement, qualification judgment, and production process optimization.

[0186] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0187] This application embodiment can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence is the theory, method, technology, and application system that uses digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results.

[0188] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A visual inspection method for surface defects of bearing rolling elements, characterized in that, The method includes: S1. Based on the multi-frame surface images and rotational pose information of the bearing rolling elements, construct the mapping relationship between the bearing rolling elements from the image sequence to the three-dimensional spherical surface of the rolling elements; S2. Based on the mapping relationship, subpixel-level registration is performed on the multi-frame surface images, and conformal mapping is performed on the aligned multi-frame surface images to obtain a distortion-free unfolded image of the bearing rolling element. S3. Decouple the orientation sensitivity, local contrast enhancement and color space anomaly of the distortion-free unfolded image to obtain the multidimensional defect-sensitive feature set of the bearing rolling element; S4. Based on the multidimensional defect sensitive feature set, perform multi-channel defect feature classification on the linear texture direction feature, morphological depression feature and color variation feature of the distortion-free unfolded image to obtain a list of suspicious areas of the bearing rolling element. S5. Calculate the confidence level of the defect probability in the list of suspicious areas, and perform super-resolution re-imaging on defect areas with confidence levels below a preset threshold to obtain the accurate location and classification results of defects in the bearing rolling elements. S6. The precise location of the defect and the classification results are comprehensively evaluated to obtain a surface quality assessment report of the bearing rolling element.

2. The visual inspection method for surface defects of bearing rolling elements as described in claim 1, characterized in that, The mapping relationship between the bearing rolling elements and the three-dimensional spherical surface of the rolling elements, based on multi-frame surface images and rotational pose information of the bearing rolling elements, includes: From the multi-frame surface images of the bearing rolling elements, key feature frames that meet the requirements of clarity and view coverage are selected to obtain the key feature point set of the bearing rolling elements. Based on the rotational pose information of the bearing rolling elements, spatiotemporal consistency tracking is performed on a continuous set of key feature points to obtain the trajectory tracking chain of the key feature point set. The trajectory tracking chain is back-projected onto the candidate space curve of the three-dimensional sphere of the rolling body, and the candidate space curve is optimized and fitted to obtain the unique spatial position of the key feature point set. Using the unique spatial location as the control point, progressive surface parameterization is employed to construct the mapping relationship between the bearing rolling element and the three-dimensional spherical surface of the rolling element from the image sequence.

3. The visual inspection method for surface defects of bearing rolling elements as described in claim 1, characterized in that, The process of performing subpixel-level registration on the multi-frame surface images based on the mapping relationship, and performing conformal mapping on the aligned multi-frame surface images to obtain a distortion-free unfolded image of the bearing rolling element includes: Based on the mapping relationship, dense matching is performed on the multi-frame surface images to obtain the sub-pixel level displacement field of the bearing rolling element; Local deformation correction is performed on the subpixel-level displacement field to obtain the global deformation field of the bearing rolling element; Based on the global deformation field, the multi-frame surface images are interpolated and resampled to obtain a high-precision aligned image of the bearing rolling elements; Based on the three-dimensional spherical surface of the rolling element, the high-precision aligned image is divided into spherical triangular meshes to obtain the spherical parametric mesh system of the bearing rolling element. Based on the spherical parametric mesh system, conformal mapping is performed on the texture information in the high-precision aligned image to obtain a distortion-free unfolded image of the bearing rolling element.

4. The visual inspection method for surface defects of bearing rolling elements as described in claim 3, characterized in that, The step of performing conformal mapping on the texture information in the high-precision aligned image according to the spherical parametric mesh system to obtain a distortion-free unfolded image of the bearing rolling element includes: Texture extraction is performed on the spherical triangular facets in the high-precision aligned image to obtain local texture blocks of the bearing rolling elements; Based on the adjacent topological relationships and geometric features of the spherical triangular facets, conformal mapping is performed on the local texture blocks to obtain the planar texture blocks of the bearing rolling elements; Based on the mesh density distribution of the spherical parametric mesh system, the planar texture blocks are divided into fusion regions to obtain a multi-level fusion strategy for the bearing rolling elements; The multi-level fusion strategy is executed to progressively stitch the planar texture blocks and perform gradient domain fusion during the stitching process to obtain a distortion-free unfolded image of the bearing rolling elements.

5. The visual inspection method for surface defects of bearing rolling elements as described in claim 1, characterized in that, The decoupling of orientation sensitivity, local contrast enhancement, and color space anomalies in the distortion-free unfolded image yields a multidimensional defect-sensitive feature set for the bearing rolling elements, including: Based on the gray-level gradient distribution of the distortion-free unfolded image, the texture direction of the distortion-free unfolded image is evaluated, and directional filtering is performed based on the evaluated main texture direction to obtain the directional enhancement feature map of the bearing rolling element. The distortion-free unfolded image is divided into multi-scale local regions. The gray-level distribution range and standard deviation within each region are statistically analyzed at each scale. Based on the statistical results, the original gray-level values ​​of the distortion-free unfolded image are subjected to background texture suppression to obtain the local contrast depth map of the bearing rolling element. Based on the Mahalanobis distance between the pixel color components in the distortion-free unfolded image and the preset normal color template, the pixels in the distortion-free unfolded image are subjected to region clustering growth to obtain the color abnormality region map of the bearing rolling element. The directional enhancement feature map, the local contrast depth map, and the color abnormality region map are fused to obtain a multidimensional defect-sensitive feature set for the bearing rolling element.

6. The visual inspection method for surface defects of bearing rolling elements as described in claim 5, characterized in that, The process involves evaluating the texture direction of the distortion-free unfolded image based on its grayscale gradient distribution, and performing directional filtering based on the evaluated main texture direction to obtain the directional enhancement feature map of the bearing rolling element, including: Orientation statistics are performed on the undistorted unfolded image to obtain the main texture orientation field of the bearing rolling element; Based on the main texture direction field, the direction-selective filter kernel of the distortion-free unfolded image is adaptively generated to obtain the filter kernel of the bearing rolling element; Based on the filter kernel, the distortion-free unfolded image is convolutionally filtered to obtain the initial response map of the bearing rolling element; Non-maximum suppression is applied to the initial response map to obtain the directional enhancement feature map of the bearing rolling element.

7. The visual inspection method for surface defects of bearing rolling elements as described in claim 1, characterized in that, Based on the multidimensional defect-sensitive feature set, the linear texture direction features, morphological depression features, and color variation features of the distortion-free unfolded image are classified using multi-channel defect features to obtain a list of suspicious regions of the bearing rolling elements, including: Based on the multidimensional defect-sensitive feature set, structural tensor analysis is performed on the linear texture direction features of the distortion-free unfolded image to obtain the linear structural skeleton diagram of the bearing rolling element. Morphological filling is performed on the linear structural skeleton diagram to obtain the candidate region for crack defects in the bearing rolling element; Based on the multidimensional defect-sensitive feature set, local extreme points are detected for the morphological concavity features of the distortion-free unfolded image, and region growth is performed with each extreme point as the center to obtain the candidate concavity region of the bearing rolling element. The candidate depression region is reconstructed in three dimensions to obtain the candidate pit defect region of the bearing rolling element; Based on the multidimensional defect-sensitive feature set, color variation features of the distortion-free unfolded image are clustered in color saturation space to obtain candidate areas for burn defects of the bearing rolling elements. Spatial topological relationship analysis is performed on the candidate areas of crack defects, pit defects, and burn defects to obtain a list of suspicious areas of the bearing rolling elements.

8. The visual inspection method for surface defects of bearing rolling elements as described in claim 1, characterized in that, The process of assessing the confidence level of the defect probability in the list of suspicious areas and performing super-resolution re-imaging on defect areas with a confidence level below a preset threshold to obtain the precise location and classification results of defects in the bearing rolling elements includes: Extract the response intensity and spatial consistency features of the list of suspicious regions on the multidimensional defect-sensitive feature set to obtain the confidence evaluation data of the bearing rolling elements; Based on the confidence assessment data, all regions in the list of suspicious regions are divided into high-confidence regions and low-confidence regions, and the low-confidence regions are classified as the targets to be re-inspected for the bearing rolling elements; Based on the position information of the target to be re-inspected in the distortion-free unfolded image, the original image patch set of the target to be re-inspected in the original multi-frame surface image sequence is back-located. High-frequency information fusion is performed on the original image patch set to obtain a super-resolution local image of the bearing rolling element; Feature extraction and analysis are performed on the orientation sensitivity, local contrast enhancement, and color space anomalies of the super-resolution local image to obtain the resampling feature description of the bearing rolling element; The re-sampling feature description is compared and corrected with the original feature description of the corresponding region in the list of suspicious regions to obtain the accurate location and classification result of the defect of the bearing rolling element.

9. The visual inspection method for surface defects of bearing rolling elements as described in claim 8, characterized in that, The step of fusing high-frequency information from the original image patch set to obtain a super-resolution local image of the bearing rolling element includes: Based on the phase correlation method, the original image block set is finely aligned to obtain the sub-image block sequence of the bearing rolling element; The grayscale observations of the sub-image block sequence are arranged and collected to obtain the temporal grayscale observation set of the bearing rolling element; By performing iterative back projection on the set of grayscale observations, an initial high-resolution image of the bearing rolling element is obtained; Noise suppression is applied to the initial high-resolution image to obtain a super-resolution local image of the bearing rolling element.

10. The visual inspection method for surface defects of bearing rolling elements as described in claim 1, characterized in that, The process of comprehensively evaluating the precise location of the defects and the classification results to obtain a surface quality assessment report for the bearing rolling elements includes: From the precise location of the defects and the classification results, the defect type, precise boundary coordinates and coverage area of ​​the bearing rolling element are extracted to obtain the defect attribute set of the bearing rolling element. Based on the preset allowable standards for the size and quantity of various defects, the defect attribute set is classified into levels to obtain the defect level list of the bearing rolling elements; The defect attribute set and the defect level list are rearranged and reorganized to obtain the surface quality assessment report of the bearing rolling elements.

Images