A visual assessment method for the surface quality of bearing rolling elements

By using a composite mapping of polar coordinates and the spectrum and a multi-stage unfolded network, the problem of separating local defect features under complex imaging conditions of bearing rolling element surfaces was solved, improving the accuracy of assessment, reducing interference from illumination and surface reflection, and achieving reliable defect assessment.

CN123118567APending Publication Date: 2026-11-17SHANDONG GUANXIAN XILU BEARING CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611370086.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-04
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing visual assessment methods struggle to separate localized minute defect features under complex imaging conditions on the rolling element surface of bearings, and the assessment results are easily affected by changes in lighting and surface reflections, resulting in insufficient reliability.

Method used

The image is transformed to the frequency space in logarithmic polar coordinates by a composite polar coordinate and spectral mapping. Local defect features and global periodic textures are extracted by a multi-stage unfolding network. Confidence weights are calculated by combining spatial location stability and amplitude growth consistency, and the low-frequency basis function convolution kernel is updated to generate a structured representation of the defects.

Benefits of technology

It improves the accuracy of visual assessment of bearing rolling element surface quality, reduces interference from rotation, dimensional changes and curved surface imaging, and enables reliable separation and assessment of local defects.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123118567A_ABST
    Figure CN123118567A_ABST
Patent Text Reader

Abstract

This invention provides a visual assessment method for the surface quality of bearing rolling elements. It involves acquiring images of the bearing rolling element surface and performing a composite mapping between polar coordinates and the frequency spectrum to transform the image into a log-polar frequency space, generating an initial feature tensor. This initial feature tensor is then input into a multi-stage unfolded network. Local defect features and global periodic textures are extracted using linear reconstruction modules and self-referenced thresholding activation modules at each stage. Based on the incremental map generated by the difference between feature maps from adjacent stages, the output weights of the high-frequency sparse kernel and the low-frequency basis function are adaptively adjusted to obtain multi-scale features. The final frequency-domain sparse feature map and the intermediate frequency-domain sparse feature map are then reconstructed into corresponding log-polar space sparse response maps and input into an assessment module. Confidence weights are calculated by combining spatial position stability and amplitude growth consistency. When the weights meet the threshold, a structured representation of the defects is output, and the low-frequency basis function convolution kernel is updated using global periodic information feedback.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application belongs to the field of evaluation, and in particular relates to a visual evaluation method for the surface quality of bearing rolling elements. Background Technology

[0002] As a core component in mechanical equipment, the surface quality of bearing rolling elements directly affects the operational stability, reliability, and service life of the mechanical system. Machine vision-based non-destructive testing technology has been widely applied to assess the surface quality of bearing rolling elements. In actual image acquisition, bearing rolling elements typically possess highly reflective curved surfaces, and their surfaces are prone to periodic textures formed by machining processes such as turning and grinding. This results in images with complex background information, significant differences in feature scale, and difficulty in distinguishing between local defects and texture interference. Existing visual assessment methods largely rely on conventional deep convolutional neural networks for spatial domain feature extraction or employ basic frequency domain filtering for texture suppression. However, under complex imaging conditions of rolling element surfaces, conventional feature extraction methods easily misjudge interference caused by machining textures, lighting variations, or surface reflections as surface defects. Furthermore, they still have limitations in dealing with image deformation caused by changes in curvature, rotation angle, and viewing angle.

[0003] When using existing network models to process the aforementioned complex surface images, the models are typically limited by a single convolutional receptive field or a fixed activation mechanism, making it difficult to fully preserve local, minute, and realistic defect features while suppressing global periodic textures. As network depth increases, high-frequency details corresponding to minute defects are easily smoothed or attenuated during multi-layer feature transfer; while shallow networks struggle to extract sufficient global texture information, failing to reliably distinguish between periodic backgrounds and irregular defects. Furthermore, in the quality assessment stage, existing methods often employ fixed threshold judgments or feature mappings, failing to fully utilize the variation patterns and spatial stability of defect features across different network stages. This results in assessment results being easily affected by local noise, reflective areas, or changes in lighting and shadow. Therefore, how to separate and extract local sparse defect features under complex periodic background textures and curved surface imaging conditions, and combine this with cross-stage feature change information to achieve reliable assessment of the surface quality of bearing rolling elements, is a pressing technical problem in this field. Summary of the Invention

[0004] To address the problems in existing technologies where rolling element surface images are affected by surface reflections, periodic processing textures, and viewing angle changes, leading to difficulties in consistently extracting localized minute defect features, easy misjudgment of evaluation results, and insufficient reliability, the first aspect of this invention provides a visual evaluation method for the surface quality of bearing rolling elements, comprising the following steps: Based on the acquired images of the bearing rolling element surface, a composite mapping of polar coordinates and frequency spectrum is performed to transform the image to the frequency space in logarithmic polar coordinates to obtain the initial feature tensor. The initial feature tensor is input into a multi-stage unfolded network. Each stage of the multi-stage unfolded network consists of a linear reconstruction module and a self-referenced thresholding activation module. Local defect features and global periodic textures are extracted based on the hybrid convolution operator embedded in the linear reconstruction module. The feature map output from the previous stage is differentially compared with the feature map output from the stage before that to obtain an incremental map. The incremental map is used as a control signal to adjust the weight ratio of the high-frequency sparse kernel output and the low-frequency basis function output in the hybrid convolution operator to obtain features at each scale. The final frequency domain sparse feature map and the intermediate frequency domain sparse feature map are reconstructed into a final log-polar coordinate space sparse response map and an intermediate log-polar coordinate space sparse response map, respectively. The final log-polar coordinate space sparse response map and the intermediate log-polar coordinate space sparse response map are input into the evaluation module. The confidence weight is obtained based on the spatial location stability and amplitude growth consistency of the features. When the confidence weight reaches a preset threshold, the structured representation of the defect is obtained. Based on the global periodicity information calculated by the evaluation module, the radial frequency parameters and angular phase parameters of the low-frequency basis function convolution kernel are updated to generate the updated low-frequency basis function convolution kernel.

[0005] Optionally, the step of performing a polar coordinate and spectral composite mapping based on the acquired bearing rolling element surface image to transform the image to a frequency space in logarithmic polar coordinates to obtain an initial feature tensor includes: The coordinates of the center of the rolling element region in the acquired image of the bearing rolling element surface are determined as the reference origin; Calculate the straight-line distance of each pixel in the surface image from the reference origin and the angle between the pixel and the reference horizontal ray; After adding a preset positive bias constant to the calculated straight-line distances, logarithmic functions are applied to each one to perform logarithmic operations, and the image coordinates are reset to a mapping grid composed of logarithmic distances and included angles. Perform a two-dimensional discrete Fourier transform on the image pixel array on the mapped grid, extract the frequency domain numerical distribution results, and reconstruct the initial feature tensor.

[0006] Optionally, the step of extracting local defect features and global periodic textures based on the hybrid convolution operator embedded in the linear reconstruction module includes: Construct a 3×3 convolutional kernel, assign positive weights to the parameters at the center position, and assign negative weights to the parameters at the surrounding positions to form a high-frequency sparse kernel; A two-dimensional sine wave array with a preset receptive field size is constructed as a low-frequency basis function; The high-frequency sparse kernel and the low-frequency basis function are used to perform independent convolution extraction on the input feature tensor in parallel while keeping the spatial dimension unchanged. The two independent extraction results are then merged and assembled according to the network channel dimension for output.

[0007] Optionally, the step of obtaining an incremental map by differencing the feature map output from the previous stage with the feature map output from the stage before that includes: Obtain the feature map output from the previous stage and the feature map output from the stage before that. Replace any missing feature map from the stage before that with a zero matrix of the same dimension. Obtain the spatial row and column dimension information of the two and align them one by one. Subtract the pixel value of the corresponding pixel at the same position in the previous stage output feature map from the pixel value at each position in the previous stage output feature map to obtain a difference value matrix with the same dimension. Determine the value of each item in the difference matrix, modify all data less than zero to zero, retain all data greater than zero, and output the incremental graph accordingly.

[0008] Optionally, adjusting the weight ratio of the high-frequency sparse kernel and the low-frequency basis function within the hybrid convolution operator using the incremental map as a control signal includes: Perform mean projection along the channel dimension on the incremental graph tensor to obtain a two-dimensional incremental graph matrix; The total number of non-zero pixels in the two-dimensional incremental map matrix is ​​counted to obtain the variable of the total number of non-zero pixels in the current stage. Divide the total number of non-zero pixels by the total number of pixels in the entire image to calculate the activation coverage ratio in the range of 0 to 1; Subtract the activation coverage ratio from 1 to obtain the high-frequency ratio coefficient, and multiply the high-frequency ratio coefficient by the output feature matrix of the high-frequency sparse kernel; The activation coverage ratio is used as the low-frequency matching coefficient, and the low-frequency matching coefficient is multiplied by the output feature matrix of the low-frequency basis function to determine the calculation weights of the high-frequency sparse kernel and the corresponding branch output of the low-frequency basis function, respectively.

[0009] Optionally, obtaining the confidence weight based on the spatial location stability and amplitude growth consistency of the feature includes: The total number of times non-zero features appear at the same grid coordinate node in the computational output of each level of the multi-stage unfolded network is added together, and the feature regions whose total number of additions exceeds half the value of the total number of layers are determined. Tracing the set of associated pixel coordinates that represent the overall incremental trend of all consecutive non-zero features in the hierarchical results at each stage; All associated points in the final log-polar coordinate space sparse response map that fall within the feature region and belong to the associated coordinate pixel set are assigned a constant value of 1 as a confidence weight, and corresponding points that fall in other positions are assigned a constant value of 0 as a confidence weight.

[0010] Optionally, updating the radial frequency parameters and angular phase parameters of the low-frequency basis function convolution kernel based on the global periodicity information calculated by the evaluation module to generate an updated low-frequency basis function convolution kernel includes: Extract the low-frequency branch response map from the low-frequency branch response cache, and generate a two-dimensional low-frequency texture response map in logarithmic polar coordinate space based on the low-frequency branch response map; Two-dimensional autocorrelation and peak finding operations are performed on the two-dimensional low-frequency texture response map to obtain global periodic information; Extract the logarithmic distance period value and the angular period phase offset value parameters from the global periodic information; Based on the logarithmic distance period value and the angular period phase offset value, update the radial frequency parameter and angular phase parameter in the low-frequency basis function convolution kernel; An updated low-frequency basis function convolution kernel is generated based on the updated radial frequency parameters and angular phase parameters.

[0011] A second aspect of the present invention provides a visual assessment system for the surface quality of bearing rolling elements, comprising the following modules: The acquisition module is used to perform polar coordinate and frequency spectrum composite mapping based on the acquired bearing rolling element surface image, transforming the image to the frequency space in logarithmic polar coordinates to obtain the initial feature tensor; An extraction module is used to input the initial feature tensor into a multi-stage unfolding network. Each stage of the multi-stage unfolding network consists of a linear reconstruction module and a self-referenced thresholding activation module. Local defect features and global periodic textures are extracted based on the hybrid convolution operator embedded in the linear reconstruction module. An incremental map is obtained by differentiating the feature map output from the previous stage with the feature map output from the stage before that. The incremental map is used as a control signal to adjust the weight ratio of the high-frequency sparse kernel output and the low-frequency basis function output in the hybrid convolution operator to obtain features at each scale. The input module is used to reconstruct the final frequency domain sparse feature map and the intermediate frequency domain sparse feature map into a final log-polar coordinate space sparse response map and an intermediate log-polar coordinate space sparse response map, respectively. The final log-polar coordinate space sparse response map and the intermediate log-polar coordinate space sparse response map are then input to the evaluation module. Confidence weights are obtained based on the spatial location stability and amplitude growth consistency of the features. When the confidence weights reach a preset threshold, a structured representation of the defect is obtained. Based on the global periodicity information calculated by the evaluation module, the radial frequency parameters and angular phase parameters of the low-frequency basis function convolution kernel are updated to generate an updated low-frequency basis function convolution kernel.

[0012] Furthermore, the step of performing a polar coordinate and spectral composite mapping on the acquired bearing rolling element surface image to transform the image to a frequency space in logarithmic polar coordinates to obtain an initial feature tensor includes: The coordinates of the center of the rolling element region in the acquired image of the bearing rolling element surface are determined as the reference origin; Calculate the straight-line distance of each pixel in the surface image from the reference origin and the angle between the pixel and the reference horizontal ray; After adding a preset positive bias constant to the calculated straight-line distances, logarithmic functions are applied to each one to perform logarithmic operations, and the image coordinates are reset to a mapping grid composed of logarithmic distances and included angles. Perform a two-dimensional discrete Fourier transform on the image pixel array on the mapped grid, extract the frequency domain numerical distribution results, and reconstruct the initial feature tensor.

[0013] Further, the step of extracting local defect features and global periodic textures based on the hybrid convolution operator embedded in the linear reconstruction module includes: Construct a 3×3 convolutional kernel, assign positive weights to the parameters at the center position, and assign negative weights to the parameters at the surrounding positions to form a high-frequency sparse kernel; A two-dimensional sine wave array with a preset receptive field size is constructed as a low-frequency basis function; The high-frequency sparse kernel and the low-frequency basis function are used to perform independent convolution extraction on the input feature tensor in parallel while keeping the spatial dimension unchanged. The two independent extraction results are then merged and assembled according to the network channel dimension for output.

[0014] Further, the step of obtaining an incremental map by differencing the feature map output from the previous stage with the feature map output from the stage before that includes: Obtain the feature map output from the previous stage and the feature map output from the stage before that. Replace any missing feature map from the stage before that with a zero matrix of the same dimension. Obtain the spatial row and column dimension information of the two and align them one by one. Subtract the pixel value of the corresponding pixel at the same position in the previous stage output feature map from the pixel value at each position in the previous stage output feature map to obtain a difference value matrix with the same dimension. Determine the value of each item in the difference matrix, modify all data less than zero to zero, retain all data greater than zero, and output the incremental graph accordingly.

[0015] Further, the step of using the incremental map as a control signal to adjust the weight ratio of the high-frequency sparse kernel and the low-frequency basis function within the hybrid convolution operator includes: Perform mean projection along the channel dimension on the incremental graph tensor to obtain a two-dimensional incremental graph matrix; The total number of non-zero pixels in the two-dimensional incremental map matrix is ​​counted to obtain the variable of the total number of non-zero pixels in the current stage. Divide the total number of non-zero pixels by the total number of pixels in the entire image to calculate the activation coverage ratio in the range of 0 to 1; Subtract the activation coverage ratio from 1 to obtain the high-frequency ratio coefficient, and multiply the high-frequency ratio coefficient by the output feature matrix of the high-frequency sparse kernel; The activation coverage ratio is used as the low-frequency matching coefficient, and the low-frequency matching coefficient is multiplied by the output feature matrix of the low-frequency basis function to determine the calculation weights of the high-frequency sparse kernel and the corresponding branch output of the low-frequency basis function, respectively.

[0016] Furthermore, the step of obtaining the confidence weight based on the spatial location stability and amplitude growth consistency of the feature includes: The total number of times non-zero features appear at the same grid coordinate node in the computational output of each level of the multi-stage unfolded network is added together, and the feature regions whose total number of additions exceeds half the value of the total number of layers are determined. Tracing the set of associated pixel coordinates that represent the overall incremental trend of all consecutive non-zero features in the hierarchical results at each stage; All associated points in the final log-polar coordinate space sparse response map that fall within the feature region and belong to the associated coordinate pixel set are assigned a constant value of 1 as a confidence weight, and corresponding points that fall in other positions are assigned a constant value of 0 as a confidence weight.

[0017] Further, the step of updating the radial frequency parameters and angular phase parameters of the low-frequency basis function convolution kernel based on the global periodicity information calculated by the evaluation module, and generating the updated low-frequency basis function convolution kernel, includes: Extract the low-frequency branch response map from the low-frequency branch response cache, and generate a two-dimensional low-frequency texture response map in logarithmic polar coordinate space based on the low-frequency branch response map; Two-dimensional autocorrelation and peak finding operations are performed on the two-dimensional low-frequency texture response map to obtain global periodic information; Extract the logarithmic distance period value and the angular period phase offset value parameters from the global periodic information; Based on the logarithmic distance period value and the angular period phase offset value, update the radial frequency parameter and angular phase parameter in the low-frequency basis function convolution kernel; An updated low-frequency basis function convolution kernel is generated based on the updated radial frequency parameters and angular phase parameters.

[0018] The present invention has the following beneficial effects: By using a composite mapping of polar coordinates and the spectrum, the image is transformed to the logarithmic polar coordinate frequency space, reducing the interference caused by the rotation of the rolling body, scale changes, and surface imaging. By using a multi-stage lightweight network expansion, and by adaptively adjusting the output weights of the high-frequency sparse kernel and the low-frequency basis function using the stage difference incremental map, the separation and extraction of local defect features and global periodic textures can be achieved. By combining spatial position stability and amplitude growth consistency to calculate confidence weights, the reliability of defect structure representation is improved. Furthermore, global periodic information is used to update the low-frequency basis function convolution kernel, forming a closed-loop optimization mechanism, thereby improving the accuracy of visual assessment of bearing rolling element surface quality. Attached Figure Description

[0019] Figure 1 A flowchart of a visual assessment method for the surface quality of rolling elements in a bearing; Figure 2 A schematic diagram of the overall structure of a multi-stage unfolding network; Figure 3 The flowchart shows the internal computation process of the Kth stage of a multi-stage unfolding network. Detailed Implementation

[0020] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.

[0021] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0022] In the description of this application and the appended claims, the terms "first," "second," "third," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0023] References to "one embodiment" or "some embodiments" as described in this application specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in another embodiment," "in yet another embodiment," etc., appearing in different parts of this application specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof all mean "including but not limited to," unless otherwise specifically emphasized.

[0024] Example 1 In Embodiment 1 of the present invention, a visual assessment method for the surface quality of bearing rolling elements is provided, such as... Figure 1 As shown, it includes the following steps: S1. Based on the acquired image of the bearing rolling element surface, a composite mapping of polar coordinates and spectrum is performed to transform the image to the frequency space in logarithmic polar coordinates to obtain the initial feature tensor.

[0025] Grayscale images of the bearing rolling element surface are acquired using an industrial area scan camera and a ring-shaped shadowless light source. The rolling element regions are segmented from the grayscale images, and their boundaries are extracted. A circle fitting method is used to determine the coordinates of the center of each rolling element region. A Cartesian coordinate system is established with this center as the origin. The number of radial and angular sampling points in logarithmic polar coordinates is set, mapping the two-dimensional pixel coordinates in the Cartesian coordinate system to radial logarithmic coordinates and angular coordinates in the logarithmic polar coordinate system.

[0026] The pixel value at the new coordinate point is calculated using a bilinear interpolation algorithm to complete the logarithmic polar coordinate transformation. A two-dimensional discrete Fourier transform is then performed on the transformed image to extract the amplitude and phase spectra of the frequency domain result. The amplitude and phase spectra are then stacked and concatenated along the feature channel dimension to construct an initial feature tensor with two channels.

[0027] In one embodiment, the step of performing a polar coordinate and spectral composite mapping based on the acquired bearing rolling element surface image to transform the image to a frequency space in logarithmic polar coordinates to obtain an initial feature tensor includes: The coordinates of the center of the rolling element region in the acquired image of the bearing rolling element surface are determined as the reference origin; Calculate the straight-line distance of each pixel in the surface image from the reference origin and the angle between the pixel and the reference horizontal ray; After adding a preset positive bias constant to the calculated straight-line distances, logarithmic functions are applied to each one to perform logarithmic operations, and the image coordinates are reset to a mapping grid composed of logarithmic distances and included angles. Perform a two-dimensional discrete Fourier transform on the image pixel array on the mapped grid, extract the frequency domain numerical distribution results, and reconstruct the initial feature tensor.

[0028] The acquired bearing rolling element surface image resolution is set to a 1024×1024 pixel matrix. First, the grayscale image is processed by median filtering to reduce isolated noise. Then, the Otsu threshold segmentation method is used to separate the rolling element region from the background region, resulting in a binary segmented image. Closing operations are performed on the binary segmented image to fill holes inside the rolling element region, and opening operations are performed to remove small-area noise connected components in the background. Connected component marking is performed on the processed binary segmented image, and the connected component with the largest area is selected as the rolling element target region. The boundary pixel set of the rolling element target region is extracted, and the least squares circle fitting method is used to calculate the center coordinates of the corresponding circles. The calculated center coordinates are then... It serves as the reference origin of the Cartesian coordinate system.

[0029] Traverse the coordinates of each pixel in the original image Calculate the straight-line distance from the origin using the Euclidean distance formula. Starting from the reference origin and using a horizontal ray to the right as the reference direction, a binary arctangent function is used to calculate the polar angle of the pixel relative to the reference direction based on the difference in horizontal and vertical coordinates of the pixel relative to the reference origin. .

[0030] Set the logarithmic base variable a = 2 and the bias constant 1, and perform the logarithmic transformation operation. Generate logarithmic distance. Using bilinear interpolation, the discrete pixels originally in Cartesian coordinates are resampled onto a logarithmic polar coordinate grid (ρ, θ) with a size of 512 × 360, forming a polar coordinate image tensor, where the angular sampling point index j is related to the radian angle. Between , The number of angular sampling points, for example =360.

[0031] Perform a two-dimensional discrete Fourier transform on the 512×360 pixel array to calculate the frequency domain matrix. And extract the amplitude spectrum respectively. Phase spectrum The two numerical matrices, amplitude spectrum and phase spectrum, each 512×360 in size, are concatenated and combined along the channel dimension to reconstruct a three-dimensional data block with a tensor shape of 2×512×360, which serves as the initial feature tensor input to the network.

[0032] Before inputting into the multi-stage unfolded network, the initial feature tensors of the two channels are converted into input feature tensors with a preset number of channels through a 1×1 convolutional layer, and this input feature tensor is recorded as the initial input feature map. The preset number of channels is, for example, 32.

[0033] S2, the initial feature tensor is input into a multi-stage unfolded network. Each stage of the multi-stage unfolded network consists of a linear reconstruction module and a self-referenced thresholding activation module. Local defect features and global periodic textures are extracted based on the hybrid convolution operator embedded in the linear reconstruction module. The feature map output from the previous stage is differentially compared with the feature map output from the stage before that to obtain an incremental map. The incremental map is used as a control signal to adjust the weight ratio of the high-frequency sparse kernel output and the low-frequency basis function output in the hybrid convolution operator to obtain features at each scale.

[0034] A multi-stage unfolded network is constructed, comprising a feature input mapping receiving module, multiple cascaded feature processing unfolding stages, and a stage output reconstruction unit. An evaluation module is then connected to the multi-stage unfolded network. Each feature processing unfolding stage includes a linear reconstruction module and a self-referenced thresholding activation module. The linear reconstruction module performs high-frequency and low-frequency dual-branch feature extraction, cross-stage differential increment acquisition, weight allocation control, and channel fusion mapping. The self-referenced thresholding activation module performs thresholding activation and denoising on the fused features. The stage output reconstruction unit converts the frequency domain sparse feature map into a logarithmic polar coordinate space sparse response map. The overall structure of the multi-stage unfolded network is as follows: Figure 2 As shown.

[0035] Set the total number of network stages to K, and obtain the initial input feature map by mapping the initial feature tensor to the input. Input is fed into the first stage. In any k-th stage, the linear reconstruction module uses the feature map of the (k-1)-th stage as the convolution input, and extracts high-frequency local defect features and low-frequency periodic texture features in parallel through high-frequency sparse convolution kernels and low-frequency basis function convolution kernels; at the same time, an incremental map is generated based on the pixel-by-pixel difference matrix between the feature map of the (k-1)-th stage and the feature map of the (k-2)-th stage, and the difference matrix is ​​truncated to a negative value, so that the incremental map represents a positive activation increment.

[0036] The incremental map tensor is subjected to mean projection along the channel dimension to obtain a two-dimensional incremental map matrix. The activation coverage ratio is calculated based on the ratio of the number of non-zero pixels in the two-dimensional incremental map matrix to the total number of pixels in the image. The activation coverage ratio is subtracted from 1 to obtain a high-frequency matching coefficient, which is then used as a low-frequency matching coefficient. The high-frequency matching coefficient is multiplied by the first feature matrix output by the high-frequency sparse convolution kernel, and the low-frequency matching coefficient is multiplied by the second feature matrix output by the low-frequency basis function convolution kernel. The weighted first and second feature matrices are concatenated along the channel dimension to obtain a concatenated feature matrix. This concatenated feature matrix is ​​then converted into a fusion feature matrix with a preset number of channels, consistent with the number of channels in the input feature map for this stage, through a 1×1 convolution or linear fusion mapping layer.

[0037] The fused feature matrix is ​​fed into a self-referenced thresholding activation module to calculate its local mean and global variance within its spatial neighborhood, generating a threshold matrix for each spatial location. A soft thresholding function is used to set elements with absolute values ​​less than the corresponding threshold to zero, and elements with values ​​greater than the corresponding threshold are subtracted from the threshold while keeping their signs unchanged, thus obtaining the frequency domain sparse feature map for the current stage. And serve as input for the next stage.

[0038] The stage output reconstruction unit performs a process on the frequency domain sparse feature map. Channel projection is performed to obtain amplitude response and phase response channels. Based on these channels, the complex spectrum is reconstructed, and an inverse two-dimensional discrete Fourier transform is executed to obtain the log-polar coordinate sparse response map for the current stage. This process is iterated until all stages are completed, outputting the final frequency domain sparse feature map, intermediate frequency domain sparse feature maps, and the corresponding reconstructed log-polar coordinate sparse response map. These are then used by the evaluation module for spatial localization and defect structuring calculations.

[0039] In one embodiment, the step of extracting local defect features and global periodic textures based on the hybrid convolution operator embedded in the linear reconstruction module includes: Construct a 3×3 convolutional kernel, assign positive weights to the parameters at the center position, and assign negative weights to the parameters at the surrounding positions to form a high-frequency sparse kernel; A two-dimensional sine wave array with a preset receptive field size is constructed as a low-frequency basis function; The high-frequency sparse kernel and the low-frequency basis function are used to perform independent convolution extraction on the input feature tensor in parallel while keeping the spatial dimension unchanged. The two independent extraction results are then merged and assembled according to the network channel dimension for output.

[0040] In the first stage of the network, a 3×3 two-dimensional convolution weight matrix is ​​initialized as a high-frequency sparse kernel. The positive weight coefficient at the center element point (2,2) is set to a constant 8, and the eight adjacent coordinate positions in the outer ring are all assigned a constant -1. This high-frequency operator is used to differentially strengthen the boundary of abrupt signal signals such as tiny pits and scratches on the bearing surface.

[0041] Initialize a receptive field mesh matrix with a size of 11×11 pixels, using the equation A two-dimensional sinusoidal array is constructed as a low-frequency basis function, with the amplitude constant A=1 and the angular frequency constants along the two directions of the logarithmic polar coordinate frequency grid set. and And the initial phase offset value φ, where φ=0, is used to characterize the low-frequency periodic components corresponding to a large range of machining tool marks in the logarithmic polar coordinate frequency characteristic tensor.

[0042] The parallel dual-branch structure reads the number of channels fed in during the pre-processing stage. ,like The input feature tensor with a stride of 32 is used to perform convolution extraction with a stride of 1 and padding of 1 by calling a high-frequency sparse kernel, and convolution extraction with a stride of 1 and padding of 5 by calling a low-frequency sine wave basis function. This padding mechanism keeps the length and width dimensions of the two output tensors unchanged from the original input size of 512×360.

[0043] After extraction, obtain with High-frequency branch response diagram of the channel and The low-frequency branch response diagram of the channel, for example =32、 =32, the channel concatenation function is called to merge matrix blocks along the channel axis dim=1 of the feature tensor to obtain a hybrid assembled feature map with an overall dimension of 64×512×360. Then, the hybrid assembled feature map is converted into a stage feature map of 32×512×360 through a 1×1 convolutional layer. The stage feature map is then fed into the self-reference thresholding activation module. After thresholding activation and denoising, it is used as the output of the current stage and passed to the next stage. At the same time, the low-frequency branch response map of the current stage is established with the stage number and recorded in the low-frequency branch response cache.

[0044] In one embodiment, the step of obtaining an incremental map by differencing the feature map output from the previous stage with the feature map output from the stage before that includes: Obtain the feature map output from the previous stage and the feature map output from the stage before that. Replace any missing feature map from the stage before that with a zero matrix of the same dimension. Obtain the spatial row and column dimension information of the two and align them one by one. Subtract the pixel value of the corresponding pixel at the same position in the previous stage output feature map from the pixel value at each position in the previous stage output feature map to obtain a difference value matrix with the same dimension. Determine the value of each item in the difference matrix, modify all data less than zero to zero, retain all data greater than zero, and output the incremental graph accordingly.

[0045] When the current processing step is at the k-th stage (k≥1) of the multi-stage unfolded network, read the 3D feature map matrix output from the (k-1)-th stage. and the three-dimensional feature map matrix generated in the (k-2)th stage. .in, The initial input feature map is obtained by mapping the initial feature tensor to the input; when k=1, the generated feature map is obtained by mapping the initial feature tensor to the input tensor. Equal-dimensional zero matrix substitution and calculate and The difference between them; when k=2, read the output of the first stage. and the initial input feature map Perform the difference. The zero matrix size is preset to channel C × height H × width W, for example 32 × 512 × 360 with all elements assigned the value 0.0. The channel C is consistent with the number of channels in the current stage feature map participating in the difference.

[0046] Read the feature map output from stage k-1 and the feature map output in the (k-2)th stage The and Having the same number of channels and spatial dimensions; reading In coordinates eigenvalues ​​at and read At the same coordinates eigenvalues ​​at ,Will minus Obtain the algebraic difference at the corresponding position; after traversing all channels and all spatial positions, construct a difference value matrix D with the same dimension as the stage feature map.

[0047] The nonlinear truncation function is called to read all the values ​​contained in the difference value matrix D. If the value is detected to be less than or equal to 0.0, the value is set to a constant of 0.0; otherwise, if the difference of the operation is greater than 0.0, the value is retained. This nonlinear truncation mechanism is used to filter and attenuate noise and reconstruct an incremental graph tensor representing the positive activation amount of the stage features.

[0048] In one embodiment, adjusting the weight ratio of the high-frequency sparse kernel and the low-frequency basis function within the hybrid convolution operator using the incremental map as a control signal includes: Perform mean projection along the channel dimension on the incremental graph tensor to obtain a two-dimensional incremental graph matrix; The total number of non-zero pixels in the two-dimensional incremental map matrix is ​​counted to obtain the variable of the total number of non-zero pixels in the current stage. Divide the total number of non-zero pixels by the total number of pixels in the entire image to calculate the activation coverage ratio in the range of 0 to 1; Subtract the activation coverage ratio from 1 to obtain the high-frequency ratio coefficient, and multiply the high-frequency ratio coefficient by the output feature matrix of the high-frequency sparse kernel; The activation coverage ratio is used as the low-frequency matching coefficient, and the low-frequency matching coefficient is multiplied by the output feature matrix of the low-frequency basis function to determine the calculation weights of the high-frequency sparse kernel and the corresponding branch output of the low-frequency basis function, respectively.

[0049] Obtain the increment whose dimensions remain at 512×360 after the above truncation. Figure 2 A dimensional matrix containing a constant number of pixels across the entire image. =184320. An integer accumulator counter is initialized to zero. All feature values ​​are read within the two-dimensional matrix domain. Each time an absolute value exceeds 10... -6 For each non-zero pixel, increment the value by one. After the statistics are completed, output a variable reflecting the total number of non-zero pixels in the active coverage area. Local sparse activation leads to Taking a count of 27648 as an example, calculate the activation coverage ratio. =0.15. Because For values ​​less than or equal to the entire graph size, the range of p is distributed within the closed interval [0.0, 1.0].

[0050] The high-frequency matching coefficient is 1-p=0.85, and the low-frequency matching coefficient is p=0.15. The high-frequency matching coefficient 0.85 is multiplied by the first feature matrix output by the high-frequency sparse kernel, and the low-frequency matching coefficient 0.15 is multiplied by the second feature matrix output by the low-frequency sinusoidal basis function. Through this method, when the non-zero activations in the incremental map are locally sparsely distributed, the activation coverage ratio p is small, and the high-frequency matching coefficient 1-p is large, which is beneficial for enhancing the local defect response. When periodic background textures cause large-scale activations, the activation coverage ratio p increases, and the low-frequency matching coefficient p increases accordingly, which is beneficial for increasing the proportion of low-frequency texture modeling, thereby reducing the interference of periodic background textures on defect identification.

[0051] S3, the final frequency domain sparse feature map and the intermediate frequency domain sparse feature map are reconstructed into a final log-polar coordinate space sparse response map and an intermediate log-polar coordinate space sparse response map, respectively. The final log-polar coordinate space sparse response map and the intermediate log-polar coordinate space sparse response map are input to the evaluation module. The confidence weight is obtained based on the spatial position stability and amplitude growth consistency of the features. When the confidence weight reaches a preset threshold, the structured representation of the defect is obtained. Based on the global periodic information calculated by the evaluation module, the radial frequency parameters and angular phase parameters of the low-frequency basis function convolution kernel are updated to generate the updated low-frequency basis function convolution kernel.

[0052] The evaluation module receives the final log-polar coordinate spatial sparse response map and all intermediate log-polar coordinate spatial sparse response maps. The final log-polar coordinate spatial sparse response map and the intermediate log-polar coordinate spatial sparse response maps are both obtained by processing the corresponding frequency domain sparse feature maps through the frequency domain to spatial response reconstruction unit.

[0053] First, the log-polar sparse response maps of each stage are projected along the channel dimension to obtain the corresponding two-dimensional stage response maps. The cumulative number of non-zero activations at the same spatial location in the two-dimensional stage response map is counted along the network stage dimension to determine the feature regions that satisfy the spatial location stability condition. Then, the cross-stage pixel amplitude sequence at that spatial location is extracted. Based on the incremental trend of the cross-stage pixel amplitude sequence as the network stage increases, the associated coordinate pixel set that satisfies the amplitude growth consistency condition is determined. For associated points within the final log-polar sparse response map that are simultaneously located in the feature region and belong to the associated coordinate pixel set, their confidence weight is assigned a value of 1; for points that do not meet the above conditions, their confidence weight is assigned a value of 0, thus obtaining a binarized confidence weight map.

[0054] In the confidence weight matrix, pixels with a confidence threshold greater than or equal to a preset threshold are assigned a value of 1, while the remaining pixels are assigned a value of 0, resulting in a binary defect mask. Connectivity labeling is performed on the binary defect mask, merging adjacent pixels with a value of 1 into the same real defect region, and extracting the closed contour of each real defect region. For each closed contour, the total number of pixels contained within it is counted as the defect area; simultaneously, the minimum and maximum x-coordinates, minimum and maximum y-coordinates of all boundary pixels in the closed contour are obtained, and the coordinates of the four vertices of the circumscribed rectangle are determined accordingly. The width w and height h of the circumscribed rectangle are then calculated.

[0055] The aspect ratio R = max(w,h) / min(w,h) is calculated based on the width w and height h of the circumscribed rectangle, and the defect shape category is determined based on the aspect ratio R. The preset first aspect ratio threshold is preferably 1.5, and the preset second aspect ratio threshold is preferably 3.0. When R is less than 1.5, it indicates that the dimensions of the defect area are similar in both directions, and the defect shape category is determined to be a point-like or near-circular defect. When R is greater than or equal to 1.5 and less than 3.0, it indicates that the defect area extends in a certain direction but does not form a distinct elongated shape, and the defect shape category is determined to be a patch-like defect. When R is greater than or equal to 3.0, it indicates that the defect area extends significantly in a single direction, and the defect shape category is determined to be a linear scratch defect. Finally, the vertex coordinates of the circumscribed rectangle corresponding to each real defect area, the defect area, and the shape category are combined to output a structured representation of the defect.

[0056] The low-frequency branch response map corresponding to the last network stage is extracted from the low-frequency branch response cache and used as a low-frequency texture feature map for periodic analysis. A low-frequency complex spectrum is generated based on the low-frequency texture feature map, and an inverse two-dimensional discrete Fourier transform is performed on the low-frequency complex spectrum to obtain a low-frequency texture response map in logarithmic polar coordinate space. Then, mean projection is performed on the low-frequency texture response map along the channel dimension to obtain a two-dimensional low-frequency texture response map. Two-dimensional autocorrelation and peak finding processing are performed on the two-dimensional low-frequency texture response map to extract global periodicity information. This global periodicity information is converted into radial frequency parameters and angular phase parameters, and the converted radial frequency parameters and angular phase parameters are used to update the low-frequency basis function convolution kernel.

[0057] In one embodiment, obtaining the confidence weight based on the spatial location stability and amplitude growth consistency of the feature includes: The total number of times non-zero features appear at the same grid coordinate node in the computational output of each level of the multi-stage unfolded network is added together, and the feature regions whose total number of additions exceeds half the value of the total number of layers are determined. Tracing the set of associated pixel coordinates that represent the overall incremental trend of all consecutive non-zero features in the hierarchical results at each stage; All associated points in the final log-polar coordinate space sparse response map that fall within the feature region and belong to the associated coordinate pixel set are assigned a constant value of 1 as a confidence weight, and corresponding points that fall in other positions are assigned a constant value of 0 as a confidence weight.

[0058] The entire multi-stage unfolded network is assumed to contain computational levels of depth K=10. The multi-channel log-polar sparse response map obtained from the reconstruction at each stage is then plotted. Perform maximum projection along the channel dimension to obtain a two-dimensional stage response map. For any log-polar grid point (ρ, θ), sequentially read the corresponding two-dimensional stage response maps for stages 1 to 10. The response value at that grid point; when Greater than the preset activation threshold If the activation occurs at a certain time, the phase is counted as a non-zero activation; otherwise, it is not counted. The results of the 10 phases are accumulated to obtain the cumulative number of non-zero activations C(ρ,θ) for the grid point. A spatial location stability judgment threshold is set. When the value of C(ρ,θ) is greater than the threshold limit K / 2=5 times, the coordinates of this point are included in the feature region.

[0059] Within this feature region, for any grid point coordinate (ρ, θ), extract its continuous amplitude vector from stage 1 to stage 10. When the cross-stage magnitude vector satisfies >0.3, and there are at least four conditions in the difference of response values ​​between adjacent stages that satisfy the condition. When the positive increment term is greater than 0.05, the grid point is determined to meet the amplitude growth consistency condition, and the set of grid points that meet the condition is determined as the associated coordinate pixel set.

[0060] Write a constant 1.0 to the matrix position (ρ, θ) that is simultaneously located in the above feature region and belongs to the dual determination of the associated coordinate pixel set; write a constant 0.0 to the grid position that does not meet the above conditions, and obtain the binary spatial confidence weight map.

[0061] Optionally, when it is necessary to output the defect location in the original image coordinates, according to inverse function ,as well as , This maps the defect contour coordinates in logarithmic polar coordinate space back to the original image coordinate system; where a is the logarithmic base and b is a preset positive bias constant. The coordinates are the reference origin.

[0062] In one embodiment, updating the radial frequency parameters and angular phase parameters of the low-frequency basis function convolution kernel based on the global periodicity information calculated by the evaluation module, and generating an updated low-frequency basis function convolution kernel, includes: Extract the low-frequency branch response map from the low-frequency branch response cache, and generate a two-dimensional low-frequency texture response map in logarithmic polar coordinate space based on the low-frequency branch response map; Two-dimensional autocorrelation and peak finding operations are performed on the two-dimensional low-frequency texture response map to obtain global periodic information; Extract the logarithmic distance period value and the angular period phase offset value parameters from the global periodic information; Based on the logarithmic distance period value and the angular period phase offset value, update the radial frequency parameter and angular phase parameter in the low-frequency basis function convolution kernel; An updated low-frequency basis function convolution kernel is generated based on the updated radial frequency parameters and angular phase parameters.

[0063] The low-frequency branch response map corresponding to the last network stage is extracted from the low-frequency branch response cache and used as a low-frequency texture feature map for periodic analysis. A low-frequency complex spectrum is generated based on the low-frequency texture feature map, and an inverse two-dimensional discrete Fourier transform is performed on the low-frequency complex spectrum to obtain a low-frequency texture response map in logarithmic polar coordinate space; then, mean projection is performed on the low-frequency texture response map along the channel dimension to obtain a two-dimensional low-frequency texture response map.

[0064] A two-dimensional autocorrelation operation is performed on the two-dimensional low-frequency texture response map to obtain a two-dimensional autocorrelation metric spectrum. This spectrum represents the repetitive distribution of low-frequency textures along the logarithmic distance and angular directions. Peak-finding processing is performed on the spectrum to extract the positions of the main peaks (excluding the central peak), and the radial grid offset of these main peak positions relative to the central peak is determined as the logarithmic distance periodicity. The angular grid index offset of the main peak position relative to the central peak is determined as the angular grid index difference. .

[0065] For example, when the extracted logarithmic distance periodic value When =15, it indicates that the low-frequency texture repeats its structure every 15 grid units in the logarithmic distance direction; simultaneously, the angular grid index difference is... according to The numerical parameter of the angular periodic phase offset converted into radians, where This represents the number of angular sampling points. For example, when converted to... =0.087 radians corresponds to approximately 5°.

[0066] Based on the conversion relationship between frequency and period, using Calculate the radial harmonic angular frequency variable. For example, when When =15, the calculation yields =2π / 15≈0.418. The radial harmonic angular frequency variable is used as the updated radial frequency parameter. The numerical parameter Δθ of the angular periodic phase offset is used as the updated angular phase parameter φ'.

[0067] Update the frequency feature grid defined in logarithmic polar coordinates using the parameters described above. Low-frequency basis function convolution kernels. For example, constructing low-frequency basis functions based on the updated parameters. Where A is the amplitude parameter, For the updated radial frequency parameters, The preset angular frequency parameter remains unchanged, and φ' is the angular phase parameter determined based on the numerical parameter Δθ of the angular periodic phase offset.

[0068] An updated low-frequency basis function convolution kernel is generated based on the updated radial frequency parameters and angular phase parameters. When the initial feature tensor of subsequent target workpieces of the same specification is fed into the multi-stage unfolding network, the low-frequency branch of the linear reconstruction module calls the updated low-frequency basis function convolution kernel for convolution extraction, so as to improve the modeling adaptability of the low-frequency branch to the background texture of the processing structure and reduce the interference of periodic background texture on defect identification.

[0069] Example 2 Embodiment 2 of the present invention proposes a visual assessment system for the surface quality of bearing rolling elements, comprising the following modules: The acquisition module is used to perform polar coordinate and frequency spectrum composite mapping based on the acquired bearing rolling element surface image, transforming the image to the frequency space under logarithmic polar coordinates to obtain the initial feature tensor; An extraction module is used to input the initial feature tensor into a multi-stage unfolding network. Each stage of the multi-stage unfolding network consists of a linear reconstruction module and a self-referenced thresholding activation module. Local defect features and global periodic textures are extracted based on the hybrid convolution operator embedded in the linear reconstruction module. An incremental map is obtained by differentiating the feature map output from the previous stage with the feature map output from the stage before that. The incremental map is used as a control signal to adjust the weight ratio of the high-frequency sparse kernel output and the low-frequency basis function output in the hybrid convolution operator to obtain features at each scale. The input module is used to reconstruct the final frequency domain sparse feature map and the intermediate frequency domain sparse feature map into a final log-polar coordinate space sparse response map and an intermediate log-polar coordinate space sparse response map, respectively. The final log-polar coordinate space sparse response map and the intermediate log-polar coordinate space sparse response map are then input to the evaluation module. Confidence weights are obtained based on the spatial location stability and amplitude growth consistency of the features. When the confidence weights reach a preset threshold, a structured representation of the defect is obtained. Based on the global periodicity information calculated by the evaluation module, the radial frequency parameters and angular phase parameters of the low-frequency basis function convolution kernel are updated to generate an updated low-frequency basis function convolution kernel.

[0070] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A method of visually assessing the surface quality of a bearing rolling element, characterized in that, Includes the following steps: Based on the acquired images of the bearing rolling element surface, a composite mapping of polar coordinates and spectrum is performed to transform the image to the frequency space in logarithmic polar coordinates to obtain the initial feature tensor. The initial feature tensor is input into a multi-stage unfolded network. Each stage of the multi-stage unfolded network consists of a linear reconstruction module and a self-referenced thresholding activation module. Local defect features and global periodic textures are extracted based on the hybrid convolution operator embedded in the linear reconstruction module. The feature map output from the previous stage is differentially compared with the feature map output from the stage before that to obtain an incremental map. The incremental map is used as a control signal to adjust the weight ratio of the high-frequency sparse kernel output and the low-frequency basis function output in the hybrid convolution operator to obtain features at each scale. The final frequency domain sparse feature map and the intermediate frequency domain sparse feature map are reconstructed into a final log-polar coordinate space sparse response map and an intermediate log-polar coordinate space sparse response map, respectively. The final log-polar coordinate space sparse response map and the intermediate log-polar coordinate space sparse response map are input into the evaluation module. The confidence weight is obtained based on the spatial location stability and amplitude growth consistency of the features. When the confidence weight reaches a preset threshold, the structured representation of the defect is obtained. Based on the global periodicity information calculated by the evaluation module, the radial frequency parameters and angular phase parameters of the low-frequency basis function convolution kernel are updated to generate the updated low-frequency basis function convolution kernel.

2. The method of claim 1, wherein, The process involves performing a composite polar coordinate and spectral mapping on the acquired bearing rolling element surface image to transform the image into a frequency space in logarithmic polar coordinates, resulting in an initial feature tensor. This includes: The coordinates of the center of the rolling element region in the acquired image of the bearing rolling element surface are determined as the reference origin; Calculate the straight-line distance of each pixel in the surface image from the reference origin and the angle between the pixel and the reference horizontal ray; After adding a preset positive bias constant to the calculated straight-line distances, logarithmic functions are applied to each one to perform logarithmic operations, and the image coordinates are reset to a mapping grid composed of logarithmic distances and included angles. Perform a two-dimensional discrete Fourier transform on the image pixel array on the mapped grid, extract the frequency domain numerical distribution results, and reconstruct the initial feature tensor.

3. The method of claim 1, wherein, The step of extracting local defect features and global periodic textures based on the hybrid convolution operator embedded in the linear reconstruction module includes: Construct a 3×3 convolutional kernel, assign positive weights to the parameters at the center position, and assign negative weights to the parameters at the surrounding positions to form a high-frequency sparse kernel; A two-dimensional sine wave array with a preset receptive field size is constructed as a low-frequency basis function; The high-frequency sparse kernel and the low-frequency basis function are used to perform independent convolution extraction on the input feature tensor in parallel while keeping the spatial dimension unchanged. The two independent extraction results are then merged and assembled according to the network channel dimension for output.

4. The method of claim 1, wherein, The step of obtaining an incremental map by differencing the feature map output from the previous stage with the feature map output from the stage before that includes: Obtain the feature map output from the previous stage and the feature map output from the stage before that. Replace any missing feature map from the stage before that with a zero matrix of the same dimension. Obtain the spatial row and column dimension information of the two and align them one by one. Subtract the pixel value of the corresponding pixel at the same position in the previous stage output feature map from the pixel value at each position in the previous stage output feature map to obtain a difference value matrix with the same dimension. Determine the value of each item in the difference matrix, modify all data less than zero to zero, retain all data greater than zero, and output the incremental graph accordingly.

5. The method of claim 4, wherein, The step of using the incremental map as a control signal to adjust the weight ratio of the high-frequency sparse kernel and the low-frequency basis function within the hybrid convolution operator includes: Perform mean projection along the channel dimension on the incremental graph tensor to obtain a two-dimensional incremental graph matrix; The total number of non-zero pixels in the two-dimensional incremental map matrix is ​​counted to obtain the variable of the total number of non-zero pixels in the current stage. Divide the total number of non-zero pixels by the total number of pixels in the entire image to calculate the activation coverage ratio in the range of 0 to 1; Subtract the activation coverage ratio from 1 to obtain the high-frequency ratio coefficient, and multiply the high-frequency ratio coefficient by the output feature matrix of the high-frequency sparse kernel; The activation coverage ratio is used as the low-frequency matching coefficient, and the low-frequency matching coefficient is multiplied by the output feature matrix of the low-frequency basis function to determine the calculation weights of the high-frequency sparse kernel and the corresponding branch output of the low-frequency basis function, respectively.

6. The method of claim 1, wherein, The confidence weights derived from the spatial location stability and amplitude growth consistency of the features include: The total number of times non-zero features appear at the same grid coordinate node in the computational output of each level of the multi-stage unfolded network is added together, and the feature regions whose total number of additions exceeds half the value of the total number of layers are determined. Tracing the set of associated pixel coordinates that represent the overall incremental trend of all consecutive non-zero features in the hierarchical results at each stage; All associated points in the final log-polar coordinate space sparse response map that fall within the feature region and belong to the associated coordinate pixel set are assigned a constant value of 1 as a confidence weight, and corresponding points that fall in other positions are assigned a constant value of 0 as a confidence weight.

7. The method of claim 1, wherein, The step of updating the radial frequency parameters and angular phase parameters of the low-frequency basis function convolution kernel based on the global periodicity information calculated by the evaluation module, and generating the updated low-frequency basis function convolution kernel, includes: Extract the low-frequency branch response map from the low-frequency branch response cache, and generate a two-dimensional low-frequency texture response map in logarithmic polar coordinate space based on the low-frequency branch response map; Two-dimensional autocorrelation and peak finding operations are performed on the two-dimensional low-frequency texture response map to obtain global periodic information; Extract the logarithmic distance period value and the angular period phase offset value parameters from the global periodic information; Based on the logarithmic distance period value and the angular period phase offset value, update the radial frequency parameter and angular phase parameter in the low-frequency basis function convolution kernel; An updated low-frequency basis function convolution kernel is generated based on the updated radial frequency parameters and angular phase parameters.

8. A visual assessment system for the surface quality of bearing rolling elements, characterized in that, Includes the following modules: The acquisition module is used to perform polar coordinate and frequency spectrum composite mapping based on the acquired bearing rolling element surface image, transforming the image to the frequency space under logarithmic polar coordinates to obtain the initial feature tensor; An extraction module is used to input the initial feature tensor into a multi-stage unfolding network. Each stage of the multi-stage unfolding network consists of a linear reconstruction module and a self-referenced thresholding activation module. Local defect features and global periodic textures are extracted based on the hybrid convolution operator embedded in the linear reconstruction module. An incremental map is obtained by differentiating the feature map output from the previous stage with the feature map output from the stage before that. The incremental map is used as a control signal to adjust the weight ratio of the high-frequency sparse kernel output and the low-frequency basis function output in the hybrid convolution operator to obtain features at each scale. The input module is used to reconstruct the final frequency domain sparse feature map and the intermediate frequency domain sparse feature map into a final log-polar coordinate space sparse response map and an intermediate log-polar coordinate space sparse response map, respectively. The final log-polar coordinate space sparse response map and the intermediate log-polar coordinate space sparse response map are then input to the evaluation module. Confidence weights are obtained based on the spatial location stability and amplitude growth consistency of the features. When the confidence weights reach a preset threshold, a structured representation of the defect is obtained. Based on the global periodicity information calculated by the evaluation module, the radial frequency parameters and angular phase parameters of the low-frequency basis function convolution kernel are updated to generate an updated low-frequency basis function convolution kernel.

9. The system of claim 8, wherein, The process involves performing a composite polar coordinate and spectral mapping on the acquired bearing rolling element surface image to transform the image into a frequency space in logarithmic polar coordinates, resulting in an initial feature tensor. This includes: The coordinates of the center of the rolling element region in the acquired image of the bearing rolling element surface are determined as the reference origin; Calculate the straight-line distance of each pixel in the surface image from the reference origin and the angle between the pixel and the reference horizontal ray; After adding a preset positive bias constant to the calculated straight-line distances, logarithmic functions are applied to each one to perform logarithmic operations, and the image coordinates are reset to a mapping grid composed of logarithmic distances and included angles. Perform a two-dimensional discrete Fourier transform on the image pixel array on the mapped grid, extract the frequency domain numerical distribution results, and reconstruct the initial feature tensor.

10. The system of claim 8, wherein, The step of extracting local defect features and global periodic textures based on the hybrid convolution operator embedded in the linear reconstruction module includes: Construct a 3×3 convolutional kernel, assign positive weights to the parameters at the center position, and assign negative weights to the parameters at the surrounding positions to form a high-frequency sparse kernel; A two-dimensional sine wave array with a preset receptive field size is constructed as a low-frequency basis function; The high-frequency sparse kernel and the low-frequency basis function are used to perform independent convolution extraction on the input feature tensor in parallel while keeping the spatial dimension unchanged. The two independent extraction results are then merged and assembled according to the network channel dimension for output.