A method and system for intelligent visual inspection of the entire surface of a polished section of a crystalline silicon substrate

Abstract

This invention provides an intelligent visual inspection method and system for the entire surface of a polished section of a crystalline silicon substrate, belonging to the field of visual inspection technology. The method includes: Step 2, constructing a three-dimensional feature evaluation field covering the entire surface of the crystalline silicon substrate based on the spatial coordinates and joint feature vectors of three typical points; determining multiple homogeneous feature blocks by performing curvature analysis and gradient calculation on the virtual surface of the three-dimensional feature evaluation field; and then performing numerical simulation in conjunction with illumination and material parameters to calculate the deviation between theoretical and measured imaging parameters, thereby obtaining the dynamic compensation coefficients for each corresponding homogeneous feature block. This invention distinguishes between microbubble defects and surface pseudo-defects in the polished section of the crystalline silicon substrate through multi-region feature extraction, three-dimensional feature evaluation field construction and dynamic compensation, and two-dimensional and three-dimensional feature fusion classification, achieving defect location and size measurement, while simultaneously completing quality grading and generating inspection reports, thus improving the accuracy, stability, and intelligence level of the full-surface visual inspection of the polished section of the crystalline silicon substrate.

Description

Technical Field

[0001] This invention relates to the field of visual inspection technology, and in particular to an intelligent visual inspection method and system for the entire surface of a polished section of a crystalline silicon substrate. Background Technology

[0002] In the surface quality inspection of the polished section of photovoltaic crystalline silicon panels, the application of machine vision methods is affected by the structural characteristics of the crystalline silicon panels themselves. Due to the differences in optical properties in different areas of the polished crystalline silicon panels, such as the different reflective properties between the central area and the edge transition area of ​​the panel, and the existence of light transmission and scattering phenomena in the subsurface area, the detection effect of detection methods based on a unified imaging model or a fixed grayscale threshold may vary when processing different areas.

[0003] Taking the inspection of single-crystal silicon polished wafers as an example, in actual production line operation, the image grayscale distribution in the central area of ​​the wafer surface is relatively stable, while the chamfered edge area may produce local reflections under specific light source angles due to the curved surface transition. This causes the grayscale statistical characteristics of this area to be inconsistent with those of the central area. If the same grayscale gradient threshold is used for defect segmentation, the normal texture of the edge area may be identified as abnormal, while microbubbles with low grayscale contrast in the central area or subsurface may be ignored during the segmentation process. In addition, due to the influence of the light penetration depth, the imaging characteristics of the subsurface area differ from those of the surface layer. Traditional algorithms based on surface grayscale information have limited sensitivity to the subtle structural changes in this area, which may affect the identification effect of subsurface microbubbles. The inconsistency of features between the above-mentioned areas limits the existing detection methods in distinguishing between real microbubble defects and surface pseudo-defects, which may affect the accuracy of defect localization and the stability of detection results. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide an intelligent visual inspection method and system for the entire surface of a polished section of a crystalline silicon board, which meets the accuracy requirements for online full surface quality inspection of the polished section of the crystalline silicon board.

[0005] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows: A first aspect is a method for intelligent visual inspection of the entire surface of a polished section of a crystalline silicon substrate, the method comprising: Step 1: Select typical points in the three regions of the preprocessed image: the geometric center of the plate surface, the edge transition, and the subsurface depth. Perform gray-level gradient statistics and texture spectrum analysis on the local image blocks of each typical point to obtain the joint feature vector representing the surface and subsurface structure. Step 2: Based on the spatial coordinates and joint feature vectors of three typical points, a three-dimensional feature evaluation field covering the entire surface of the silicon wafer is constructed. Multiple homogeneous feature blocks are determined by curvature analysis and gradient calculation of the virtual surface of the three-dimensional feature evaluation field. Then, numerical simulation is performed in combination with illumination and material parameters to calculate the deviation between theoretical and measured imaging parameters and obtain the dynamic compensation coefficients for each homogeneous feature block. Step 3: Use dynamic compensation coefficients to perform adaptive threshold segmentation on the preprocessed image to obtain a binarized image, and extract the two-dimensional geometric features of each candidate defect region. Map the contour point set of each candidate defect region to the virtual surface of the three-dimensional feature evaluation field to obtain the spatial coordinates and centroid position of each contour point on the virtual surface. Calculate the two principal curvature direction vectors at the corresponding centroid point, derive the geometric anisotropy index of the corresponding defect region, and obtain the surface geometric feature parameters. Step 4: After fusing the two-dimensional geometric features and the surface geometric features parameters, input them into the classification model to distinguish between microbubble defects and surface pseudo-defects, obtain the position coordinates and geometric dimensions of the microbubbles, and perform statistical evaluation of the microbubble defects according to the quality grading standards to generate an inspection report.

[0006] Secondly, a full-surface intelligent visual inspection system for the polished section of a crystalline silicon substrate includes: The feature extraction module is used to select typical points in three regions on the preprocessed image: the geometric center of the plate surface, the edge transition, and the depth of the subsurface. Gray-level gradient statistics and texture spectrum analysis are performed on the local image blocks of each typical point to obtain a joint feature vector representing the surface and subsurface structure. The dynamic compensation coefficient calculation module is used to construct a three-dimensional feature evaluation field covering the entire surface of the silicon wafer based on the spatial coordinates and joint feature vectors of three typical points. By performing curvature analysis and gradient calculation on the virtual surface of the three-dimensional feature evaluation field, multiple homogeneous feature blocks are determined. Then, combined with lighting and material parameters, numerical simulation is performed to calculate the deviation between theoretical and measured imaging parameters, and the dynamic compensation coefficients of each homogeneous feature block are obtained. The defect region segmentation module is used to perform adaptive threshold segmentation on the preprocessed image using dynamic compensation coefficients to obtain a binarized image, and extract the two-dimensional geometric features of each candidate defect region. The contour point set of each candidate defect region is mapped to the virtual surface of the three-dimensional feature evaluation field to obtain the spatial coordinates and centroid position of each contour point on the virtual surface. The two principal curvature direction vectors at the corresponding centroid point are calculated to derive the geometric anisotropy index of the corresponding defect region and obtain the surface geometric feature parameters. The quality assessment module is used to integrate two-dimensional geometric features and surface geometric features into the classification model, distinguish between microbubble defects and surface pseudo-defects, obtain the position coordinates and geometric dimensions of microbubbles, and perform statistical evaluation of microbubble defects according to quality grading standards to generate an inspection report.

[0007] Thirdly, a computer-readable storage medium storing a program that, when executed by a processor, implements the method.

[0008] The above-described solution of the present invention has at least the following beneficial effects: By extracting and analyzing local features from multiple typical points, the differences in surface and subsurface structures in different regions are accurately characterized, improving the ability to describe the features of complex plate areas. A three-dimensional feature evaluation field is constructed based on the spatial coordinates and feature vectors of typical points, enabling continuous and comprehensive characterization of the crystalline silicon plate from the surface to the subsurface, effectively addressing the shortcomings of relying solely on local grayscale information and poor regional adaptability. Curvature analysis and gradient calculation are performed on the three-dimensional feature evaluation field to divide it into homogeneous feature blocks. Dynamic compensation coefficients are obtained by combining theoretical and measured imaging deviations, achieving adaptive imaging compensation and reducing the impact of uneven illumination, edge reflections, etc. Interference caused by regional characteristic differences is mitigated, improving the stability and accuracy of defect segmentation. Candidate defect regions are mapped from two-dimensional images to a three-dimensional feature evaluation field. Combining two-dimensional and surface geometric features, and introducing principal curvature and geometric anisotropy indices, this effectively distinguishes between real microbubble defects and surface pseudo-defects, reducing false detection rates and improving the detection capability for low-contrast, subsurface fine defects. This enables intelligent detection throughout the entire process, from defect segmentation, feature extraction, and defect classification to quality assessment and report generation, ensuring accurate defect location and reliable dimensional measurement, meeting the accuracy and stability requirements of online full-surface quality inspection in the polishing section of crystalline silicon panels. Attached Figure Description

[0009] Figure 1 This is a schematic flowchart of an intelligent visual inspection method for the entire surface of a polished section of a crystalline silicon substrate, provided by an embodiment of the present invention.

[0010] Figure 2 This is a schematic diagram of a full-surface intelligent visual inspection system for the polishing section of a crystalline silicon substrate, provided by an embodiment of the present invention. Detailed Implementation

[0011] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0012] like Figure 1 As shown, an embodiment of the present invention proposes a full-surface intelligent visual inspection method for a polished section of a crystalline silicon substrate, the method comprising the following steps: Step 1: Select typical points in the three regions of the preprocessed image: the geometric center of the plate surface, the edge transition, and the subsurface depth. Perform gray-level gradient statistics and texture spectrum analysis on the local image blocks of each typical point to obtain the joint feature vector representing the surface and subsurface structure. Step 2: Based on the spatial coordinates and joint feature vectors of three typical points, a three-dimensional feature evaluation field covering the entire surface of the silicon wafer is constructed. Multiple homogeneous feature blocks are determined by curvature analysis and gradient calculation of the virtual surface of the three-dimensional feature evaluation field. Then, numerical simulation is performed in combination with illumination and material parameters to calculate the deviation between theoretical and measured imaging parameters and obtain the dynamic compensation coefficients for each homogeneous feature block. Step 3: Use dynamic compensation coefficients to perform adaptive threshold segmentation on the preprocessed image to obtain a binarized image, and extract the two-dimensional geometric features of each candidate defect region. Map the contour point set of each candidate defect region to the virtual surface of the three-dimensional feature evaluation field to obtain the spatial coordinates and centroid position of each contour point on the virtual surface. Calculate the two principal curvature direction vectors at the corresponding centroid point, derive the geometric anisotropy index of the corresponding defect region, and obtain the surface geometric feature parameters. Step 4: After fusing the two-dimensional geometric features and the surface geometric features parameters, input them into the classification model to distinguish between microbubble defects and surface pseudo-defects, obtain the position coordinates and geometric dimensions of the microbubbles, and perform statistical evaluation of the microbubble defects according to the quality grading standards to generate an inspection report.

[0013] In this embodiment of the invention, by extracting and analyzing local features at multiple typical points, the differences in surface and subsurface structures in different regions are accurately characterized, improving the ability to describe the features of complex plate areas. A three-dimensional feature evaluation field is constructed based on the spatial coordinates and feature vectors of typical points, enabling continuous and comprehensive characterization of the crystalline silicon plate from the surface to the subsurface, effectively addressing the shortcomings of relying solely on local grayscale information and poor regional adaptability. By performing curvature analysis and gradient calculation on the three-dimensional feature evaluation field to divide it into homogeneous feature blocks, and combining the theoretical imaging and measured imaging deviations to obtain dynamic compensation coefficients, adaptive imaging compensation is achieved, reducing the impact of uneven illumination and edge distortion. Interference caused by edge reflection and regional characteristic differences is mitigated, improving the stability and accuracy of defect segmentation. Candidate defect regions are mapped from two-dimensional images to a three-dimensional feature evaluation field. Combining two-dimensional and surface geometric features, and introducing principal curvature and geometric anisotropy indices, this effectively distinguishes between real microbubble defects and surface pseudo-defects, reducing false detection rates and improving the detection capability for low-contrast, subsurface, and minute defects. This enables intelligent detection throughout the entire process, from defect segmentation, feature extraction, and defect classification to quality assessment and report generation, ensuring accurate defect location and reliable dimensional measurement, meeting the accuracy and stability requirements of online full-surface quality inspection in the polishing section of crystalline silicon wafers.

[0014] In another preferred embodiment of the present invention, the process of obtaining the preprocessed image is as follows: Step 001: Perform flat-field correction on the original image data based on pixel response characteristics. This is done by acquiring a bright-field image under uniform illumination and a dark-field image after turning off the light source, and calculating the response characteristic parameters of each pixel. The gray values ​​of each pixel in the original image data are then linearly corrected using these response characteristic parameters to eliminate the non-uniformity of photoelectric response between pixels, resulting in a flat-field corrected image. Specifically, this includes: first, acquiring a bright-field image under uniform illumination conditions, and then acquiring a dark-field image under conditions with the light source turned off. For each pixel in the image, the gray value at the corresponding position in the bright-field image is subtracted from the gray value at the corresponding position in the dark-field image to obtain the response characteristic parameters of that pixel. For each pixel in the original image data, the gray value of the original image is subtracted from the gray value of the corresponding pixel in the dark-field image, and then divided by the response characteristic parameters of that pixel to complete linear gray-scale correction, eliminating the problem of non-uniform photoelectric response between pixels of the sensor, and obtaining a flat-field corrected image.

[0015] Step 002: Perform column-wise correction on the flat-field correction image based on the subtraction of gray-level averages. Calculate the gray-level average of each column of pixels in the flat-field correction image. Subtract the gray-level average of the corresponding column from the gray-level of each column of pixels in the flat-field correction image, and then superimpose a set ratio of the gray-level average of the entire image to eliminate the lateral gray-level distribution difference caused by uneven lighting, thus obtaining the corrected image. Specifically, this includes: calculating the gray-level average of all pixels in each column of the flat-field correction image, subtracting the corresponding gray-level average of the column from the gray-level of each pixel in each column of the flat-field correction image, and then superimposing the calculation result on 0.1 times the gray-level average of the entire flat-field correction image. This set ratio is used to pull the corrected gray-level back to the normal image gray-level range, avoiding the overall gray-level being too low due to the subtraction of column averages, thereby eliminating the lateral gray-level distribution difference caused by uneven lighting. After completing the column-wise correction, the corrected image is obtained.

[0016] Step 003 involves edge detection on the corrected image, extracting the contour edge point set of the silicon wafer in the corrected image, and performing linear fitting on the contour edge point set to obtain the linear equations constituting the edges of the wafer. The intersection points of adjacent linear equations are then solved to obtain the coordinates of the four corner points of the silicon wafer. Specifically, this includes: performing edge detection on the corrected image to extract the contour edge point set of the silicon wafer region; performing linear fitting on the edge point set using a general linear equation, Ax + By + C = 0, where x and y are the abscissa and ordinate of the edge points, and A, B, and C are the linear coefficients obtained through fitting; substituting the coordinates of the edge point set into the equations and completing the fitting to obtain the linear equations constituting the four edges of the silicon wafer; and solving the simultaneous equations of adjacent linear equations to calculate the intersection points between the lines, using these intersection points as the four corner points of the silicon wafer, and obtaining the precise coordinates of the four corner points.

[0017] Step 004: Construct an affine transformation matrix based on the coordinates of the four corner points. Use the affine transformation matrix to perform geometric distortion correction on the corrected image, extract the entire surface area of ​​the silicon wafer to be detected, and perform median filtering on the image of the entire surface area of ​​the silicon wafer to remove isolated noise points and obtain a preprocessed image. Specifically, this includes: establishing a one-to-one mapping relationship between the actual coordinates of the four corner points of the silicon wafer obtained in step 003 and the coordinates of the four corner points of a preset standard rectangle. The coordinates of the four corner points of the preset standard rectangle are set as the coordinates of the upper left corner start point, the upper right corner end point, the lower right corner end point, and the lower left corner start point, respectively, to define the standard position and size of the regular rectangle; using this correspondence, solve and construct an affine transformation matrix. This affine transformation matrix is ​​used to describe the coordinate transformation rules of the image from the distorted state to the standard rectangle state, and can simultaneously complete the translation, rotation, scaling, and shearing distortion correction of the image. The original x-coordinate and y-coordinate of each pixel in the corrected image are multiplied by the corresponding element in the affine transformation matrix, and the results of similar operations are added together to obtain the target x-coordinate and target y-coordinate of the pixel under standard geometric shape. This coordinate transformation is then performed point-by-point on all pixels in the corrected image to complete the overall geometric distortion correction, restoring the crystalline silicon plate area from a distorted shape to a regular standard rectangular shape. The entire surface area to be detected, containing only the crystalline silicon plate, is then precisely extracted from the geometrically corrected image, eliminating background interference for subsequent detection. Median filtering is applied to the extracted image of the entire surface area to be detected on the crystalline silicon plate. A fixed-size square neighborhood window is selected centered on each pixel to be processed. All pixels within the window are traversed, and their corresponding gray values ​​are obtained. These gray values ​​are arranged in ascending order, and the gray value in the middle position after sorting is used to replace the original gray value of the center pixel. This filtering method effectively removes isolated noise and impulse interference from the image, while preserving the true shape of the edge contour, fine structure, and defects on the surface of the silicon wafer, ultimately resulting in a pre-processed image with uniform grayscale, regular geometry, and low noise interference.

[0018] This embodiment, through pixel response flat-field correction combining bright and dark fields, can eliminate systematic errors caused by inconsistencies in the photoelectric response of the image sensor itself, improving the grayscale consistency and reliability of the original image. By employing a correction method that subtracts the column-wise grayscale mean and superimposes the overall mean ratio, it can specifically eliminate lateral grayscale deviations caused by uneven illumination, avoiding the impact of excessively bright or dark areas on defect segmentation and feature extraction, and improving the uniformity of grayscale distribution across the entire image. Through edge detection, line fitting, and corner point solving, it can accurately locate the true boundaries and geometric corners of the silicon wafer, eliminating interference from image background and irrelevant areas. Geometric distortion correction based on affine transformation can eliminate geometric deformations generated during shooting or transmission, ensuring the accurate and true position and size information of the silicon wafer surface. Combined with median filtering, it effectively improves the quality of the preprocessed image while preserving defect features and filtering out isolated noise points.

[0019] In a preferred embodiment of the present invention, step 1 includes: Step 100: Based on the differences in optical reflection characteristics of different regions on the surface of the silicon wafer, determine the center point located in the geometric center region of the wafer, the edge point located in the transition region of the wafer edge, and the depth point located in the subsurface depth feature region in the preprocessed image. Specifically, this includes: based on the differences in optical reflection characteristics and structural features of different regions on the surface of the silicon wafer in the preprocessed image, first extract the geometric contour coordinates of the silicon wafer. Take the average of the abscissa of all points in the horizontal direction and the average of the ordinate of all points in the vertical direction. The position corresponding to the average of the abscissa and ordinate is the geometric center of the wafer. This position is determined as the center point used to characterize the overall uniform region of the wafer. Then, along the edge of the silicon wafer, shrink inward by a fixed distance to select the transition zone connecting the edge and the inner region as the edge point. This position can accurately reflect the reflective characteristics and structural transition changes of the edge region of the wafer. At the same time, select a region on the surface of the silicon wafer that is a certain distance from the geometric center and can reflect the characteristics of subsurface light transmission and scattering. The position in this region with stable gray-scale changes and typical structural features is determined as the depth point, thus completing the selection of three types of typical points.

[0020] Step 101: Extract local image blocks of a first preset size centered on the center point, extract local image blocks of a second preset size centered on the edge points, and extract local image blocks of a third preset size centered on the depth points. Perform gradient calculation on each local image block. The gradient calculation uses horizontal and vertical convolution kernels to convolve each local image block, obtaining the horizontal and vertical gradient components of each pixel within each local image block. Calculate the gradient magnitude and gradient direction of each pixel based on the horizontal and vertical gradient components, and statistically analyze the gradient magnitude histogram and gradient direction histogram of each local image block as the gradient feature components of each local image block. Specifically, this includes: extracting local image blocks of a first preset size centered on the center point. In the preprocessed image, square local image patches with a side length of 30 pixels are cropped. Square local image patches with a side length of 25 pixels are cropped centered on edge points, and square local image patches with a side length of 20 pixels are cropped centered on depth points. Gradient calculation is performed on each local image patch. A horizontal convolution kernel is used to perform pixel-by-pixel convolution on the local image patch. The elements within the convolution kernel are multiplied sequentially by the corresponding pixel grayscale values ​​of the local image patch, and then summed to obtain the horizontal gradient component of each pixel. Similarly, a vertical convolution kernel is used to perform pixel-by-pixel convolution on the local image patch. The elements within the convolution kernel are multiplied sequentially by the corresponding pixel grayscale values ​​of the local image patch, and then summed to obtain the vertical gradient component of each pixel. For each pixel, the horizontal gradient component is squared, and the vertical gradient component is squared. The two squared results are added together, and then the square root of the sum is performed to obtain the gradient magnitude of that pixel. Divide the vertical gradient component by the horizontal gradient component, and perform an arctangent operation on the result to obtain the gradient direction of the pixel. Traverse all pixels within the local image patch, counting the number of pixels in each gradient magnitude interval to form a gradient magnitude histogram, and counting the number of pixels in each gradient direction angle interval to form a gradient direction histogram. Combine the gradient magnitude histogram and the gradient direction histogram as the gradient feature components of the local image patch.

[0021] Step 102: Calculate the gray-level co-occurrence matrix (GLCM) for each local image block. Generate GLCMs by selecting a preset pixel spacing and multiple angle directions. Extract contrast, correlation, energy, and inverse difference moment texture statistics from each GLCM. Mean the extracted texture statistics from each angle direction and use them as texture feature components for each local image block. Specifically, this includes: calculating the GLCM for each local image block, selecting a pixel spacing of two pixels, and generating corresponding GLCMs at four angles: horizontal, 45-degree diagonal, vertical, and 135-degree diagonal. At each angle, sequentially traverse each pixel within the local image block, starting from the current pixel, and search for corresponding neighboring pixels according to the corresponding angle and the distance between two pixels. Count the number of times each gray-level combination occurs, and arrange the statistical results according to gray-level to form the GLCM for that angle. For each gray-level co-occurrence matrix, four types of texture statistics are calculated: contrast, correlation, energy, and inverse difference moment. Contrast is obtained by weighted summation of the squares of gray-level differences at different locations within a local image patch, characterizing the significance of gray-level differences within that patch. Correlation is calculated by multiplying each element in the gray-level co-occurrence matrix by its corresponding gray level, and then performing a linear operation using the mean and standard deviation of the gray levels. Specifically, the operation first calculates the product of each element's value and its corresponding two gray levels. The product of the two gray levels is subtracted from the product of their respective means. The sum of all the calculation results is then divided by the product of the standard deviations of the two gray levels. This quantifies the linear correlation of gray level distributions within an image patch and is used to characterize the degree of correlation of gray level distributions within a local image patch. The energy is obtained by summing the squares of all elements in the gray-level co-occurrence matrix and is used to characterize the uniformity of gray level distributions within a local image patch. The inverse difference moment is obtained by weighted summation of the reciprocals of the gray level differences and is used to characterize the smoothness and stability of textures within a local image patch. The average contrast value is obtained by summing the contrast values ​​of the same local image patch at four different angles and then dividing by the number of angles. The average correlation value is obtained by summing the correlation values ​​at four angles and then dividing by the number of angles. The average energy value is obtained by summing the energy values ​​at four angles and then dividing by the number of angles. The average inverse moment value is obtained by summing the inverse moment values ​​at four angles and then dividing by the number of angles. These four types of averaged texture statistics are combined in order as the texture feature components of the local image patch.

[0022] Step 103 involves fusing the gradient feature components and texture feature components corresponding to each local image patch at the feature layer to form a joint feature vector. Specifically, this includes concatenating and combining the gradient feature components and texture feature components corresponding to each local image patch in a fixed order. First, all statistical data of the gradient magnitude histogram and gradient direction histogram are arranged in sequence. Then, the mean contrast, mean correlation, mean energy, and mean inverse difference moment are arranged in sequence. All feature data are then concatenated into a continuous data sequence in chronological order to complete the data fusion at the feature level and form a joint feature vector that can simultaneously represent the gradient information and texture information of the local region.

[0023] This embodiment selects typical points for the three regions with the most significant differences in optical properties—the center, edge transition, and subsurface depth of the silicon substrate—to capture the true surface and subsurface structural features of different regions, solving the problem of inaccurate feature description caused by uneven regional optical properties. By extracting local image patches by region and calculating gradient magnitude and gradient direction histograms, the gray-level variation and structural orientation of the silicon substrate surface can be effectively characterized, enhancing the ability to describe subtle defects and regional transition features. By using a multi-angle gray-level co-occurrence matrix to extract and average texture statistics, the texture distribution pattern of the silicon substrate surface can be stably reflected, reducing the randomness of single-angle texture features and improving the ability to distinguish between polished surface texture and defect texture. By fusing gradient features and texture features at the feature layer to form a joint feature vector, regional characteristics can be characterized simultaneously from two dimensions: gray-level variation and texture distribution.

[0024] In a preferred embodiment of the present invention, step 2 includes: Step 200: Based on the spatial coordinates of the center point, edge point, and depth point, the joint feature vector corresponding to each point is used as the feature value at the corresponding spatial location. A weight function is constructed based on the spatial distance between each point. The feature values ​​at each spatial location on the entire surface are interpolated by weighted summation to construct a three-dimensional feature evaluation field covering the entire surface of the silicon wafer. The three-dimensional feature evaluation field represents the spatial distribution of the comprehensive optical properties and structural information of the silicon wafer from the surface to the subsurface in the form of a continuous virtual surface. Specifically, this includes: based on the spatial coordinates of the center point, edge point, and depth point in the preprocessed image, the joint feature vector corresponding to the three points is used as the known feature value at these three spatial locations; for any spatial location to be calculated on the surface of the silicon wafer, the straight-line distance from this location to the center point, edge point, and depth point is calculated respectively. The reciprocal of the distance is used as the basic weight, with a larger weight for closer distances and a smaller weight for farther distances. Each basic weight is then divided by the sum of all basic weights to complete weight normalization, resulting in the final weight function used for interpolation calculation. For each spatial location to be calculated on the surface of the crystalline silicon substrate, the feature values ​​of the center point, edge point, and depth point are multiplied by the corresponding weight value. The three sets of product results are then added together, and the feature value of the location is obtained by weighted summation. This process is repeated for all spatial locations on the entire surface of the crystalline silicon substrate, and feature interpolation calculation is performed point by point. This constructs a three-dimensional feature evaluation field that can completely cover the entire surface of the crystalline silicon substrate. This three-dimensional feature evaluation field is presented in the form of a continuous and smooth virtual surface. Each spatial coordinate point of the surface corresponds to a set of feature parameters. These feature parameters integrate information such as gray-level gradient and texture distribution at the corresponding location, and can uniformly and continuously characterize the comprehensive optical properties of the crystalline silicon substrate from the surface to the subsurface, including surface reflectivity, subsurface light transmission and scattering characteristics, and differences in internal structure in different regions. It clearly presents the spatial distribution law of these optical and structural information on the entire surface of the crystalline silicon substrate, and achieves accurate characterization of the global properties of the crystalline silicon substrate.

[0025] Step 201 involves discretizing the virtual surface of the 3D feature evaluation field to obtain the normal vector and principal curvature values ​​at each discrete point on the virtual surface. Gaussian curvature and mean curvature are calculated based on the principal curvature values ​​at each discrete point. A curvature threshold is set to divide the virtual surface into multiple homogeneous feature blocks with continuous curvature changes. Specifically, this includes: discretizing the continuous virtual surface corresponding to the 3D feature evaluation field using a uniform grid; decomposing the entire virtual surface into a densely packed set of regularly arranged discrete points with a fixed step size of 0.5 pixels to ensure that each discrete point can completely represent the local shape of the surface. For each discrete point, adjacent surface points within a radius of 1.5 pixels centered on that point are selected as a local neighborhood. Difference operations are performed using the spatial coordinates of each point within the neighborhood. Specifically, three non-collinear surface points within the neighborhood are selected, and the spatial coordinate difference between these three points is calculated. The normal vector at that discrete point is obtained through a vector cross product. The direction of the cross product result is the direction of the normal vector. The magnitude of the cross product result is normalized to obtain the unit normal vector. After obtaining the normal vector, the coefficients of the first and second fundamental forms of the virtual surface at that point are calculated based on the surface coordinates and the normal vector in the local neighborhood. The coefficients of the first fundamental form are obtained by multiplying and adding the partial derivatives of the coordinates of the surface points in the neighborhood. Specifically, surface points along two orthogonal directions in the neighborhood are selected, the first-order partial derivatives of the coordinates with respect to the parameters are calculated, the partial derivatives in the same direction are squared and added to obtain the two principal coefficients of the first fundamental form, and the partial derivatives in the two orthogonal directions are multiplied to obtain the cross coefficients. The coefficients of the second fundamental form are obtained by the dot product of the normal vector and the second-order partial derivatives of the coordinates. First, the second-order partial derivatives of the coordinates with respect to the parameters are calculated, and then the second-order partial derivatives are multiplied by the normal vector to obtain the three coefficients of the second fundamental form.

[0026] Subsequently, the generalized eigenvalues ​​of the second fundamental form relative to the first fundamental form are solved. Specifically, the coefficient matrices of the second and first fundamental forms are constructed first. The first fundamental form coefficient matrix is ​​a 2×2 matrix composed of the three coefficients of the first fundamental form calculated previously. Specifically, the two elements on the main diagonal of the matrix are the two principal coefficients of the first fundamental form, and the two elements off the main diagonal are the cross coefficients of the first fundamental form. The second fundamental form coefficient matrix is ​​also a 2×2 matrix composed of the three coefficients of the second fundamental form calculated previously. The construction method is the same as that of the first fundamental form coefficient matrix, with the two elements on the main diagonal being the two principal coefficients of the second fundamental form, and the two elements off the main diagonal being the cross coefficients of the second fundamental form. After constructing the two coefficient matrices, the generalized eigenvalue equation is solved. The generalized eigenvalue equation is given by Bv = λAv, where B is the coefficient matrix of the second fundamental form, A is the coefficient matrix of the first fundamental form, λ is the eigenvalue to be solved, and v is the corresponding eigenvector. Solving this equation yields two eigenvalues, which are the two mutually orthogonal principal curvature values ​​at that point. Subsequent calculations are performed using the two principal curvature values ​​at each discrete point. The two principal curvature values ​​are directly multiplied to obtain the Gaussian curvature of that point; the two principal curvature values ​​are added together and then divided by two to obtain the average curvature of that point. Based on the actual structural characteristics of the polished surface of the silicon wafer, the Gaussian curvature threshold is set to 0.02 and the average curvature threshold is set to 0.05. Regions on the virtual surface with Gaussian curvature in the range of 0 to 0.02 and average curvature in the range of 0 to 0.05, and with continuous and smooth curvature changes between adjacent points, are merged into the same region. Specifically, all discrete points are traversed, and discrete points that meet the curvature threshold requirements and whose curvature difference between adjacent points is less than 0.005 are grouped into the same region. Finally, multiple homogeneous feature blocks with stable curvature changes and optical reflection characteristics and internal structural features are formed, realizing adaptive partitioning of the entire surface of the silicon wafer according to the real structure and optical characteristics.

[0027] Step 202: For each homogeneous feature block, obtain the corresponding illumination incident angle and material optical parameters based on the block's position on the virtual surface. Establish a theoretical imaging model for the block, solve for the theoretical final imaging parameters, and use the average actual imaging parameters of all pixels within the block as the measured imaging parameters. Calculate the difference between the theoretical final imaging parameters and the measured imaging parameters as the imaging parameter deviation for the corresponding block. Specifically, for each homogeneous feature block, determine the corresponding illumination incident angle, light intensity, and material optical parameters such as the refractive index and reflectivity of the silicon substrate based on the block's spatial position on the virtual surface. Substitute these parameters into the optical imaging formula; the theoretical imaging formula is... Iideal = I0 × R × k, where Iideal is the ideal imaging grayscale value, I0 is the incident light intensity, and R is a combined coefficient of the material's reflectivity and refractive index, fixed at 0.85, because the typical refractive index of a polished silicon substrate is 1.55 and the typical surface reflectivity is 0.55. The two are first multiplied to obtain the basic optical coefficient, and then corrected for slight optical absorption loss during actual imaging. Specifically, during actual imaging, light passing through the imaging lens, air medium, and the surface of the silicon substrate will experience slight absorption loss. The loss coefficient is calculated through actual measurement. The specific calculation method is as follows: Under an ideal environment with no loss, the amount of light passing through the same imaging lens, air medium, and silicon substrate surface is measured. The light intensity behind the silicon substrate surface is measured, and then the light intensity after passing through the aforementioned medium under the same conditions in an actual imaging scenario is measured. The light intensity in the actual scenario is subtracted from the light intensity in the ideal environment to obtain the light intensity loss. The light intensity loss is then divided by the light intensity in the ideal environment to calculate the loss coefficient corresponding to this type of absorption loss. This loss coefficient is subtracted from the basic optical coefficient to obtain the corrected comprehensive coefficient. This value not only conforms to the inherent optical characteristics of the polished silicon substrate but also accurately reflects the light reflection law under the combined effect of reflectivity and refractive index, ensuring the accuracy of theoretical imaging. k is a cosine illumination angle correction factor with a fixed value of 0.92. Because the polished surface of the silicon substrate has weak specular reflection characteristics, the illumination incident angle is usually controlled at 30°. The cosine of 30° is approximately 0.866. This is corrected for the optical loss of the actual imaging system. Specifically, in an actual imaging system, light experiences slight optical loss as it passes through the illumination source and imaging lens. This loss is calculated through actual measurements. The calculation involves measuring the light transmission efficiency at a 30° incident angle in an ideal environment with no system loss, and then measuring the light transmission efficiency in the actual imaging system at the same incident angle. Subtracting the transmission efficiency in the actual system from the transmission efficiency in the ideal environment yields the loss coefficient corresponding to the system's optical loss. Adding this loss coefficient to the 0.866 corresponding to the cosine of 30° gives the corrected illumination angle correction factor, which is assumed to be 0.92 is a value that can compensate for the grayscale attenuation caused by the illumination angle and also conforms to the optical characteristics of the actual imaging scene, thus establishing a theoretical imaging model suitable for this area.

[0028] Numerical iterative calculations are performed using a theoretical imaging model. The specific calculation process is as follows: First, the determined incident illumination intensity, the comprehensive coefficient R, and the correction factor k are substituted into the theoretical imaging formula to obtain the initial ideal imaging grayscale value. Then, based on the initial value and combined with the optical characteristics corresponding to the spatial location of the block, the parameters are adjusted iteratively. After each iteration, the deviation between the current ideal value and the theoretical optimal value is calculated. If the deviation is greater than a preset small threshold (0.001), the parameters are adjusted and the calculation is repeated until the deviation is less than the threshold, at which point the iteration stops. The ideal imaging grayscale value obtained at this point is the theoretical final imaging parameter corresponding to the block under ideal imaging conditions of distortion-free, noise-free, and uniform illumination. Simultaneously, all images within this homogeneous block are extracted from the preprocessed image. The actual imaging parameters of a pixel are obtained by summing the actual imaging parameters of all pixels within the block, dividing the sum by the total number of pixels in the block, and using the average value as the measured imaging parameter of that block. The difference between the measured imaging parameter and the theoretical final imaging parameter is the imaging parameter deviation corresponding to the homogeneous feature block. This deviation is used to quantify the objective difference between the actual imaging result and the ideal imaging result. The larger the absolute value of the deviation, the more obvious the deviation between the actual imaging and the ideal imaging. If the deviation is positive, it means that the actual imaging grayscale is lower than the ideal imaging grayscale. If the deviation is negative, it means that the actual imaging grayscale is higher than the ideal imaging grayscale. By using the specific value of the deviation, the magnitude and direction of the imaging error of the block can be accurately determined.

[0029] Step 203: After normalizing the imaging parameter deviations of each homogeneous feature block, the deviations are associated with the spatial position information of the corresponding blocks on the virtual surface to form a set of dynamic compensation coefficients corresponding to each homogeneous feature block on the entire surface. Specifically, this includes: mapping the imaging parameter deviations corresponding to all homogeneous feature blocks to a standard numerical range of zero to one. The specific calculation method is to subtract the minimum value among all imaging parameter deviations from the current imaging parameter deviation, and then divide the result by the difference between the maximum and minimum values ​​among all imaging parameter deviations to complete the normalization process and eliminate the influence of inconsistent deviation value ranges between different homogeneous feature blocks; binding the normalized imaging parameter deviations one-to-one with the spatial position and area information of the corresponding homogeneous feature blocks on the virtual surface to form a set of dynamic compensation coefficients corresponding one-to-one with each homogeneous feature block on the entire surface of the crystalline silicon substrate.

[0030] This embodiment constructs a three-dimensional feature evaluation field for the entire surface based on spatial distance weighted interpolation, which can accurately characterize the optical and structural differences in different regions of a crystalline silicon substrate, solving the problem of not being able to uniformly describe surface and subsurface features, and providing a foundation for global adaptive detection. Based on virtual surface curvature analysis, it divides the feature homogeneous blocks, which can adaptively partition the substrate according to its actual structural characteristics, effectively distinguishing different characteristic regions such as the center, edge, and subsurface. By comparing the theoretical imaging model with the actual imaging parameters to calculate the deviation, the imaging error in different regions can be accurately quantified, providing a reliable basis for eliminating interference such as uneven illumination and edge reflection. It generates dynamic compensation coefficients bound to spatial position to achieve differentiated compensation for different regions, improving the accuracy and stability of subsequent defect segmentation and reducing the probability of false defects and missed detections.

[0031] In a preferred embodiment of the present invention, a dynamic compensation coefficient is used to perform adaptive threshold segmentation on the preprocessed image to obtain a binarized image, and the two-dimensional geometric features of each candidate defect region are extracted. The contour point set of each candidate defect region is mapped to a virtual surface of a three-dimensional feature evaluation field, and the spatial coordinates and centroid positions of each contour point on the virtual surface are obtained, including: Step 300a1: Based on the dynamic compensation coefficients of each homogeneous feature block, adaptive threshold segmentation is performed on the preprocessed image. The binarization threshold of the corresponding pixel is adjusted according to the dynamic compensation coefficient of the homogeneous feature block where each pixel is located. Pixels in the preprocessed image with gray values ​​lower than the adjusted threshold are marked as foreground pixels, and pixels with gray values ​​higher than the adjusted threshold are marked as background pixels, thus obtaining a binarized image. Specifically, this includes: setting a preset basic binarization threshold of 60, because after gray-level correction and noise suppression, the gray values ​​of the effective defect area are generally concentrated between 30 and 60, while the gray values ​​of the normal board area are concentrated between 80 and 120. Setting the basic binarization threshold to 60 can effectively distinguish between the defect area and the normal board area; firstly, the dynamic compensation coefficient corresponding to the homogeneous feature block where each pixel is located in the image is read, and the dynamic compensation coefficient is multiplied by the preset basic binarization threshold to obtain the adjusted binarization threshold corresponding to the current pixel. Multiplication is used because the dynamic compensation coefficient is at 0 after normalization. The values ​​between 1 and 1 correspond to the imaging parameter deviation of the corresponding homogeneous feature block. The larger the deviation, the closer the compensation coefficient is to 1; the smaller the deviation, the closer the compensation coefficient is to 0. The base threshold of 60 is the benchmark for distinguishing between defects and normal areas. Multiplying the two values ​​allows for adaptive adjustment of the threshold. For areas with large imaging deviations and prone to false defects (such as edge blocks), the compensation coefficient is close to 1, and the adjusted threshold is close to 60, maintaining the benchmark's distinguishing ability. For areas with small imaging deviations and uniform gray-level distribution (such as central blocks), the compensation coefficient is smaller, and the adjusted threshold is lower than 60. This avoids misjudging slight gray-level fluctuations as defects and prevents real defects from being missed due to high gray-level. It ensures that the thresholds of different areas can fit their own optical characteristics, improving segmentation accuracy. The relationship between the gray-level value and the adjusted threshold is judged pixel by pixel. Pixels with gray-level values ​​lower than the adjusted threshold are marked as foreground pixels, and pixels with gray-level values ​​higher than the adjusted threshold are marked as background pixels. After all marking is completed, a binary image containing only foreground and background pixels is obtained.

[0032] Step 300a2 involves performing connected component labeling on the binarized image, merging adjacent foreground pixels into the same connected region as candidate defect regions, extracting the contour point set of each candidate defect region, and calculating the area, perimeter, and ratio of the major axis to the minor axis of the minimum enclosing ellipse of each candidate defect region as two-dimensional geometric features of each candidate defect region. Specifically, this includes: performing connected component labeling on the obtained binarized image, using the four-neighbor connectivity rule, traversing all foreground pixels in the image, merging foreground pixels that are directly adjacent in the four directions (up, down, left, and right) and belong to the same cluster into the same connected region, and treating each independent connected region as a candidate defect region; sequentially extracting the outer contour point set of each candidate defect region, counting the total number of all foreground pixels within the contour range, and directly using this as the area of ​​the candidate defect region; sequentially traversing each adjacent contour point along the outer contour of the candidate defect region, calculating the straight-line distance between the two points segment by segment and accumulating the results to obtain the perimeter of the candidate defect region. For each candidate defect region, a minimum bounding ellipse fitting is performed. The specific fitting process is as follows: First, extract the two-dimensional coordinates of all foreground pixels within the candidate defect region. Assume there are n foreground pixels in the region, with pixel coordinates (x1, y1), (x2, y2), ..., (xn, yn). Calculate the mean of all pixel coordinates in the region, including the mean of the horizontal and vertical coordinates, and use this mean coordinate as the centroid of the candidate defect region. Based on the centroid, calculate the covariance matrix of the region. The covariance matrix is ​​a 2×2 matrix. Specifically, the element in the first row and first column is the sum of the squares of the differences between the horizontal coordinates of all pixels and the mean horizontal coordinate, divided by the total number of pixels. The element in the first row and second column (or second row and first column) is the sum of the products of the differences between the horizontal coordinates of all pixels and the mean horizontal coordinate, and the differences between the vertical coordinates of all pixels and the mean vertical coordinate, divided by the total number of pixels. The element in the second row and second column is the sum of the squares of the differences between the vertical coordinates of all pixels and the mean vertical coordinate, divided by the total number of pixels. This yields the complete covariance matrix.

[0033] The eigenvalues ​​and eigenvectors of the covariance matrix are solved by first constructing the characteristic equation |M-eI|=0, where M is the covariance matrix calculated above, e is the eigenvalue to be solved, and I is a 2×2 identity matrix. Expanding the characteristic equation yields a quadratic equation, and the two eigenvalues ​​e1 and e2 (e1>e2) are calculated using the quadratic formula. Then, each eigenvalue is substituted into the equation (M-eI)q=0 (q is the eigenvector), and the corresponding eigenvectors q1 (corresponding to e1) and q2 (corresponding to e2) are obtained. The direction of the eigenvector q1 corresponding to the larger eigenvalue e1 is the major axis direction of the ellipse, and the direction of the eigenvector q2 corresponding to the smaller eigenvalue e2 is the minor axis direction of the ellipse. Then, taking the centroid of the candidate defect region as the center of the ellipse and the square root of the eigenvalue as the semi-axis length reference, the lengths of the major and minor semi-axis of the ellipse and the rotation angle are gradually adjusted until the ellipse can completely enclose all pixels in the candidate defect region with the smallest area, thus completing the minimum bounding ellipse fitting. After fitting, the lengths of the major and minor axes of the ellipse are measured respectively. The ratio of the major axis to the minor axis is obtained by dividing the length of the major axis by the length of the minor axis. The area, perimeter, and ratio of the major axis to the minor axis are used as the two-dimensional geometric features of each candidate defect region.

[0034] Step 300a3: Based on the two-dimensional coordinates of each pixel in the preprocessed image of the candidate defect region contour point set, and combined with the virtual surface expression of the three-dimensional feature evaluation field, calculate the spatial coordinates of each contour point on the virtual surface through spatial mapping relationship to obtain the spatial coordinate set of each contour point on the virtual surface; calculate the mean of the virtual surface spatial coordinates of each candidate defect region contour point set to obtain the spatial coordinates of the centroid position of each candidate defect region on the virtual surface. Specifically, this includes: based on the two-dimensional coordinates (denoted as x, y) of each pixel in the preprocessed image of the candidate defect region contour point set, and combined with the virtual surface expression corresponding to the three-dimensional feature evaluation field, substitute the two-dimensional image coordinates into the elevation calculation formula of the virtual surface through spatial mapping relationship. The elevation calculation formula is: z = a×x² + b×y² + c×x×y + d×x + u×y + f, where a, b, and c are the coordinates of each pixel in the preprocessed image of the candidate defect region contour point set, respectively. c, d, u, and f are the surface coefficients obtained by fitting the three-dimensional feature evaluation field, x and y are the two-dimensional coordinates of the pixels, and z is the elevation value corresponding to the coordinates on the virtual surface. Substituting the x and y coordinates of each contour point into the formula, the corresponding z value is calculated. Combined with the original two-dimensional coordinates, the three-dimensional spatial coordinates (x, y, z) of each contour point on the virtual surface are obtained, thus obtaining the complete set of surface spatial coordinates for each candidate defect region. The abscissas of all points in the surface spatial coordinate set corresponding to each candidate defect region are summed and divided by the total number of points to obtain the mean abscissa of the centroid position. The ordinates of all points are summed and divided by the total number of points to obtain the mean ordinate of the centroid position. The elevation coordinates of all points are summed and divided by the total number of points to obtain the mean elevation coordinate of the centroid position. The three sets of mean values ​​together constitute the spatial coordinates of the centroid of the candidate defect region on the virtual surface.

[0035] This embodiment uses adaptive threshold segmentation based on dynamic compensation coefficients. It can automatically adjust the binarization threshold according to the optical characteristics of homogeneous blocks with different features, effectively eliminating false defects caused by uneven illumination, edge reflection, and subsurface interference, and improving the accuracy of defect segmentation. By extracting candidate defect regions through connected component labeling and calculating two-dimensional geometric features such as area, perimeter, and aspect ratio, it is possible to initially distinguish real defects from interference points in terms of morphology. By mapping the defect contour from two-dimensional image coordinates to the virtual surface space of the three-dimensional feature evaluation field, it is possible to accurately locate the defect in three-dimensional space and obtain the true spatial centroid and distribution location of the defect.

[0036] In a preferred embodiment of the present invention, the two principal curvature direction vectors at the corresponding centroid point are calculated to derive the geometric anisotropy index of the corresponding defect region, thereby obtaining the surface geometric feature parameters, including: Step 300b1: Using the spatial coordinates of the centroid of each candidate defect region on the virtual surface as the base point, construct a local parameter domain on the virtual surface and calculate the coefficient matrices of the first and second basic forms of the virtual surface at the base point. Specifically, this includes: using the three-dimensional spatial coordinates corresponding to the centroid of each candidate defect region on the virtual surface as the calculation base point, constructing a local parameter domain on the virtual surface centered on this base point. That is, extending outward from the base point along the two inherent mutually perpendicular parameter directions of the virtual surface to form a square local analysis region. This region covers a 5×5 pixel area of ​​the surface around the base point, ensuring that the local geometry of the surface at the defect center position can be completely and stably reflected. The shape, curvature trend, and spatial distribution characteristics of the virtual surface are analyzed. Based on the constructed local parameter domain, the core objective of subsequent calculations is to solve for the first and second fundamental form coefficient matrices of the virtual surface at the base point. The first fundamental form coefficient matrix is ​​obtained by multiplying the first-order partial derivatives of the surface position vector in two mutually perpendicular parameter directions. Specifically, it contains three independent components, calculated by multiplying the first-order partial derivatives in the first parameter direction by themselves, multiplying the first-order partial derivatives in the two parameter directions by each other, and multiplying the first-order partial derivatives in the second parameter direction by themselves. These components are used to quantitatively describe the intrinsic metric properties of the virtual surface in the local parameter domain, such as length, distance, angle, and area. The first fundamental formal coefficient matrix is ​​only related to the tangent vectors in the parameter directions and is unaffected by the curvature and normal changes of the surface space. Therefore, it can stably reflect the inherent geometric relationships of the surface under conditions where there is no intrinsic deformation such as stretching or compression, and only reflects the intrinsic geometric structure of the surface itself, unaffected by external spatial attitude interference. The second fundamental formal coefficient matrix first obtains the unit normal vector by performing a cross product operation on the first-order partial derivatives of the two parameter directions. Then, it performs a dot product operation on the second-order partial derivatives and mixed second-order partial derivatives of the position vector in the parameter directions and takes the opposite value, forming three independent components. These components are used to quantitatively describe the degree of curvature, curvature direction, concavity and convexity characteristics, and spatial variation of the normal vector of the virtual surface in the local parameter domain. The variation law directly reflects the local deformation characteristics of the surface in the defect area. At the same time, the specific division rules of the local parameterized mesh are clearly defined. With the base point as the center, it is uniformly expanded along two orthogonal parameter directions. A fixed parameter step size of 0.5 pixels and a fixed mesh size of 11×11 are used for regular arrangement. The mesh range exactly covers the surface area of ​​2.5 pixels around the base point, ensuring that the mesh nodes are evenly distributed on the local surface and can accurately capture the local geometric changes of the surface. The partial derivative of the position vector is solved by the central difference method. It is calculated by dividing the coordinate difference of the adjacent positions of the node by twice the parameter step size (i.e., 0.5×2=1 pixel). This forms a complete, unified and reproducible calculation process.

[0037] Step 300b21: Using the spatial coordinates of the centroid of each candidate defect region on the virtual surface as the base point, a local parameterized mesh with a preset step size and a preset number of meshes is constructed on the virtual surface along the parameter direction with the corresponding base point as the origin. The local parameterized mesh covers the surface region within a preset radius centered on the corresponding base point. The partial derivatives of the position vectors at each node in the local parameterized mesh with respect to each parameter direction are calculated using the difference method. Specifically, this includes: using the spatial coordinates of the centroid of each candidate defect region on the virtual surface as the base point, and constructing a local parameterized mesh on the virtual surface along two mutually perpendicular parameter directions with the base point as the origin. The grid is set to a fixed step size of 0.5 pixels, with a total grid size of 11×11, covering a curved area with a radius of 2.5 pixels centered at the base point. Each node within the locally parameterized grid is traversed, and its spatial coordinates are obtained. The first partial derivative of the position coordinates with respect to each parameter direction is calculated using a difference method. Specifically, the first partial derivative with respect to the first parameter direction is calculated by taking the difference in position coordinates between the node and its two adjacent nodes along that parameter direction, and then dividing by twice the fixed step size. Similarly, the first partial derivative with respect to the second parameter direction is calculated by taking the difference in position coordinates between the node and its two adjacent nodes along that parameter direction, and then dividing by twice the fixed step size. Through these calculations, the first partial derivatives of the position coordinates of all nodes within the locally parameterized grid with respect to both parameter directions are obtained.

[0038] Step 300b22: Based on the partial derivatives of the position vector with respect to each parameter, calculate the first type of fundamental quantity of the virtual surface at the corresponding base point, including each element of the first fundamental form coefficient matrix. The first fundamental form coefficient matrix is ​​used to describe the metric properties of the virtual surface at the base point. Based on the partial derivatives of the position vector with respect to each parameter and the second-order partial derivatives of the position vector at each node in the local parameterized mesh along the parameter directions, calculate the second type of fundamental quantity of the virtual surface at the corresponding base point, including each element of the second fundamental form coefficient matrix. The second fundamental form coefficient matrix is ​​used to describe the curvature properties of the virtual surface at the base point, specifically including: calculating the first type of fundamental quantity of the virtual surface based on the first-order partial derivatives of the position coordinates at the base point with respect to the two parameter directions. The elements of a basic formal coefficient matrix are given. This matrix is ​​a symmetric matrix with two rows and two columns. The element in the upper left corner of the matrix is ​​obtained by multiplying the first partial derivative of the first parameter direction by itself, that is, multiplying the three coordinate components of the partial derivative by itself and then adding them together. The elements in the upper right and lower left corners of the matrix are obtained by multiplying the first partial derivatives of the two parameter directions by themselves, that is, multiplying the corresponding coordinate components of the two partial derivatives and then adding them together. The element in the lower right corner of the matrix is ​​obtained by multiplying the first partial derivative of the second parameter direction by itself, that is, multiplying the three coordinate components of the partial derivative by itself and then adding them together. The resulting first basic formal coefficient matrix is ​​used to describe the metric properties of the virtual surface at the base point, such as length and angle.

[0039] Step 300b23: Establish the generalized eigenvalue equation of the second fundamental form coefficient matrix relative to the first fundamental form coefficient matrix. Solve the generalized eigenvalue equation to obtain two eigenvalues, which are then used as the principal curvature values ​​of the virtual surface at the corresponding base points. Two eigenvectors corresponding to these eigenvalues ​​are also obtained, and these two eigenvectors are used as the principal curvature direction vectors of the virtual surface at the corresponding base points. Specifically, this includes: establishing the generalized eigenvalue equation of the second fundamental form coefficient matrix (denoted as B) relative to the first fundamental form coefficient matrix (denoted as A), with the equation in the form B × = ×A× (in For eigenvalues, (where B is the eigenvector); the specific process of solving this generalized eigenvalue equation is to construct the equation |B- The equation ×A|=0, when expanded, yields a quadratic equation. The two eigenvalues ​​are then calculated using the quadratic formula. , These two eigenvalues ​​are the two principal curvature values ​​of the virtual surface at the corresponding base point (representing the curvature values ​​at which the surface has the maximum and minimum curvature at that point, respectively); , Substitute them into the equation (B-λ×A)× =0, and the eigenvectors corresponding to the two eigenvalues ​​are obtained by solving the problem. , These two eigenvectors are the principal curvature direction vectors of the virtual surface at the corresponding base point (pointing to the directions of maximum and minimum curvature, respectively).

[0040] Step 300b3: Based on the spatial angle relationship between the two principal curvature direction vectors and the ratio of the two principal curvature values, construct the anisotropic tensor of the virtual surface at the corresponding base point. Calculate the ratio of the eigenvalues ​​of the anisotropic tensor as the geometric anisotropy index of the corresponding defect region. Specifically, this includes: first, calculating the spatial angle between the two principal curvature direction vectors; performing a dot product operation on the two principal curvature direction vectors; then dividing by the product of the magnitudes of the two vectors to obtain the cosine of the angle, thereby determining the orthogonality of the two principal directions (ideally θ = 90°); combining the two... The ratio of principal curvature values ​​is used to construct the anisotropic tensor of the virtual surface at the corresponding base point. That is, the tensor is formed by using the two principal curvature direction vectors as the orthogonal basis and the two principal curvature values ​​as the diagonal elements of the tensor. The ratio of eigenvalues ​​of this anisotropic tensor is calculated by dividing the larger eigenvalue of the tensor by the smaller eigenvalue. The result is the geometric anisotropy index of the corresponding defect region. The larger the index value, the more significant the difference in bending characteristics of the surface in different directions of the defect region (e.g., the index of linear defects is much larger than that of point defects).

[0041] Step 300b4 combines the geometric anisotropy index with the principal curvature values ​​and normal vector directions of the virtual surface at the corresponding base point to form the surface geometric feature parameters of the corresponding defect region. Specifically, this includes: integrating the geometric anisotropy index, the two principal curvature values, and the normal vector direction (x, y, and z components of the unit normal vector) to form the surface geometric feature parameters of the corresponding defect region. This involves organizing the core information of the principal curvature values, including the specific values ​​of the two principal curvature values ​​and their positive or negative sign (positive values ​​indicate convexity of the surface, and negative values ​​indicate concaveness); clarifying the quantitative information of the normal vector direction, i.e., recording the x, y, and z component values ​​of the unit normal vector to characterize the spatial orientation of the surface at the base point; and integrating the geometric anisotropy index, the two principal curvature values, and the normal vector direction (x, y, and z components of the unit normal vector) into an ordered set of parameters to form a complete surface geometric feature parameter, comprehensively characterizing the three-dimensional geometric properties of the defect region.

[0042] This embodiment, by constructing a locally parameterized mesh and accurately calculating the first and second fundamental form coefficient matrices of the surface, can precisely characterize the local geometric properties of the defect region surface from both metric and curvature dimensions. Solving the generalized eigenvalue equation yields the principal curvature values ​​and principal curvature direction vectors, which can clearly identify the core curvature features of the defect region surface. Combining the anisotropic tensor to calculate the geometric anisotropy index can quantitatively distinguish the surface morphological differences of different types of defects, such as the significant difference in anisotropic indices between linear defects and point defects. Combining the geometric anisotropy index, principal curvature values, and normal vector directions to form surface geometric feature parameters can comprehensively characterize the three-dimensional geometric features of the defect region, overcoming the limitations of two-dimensional geometric features.

[0043] In a preferred embodiment of the present invention, step 4 includes: Step 400: Concatenate the two-dimensional geometric features of each candidate defect region with the corresponding surface geometric feature parameters to construct a multi-dimensional fused feature vector for each candidate defect region. The two-dimensional geometric features include the defect area, defect perimeter, and the ratio of the major axis to the minor axis of the minimum enclosing ellipse. The surface geometric feature parameters include the anisotropy index, principal curvature value, and normal vector direction. Specifically, this involves concatenating the two-dimensional geometric features and surface geometric feature parameters corresponding to each candidate defect region in a fixed order to construct a unified and standardized multi-dimensional fused feature vector. The two-dimensional geometric features include three items: the defect region area, the contour perimeter, and the ratio of the major axis to the minor axis of the minimum enclosing ellipse. These are mainly used to describe the overall shape, size, and extension characteristics of the defect from a planar perspective. The surface geometric feature parameters include six items: the geometric anisotropy index, the principal curvature value, and the normal vector direction. The three spatial coordinate components of the number, two sets of principal curvature values, and normal vector are used to describe the curvature, anisotropy, and spatial orientation attributes of the defect region from a three-dimensional surface perspective. The order of arranging the two-dimensional geometric features first, followed by the surface geometric feature parameters, is adopted because the two-dimensional geometric features can intuitively reflect the external morphological differences of the defect, serving as the basis for classification and recognition. The surface geometric features are then used to further distinguish the three-dimensional geometric differences between real defects and pseudo-defects such as those caused by lighting or reflection. This arrangement of basic appearance first, followed by deep geometry, conforms to the feature extraction and learning logic of the classification model, improving model training stability and classification accuracy. All the above feature values ​​are concatenated in a fixed order to form an ordered one-dimensional vector, which serves as the multi-dimensional fusion feature vector for the current candidate defect region, realizing the joint expression of two-dimensional morphological information and three-dimensional surface geometric information.

[0044] Step 401: The multidimensional fused feature vectors of each candidate defect region are used as input data and input into the pre-trained classification model. The output category labels of each candidate defect region are then used. The category labels include microbubble defects and surface pseudo-defects. Specifically, to meet the requirements of lightweight and high-precision defect detection in the polishing section of crystalline silicon plates, an improved MLP (Multilayer Perceptron) is selected as the basic classification model of a lightweight deep neural network. The specific architecture design is as follows: the number of nodes in the input layer is perfectly matched with the dimension of the multidimensional fused feature vector (e.g., if the dimension of the fused feature vector is 64, then the input layer is set with 64 nodes), ensuring that the feature vector is input into the model without loss; the hidden layer is set with two fully connected layers, and the number of nodes in the first layer is equal to the number of nodes in the feature vector. The first layer has twice the feature vector dimension (64-dimensional features corresponding to 128 nodes), while the second layer has half the number of nodes (64-dimensional features corresponding to 32 nodes). Both layers use the ReLU activation function to mine deep correlations between features through non-linear transformation. At the same time, a Dropout layer (with a dropout rate of 0.2) is added after each fully connected layer to reduce the risk of model overfitting. The output layer has two nodes, corresponding to microbubble defects and surface pseudo-defects, respectively. The Softmax activation function is used to map the output value to a probability value in the interval [0, 1]. The sum of the probability values ​​of the two nodes is 1, which directly represents the probability of the candidate defect region belonging to the corresponding category, thus completing the overall architecture of the binary classification model.

[0045] To ensure the effectiveness and generalization ability of the model training, a dataset was constructed and preprocessed according to the principles of full coverage, standardization, and hierarchical partitioning. Actual production samples from the full surface inspection of the polished section of a silicon wafer were collected, covering different polishing pressures, polishing fluid concentrations, illumination angles, and other production conditions, as well as samples from different optical characteristic areas such as the edges, centers, and corners of the silicon wafer. At least 5000 samples of microbubble defects and at least 8000 samples of surface pseudo-defects (including edge reflections, uneven illumination, texture noise, and water stains) were included to ensure that the samples cover all typical situations in the actual inspection scenario. For each sample, its two-dimensional geometric features (defect area, perimeter, roundness, aspect ratio) and surface geometric features (curvature of the defect area) were extracted. The two types of features (value, curvature variance, and normal vector deviation) are concatenated in dimensional order to form a multi-dimensional fusion feature vector consistent with the dimension of the model input layer. All fusion feature vectors are randomly divided into training set (70%), validation set (20%), and test set (10%) in a ratio of 7:2:1. During the division process, the proportion of each category of samples in different sets is kept consistent with the total sample (i.e., the proportion of microbubble defects and surface pseudo-defects is kept consistent in the training set, validation set, and test set). The Min-Max normalization method is used to process all fusion feature vectors, mapping all feature values ​​to the interval [-1, 1] to eliminate the difference in the dimensions of different feature dimensions and avoid the single-dimensional feature value from dominating the model training process.

[0046] With the goals of minimizing loss, maximizing accuracy, and optimizing generalization ability, model training and hyperparameter optimization were conducted. The cross-entropy loss function was used as the loss metric for model training (adapting to loss calculation in binary classification scenarios). The Adam optimizer was used to iteratively update model parameters. The initial learning rate was set to 0.001, and the weight decay coefficient was set to 0.0001 to constrain the parameter size. A dynamic learning rate decay strategy was implemented: after every 10 training epochs, the learning rate decreased to 0.9 of the current value, until it stopped decaying when it reached 1e-5. The batch size was set to 32, meaning 32 sets of feature vectors were input for parameter updates each time. The training set was input into the model in batches for iterative training. After each full training epoch (traversing all training set samples), the validation set was immediately input into the model to calculate the validation. The model is trained on the following parameters: classification accuracy and loss value of the validation set; early stopping is used to prevent overfitting. Training is terminated immediately when the loss value on the validation set does not decrease for 5 consecutive rounds and the classification accuracy does not improve, thus avoiding overfitting of the model to the features of the training set; after training is terminated, the test set is input into the model for performance verification, and the three core indicators of classification accuracy, precision, and recall are calculated; if the indicators do not reach 99% or above, the hyperparameters such as the number of hidden layer nodes (±10%), the initial learning rate (1e-4 to 1e-3), and the batch size (16 / 32 / 64) are adjusted, and training is restarted until the model achieves a classification accuracy ≥99%, precision ≥99%, and recall ≥99% on the test set; finally, the trained model structure and parameter file are saved as a pre-trained model for silicon wafer defect detection.

[0047] Based on the pre-trained classification model, candidate defect regions are accurately classified. This involves extracting multi-dimensional fused feature vectors from each candidate defect region, then strictly standardizing them according to the Min-Max normalization rule of the training set to ensure the distribution of input features is consistent with the training set. The pre-processed feature vectors are then input into the pre-trained model, passed through the input layer to two fully connected layers for non-linear extraction and mapping of deep features. Finally, the classification result of the fused features is normalized using the Softmax activation function in the output layer, ultimately outputting that the region belongs to a microbubble defect. The probability values ​​of surface pseudo-defects are determined. A probability threshold of 0.95 is set. If the probability value of a microbubble defect is ≥0.95, the area is labeled as a microbubble defect. If the probability value of a surface pseudo-defect is ≥0.95, the area is labeled as a surface pseudo-defect. If both probability values ​​are below 0.95, the area is considered a suspected defect area and is temporarily labeled as a surface pseudo-defect. The probability distribution data of this area is separately marked in the final inspection report for manual review. The category labels of all candidate defect areas are summarized to form a list of classification results corresponding to the location and size of the defect area.

[0048] Step 402: Based on the category labels, candidate defect regions belonging to microbubble defects are screened, and the position coordinates of each microbubble defect region in the original image are extracted. The geometric dimensions of each microbubble defect region are calculated based on its contour point set, obtaining the position coordinates and geometric dimensions of each microbubble defect. Specifically, this includes: screening all candidate defect regions based on the category labels output by the classification model, retaining regions labeled as microbubble defects, and directly removing interference regions determined to be surface pseudo-defects, thus completing the purification and denoising of the real defects; for each retained microbubble defect region, extracting its complete two-dimensional coordinate information in the original detection image, including the coordinate set of all pixels on the outer contour of the defect and the coordinates of the defect centroid in the image, achieving precise localization of the microbubble defect; and counting the total number of all foreground pixels within the contour enclosure as the actual surface of the defect. The process involves iterating through the contour points in a clockwise or counterclockwise order, calculating the straight-line distance between adjacent points segment by segment, and summing all segment distances to obtain the perimeter of the defect. A minimum enclosing ellipse fitting calculation is then performed on the microbubble defect area. First, the two-dimensional coordinates of all foreground pixels within the defect area are extracted. By calculating the average of the horizontal and vertical coordinates of all pixels, this average coordinate is determined as the center position of the minimum enclosing ellipse. Next, using this center as a reference, the covariance matrix of the pixel coordinates is calculated, and the eigenvalues ​​and corresponding eigenvectors of the matrix are solved. The direction of the eigenvectors is used as the initial direction of the major and minor axes of the ellipse, and the square root of the eigenvalues ​​is used as the initial length of the major and minor axes. Subsequently, based on the premise of completely enclosing all pixels in the defect area, the lengths of the major and minor axes of the ellipse are adjusted incrementally in steps of a single pixel, while simultaneously adjusting the axis direction by 1° increments. The step size is optimized by rotation, and the iteration is repeated until the ellipse can fully cover all pixels of the defect area and minimize the overall area. This completes the accurate fitting of the minimum bounding ellipse, and the actual length of the major axis of the fitted ellipse is used as the equivalent maximum size of the microbubble defect. Finally, combined with the aforementioned detection and analysis results, the centroid coordinates, the complete coordinates of the contour pixels, the defect area obtained by pixel statistics, the actual perimeter obtained by accumulating the distances calculated segment by segment of the contour, and the equivalent maximum size obtained by ellipse fitting are extracted and output for each microbubble defect.

[0049] Step 403: Based on the preset quality grading standards for crystalline silicon panels, statistically analyze the number, maximum size, and distribution density of microbubble defects on the entire crystalline silicon panel. Compare the statistical results with the quality grading standards to determine the corresponding quality grade of the crystalline silicon panel. Generate a test report containing a defect distribution cloud map, quality grade determination results, and defect statistics. Specifically, this includes: performing global statistics and grade determination on all microbubble defects on the entire crystalline silicon panel according to the preset quality grading standards. The preset quality grading standards are as follows: No microbubble defects are Grade 1 (Good); Total number of defects not exceeding three and maximum size not exceeding 0.2 mm are Grade 2 (Good); Total number of defects between 4 and 8, or maximum size between 0.2 mm and 0.5 mm are Grade 3 (Defective); Total number of defects exceeding 8, or maximum size exceeding... A defect of 0.5 mm is considered non-conforming. Statistical indicators include the total number of microbubble defects, the maximum value among the equivalent maximum sizes of all defects, and the defect distribution density, where the defect distribution density is equal to the total number of defects divided by the total area of ​​the effective inspection area of ​​the crystalline silicon board. The three statistical indicators are compared with the quality grading standards item by item, and the final grade is determined according to the most stringent indicator to determine the quality grade corresponding to the current crystalline silicon board. A defect distribution cloud map is generated based on the centroid coordinates of all microbubble defects to intuitively show the distribution location and density of defects on the surface of the crystalline silicon board. The defect distribution cloud map, quality grade determination results, total number of defects, maximum size, average size, distribution density, and other data are integrated to generate a complete crystalline silicon board inspection report that includes defect detection results, geometric parameters, quality evaluation, and distribution visualization, for production judgment and process traceability.

[0050] This embodiment cascades two-dimensional geometric features with three-dimensional curved surface geometric features to form a multi-dimensional fusion feature. It can simultaneously utilize the planar morphology of defects and three-dimensional curved surface information to improve the accuracy of distinguishing between microbubble defects and pseudo-defects. Based on a pre-trained classification model, it achieves automatic discrimination, which can quickly and stably eliminate pseudo-defects caused by interference from lighting, reflection, etc., thereby improving the accuracy and robustness of defect detection. It performs precise positioning, size calculation, and global statistical analysis of microbubble defects, and automatically completes quality grading according to standards, realizing full automation from defect detection to quality judgment. It generates a complete inspection report containing distribution cloud maps, statistical data, and grading results, providing intuitive and reliable basis for the production control, process optimization, and factory acceptance of crystalline silicon boards.

[0051] like Figure 2 As shown, embodiments of the present invention also provide a full-surface intelligent visual inspection system for the polished section of a crystalline silicon substrate, comprising: The feature extraction module is used to select typical points in three regions on the preprocessed image: the geometric center of the plate surface, the edge transition, and the depth of the subsurface. Gray-level gradient statistics and texture spectrum analysis are performed on the local image blocks of each typical point to obtain a joint feature vector representing the surface and subsurface structure. The dynamic compensation coefficient calculation module is used to construct a three-dimensional feature evaluation field covering the entire surface of the silicon wafer based on the spatial coordinates and joint feature vectors of three typical points. By performing curvature analysis and gradient calculation on the virtual surface of the three-dimensional feature evaluation field, multiple homogeneous feature blocks are determined. Then, combined with lighting and material parameters, numerical simulation is performed to calculate the deviation between theoretical and measured imaging parameters, and the dynamic compensation coefficients of each homogeneous feature block are obtained. The defect region segmentation module is used to perform adaptive threshold segmentation on the preprocessed image using dynamic compensation coefficients to obtain a binarized image, and extract the two-dimensional geometric features of each candidate defect region. The contour point set of each candidate defect region is mapped to the virtual surface of the three-dimensional feature evaluation field to obtain the spatial coordinates and centroid position of each contour point on the virtual surface. The two principal curvature direction vectors at the corresponding centroid point are calculated to derive the geometric anisotropy index of the corresponding defect region and obtain the surface geometric feature parameters. The quality assessment module is used to integrate two-dimensional geometric features and surface geometric features into the classification model, distinguish between microbubble defects and surface pseudo-defects, obtain the position coordinates and geometric dimensions of microbubbles, and perform statistical evaluation of microbubble defects according to quality grading standards to generate an inspection report.

[0052] It should be noted that this system is a system corresponding to the above method. All implementation methods in the above method embodiments are applicable to this embodiment and can achieve the same technical effect.

[0053] Embodiments of the present invention also provide a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the method described above. All implementations in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0054] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for intelligent visual inspection of the entire surface of a polished section of a crystalline silicon substrate, characterized in that, The method includes: Step 1: Select typical points in the three regions of the preprocessed image: the geometric center of the plate surface, the edge transition, and the subsurface depth. Perform gray-level gradient statistics and texture spectrum analysis on the local image blocks of each typical point to obtain the joint feature vector representing the surface and subsurface structure. Step 2: Based on the spatial coordinates and joint feature vectors of three typical points, a three-dimensional feature evaluation field covering the entire surface of the silicon wafer is constructed. Multiple homogeneous feature blocks are determined by curvature analysis and gradient calculation of the virtual surface of the three-dimensional feature evaluation field. Then, numerical simulation is performed in combination with illumination and material parameters to calculate the deviation between theoretical and measured imaging parameters and obtain the dynamic compensation coefficients for each homogeneous feature block. Step 3: Use dynamic compensation coefficients to perform adaptive threshold segmentation on the preprocessed image to obtain a binarized image, and extract the two-dimensional geometric features of each candidate defect region. Map the contour point set of each candidate defect region to the virtual surface of the three-dimensional feature evaluation field to obtain the spatial coordinates and centroid position of each contour point on the virtual surface. Calculate the two principal curvature direction vectors at the corresponding centroid point, derive the geometric anisotropy index of the corresponding defect region, and obtain the surface geometric feature parameters. Step 4: After fusing the two-dimensional geometric features and the surface geometric features parameters, input them into the classification model to distinguish between microbubble defects and surface pseudo-defects, obtain the position coordinates and geometric dimensions of the microbubbles, and perform statistical evaluation of the microbubble defects according to the quality grading standards to generate an inspection report.

2. The intelligent visual inspection method for the entire surface of the polished section of a crystalline silicon substrate according to claim 1, characterized in that, The process of obtaining the preprocessed image is as follows: Flat-field correction based on pixel response characteristics is performed on the original image data. By acquiring a bright-field image under a uniform light field and a dark-field image after turning off the light source, the response characteristic parameters of each pixel are calculated. The gray values ​​of each pixel in the original image data are linearly corrected using the response characteristic parameters to eliminate the non-uniformity of photoelectric response between pixels and obtain a flat-field corrected image. The flat field correction image is subjected to column correction based on the subtraction of gray-level mean values. The gray-level mean value of each column of pixels in the flat field correction image is calculated. The gray-level value of each column of pixels in the flat field correction image is subtracted from the gray-level mean value of the corresponding column, and then the gray-level mean value of the whole image is superimposed by a set ratio value to eliminate the difference in horizontal gray-level distribution caused by uneven lighting, and the corrected image is obtained. Edge detection is performed on the corrected image to extract the contour edge point set of the silicon wafer in the corrected image. The contour edge point set is then fitted with a straight line to obtain the straight line equation that constitutes the edge of the wafer. The intersection point of adjacent straight line equations is solved to obtain the coordinates of the four corner points of the silicon wafer. An affine transformation matrix is ​​constructed based on the coordinates of the four corner points. The affine transformation matrix is ​​used to correct the geometric distortion of the calibrated image. The entire surface area of ​​the silicon wafer to be detected is extracted. The image of the entire surface area of ​​the silicon wafer is then processed by median filtering to remove isolated noise points and obtain the preprocessed image.

3. The intelligent visual inspection method for the entire surface of the polished section of a crystalline silicon substrate according to claim 2, characterized in that, Step 1 includes: Based on the differences in optical reflection characteristics of different regions on the surface of the crystalline silicon board, the center point located in the geometric center region of the board surface, the edge point located in the transition region of the board surface, and the depth point located in the subsurface depth feature region are determined in the preprocessed image. A local image patch of a first preset size is extracted centered on the center point, a local image patch of a second preset size is extracted centered on the edge point, and a local image patch of a third preset size is extracted centered on the depth point. Gradient calculation is performed on each local image patch. The gradient calculation uses horizontal and vertical convolution kernels to convolve each local image patch to obtain the horizontal and vertical gradient components of each pixel in each local image patch. The gradient magnitude and gradient direction of each pixel are calculated based on the horizontal and vertical gradient components. The gradient magnitude histogram and gradient direction histogram of each local image patch are statistically analyzed as the gradient feature components of each local image patch. The gray-level co-occurrence matrix is ​​calculated for each local image block. The gray-level co-occurrence matrix is ​​generated by selecting the preset pixel spacing and multiple angle directions. Contrast, correlation, energy and inverse difference moment texture statistics are extracted from each gray-level co-occurrence matrix. The texture statistics extracted from each angle direction are averaged and used as the texture feature components of each local image block. The gradient feature components and texture feature components corresponding to each local image patch are fused at the feature layer to form a joint feature vector.

4. The intelligent visual inspection method for the entire surface of the polished section of a crystalline silicon substrate according to claim 3, characterized in that, Step 2 includes: Based on the spatial coordinates of the center point, edge point, and depth point, the joint feature vector corresponding to each point is used as the feature value at the corresponding spatial location. A weight function is constructed based on the spatial distance between each point. The feature values ​​at each spatial location on the entire surface are interpolated by weighted summation to construct a three-dimensional feature evaluation field covering the entire surface of the crystalline silicon board. The three-dimensional feature evaluation field represents the spatial distribution of the comprehensive optical properties and structural information of the crystalline silicon board from the surface layer to the subsurface layer in the form of a continuous virtual surface. The virtual surface of the three-dimensional feature evaluation field is discretized to obtain the normal vector and principal curvature value at each discrete point on the virtual surface. The Gaussian curvature and average curvature are calculated based on the principal curvature value at each discrete point. By setting a curvature threshold, the virtual surface is divided into multiple homogeneous feature blocks with continuous curvature changes. For each homogeneous feature block, the corresponding illumination incident angle and material optical parameters are obtained according to the position of the corresponding block on the virtual surface. The theoretical imaging model of the block is established, the theoretical final imaging parameters of the corresponding block are solved, the average value of the actual imaging parameters of all pixels in the corresponding block is used as the measured imaging parameters, and the difference between the theoretical final imaging parameters and the measured imaging parameters is calculated as the imaging parameter deviation of the corresponding block. After normalizing the imaging parameter deviation of each homogeneous feature block, it is correlated with the spatial position information of the corresponding block on the virtual surface to form a set of dynamic compensation coefficients corresponding to each homogeneous feature block of the entire surface.

5. The intelligent visual inspection method for the entire surface of the polished section of a crystalline silicon substrate according to claim 4, characterized in that, Adaptive thresholding is performed on the preprocessed image using dynamic compensation coefficients to obtain a binarized image. Two-dimensional geometric features of each candidate defect region are extracted, and the contour point set of each candidate defect region is mapped to a virtual surface of a three-dimensional feature evaluation field. The spatial coordinates and centroid positions of each contour point on the virtual surface are obtained, including: Based on the dynamic compensation coefficient of each feature homogeneous block, the preprocessed image is subjected to adaptive threshold segmentation. The binarization threshold of the corresponding pixel is adjusted according to the dynamic compensation coefficient of the feature homogeneous block where each pixel is located. Pixels in the preprocessed image with gray values ​​lower than the adjusted threshold are marked as foreground pixels, and pixels with gray values ​​higher than the adjusted threshold are marked as background pixels, thus obtaining a binarized image. Connected component labeling is performed on the binarized image, and adjacent foreground pixels are merged into the same connected region as candidate defect regions. The contour point set of each candidate defect region is extracted, and the area, perimeter, and ratio of the major axis to the minor axis of the minimum enclosing ellipse of each candidate defect region are calculated as the two-dimensional geometric features of each candidate defect region. Based on the two-dimensional coordinates of each pixel in the preprocessed image of the contour point set of each candidate defect region, and combined with the virtual surface expression of the three-dimensional feature evaluation field, the spatial coordinates of each contour point on the virtual surface are calculated through spatial mapping relationship to obtain the set of spatial coordinates of each contour point on the virtual surface; the mean value of the virtual surface spatial coordinates of the contour point set of each candidate defect region is calculated to obtain the spatial coordinates of the centroid position of each candidate defect region on the virtual surface.

6. The intelligent visual inspection method for the entire surface of a polished section of a crystalline silicon substrate according to claim 5, characterized in that, Calculate the two principal curvature direction vectors at the corresponding centroid point, derive the geometric anisotropy index of the corresponding defect region, and obtain the surface geometric characteristic parameters, including: Using the spatial coordinates of the centroid positions of each candidate defect region on the virtual surface as the base points, a local parameter domain is constructed on the virtual surface. The coefficient matrices of the first and second basic forms of the virtual surface at the base points are calculated. The generalized eigenvalue problem of the coefficient matrix of the second basic form relative to the coefficient matrix of the first basic form is solved to obtain two eigenvalues ​​and corresponding eigenvectors. The two eigenvalues ​​are used as the principal curvature values ​​of the virtual surface at the corresponding base points, and the two eigenvectors are used as the principal curvature direction vectors of the virtual surface at the corresponding base points. Based on the spatial angle relationship between the two principal curvature direction vectors and the ratio of the two principal curvature values, an anisotropic tensor of the virtual surface at the corresponding base point is constructed, and the ratio of the eigenvalues ​​of the anisotropic tensor is calculated as the geometric anisotropic index of the corresponding defect region. The geometric anisotropy index is combined with the principal curvature value and normal vector direction of the virtual surface at the corresponding base point to form the surface geometric feature parameters of the corresponding defect region.

7. The intelligent visual inspection method for the entire surface of a polished section of a crystalline silicon substrate according to claim 6, characterized in that, Solving the generalized eigenvalue problem of the coefficient matrix of the second fundamental form relative to the coefficient matrix of the first fundamental form yields two eigenvalues ​​and their corresponding eigenvectors. The two eigenvalues ​​are used as the principal curvature values ​​of the virtual surface at the corresponding base points, and the two eigenvectors are used as the principal curvature direction vectors of the virtual surface at the corresponding base points, including: Using the spatial coordinates of the centroid of each candidate defect region on the virtual surface as the base point, a local parameterized mesh with a preset step size and a preset number of meshes is constructed on the virtual surface with the corresponding base point as the origin along the parameter direction. The local parameterized mesh covers the surface region with the corresponding base point as the center and a preset radius. The partial derivatives of the position vectors at each node in the local parameterized mesh with respect to each parameter direction are calculated by the difference method. Based on the partial derivatives of the position vector with respect to each parameter, calculate the first type of fundamental quantity of the virtual surface at the corresponding base point, including each element of the first fundamental form coefficient matrix, which is used to describe the metric properties of the virtual surface at the base point; based on the partial derivatives of the position vector with respect to each parameter and the second-order partial derivatives of the position vector at each node in the local parameterized mesh along the parameter direction, calculate the second type of fundamental quantity of the virtual surface at the corresponding base point, including each element of the second fundamental form coefficient matrix, which is used to describe the bending properties of the virtual surface at the base point; Establish a generalized eigenvalue equation for the coefficient matrix of the second fundamental form relative to the coefficient matrix of the first fundamental form. By solving the generalized eigenvalue equation, two eigenvalues ​​are obtained. These two eigenvalues ​​are used as the principal curvature values ​​of the virtual surface at the corresponding base point, and two eigenvectors corresponding to the two eigenvalues ​​are obtained. These two eigenvectors are used as the principal curvature direction vectors of the virtual surface at the corresponding base point.

8. The intelligent visual inspection method for the entire surface of a polished section of a crystalline silicon substrate according to claim 7, characterized in that, Step 4 includes: The two-dimensional geometric features of each candidate defect region are concatenated with the corresponding surface geometric feature parameters to construct a multi-dimensional fusion feature vector for each candidate defect region. The two-dimensional geometric features include the defect area, defect perimeter, and the ratio of the major axis to the minor axis of the minimum enclosing ellipse. The surface geometric feature parameters include the anisotropy index, principal curvature value, and normal vector direction. The multidimensional fusion feature vectors of each candidate defect region are used as input data and fed into a pre-trained classification model to output the category label of each candidate defect region. The category label includes microbubble defects and surface pseudo-defects. Based on the category label, candidate defect regions belonging to microbubble defects are filtered, the position coordinates of each microbubble defect region in the original image are extracted, and the geometric dimensions of each microbubble defect region are calculated based on the contour point set of each microbubble defect region to obtain the position coordinates and geometric dimensions of each microbubble defect. Based on the preset quality grading standard for crystalline silicon boards, the number, maximum size, and distribution density of microbubble defects on the entire crystalline silicon board are statistically analyzed. The statistical results are compared with the quality grading standard to determine the quality grade of the corresponding crystalline silicon board, and an inspection report is generated that includes a defect distribution cloud map, quality grade determination results, and defect statistics.

9. A full-surface intelligent visual inspection system for a polished section of a crystalline silicon substrate, wherein the system implements the method as described in any one of claims 1 to 8, characterized in that, include: The feature extraction module is used to select typical points in three regions on the preprocessed image: the geometric center of the plate surface, the edge transition, and the depth of the subsurface. Gray-level gradient statistics and texture spectrum analysis are performed on the local image blocks of each typical point to obtain a joint feature vector representing the surface and subsurface structure. The dynamic compensation coefficient calculation module is used to construct a three-dimensional feature evaluation field covering the entire surface of the silicon wafer based on the spatial coordinates and joint feature vectors of three typical points. By performing curvature analysis and gradient calculation on the virtual surface of the three-dimensional feature evaluation field, multiple homogeneous feature blocks are determined. Then, combined with lighting and material parameters, numerical simulation is performed to calculate the deviation between theoretical and measured imaging parameters, and the dynamic compensation coefficients of each homogeneous feature block are obtained. The defect region segmentation module is used to perform adaptive threshold segmentation on the preprocessed image using dynamic compensation coefficients to obtain a binarized image, and extract the two-dimensional geometric features of each candidate defect region. The contour point set of each candidate defect region is mapped to the virtual surface of the three-dimensional feature evaluation field to obtain the spatial coordinates and centroid position of each contour point on the virtual surface. The two principal curvature direction vectors at the corresponding centroid point are calculated to derive the geometric anisotropy index of the corresponding defect region and obtain the surface geometric feature parameters. The quality assessment module is used to integrate two-dimensional geometric features and surface geometric features into the classification model, distinguish between microbubble defects and surface pseudo-defects, obtain the position coordinates and geometric dimensions of microbubbles, and perform statistical evaluation of microbubble defects according to quality grading standards to generate an inspection report.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a program that, when executed by a processor, implements the method as described in any one of claims 1 to 8.

Images