A wafer edge defect detection method, system, and storage medium based on image recognition.

Abstract

This application provides a wafer edge defect detection method, system, and storage medium based on image recognition, belonging to the field of semiconductor manufacturing technology. The method includes: processing annular image data to obtain target image data; inputting the target image data into an edge defect prediction model to obtain a defect probability distribution map; identifying pixel regions with probability values ​​greater than or equal to a preset threshold as candidate defect regions based on the defect probability distribution map; extracting features from the candidate defect regions to obtain target feature vectors; inputting the target feature vectors into a defect classification model to determine the defect type and classification confidence of the candidate defect regions; selecting target defect regions from the candidate defect regions based on the defect type and / or classification confidence; acquiring three-dimensional point cloud data of the target defect regions; calculating the three-dimensional geometric parameters of the target defect regions based on the three-dimensional point cloud data; and determining the detection result of the target wafer based on the defect type and the three-dimensional geometric parameters.

Description

Technical Field

[0001] This application relates to the field of semiconductor manufacturing technology, and in particular to a wafer edge defect detection method, system, and storage medium based on image recognition. Background Technology

[0002] In the semiconductor manufacturing industry, wafers, as the fundamental carriers of integrated circuits, play a decisive role in the performance of the final product. As the semiconductor industry continues to advance towards higher process nodes, the requirements for wafer quality are becoming increasingly stringent. The wafer edge area is prone to defects such as film peeling, scratches, and particle adhesion due to its frequent contact with equipment components during manufacturing and handling, and the differences in edge cleaning distances and thin film growth layers across different processes.

[0003] If wafer edge defects are not detected and addressed in a timely manner, they can fall into the effective device area of ​​the wafer during subsequent processes, severely impacting chip performance and even causing product failure, significantly increasing production costs and reducing production yield. Traditional wafer edge defect detection methods, such as manual visual inspection, are inefficient and have low accuracy in identifying minute defects; optical inspection, while suitable for large-area inspection, is ineffective for detecting small and shallow defects and struggles to detect internal wafer defects; existing monitoring methods for wafer edge defects mostly rely on scanning programs based on known defect types, failing to promptly detect unknown new types of defects, resulting in monitoring blind spots.

[0004] Therefore, there is an urgent need for an efficient, accurate, and adaptable wafer edge defect detection method that can adapt to complex defect types. Summary of the Invention

[0005] To address the aforementioned technical problems, this application provides a wafer edge defect detection method, system, and storage medium based on image recognition.

[0006] A first aspect of this application provides a wafer edge defect detection method based on image recognition, comprising:

[0007] Obtain annular image data of the chamfered region of the target wafer, and process the annular image data to obtain the target image data;

[0008] The target image data is input into the edge defect prediction model to obtain a defect probability distribution map that characterizes the probability of defects existing at each pixel position in the chamfer region of the target wafer.

[0009] Based on the defect probability distribution map, pixel regions with probability values ​​greater than or equal to a preset threshold are identified as candidate defect regions.

[0010] Feature extraction is performed on the candidate defect region to obtain the target feature vector;

[0011] The target feature vector is input into the defect classification model to determine the defect type of the candidate defect region and its corresponding classification confidence.

[0012] Based on the defect type and / or classification confidence level, target defect regions requiring three-dimensional measurement are selected from the candidate defect regions;

[0013] Obtain three-dimensional point cloud data of the target defect region, and calculate the three-dimensional geometric parameters of the target defect region based on the three-dimensional point cloud data;

[0014] Based on the defect type and the three-dimensional geometric parameters, the detection result of the target wafer is determined.

[0015] Preferably, the annular image data includes: bright-field image data and dark-field image data;

[0016] The process of processing the annular image data to obtain the target image data includes:

[0017] Based on wafer geometric features or image feature points, the bright field image data and the dark field image data are spatially registered to obtain a spatially aligned registered bright field image and a registered dark field image.

[0018] The image portions corresponding to the wafer chamfer region are segmented from the registered bright-field image and the registered dark-field image to obtain the registered bright-field chamfer image and the registered dark-field chamfer image.

[0019] The registered bright-field beveled image and the registered dark-field beveled image are respectively subjected to ring-shaped unfolding processing to obtain a planar bright-field image and a planar dark-field image; wherein, the ring-shaped image is unfolded along the circumferential direction and mapped to a rectangular planar image, so that the angular coordinates on the ring correspond to the width direction coordinates of the rectangular plane, and the radial coordinates on the ring correspond to the height direction coordinates of the rectangular plane.

[0020] The planar bright-field image and the planar dark-field image are subjected to image fusion processing to generate a noise-reduced and enhanced image as the target image data.

[0021] This embodiment combines bright-field and dark-field image data to acquire wafer chamfer region information under different lighting conditions. The complementary image features help to more comprehensively detect defects. Spatial registration of the two images ensures consistent spatial positions. The chamfer region is segmented and expanded into a ring shape, converting complex ring data into a planar image that is easier to process, simplifying subsequent operations. Image fusion generates a noise-resistant enhanced image, which can suppress noise interference, highlight potential defect signals, and significantly improve the accuracy and reliability of defect detection, especially for detecting minute defects.

[0022] Preferably, the step of performing image fusion processing on the planar bright-field image and the planar dark-field image to generate a noise-reduced and enhanced image includes:

[0023] Calculate the pixel difference matrix between the planar bright-field image and the planar dark-field image to obtain the difference map;

[0024] The difference map is subjected to denoising filtering to obtain a filtered difference map; wherein, the denoising filtering is used to suppress noise and retain potential defect difference signals;

[0025] The filtered difference image, the planar bright field image, and the planar dark field image are fused together to generate a noise-enhanced image.

[0026] This embodiment obtains a difference map by calculating the pixel difference matrix, which highlights the differences between bright and dark field images. These differences are often related to defects. The difference map is then denoised by filtering, suppressing noise while preserving key defect difference signals, thus avoiding noise interference with defect detection. The filtered difference map is then fused with the original image to further enhance defect features, resulting in a noise-resistant enhanced image that more clearly presents defect information, improving the accuracy and stability of defect detection and reducing false positives and false negatives.

[0027] Preferably, the step of fusing the filtered difference image, the planar bright-field image, and the planar dark-field image to generate a noise-reduced and enhanced image includes:

[0028] Based on the pixel values ​​of the filtered difference map, the fusion weights of the planar bright field image and the planar dark field image at corresponding pixel positions are calculated;

[0029] The planar bright-field image and the planar dark-field image are weighted and superimposed according to the fusion weights to generate an initial fused image;

[0030] The filtered difference image is superimposed and fused with the initial fused image to generate the noise-reduced enhanced image.

[0031] This embodiment calculates fusion weights based on the filtered difference map, adaptively adjusting the contribution of bright and dark field images during fusion according to potential defect information in the images, making the fusion result more prominent in defect features. Weighted superposition generates an initial fused image, combining the advantages of both images. This image is then superimposed with the filtered difference map to further enhance defect details, ensuring that the final noise-enhanced image retains effective information from the original image while maximizing defect identification and detection accuracy, providing high-quality image data for subsequent defect analysis.

[0032] Preferably, the step of identifying pixel regions with probability values ​​greater than or equal to a preset threshold as candidate defect regions based on the defect probability distribution map includes:

[0033] The defect probability distribution map is binarized to generate a binary mask map;

[0034] Pixels in the binary mask image whose probability value is greater than or equal to a preset threshold are marked as foreground regions;

[0035] The binary mask image is subjected to morphological closing operation to obtain the target foreground region;

[0036] Based on the connected component analysis algorithm, connected components are extracted from the target foreground region to obtain a set of connected components;

[0037] Connected regions in the connected region set that are greater than or equal to the defect area threshold are selected as candidate defect regions.

[0038] This embodiment performs binarization processing on the defect probability distribution map, converting continuous probability values ​​into a binary form that facilitates analysis and highlights potential defect areas. Marking the foreground region and performing morphological closing operations fills in small holes and connects broken sections, making the defect area more complete and easier to identify accurately. Connected component analysis algorithms extract connected components, separating independent defect areas. Candidate defect areas are filtered based on area thresholds, eliminating minor interference areas caused by noise, accurately identifying areas that are truly likely to contain defects, improving the accuracy and efficiency of defect detection, and reducing subsequent invalid analysis.

[0039] Preferably, the method for setting the preset threshold includes:

[0040] Statistically analyze the global pixel probability histogram of the defect probability distribution map;

[0041] When the global pixel probability histogram has bimodal characteristics, the segmentation threshold calculated by the adaptive threshold algorithm is used as the preset threshold.

[0042] When the global pixel probability histogram has no bimodal features, the segmentation threshold calculated by the maximum inter-class variance method is used as the preset threshold.

[0043] This embodiment selects a suitable preset threshold setting method based on the characteristics of the global pixel probability histogram, exhibiting strong adaptability. When the histogram exhibits bimodal characteristics, the adaptive thresholding algorithm can automatically determine a suitable threshold based on the image's own characteristics, accurately segmenting defect regions. When there are no bimodal characteristics, the maximum inter-class variance method can effectively calculate the threshold that best distinguishes defects from the background. This flexible threshold setting method ensures accurate identification of candidate defect regions under different image features, improving the versatility and accuracy of the detection method and adapting to various complex wafer edge image detection scenarios.

[0044] Preferably, calculating the three-dimensional geometric parameters of the target defect region based on the three-dimensional point cloud data includes:

[0045] The three-dimensional point cloud data is registered with the theoretical geometric surface position in the wafer physical coordinate system obtained by mapping the position information of the target defect region in the annular image data;

[0046] The registered 3D point cloud data is denoised, and a triangular mesh surface model of the defect area is constructed.

[0047] Based on the triangular mesh surface model, the depth distribution, volume, surface area, and aspect ratio of the defect region projected onto the wafer tangent plane are calculated.

[0048] Extract the three-dimensional coordinates of the edge contour of the defect area, and calculate the average curvature and curvature fluctuation of the contour line.

[0049] This embodiment registers 3D point cloud data with the theoretical geometric surface position, accurately mapping measurement data to the actual physical coordinate system of the wafer, providing an accurate benchmark for subsequent analysis. Noise reduction and the construction of a triangular mesh surface model effectively model the defect region, facilitating in-depth analysis of its geometric characteristics. Calculating parameters such as depth distribution, volume, and surface area allows for the quantification of defect size and severity from multiple dimensions; calculating the projected aspect ratio, average curvature of the edge contour, and curvature fluctuation further describes the defect shape and edge features, comprehensively acquiring the 3D geometric information of the defect, providing rich and accurate data support for accurately assessing wafer quality and the impact of defects.

[0050] Preferably, the step of extracting the three-dimensional coordinates of the edge contour of the defect region and calculating the average curvature and curvature fluctuation of the contour line includes:

[0051] Extract the boundary vertices of the defect region from the triangular mesh surface model and connect them in spatial order to form a closed edge contour line;

[0052] Traverse each vertex on the edge contour line and calculate the discrete curvature of each vertex;

[0053] The average curvature is obtained by taking the arithmetic mean of the discrete curvatures of all vertices on the edge contour line;

[0054] Calculate the standard deviation of all discrete curvatures as the curvature fluctuation.

[0055] This embodiment extracts edge contours from a triangular mesh surface model and calculates relevant parameters, enabling detailed analysis of defect edge characteristics. The average curvature, obtained by calculating the discrete curvature of the vertices and averaging it, reflects the overall degree of curvature of the defect edge. The standard deviation of the discrete curvature is calculated as a curvature fluctuation, reflecting the changes in edge curvature. These parameters help determine the smoothness and complexity of the defect edge, further enriching the description of defect geometric features. This facilitates more accurate identification of defect types, assessment of the impact of defects on wafer performance, and improvement of the accuracy and reliability of inspection results.

[0056] A second aspect of this application provides a wafer edge defect detection system based on image recognition, comprising:

[0057] The data processing module is used to acquire annular image data of the chamfered region of the target wafer, and process the annular image data to obtain target image data;

[0058] The data prediction module is used to input the target image data into the edge defect prediction model to obtain a defect probability distribution map that characterizes the probability of defects at each pixel position in the chamfer region of the target wafer.

[0059] The defect identification module is used to identify pixel regions with probability values ​​greater than or equal to a preset threshold as candidate defect regions based on the defect probability distribution map.

[0060] The data extraction module is used to extract features from the candidate defect region to obtain the target feature vector;

[0061] The defect type module is used to input the target feature vector into the defect classification model to determine the defect type of the candidate defect region and its corresponding classification confidence.

[0062] The defect determination module is used to filter out the target defect region that needs to be measured in three dimensions from the candidate defect regions based on the defect type and / or classification confidence level.

[0063] The geometric parameter module is used to acquire the three-dimensional point cloud data of the target defect area and calculate the three-dimensional geometric parameters of the target defect area based on the three-dimensional point cloud data.

[0064] The detection result module is used to determine the detection result of the target wafer based on the defect type and the three-dimensional geometric parameters.

[0065] A third aspect of this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the image recognition-based wafer edge defect detection system described above.

[0066] A fourth aspect of this application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the computer program to implement the steps of the above-described image recognition-based wafer edge defect detection method.

[0067] The beneficial effects of the wafer edge defect detection method, system, and storage medium based on image recognition provided in this application are as follows: This application processes the annular image data of the wafer chamfer region and generates a defect probability distribution map using an edge defect prediction model, enabling rapid and accurate localization of potential defect areas, thus improving the efficiency and sensitivity of defect detection. Multiple feature extraction methods are employed to extract features from candidate defect areas, and the defect type and its classification confidence are determined based on a defect classification model, effectively improving the accuracy of defect type identification and providing a more reliable basis for subsequent process improvement and quality control. Based on the defect type and classification confidence, target defect areas requiring three-dimensional measurement are selected, and their three-dimensional point cloud data is obtained to calculate three-dimensional geometric parameters, enabling a more comprehensive understanding of the actual defect situation and providing richer and more accurate information for wafer quality assessment. This application achieves deep analysis from two-dimensional recognition to three-dimensional quantification, improving defect identification accuracy and classification precision, providing reliable data support for wafer quality judgment, helping to promptly identify potential quality risks, improve wafer production yield, and reduce manufacturing costs. Attached Figure Description

[0068] Figure 1 A schematic flowchart of a wafer edge defect detection method based on image recognition provided in an embodiment of this application;

[0069] Figure 2 This is a structural block diagram of a wafer edge defect detection system based on image recognition provided in an embodiment of this application;

[0070] Figure 3 This is a schematic block diagram of an electronic device provided in an embodiment of this application. Detailed Implementation

[0071] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0072] To make the purpose, technical solution, and advantages of this application clearer, the following will be described in conjunction with the appendix. Figure 1-3 The following is an explanation using specific examples.

[0073] Please refer to Figure 1 , Figure 1 This is a flowchart illustrating an image recognition-based wafer edge defect detection method according to an embodiment of this application. The method includes:

[0074] S101: Obtain the annular image data of the chamfered region of the target wafer, process the annular image data to obtain the target image data;

[0075] In this embodiment, an industrial camera and optical system are used to acquire annular image data of the chamfered area of ​​the target wafer. The industrial camera has a resolution of more than 5 million pixels to ensure that it can capture micron-level details. The optical system uses multi-angle light source illumination, typically including a 45° oblique light source and a vertical light source. The oblique light source can highlight the concave and convex features of the edge, while the vertical light source ensures uniform brightness of the overall image. By synchronously controlling the camera exposure and the light source flicker, motion blur is avoided. The final acquired annular image data covers the complete 360° information of the wafer chamfered area, and the ratio of pixel size to actual size also needs to be accurately calibrated.

[0076] In this embodiment, the annular image data is processed to obtain the target image data. First, image preprocessing is performed, using Gaussian filtering to remove high-frequency noise, with the filter kernel size set according to the noise intensity. Next, an adaptive threshold segmentation method is used to convert the image into a binary image, effectively distinguishing the chamfered area from the background. Then, image registration technology is used to align the annular image with the standard contour model of the wafer, eliminating positional errors caused by wafer placement deviations. Finally, distortion correction is performed, using the camera's intrinsic parameter matrix and distortion coefficients to correct radial and tangential distortions caused by the optical system, ensuring the geometric accuracy of the target image data. The processed target image data meets the requirements for subsequent detection in terms of contrast, sharpness, and geometric accuracy.

[0077] S102: Input the target image data into the edge defect prediction model to obtain a defect probability distribution map that represents the probability of defects existing at each pixel position in the chamfer region of the target wafer;

[0078] In this embodiment, the edge defect prediction model is built on deep learning and incorporates an attention mechanism module in the encoding section, enabling it to focus more intently on areas with defects. During model training, a wafer chamfer image dataset including various typical defects is used. The training process employs a cross-entropy loss function and iterates parameters using the Adam optimizer. In this embodiment, after inputting target image data into the edge defect prediction model, the output defect probability distribution map shows that the probability value of each pixel ranges from 0 to 1, accurately reflecting the likelihood of a defect at that location.

[0079] S103: Based on the defect probability distribution map, identify pixel regions with probability values ​​greater than or equal to a preset threshold as candidate defect regions.

[0080] In this embodiment, the preset threshold was obtained through extensive experimental verification. For pixel regions with a probability value greater than or equal to the threshold, this embodiment uses a region growing algorithm to perform connected component analysis, merging adjacent high-probability pixels into a complete candidate defect region, while recording basic information such as the position coordinates, area size, and bounding rectangle of each region.

[0081] S104: Extract features from the candidate defect region to obtain the target feature vector;

[0082] In this embodiment, feature extraction is performed on the candidate defect region to obtain the target feature vector. The extracted features include geometric features, texture features, and grayscale features. Geometric features are calculated using a contour extraction algorithm, and these features include the perimeter, area, circularity, and rectangularity of the candidate defect region. Texture features are obtained using a gray-level co-occurrence matrix (GLCM), extracting features such as energy, entropy, and contrast. Grayscale features include the average grayscale value, grayscale standard deviation, and maximum and minimum grayscale difference of pixels within the candidate defect region. After standardization, these features are combined to form the target feature vector.

[0083] S105: Input the target feature vector into the defect classification model to determine the defect type of the candidate defect region and its corresponding classification confidence.

[0084] In this embodiment, the defect classification model employs Support Vector Machines (SVMs), Random Forests, or Deep Neural Networks. If SVMs are used, the penalty parameters and kernel function parameters are optimized using a grid search method, with the radial basis function (RBF) selected as the kernel function. If Deep Neural Networks are used, a softmax layer is added at the end of the model to output the probabilities of each category. The training dataset includes multiple common defects, with each category containing multiple sample sizes. Data augmentation techniques (such as rotation, scaling, and adding noise) are used to expand the sample size. After model training, the target feature vectors are classified, and the output defect types include scratches, chipped edges, dents, and bulges. The classification confidence level is between 0 and 1, used to describe the reliability of the classification results.

[0085] S106: Based on defect type and / or classification confidence, filter out the target defect areas that require three-dimensional measurement from the candidate defect areas;

[0086] In this embodiment, when filtering target defect areas requiring 3D measurement based on defect type and classification confidence level, filtering rules can be set. For example, for defect types that significantly impact wafer performance, such as scratches and edge chipping, all are included in the 3D measurement range regardless of classification confidence level; for defects such as dents and protrusions, 3D measurement is performed when the classification confidence level is greater than a preset value. Simultaneously, areas that are clearly pseudo-defects, such as small, high-probability areas caused by image noise, are excluded to ensure the efficiency and accuracy of 3D measurement.

[0087] S107: Obtain the three-dimensional point cloud data of the target defect area, and calculate the three-dimensional geometric parameters of the target defect area based on the three-dimensional point cloud data;

[0088] In this embodiment, acquiring the three-dimensional point cloud data of the target defect region includes:

[0089] Determine the position coordinates of the target defect region in the target image data, and calculate the physical position of the target defect region on the wafer according to the predefined mapping relationship from the image coordinate system to the wafer physical coordinate system;

[0090] Control the 3D measurement equipment to scan at the physical location and acquire 3D point cloud data of the target defect area.

[0091] In this embodiment, when acquiring the three-dimensional point cloud data of the target defect area, the three-dimensional measurement device used can be a laser scanning microscope or a white light interferometer. The laser scanning microscope scans the target defect area point by point to generate high-density three-dimensional point cloud data, while the white light interferometer analyzes the changes in interference fringes to quickly acquire the three-dimensional topography of a large area. Appropriate measurement equipment and scanning range are selected based on the size of the defect area to ensure that the point cloud data completely covers the defect area and a certain surrounding area.

[0092] In this embodiment, the three-dimensional geometric parameters of the target defect region are calculated based on three-dimensional point cloud data. First, the three-dimensional point cloud is preprocessed to remove outliers. Then, point cloud registration is performed to align the point cloud data with the three-dimensional model of the wafer. The calculated three-dimensional geometric parameters include the maximum depth / height, length, width, volume, surface area, and slope of the defect region. The volume calculation uses triangulation, where the point cloud surface is meshed and then integrated.

[0093] S108: Determine the inspection results of the target wafer based on the defect type and three-dimensional geometric parameters.

[0094] In this embodiment, a defect evaluation standard is established based on the inspection results of the target wafer determined by defect type and three-dimensional geometric parameters. Different defect judgment thresholds are formulated according to the application scenario of the wafer. The three-dimensional geometric parameters corresponding to the defect type are compared with the judgment thresholds, and the final inspection results are output, including qualified, unqualified, and detailed defect information, providing a basis for wafer quality control and subsequent processing. The inspection results in this embodiment also include processing levels, divided into three categories: release, rework, and scrap.

[0095] As can be seen from the above, this application, by processing the annular image data of the wafer chamfer region and generating a defect probability distribution map using an edge defect prediction model, can quickly and accurately locate potential defect areas, improving the efficiency and sensitivity of defect detection. Multiple feature extraction methods are employed to extract features from candidate defect areas, and the defect type and its classification confidence are determined based on a defect classification model, effectively improving the accuracy of defect type identification and providing a more reliable basis for subsequent process improvement and quality control. Based on the defect type and classification confidence, target defect areas requiring 3D measurement are selected, and their 3D point cloud data is obtained to calculate 3D geometric parameters, enabling a more comprehensive understanding of the actual defect situation and providing richer and more accurate information for wafer quality assessment. This application achieves deep analysis from 2D recognition to 3D quantization, improving defect identification accuracy and classification precision, providing reliable data support for wafer quality judgment, helping to promptly identify potential quality risks, improve wafer production yield, and reduce manufacturing costs.

[0096] In one embodiment of this application, the ring image data includes: bright field image data and dark field image data;

[0097] The ring-shaped image data is processed to obtain the target image data, including:

[0098] Based on wafer geometric features or image feature points, spatial registration is performed on bright-field image data and dark-field image data to obtain spatially aligned registered bright-field and dark-field images.

[0099] For the registered bright-field image and the registered dark-field image, the image portion corresponding to the wafer chamfer region is segmented to obtain the registered bright-field chamfer image and the registered dark-field chamfer image;

[0100] The registered bright-field beveled image and the registered dark-field beveled image are respectively subjected to ring-shaped unfolding to obtain planar bright-field image and planar dark-field image; wherein, the ring image is unfolded along the circumference and mapped to a rectangular planar image, so that the angular coordinates on the ring correspond to the width direction coordinates of the rectangular plane, and the radial coordinates on the ring correspond to the height direction coordinates of the rectangular plane.

[0101] Image fusion processing is performed on planar bright-field images and planar dark-field images to generate noise-resistant enhanced images as target image data.

[0102] In this embodiment, the annular image data includes bright-field image data and dark-field image data. Bright-field image data and dark-field image data can present the features of the wafer chamfer area from different angles. Bright-field images can clearly show the overall structure and larger defects on the wafer surface, while dark-field images can show the tiny defects and surface scratches on the wafer surface. Fusing bright-field image data and dark-field image data can improve the comprehensiveness and accuracy of defect detection.

[0103] In this embodiment, spatial registration is performed on bright-field image data and dark-field image data based on wafer geometric features or image feature points to obtain spatially aligned registered bright-field and dark-field images. Specifically, if wafer geometric features are used for registration, geometric parameters such as the wafer's center and radius can be used to make the wafer centers of the bright-field and dark-field images coincide and have the same radius through coordinate transformation. If image feature points are used for registration, feature points are first extracted from the bright-field and dark-field image data using a scale-invariant feature transformation algorithm. These feature points can be specific protrusions, depressions, or other obvious texture structures on the wafer edge. Then, a random sampling consensus algorithm is used to match the feature points and eliminate mismatched points. Finally, a spatial transformation matrix is ​​calculated based on the matched feature points, and a geometric transformation is performed on the dark-field image data to achieve accurate alignment with the bright-field image data.

[0104] For the registered bright-field and dark-field images, the image portions corresponding to the wafer chamfer region are segmented to obtain the registered bright-field chamfer image and the registered dark-field chamfer image. During the segmentation process, the approximate range of the chamfer region is first determined in the registered image based on preset parameters such as the wafer diameter and chamfer width. Then, the Canny edge detection algorithm is used to extract the inner and outer edge contours of the wafer. The accurate boundary of the chamfer region is determined through contour analysis, and the image portion within the boundary is segmented to remove background and interference from other areas of the wafer.

[0105] The registered bright-field beveled image and the registered dark-field beveled image were respectively subjected to ring-shaped unfolding to obtain planar bright-field image and planar dark-field image. Specifically, the ring-shaped image was unfolded along the circumference and mapped onto a rectangular planar image, so that the angular coordinates on the ring corresponded to the width coordinates of the rectangular plane, and the radial coordinates on the ring corresponded to the height coordinates of the rectangular plane. For example, for a ring-shaped beveled region with radius R, an angle range of 0-360°, and a radial range of R1-R2 (R1 being the inner radius and R2 the outer radius), the number of pixels in the width direction of the unfolded rectangular planar image was determined based on the angular resolution. If each degree corresponds to 2 pixels, the width is 720 pixels. The number of pixels in the height direction was determined based on the radial length and pixel size. If the radial length is 100 μm and the pixel size is 0.5 μm / pixel, the height is 200 pixels. This unfolding method transforms the two-dimensional distribution features of the ring into rectangular planar features that are easier for subsequent processing and analysis, while maintaining the relative positional relationships and grayscale information between pixels during the unfolding process.

[0106] This embodiment performs image fusion processing on planar bright-field and dark-field images. A multi-scale transform fusion algorithm, such as wavelet transform-based fusion, can be employed. First, wavelet decomposition is performed on the two planar images to obtain low-frequency and high-frequency components. The low-frequency components reflect the overall brightness and contour information of the image and are fused using a weighted average method. The weights are determined based on the sharpness of the two images, with the image with higher sharpness receiving a larger weight. The high-frequency components reflect the edge and detail information of the image and are fused using the method of taking the largest absolute value to retain richer detail features. The fused, noise-reduced, enhanced image combines the advantages of both bright-field and dark-field images, clearly presenting the overall structure of the wafer chamfer region while highlighting the characteristics of minute defects, providing high-quality image data for subsequent defect detection.

[0107] In one embodiment of this application, image fusion processing is performed on a planar bright-field image and a planar dark-field image to generate a noise-enhanced image, including:

[0108] Calculate the pixel difference matrix between the planar bright-field image and the planar dark-field image to obtain the difference map;

[0109] The difference map is subjected to denoising filtering to obtain a filtered difference map; the denoising filtering is used to suppress noise and retain potential defect difference signals.

[0110] The filtered difference image, the planar bright field image, and the planar dark field image are fused to generate a noise-reduced and enhanced image.

[0111] In this embodiment, the process of image fusion processing of planar bright-field images and planar dark-field images to generate noise-resistant enhanced images can be divided into the following three key steps. Each step aims to fully exploit the feature differences between the two images while suppressing noise interference:

[0112] The process involves calculating the pixel difference matrix between the planar bright-field image and the planar dark-field image to obtain a difference map. Planar bright-field and planar dark-field images differ significantly in their imaging principles: bright-field images are imaged using direct light, and defect areas typically exhibit a decrease in grayscale value; dark-field images, on the other hand, utilize scattered light, and small defects show an increase in grayscale value due to light scattering. Based on this characteristic, the pixel difference matrix is ​​calculated by subtracting the grayscale values ​​pixel by pixel. That is, the grayscale value of a pixel in the difference map = the grayscale value of the corresponding pixel in the dark-field image - the grayscale value of the corresponding pixel in the bright-field image. This calculation highlights the difference in grayscale response to the same defect area in the two images. For example, a scratch may have a higher grayscale value in the dark-field image but a lower grayscale value in the bright-field image; after difference, a distinct high-grayscale area will be formed, laying the foundation for subsequent extraction of potential defect signals.

[0113] The difference image is denoised and filtered to obtain a filtered difference image. The denoising filter suppresses noise while preserving potential defect difference signals. While highlighting defect differences, the difference image also amplifies random noise in both images; therefore, a filtering algorithm with both denoising and edge preservation capabilities is needed. In practical applications, bilateral filtering or nonlocal mean filtering is often used: bilateral filtering considers pixel spatial distance and gray-level similarity to smooth noise while preserving the sharpness of defect edges; nonlocal mean filtering effectively suppresses Gaussian noise and impulse noise by weighted averaging of neighboring blocks similar to the target pixel in the image. In this embodiment, after denoising filtering, the noise energy in the resulting filtered difference image is significantly reduced, while the difference signal in the defect region still maintains a clear gray-level gradient.

[0114] A noise-enhanced image is generated by fusing a filtered difference image, a planar bright-field image, and a planar dark-field image. The fusion process employs a multi-feature weighted fusion strategy, the core of which is to assign dynamic weights to the three images. This ensures that the fusion result preserves the overall structural information of the bright-field image, the minute defect details of the dark-field image, and highlights the defect difference signals in the filtered difference image. Specifically, the three images are first normalized to grayscale. Then, the local sharpness of each image is evaluated by calculating the gradient magnitude: the gradient magnitude is higher in the overall structural regions of the planar bright-field image, higher in the minute defect regions of the planar dark-field image, and higher in the regions with significant defect differences in the filtered difference image. Based on the gradient magnitude, a weight is assigned to each pixel: at a given pixel location, the larger the gradient magnitude, the higher the weight. Finally, the pixel grayscale value of the noise-enhanced image is calculated as: Planar bright-field image pixel grayscale value × bright-field weight + Planar dark-field image pixel grayscale value × dark-field weight + Filtered difference image pixel grayscale value × difference weight. Through this dynamic weighted fusion, the noise-enhanced image can not only clearly present the overall outline of the wafer chamfer area, but also highlight the detailed features of tiny defects. At the same time, by enhancing the differential signal, the contrast between defects and the background is further improved, providing more reliable image data for subsequent defect detection and classification.

[0115] In one embodiment of this application, a filtered difference image, a planar bright-field image, and a planar dark-field image are fused to generate a noise-reduced and enhanced image, including:

[0116] Based on the pixel values ​​of the filtered difference map, calculate the fusion weights of the planar bright field image and the planar dark field image at corresponding pixel positions;

[0117] The planar bright-field image and the planar dark-field image are weighted and superimposed according to the fusion weights to generate an initial fused image;

[0118] The filtered difference image is superimposed and fused with the initial fused image to generate a noise-reduced and enhanced image.

[0119] In this embodiment, the pixel values ​​of the filtered difference map directly reflect the grayscale difference between the planar bright-field image and the planar dark-field image at that location, and this difference is often closely related to potential defects. Specifically, the pixel values ​​of the filtered difference map are first normalized, mapping them to the range of 0-1 to obtain a normalized difference matrix. Then, based on the difference signal, dynamic weight allocation is performed on the planar bright-field image and the planar dark-field image, allowing each image to play its strengths in different regions, laying the foundation for subsequent fusion.

[0120] In this embodiment, the planar bright-field image and the planar dark-field image are weighted and superimposed according to fusion weights to generate an initial fused image. Specifically, in normal areas, due to the higher weight of the bright-field image, the initial fused image can retain the clear overall outline and uniform background of the bright-field image; in areas suspected of defects, the dark-field image has a dominant weight, and the initial fused image can highlight the detailed features of minor defects in the dark-field image. This embodiment initially integrates the advantages of planar bright-field images and planar dark-field images through the above-mentioned weighted superposition, while avoiding the limitations of a single image in defect detection.

[0121] This embodiment overlays and fuses the filtered difference image with the initial fused image to generate a noise-enhanced image. The denoised defect difference signal in the filtered difference image effectively supplements and strengthens the defect features in the initial fused image. During the overlay and fusion, pixel values ​​are not simply added together; instead, adaptive gain control is employed: for areas in the initial fused image that already exhibit clear defect features, a low gain coefficient is used for overlay to avoid overexposure; for areas in the initial fused image where defect features are relatively blurry, a high gain coefficient is used for overlay to enhance the defect signal. Through this overlay process, the noise-enhanced image retains the clear overall structure and basic defect features of the initial fused image, while the enhancement of the difference signal further improves the contrast between the defect and the background, making previously blurred, minute defects easier for subsequent defect detection models to identify.

[0122] In one embodiment of this application, based on a defect probability distribution map, pixel regions with probability values ​​greater than or equal to a preset threshold are identified as candidate defect regions, including:

[0123] The defect probability distribution map is binarized to generate a binary mask map;

[0124] Pixels in the binary mask image whose probability value is greater than or equal to a preset threshold are marked as foreground regions;

[0125] Morphological closing operations are performed on the binary mask image to obtain the target foreground region;

[0126] Based on the connected component analysis algorithm, connected components are extracted from the target foreground region to obtain a set of connected components;

[0127] Connected regions in the connected region set that are greater than or equal to the defect area threshold are selected as candidate defect regions.

[0128] In this embodiment, the probability value of each pixel in the defect probability distribution map is between 0 and 1, directly reflecting the likelihood of a defect at that location. The core of the binarization process is to convert continuous probability values ​​into discrete 0s and 1s by using a preset threshold, where 1 represents a high-probability defect area and 0 represents a low-probability non-defect area. The preset threshold can be determined based on wafer manufacturing quality standards and historical inspection data. Specifically, binarization is as follows: if the probability value of a pixel in the defect probability distribution map is greater than or equal to the preset threshold, the corresponding pixel in the binary mask is 1; otherwise, it is 0. This embodiment can quickly separate high-probability areas from complex probability distributions, providing a simplified image foundation for subsequent processing.

[0129] In this embodiment, pixels with a probability value greater than or equal to a preset threshold in the binary mask image are marked as foreground regions. Here, the foreground region is the area in the binary mask image where the pixel value is 1; it represents a preliminary set of potential defects. During the marking process, the pixel coordinate information of the foreground region is recorded simultaneously to form a coordinate list for subsequent morphological processing and analysis of the region.

[0130] In this embodiment, the foreground region may include isolated points or fragmented areas caused by noise, requiring further processing to remove invalid information. Therefore, a morphological closing operation is performed on the binary mask image to obtain the target foreground region. This morphological closing operation consists of two basic operations: dilation and erosion. Dilation is performed first, followed by erosion. Its function is to fill in tiny voids within the foreground region, connect adjacent fragmented areas, and maintain the overall shape and size of the region essentially unchanged. For example, when a scratch in the defect probability distribution map is broken due to local probability fluctuations, the closing operation can expand the foreground region boundary through dilation and then shrink it to a reasonable range through erosion, ultimately connecting the broken parts into a complete region. The resulting target foreground region more realistically reflects the actual shape of the defect and reduces the problem of region fragmentation caused by uneven probability distribution.

[0131] In this embodiment, a connected component refers to a set of pixels in the target foreground region that have a pixel value of 1 and are interconnected. Pixels within the same connected component are spatially adjacent. This embodiment uses a connected component analysis algorithm to scan the target foreground region line by line, performing region growing on each unlabeled foreground pixel, grouping all pixels connected to it into the same connected component, and assigning a unique identifier. During the extraction process, basic parameters of each connected component are calculated, such as the number of pixels (area), centroid coordinates, and the size of the bounding rectangle.

[0132] This embodiment uses connected regions whose area is greater than or equal to a defect area threshold as candidate defect regions. The defect area threshold is determined based on the wafer size and the minimum tolerable defect standard. This embodiment reduces the computational load of subsequent feature extraction and classification by eliminating small, isolated regions caused by noise or image interference. Simultaneously, for connected regions whose area exceeds the threshold, their complete parameter information is preserved, providing basic data for subsequent feature extraction and defect type determination. Through the above screening process, the final candidate defect regions obtained in this embodiment include both high-probability potential defects and eliminate most invalid interference.

[0133] In one embodiment of this application, the method for setting a preset threshold includes:

[0134] A global pixel probability histogram of the statistical defect probability distribution map;

[0135] When the global pixel probability histogram has bimodal characteristics, the segmentation threshold calculated by the adaptive threshold algorithm is used as the preset threshold.

[0136] When the global pixel probability histogram has no bimodal features, the segmentation threshold calculated by the maximum inter-class variance method is used as the preset threshold.

[0137] In this embodiment, the global pixel probability histogram uses the pixel probability value as the horizontal axis and the number of pixels corresponding to that probability value as the vertical axis, visually presenting the distribution characteristics of pixel probabilities in the entire defect probability distribution map. During the statistical process, the probability values ​​are divided into 100 intervals at preset intervals, and the total number of pixels in each interval is calculated and a histogram curve is plotted. For example, if the number of pixels in a certain probability interval is significantly higher than in other intervals, it indicates that there are a large number of high-probability defect points in that area, and the histogram will show a clear peak; while if the probability distribution is relatively flat, the histogram curve will fluctuate less. By observing the shape characteristics of the histogram, the distribution pattern of defect probabilities can be determined, providing a basis for the selection of subsequent threshold algorithms.

[0138] In this embodiment, the bimodal feature refers to the presence of two distinct peaks in the histogram, corresponding to the non-defect region (low-probability peak) and the potential defect region (high-probability peak), respectively. A distinct valley exists between the two peaks, and the location of this valley is the ideal segmentation point. The adaptive threshold algorithm locates the threshold position by calculating the intra-class variance under different thresholds within the interval between the two peaks. The threshold corresponding to the minimum intra-class variance is the segmentation threshold. The adaptive threshold algorithm used in this embodiment is suitable for scenarios where the probability distributions of defective and non-defective regions differ significantly, minimizing human intervention and improving the objectivity of threshold setting.

[0139] In this embodiment, the absence of bimodal features typically manifests as a histogram with a unimodal distribution or blurred peaks. In such cases, the probability boundary between defective and non-defective regions is unclear, necessitating the use of statistical methods to find the optimal segmentation point. The core of the Otsu's method is to iterate through all possible thresholds, calculating the inter-class variance between the foreground and background regions corresponding to each threshold. When the inter-class variance reaches its maximum value, the corresponding threshold is the optimal segmentation threshold. This embodiment employs the Otsu's method, suitable for scenarios with complex probability distributions and no clear boundaries. It can find a relatively optimal segmentation point through mathematical optimization, ensuring the integrity and accuracy of candidate defective regions.

[0140] The preset threshold set by the above method in this embodiment can not only adapt to defect probability distribution maps with different probability distribution characteristics, but also ensure the matching degree between the segmentation result and the actual defect distribution.

[0141] In one embodiment of this application, the calculation of the three-dimensional geometric parameters of the target defect region based on three-dimensional point cloud data includes:

[0142] The three-dimensional point cloud data is registered with the theoretical geometric surface position in the wafer physical coordinate system obtained by mapping the position information of the target defect region in the ring image data;

[0143] The registered 3D point cloud data is denoised, and a triangular mesh surface model of the defect area is constructed.

[0144] Based on the triangular mesh surface model, the depth distribution, volume, surface area, and aspect ratio of the defect region projected onto the wafer tangent plane are calculated.

[0145] Extract the three-dimensional coordinates of the edge contour of the defect area, and calculate the average curvature and curvature fluctuation of the contour line.

[0146] In this embodiment, the 3D point cloud data is a set of 3D coordinates of the target defect area obtained by a 3D measurement device. Since its coordinate system deviates from the wafer physical coordinate system, spatial alignment is achieved through registration. Specifically, based on the pixel coordinates of the target defect area in the annular image data and the mapping relationship between the image and the wafer physical dimensions, the theoretical position of the area in the wafer physical coordinate system is calculated. An iterative nearest-point algorithm is used for registration, with the theoretical geometric surface of the wafer as a reference model. By minimizing the distance error between the 3D point cloud data and the reference model, the translation and rotation parameters of the point cloud are adjusted.

[0147] The registered 3D point cloud data is denoised to construct a triangular mesh surface model of the defect region. Outliers caused by measurement noise or environmental interference are removed through denoising. Next, a triangular mesh surface model is constructed based on the denoised 3D point cloud data. Based on this model, the depth distribution, volume, surface area, and aspect ratio of the defect region's projection onto the wafer tangent plane are calculated. The depth distribution is obtained by calculating the perpendicular distance between each vertex on the triangular mesh surface and the theoretical geometric surface of the wafer, generating a depth heatmap to visually display the depth variations of the defect. The maximum and average depths are key indicators. Volume calculation uses the tetrahedral volume integral method, dividing the space between the triangular mesh model and the theoretical wafer surface into several tetrahedra. The total defect volume is obtained by summing the volumes of each tetrahedron. The surface area is the sum of the areas of all triangles in the triangular mesh model, calculated using Heron's formula and then summed. The aspect ratio of the projection on the wafer tangent plane is obtained by first projecting the triangular mesh model onto the wafer tangent plane to obtain the two-dimensional projection profile. Then, the length and width of the profile are calculated by the minimum bounding rectangle algorithm. The ratio of the two is the projection aspect ratio, which can reflect the shape characteristics of the defect.

[0148] The three-dimensional coordinates of the defect region's edge contour are extracted, and the average curvature and curvature fluctuation of the contour line are calculated. The edge contour of the defect region is the boundary line distinguishing the defect from the normal surface. It is extracted from the triangular mesh model using an edge detection algorithm, resulting in a contour line composed of a series of three-dimensional coordinate points. The calculation of the average curvature involves first fitting the contour line piecewise, calculating the curvature of each segment using the radius of the arc, and then averaging the curvature of all segments to reflect the overall curvature of the contour line. The curvature fluctuation is obtained by calculating the standard deviation of the curvature of each segment from the average curvature, and is used to measure the irregularity of the contour line.

[0149] The three-dimensional geometric parameters calculated through the above steps in this embodiment quantify the spatial attributes of defects from multiple dimensions such as depth, volume, shape and contour features, providing accurate numerical support for the determination of wafer inspection results based on defect type and three-dimensional parameters.

[0150] In one embodiment of this application, the three-dimensional coordinates of the edge contour of the defect region are extracted, and the average curvature and its change of the contour line are calculated, including:

[0151] Extract the boundary vertices of the defect region from the triangular mesh surface model and connect them in spatial order to form a closed edge contour line;

[0152] Traverse each vertex on the edge contour line and calculate the discrete curvature of each vertex;

[0153] The average curvature is obtained by taking the arithmetic mean of the discrete curvatures of all vertices on the edge contour line.

[0154] Calculate the standard deviation of all discrete curvatures as the curvature fluctuation.

[0155] In this embodiment, the triangular mesh surface model consists of a large number of triangular faces, and the boundary vertices of the defect region refer to vertices belonging to only one triangular face. During extraction, the number of adjacent faces of all vertices in the triangular mesh is traversed, and vertices with 1 adjacent face are selected as boundary vertices. To ensure the closure of the contour lines, the boundary vertices need to be spatially sorted. This embodiment adopts a polar coordinate-based sorting method, using the centroid of the defect region as the pole, calculating the polar angle of each boundary vertex relative to the pole, and connecting the vertices in ascending order of polar angle to form a continuous closed edge contour line. If multiple independent boundary loops exist, they are extracted and marked separately, and each loop is treated as an independent edge contour line. Each vertex on the edge contour line is traversed, and the discrete curvature of each vertex is calculated. The calculation of discrete curvature is based on the slope change of the line segment formed by the vertex and its two adjacent vertices.

[0156] This embodiment can accurately extract the three-dimensional coordinates of the defect edge contour from the triangular mesh surface model, and obtain the average curvature and curvature fluctuation through statistical analysis of discrete curvature. These two parameters quantify the geometric characteristics of the contour line from two dimensions: overall trend and local change.

[0157] In one embodiment of this application, feature extraction is performed on the candidate defect region to obtain a target feature vector, including:

[0158] Calculate the geometric center coordinates of each candidate defect region and map them to the original spatial location of the annular chamfer region;

[0159] Based on the original spatial location, the candidate defect region is expanded and corrected in a ring shape to obtain the corrected defect region.

[0160] For the corrected defect area, multi-scale texture features and shape features are extracted to form the target feature vector.

[0161] In this embodiment, the geometric center coordinates of each candidate defect region are calculated and mapped to the original spatial position of the annular chamfered region. The candidate defect regions are determined in the planar image after annular unfolding, and their coordinates are two-dimensional coordinates after planar unfolding, which need to be converted to geometric center coordinates; the geometric center coordinates are calculated using the region moment method. In this embodiment, after obtaining the planar geometric center, it is mapped to the original spatial position of the annular chamfered region through the inverse transformation of annular unfolding, i.e., converted to polar coordinates. The mapping process incorporates the parameters from the annular unfolding to ensure the accuracy of the original spatial position, providing an accurate spatial reference for subsequent correction processing.

[0162] In this embodiment, during the ring-shaped unfolding process, due to the arc-shaped structure of the wafer chamfer region, defects at different radial positions in the unfolded planar image will exhibit stretching or compression deformation, especially in areas near the chamfer edge. The purpose of correction is to eliminate this deformation and restore the true shape of the defects. Specifically, based on the original spatial position of the candidate defect region, the radius of curvature of that region in the ring structure is determined. Then, an inverse geometric transformation is used to correct the deformation of the candidate region in the planar image: for the radial direction, the pixel scaling ratio is adjusted according to the radius of curvature to compensate for the stretching of the arc-shaped region; for the circumferential direction, the linear deformation caused by unfolding is corrected through angle mapping. The corrected defect region is geometrically closer to its true state on the wafer. For example, a defect that originally appears as an ellipse in the planar image will be restored to a circle or a regular polygon after correction, providing a more accurate image basis for subsequent feature extraction.

[0163] In this embodiment, the extraction of scale-based texture features employs Gaussian pyramids and local binary models. First, a Gaussian pyramid of the corrected defect region is constructed. LBP features are calculated at each scale, including the original LBP, rotation-invariant LBP, and uniform LBP. These multi-scale features are combined to form a texture feature vector. This texture feature vector effectively describes the surface roughness, texture direction, and other microscopic features of the defect. The extraction of shape features in this embodiment includes geometric features and contour features. Geometric features include the area, perimeter, circularity, and rectangularity of the region. Contour features are extracted using the contour Fourier descriptor to reflect the overall shape and main trends of change of the contour. The multi-scale texture features and shape features are sequentially concatenated to form a target feature vector. This target feature vector includes both microscopic texture and macroscopic shape information of the defect, providing rich input features for the defect classification model and ensuring the accuracy and robustness of the classification.

[0164] In this embodiment, the features of the candidate defect region are quantified into target feature vectors, which not only eliminates the geometric deformation caused by the ring unfolding, but also captures the essential features of the defect from multiple scales and dimensions, laying a high-quality data foundation for subsequent defect type identification.

[0165] In one embodiment of this application, based on defect type and / or classification confidence level, a target defect region requiring three-dimensional measurement is selected from the candidate defect regions, including:

[0166] If the defect type of the candidate defect region belongs to the preset defect type set, the candidate defect region will be directly included in the target defect region.

[0167] If the classification confidence is less than the preset confidence threshold, the candidate defect region will be included in the target defect region.

[0168] For a candidate defect region whose classification confidence reaches a preset confidence threshold, if the following conditions are met: the defect area is greater than a preset area threshold; the defect aspect ratio is greater than a preset range and the classification confidence is less than a second preset confidence threshold; the defect contour complexity is greater than a preset complexity threshold and the defect type belongs to any of the preset types that are prone to producing complex contour defects, then it is included in the target defect region.

[0169] In this embodiment, the preset defect type set includes defects that significantly affect wafer performance and require three-dimensional parameter evaluation, such as edge chipping, cracks, and deep depressions. Furthermore, the setting of the preset defect type set also needs to determine the application scenario of the wafer. For example, for wafers used in high-end chip manufacturing, the preset defect type set includes more minor but fatal defect types, while for wafers used in low-end devices, the set can be appropriately simplified.

[0170] In this embodiment, classification confidence is used to describe the reliability of the defect classification model's judgment on the candidate region type. When the confidence is below this threshold, it indicates that the model's judgment on the defect type has significant uncertainty, and there is a risk of misclassification. At this time, geometric parameters obtained through three-dimensional measurement can help verify the defect type. For example, shallow scratches and stains that are difficult to distinguish in two-dimensional images can be clearly distinguished by three-dimensional depth parameters. The adjustment of the preset confidence threshold needs to be dynamically optimized according to the model's classification accuracy. If the overall accuracy of the model is high, the threshold can be appropriately increased, and vice versa. For candidate defect regions whose classification confidence reaches the preset confidence threshold, it is further determined whether they meet any of the following conditions. If they do, they are included in the target defect region:

[0171] The defect area exceeds the preset area threshold. The preset area threshold is determined based on the wafer size and process requirements. Even if the two-dimensional features of a large defect are clear, its three-dimensional distribution will still affect the evaluation results. The actual volume and depth distribution are confirmed through three-dimensional measurement.

[0172] The defect aspect ratio is greater than a preset range and the classification confidence level is less than a second preset confidence threshold. The preset range is usually set based on the morphological characteristics of common defects. When the classification confidence level is within the range of [preset confidence threshold, second preset confidence threshold), three-dimensional measurement can be triggered based on the abnormal aspect ratio to verify the actual extension of the defect.

[0173] The defect profile complexity exceeds a preset complexity threshold and the defect type belongs to a preset category of defects prone to complex profiles. Defect profile complexity can be quantified using parameters such as profile curvature fluctuation and the number of concave and convex points. The preset complexity threshold is set based on historical data. Types of defects prone to complex profiles include irregular edge chipping and composite defects. The two-dimensional profiles of these defects cannot fully reflect their internal structure; three-dimensional measurement can reveal their three-dimensional morphology.

[0174] This embodiment, through the above-mentioned multi-level screening, ensures that the target defect area includes both high-risk and high-uncertainty defects, while avoiding redundant measurements of low-risk and well-defined defects. While ensuring detection accuracy, it significantly improves the efficiency of three-dimensional measurement, providing a precise measurement object for subsequent determination of the final detection results based on three-dimensional parameters.

[0175] In one embodiment of this application, the method for setting the preset defect type set, preset confidence threshold, preset area threshold, preset range, and preset complexity threshold includes:

[0176] Based on historical defect data, the misjudgment rate of different types of defects and the necessity weight of three-dimensional measurement are statistically analyzed.

[0177] The preset reliability threshold is dynamically adjusted based on the false positive rate: the preset reliability threshold is lowered for defect types with a false positive rate higher than the first threshold.

[0178] The preset area threshold is dynamically set according to the necessity weight of 3D measurement: the preset area threshold is lowered for defect types with a necessity weight higher than the second threshold.

[0179] For aspect ratio-sensitive defect types, the upper limit of the preset range is reduced.

[0180] In this embodiment, historical defect data includes defect type, two-dimensional image features, three-dimensional geometric parameters, final judgment results, and production traceability information. The false positive rate in this embodiment is calculated as the percentage of the number of defects of a certain type whose two-dimensional image detection results are inconsistent with the three-dimensional measurement verification results, relative to the total number of defects of that type. For example, the false positive rate for edge chipping defects is higher because the depth is difficult to determine in two-dimensional images, while the false positive rate for shallow scratches is lower. The necessity weight of three-dimensional measurement is determined based on the production impact through expert scoring. The scoring dimensions include the degree of impact of the defect on the wafer's mechanical strength and electrical performance, as well as the degree of ambiguity of the two-dimensional features. The weight value ranges from 0 to 1; a higher weight indicates that the defect of that type requires more three-dimensional measurement verification.

[0181] In this embodiment, through the above-mentioned dynamic setting method, the preset parameters are closely related to the actual characteristics and historical performance of the defects. This ensures the three-dimensional measurement coverage of high-risk and high-ambiguity defects while avoiding redundant detection of low-risk defects, making the entire screening mechanism more adaptable and accurate.

[0182] In one embodiment of this application, after inputting the target feature vector into the defect classification model to obtain the defect type and classification confidence, the method further includes:

[0183] If the classification confidence of the candidate defect region is lower than the preset review threshold or the critical size of the defect is within the preset critical range, then the feature information of the defect in the candidate defect region is extracted.

[0184] The feature information is matched with baseline features corresponding to multiple defect types or levels in the standard defect feature database to obtain matching results; the standard defect feature database contains baseline features corresponding to each defect type or level obtained through training with historical defect samples.

[0185] Based on the matching results, determine the review type or review level of defects in the candidate defect area;

[0186] Update or confirm the processing level of the target wafer based on the review type or review level.

[0187] In this embodiment, the preset review threshold is set based on the overall accuracy of the defect classification model. The preset critical range is a fuzzy interval defined for the key dimensions of different defect types based on the quality standards of wafer manufacturing. When any of the above conditions are met, more detailed feature information is extracted, including the three-dimensional geometric parameters of the defect, high-resolution texture features, and fine curvature changes of the edge contour. This information is more comprehensive than the target feature vector and can provide richer evidence for the review.

[0188] The standard defect feature database in this embodiment is constructed through training on a large number of historical defect samples. It includes typical features of various defects at different levels. For example, scratches can be divided into three levels: minor, moderate, and severe. Each level corresponds to a set of baseline feature vectors. The matching process uses weighted Euclidean distance or cosine similarity to calculate the similarity between the defect features to be reviewed and the baseline features. The weights are set according to the discriminative power of the features. The matching results include similarity scores with each defect type / level, as well as the top three candidate types / levels.

[0189] Based on the matching results, the verification type or verification level of defects in the candidate defect region is determined. When the similarity score of a certain defect type / level in the matching results is significantly higher than others, and the score exceeds the preset matching threshold, then the type / level is directly determined as the verification result; if the highest score does not exceed the threshold or there are multiple high similarity candidates, then a multi-level verification mechanism is activated, such as based on manual interpretation or by introducing additional auxiliary features, such as material composition analysis of the defect, or by using spectral data to assist in the judgment.

[0190] In this embodiment, wafer processing levels are divided into three categories: release, rework, and scrap. If the review result is consistent with the original classification result, the original processing level is confirmed; if the review result is more serious, the processing level is upgraded; if the review result is less serious, the processing level is downgraded. Through this process, this embodiment can effectively reduce improper processing caused by model misjudgment or critical defects, and improve the quality control accuracy and resource utilization of wafer production.

[0191] In one embodiment of this application, the wafer edge defect detection method based on image recognition further includes:

[0192] Based on a preset statistical sampling period, the detection results of multiple wafers are collected to obtain a defect dataset. The detection results include the defect type, location, classification confidence level, and three-dimensional geometric parameters of the target defect region identified in each wafer.

[0193] Analyze the defect dataset to calculate the trend or distribution of defect characteristics;

[0194] The trend or distribution is compared with a preset process baseline to generate a difference map;

[0195] Based on the difference map, when the trend of defect features in the difference map deviates from the process reference by more than a preset reference threshold, it is determined that the process has deviated; when the defect distribution of wafers produced by a specific device shows a regular abnormality, it is determined that the device needs maintenance.

[0196] In this embodiment, the statistical sampling period is set based on the wafer production cycle time and process stability. The defect dataset includes a unique identifier for each wafer, defect type, specific location of the defect on the wafer, classification confidence level, three-dimensional geometric parameters of the target defect region, and the equipment number that produced the wafer. Data storage uses a structured database and is associated with wafer production parameters to provide complete contextual information for subsequent analysis.

[0197] The defect feature trend analysis in this embodiment mainly focuses on the time dimension, such as statistically analyzing the proportion of various types of defects by sampling period and the average changes of key three-dimensional parameters. The data is smoothed using a moving average method to eliminate random fluctuations and highlight trend changes. The defect feature distribution analysis focuses on the spatial and device dimensions, including wafer spatial distribution and device distribution. For example, the analysis may reveal that the edge chipping defect density in a certain radial ring is three times that of other regions, or that the average depth of depression defects in wafers produced by a certain device is significantly higher than that produced by other devices.

[0198] In this embodiment, the process baseline is a baseline model constructed from defect data from historical stable production stages. This includes the defect characteristic range under normal conditions, such as the percentage range of various defect types, the mean ± 3 standard deviation range of three-dimensional parameters, a thermal baseline map of spatial distribution, and defect level deviation thresholds between equipment. The difference map is generated using a visual comparison method. The trend difference map plots time as the horizontal axis, drawing the actual trend line and the baseline trend line of defect characteristics, and marking the deviation values ​​between them. The distribution difference map is displayed through color-coded overlays. On the wafer spatial distribution map, areas with defect densities exceeding twice the baseline are marked in red, and areas below the baseline are marked in blue. Equipment distribution differences are compared in bar chart form, showing the deviation of each equipment's defect indicators from the baseline mean.

[0199] In this embodiment, the preset benchmark threshold is set based on the degree of impact of defects on the manufacturing process. The process deviation determination logic in this embodiment is as follows: if the quantity trend of a certain type of defect deviates from the benchmark and the deviation exceeds the threshold within three consecutive sampling periods, or the mean value of the key three-dimensional parameters exceeds the benchmark range for two consecutive periods, then a process deviation is determined to have occurred. The equipment maintenance determination in this embodiment requires that the defect distribution exhibits a regular anomaly. For example, if all wafers produced by a certain equipment show highly dense scratches in the circumferential angle range, and this regularity persists for five sampling periods, excluding similar phenomena in other equipment, it can be determined that the edge processing components of this equipment are worn and require maintenance.

[0200] In this embodiment, the defect detection method extends from quality control of a single wafer to monitoring the entire production process and equipment status, achieving an upgrade from passive detection to proactive prevention. This helps to promptly identify potential process defects, reduce the generation of batch defects, and improve equipment uptime and production stability.

[0201] In one embodiment of this application, determining the detection result of the target wafer based on defect type and three-dimensional geometric parameters includes:

[0202] The defect type of the target defect area is matched with the association rule base to extract the corresponding qualified threshold range. The association rule base is established based on the defect type and three-dimensional geometric parameters. For each defect type, the qualified threshold range of its corresponding three-dimensional geometric parameters is preset.

[0203] If the three-dimensional geometric parameters of the target defect area are greater than the acceptable threshold range corresponding to its defect type, it is judged as an unacceptable defect.

[0204] If the three-dimensional geometric parameters are within the threshold range, they are judged as acceptable defects;

[0205] Statistically determine the number and distribution density of non-conforming defects in all target defect areas;

[0206] If the number of non-conforming defects exceeds the wafer-level tolerance threshold, or if the geometric parameters of a single defect reach a critical value, then a wafer rejection result will be output.

[0207] In this embodiment, the association rule base is a structured database built from a large amount of historical inspection data and wafer production quality standards. Its core is to establish a one-to-one correspondence between defect types and acceptable threshold values ​​for three-dimensional geometric parameters. During the matching process, this embodiment automatically retrieves the corresponding threshold range from the rule base based on the defect type of the target defect area, providing a clear standard for subsequent defect nature determination.

[0208] In this embodiment, if the three-dimensional geometric parameters of the target defect region exceed the acceptable threshold range corresponding to its defect type, it is determined to be an unacceptable defect. This embodiment compares each parameter in the three-dimensional geometric parameters individually; if any parameter exceeds the threshold range, the defect is determined to be unacceptable. This veto-based determination method ensures that all defects exceeding the standard that affect wafer performance are accurately identified, avoiding quality risks caused by a single parameter exceeding the standard. If the three-dimensional geometric parameters are within the threshold range, it is determined to be an acceptable defect. An acceptable defect refers to a defect that, although it has defect characteristics, has three-dimensional geometric parameters that do not exceed the acceptable threshold, and has no significant impact on subsequent wafer processing and final performance. These types of defects do not require special treatment, and the wafer can proceed normally to the next process step. However, their location, type, and parameters are still recorded for subsequent process analysis and quality traceability.

[0209] In this embodiment, the wafer-level tolerance threshold refers to the maximum number of non-conforming defects allowed to exist on the entire wafer, and its value is set according to the wafer's quality grade. The critical threshold refers to an extreme parameter value that would cause the wafer to fail directly.

[0210] This embodiment makes a precise and quantitative judgment on the quality of the target wafer based on defect type and three-dimensional geometric parameters. This ensures that unqualified wafers are effectively screened and avoids misjudging qualified wafers, providing reliable test results for quality control in wafer production.

[0211] Corresponding to the image recognition-based wafer edge defect detection method in the above embodiments, Figure 2 This is a structural block diagram of a wafer edge defect detection system based on image recognition, provided as an embodiment of this application. For ease of explanation, only the parts relevant to the embodiment of this application are shown. References Figure 2 The image recognition-based wafer edge defect detection system 20 includes: a data processing module 21, a data prediction module 22, a defect identification module 23, a data extraction module 24, a defect type module 25, a defect determination module 26, a geometric parameter module 27, and a detection result module 28.

[0212] The data processing module 21 is used to acquire the annular image data of the chamfered region of the target wafer, process the annular image data, and obtain the target image data.

[0213] Data prediction module 22 is used to input target image data into edge defect prediction model to obtain a defect probability distribution map that characterizes the probability of defects at each pixel position in the chamfer region of the target wafer;

[0214] The defect identification module 23 is used to identify pixel regions with probability values ​​greater than or equal to a preset threshold as candidate defect regions based on the defect probability distribution map.

[0215] Data extraction module 24 is used to extract features from candidate defect regions to obtain target feature vectors;

[0216] The defect type module 25 is used to input the target feature vector into the defect classification model to determine the defect type of the candidate defect region and its corresponding classification confidence.

[0217] Defect determination module 26 is used to filter out target defect regions that require three-dimensional measurement from candidate defect regions based on defect type and / or classification confidence.

[0218] The geometric parameter module 27 is used to acquire the three-dimensional point cloud data of the target defect area and calculate the three-dimensional geometric parameters of the target defect area based on the three-dimensional point cloud data.

[0219] The detection result module 28 is used to determine the detection result of the target wafer based on the defect type and three-dimensional geometric parameters.

[0220] See Figure 3 , Figure 3 This is a schematic block diagram of an electronic device provided according to an embodiment of this application. Figure 3 The electronic device 300 in this embodiment may include one or more processors 301, one or more input devices 302, one or more output devices 303, and one or more memories 304. The processors 301, input devices 302, output devices 303, and memories 304 communicate with each other via a communication bus 305. The memories 304 store computer programs, including program instructions. The processors 301 execute the program instructions stored in the memories 304. Specifically, the processors 301 are configured to invoke the program instructions to perform the functions of the modules in the aforementioned device embodiments, for example... Figure 2 The functions of the data processing module 21, data prediction module 22, defect identification module 23, data extraction module 24, defect type module 25, defect determination module 26, geometric parameter module 27, and detection result module 28 are shown.

[0221] It should be understood that, in the embodiments of this application, the processor 301 may be a central processing unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor.

[0222] Input device 302 may include a touchpad, a fingerprint sensor (for collecting the user's fingerprint information and fingerprint orientation information), a microphone, etc., and output device 303 may include a display (LCD, etc.), a speaker, etc.

[0223] The memory 304 may include read-only memory and random access memory, and provides instructions and data to the processor 301. A portion of the memory 304 may also include non-volatile random access memory. For example, the memory 304 may also store device type information.

[0224] In specific implementations, the processor 301, input device 302, and output device 303 described in the embodiments of this application can execute the implementation methods described in any embodiment of the wafer edge defect detection method based on image recognition provided in the embodiments of this application, or they can execute the implementation methods of the electronic devices described in the embodiments of this application, which will not be repeated here.

[0225] In another embodiment of this application, a computer-readable storage medium is provided. This computer-readable storage medium stores a computer program, which includes program instructions. When executed by a processor, the program instructions implement all or part of the processes in the methods described above. Alternatively, the computer program can instruct related hardware to complete the process. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include any entity or device capable of carrying computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.

[0226] The computer-readable storage medium can be an internal storage unit of the electronic device in any of the foregoing embodiments, such as a hard disk or memory of the electronic device. The computer-readable storage medium can also be an external storage device of the electronic device, such as a plug-in hard disk, smart media card (SMC), secure digital card (SD), flash card, etc., equipped on the electronic device. Furthermore, the computer-readable storage medium can include both internal and external storage units of the electronic device. The computer-readable storage medium is used to store computer programs and other programs and data required by the electronic device. The computer-readable storage medium can also be used to temporarily store data that has been output or will be output.

[0227] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this application.

[0228] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the electronic devices and units described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0229] In the several embodiments provided in this application, it should be understood that the disclosed electronic devices and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. In addition, the mutual coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces or units, or it may be an electrical, mechanical, or other form of connection.

[0230] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of the embodiments of this application, depending on actual needs.

[0231] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0232] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and these modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A wafer edge defect detection method based on image recognition, characterized by, The method comprises: obtaining annular image data of a target wafer chamfer region, processing the annular image data to obtain target image data; inputting the target image data into an edge defect prediction model to obtain a defect probability distribution map representing the probability of defects existing at each pixel position of the target wafer chamfer region; based on the defect probability distribution map, identifying a pixel region with a probability value greater than or equal to a preset threshold as a candidate defect region; performing feature extraction on the candidate defect region to obtain a target feature vector; inputting the target feature vector into a defect classification model to determine the defect type of the candidate defect region and the corresponding classification confidence; based on the defect type and / or classification confidence, screening a target defect region requiring three-dimensional measurement from the candidate defect region; obtaining three-dimensional point cloud data of the target defect region, and calculating three-dimensional geometric parameters of the target defect region based on the three-dimensional point cloud data; based on the defect type and the three-dimensional geometric parameters, determining the detection result of the target wafer. 2.The wafer edge defect detection method based on image recognition of claim 1, wherein, The annular image data comprises bright-field image data and dark-field image data; The processing of the annular image data to obtain target image data comprises: based on wafer geometric features or image feature points, performing spatial registration on the bright-field image data and the dark-field image data to obtain spatially aligned registered bright-field images and registered dark-field images; segmenting the registered bright-field images and the registered dark-field images to obtain image parts corresponding to the wafer chamfer region, to obtain registered bright-field chamfer images and registered dark-field chamfer images; performing annular unwrapping processing on the registered bright-field chamfer images and the registered dark-field chamfer images to obtain planar bright-field images and planar dark-field images; wherein the annular images are unwrapped and mapped to rectangular planar images along the circumferential direction, so that the angular coordinates on the annular images correspond to the width direction coordinates of the rectangular planar images, and the radial coordinates on the annular images correspond to the height direction coordinates of the rectangular planar images; performing image fusion processing on the planar bright-field images and the planar dark-field images to generate a noise-resistant enhanced image as the target image data. 3.The wafer edge defect detection method based on image recognition of claim 2, wherein, The image fusion processing of the planar bright-field images and the planar dark-field images to generate a noise-resistant enhanced image comprises: calculating the pixel difference matrix of the planar bright-field images and the planar dark-field images to obtain a difference image; performing denoising filtering processing on the difference image to obtain a filtered difference image; wherein the denoising filtering is used to suppress noise and retain potential defect difference signals; performing fusion processing on the filtered difference image, the planar bright-field images and the planar dark-field images to generate a noise-resistant enhanced image.

4. The wafer edge defect detection method based on image recognition according to claim 3, wherein, The fusion processing of the filtered difference image, the planar bright-field images and the planar dark-field images to generate a noise-resistant enhanced image comprises: based on the pixel values of the filtered difference image, calculating the fusion weights of the planar bright-field images and the planar dark-field images at the corresponding pixel positions; performing weighted superposition on the planar bright-field images and the planar dark-field images according to the fusion weights to generate an initial fusion image; Superimpose and fuse the filtered difference image with the initial fusion image to generate the anti-noise enhanced image.

5. The wafer edge defect detection method based on image recognition according to claim 1, wherein, The method further includes: identifying, based on the defect probability distribution map, a pixel region with a probability value greater than or equal to a preset threshold as a candidate defect region. The method further includes: performing binarization processing on the defect probability distribution map to generate a binary mask image. The method further includes: marking, in the binary mask image, a pixel with a probability value greater than or equal to the preset threshold as a foreground region. The method further includes: performing a morphological closing operation on the binary mask image to obtain a target foreground region. The method further includes: extracting, based on a connected domain analysis algorithm, a connected domain in the target foreground region to obtain a connected domain set. The method further includes: regarding a connected domain in the connected domain set that is greater than or equal to a defect area threshold as a candidate defect region.

6. The wafer edge defect detection method based on image recognition according to claim 5, wherein, The method further includes: setting the preset threshold by using the following method. The method further includes: counting a global pixel probability histogram of the defect probability distribution map. When the global pixel probability histogram has a bimodal feature, a segmentation threshold calculated by using an adaptive threshold algorithm is used as the preset threshold. When the global pixel probability histogram has no bimodal feature, a segmentation threshold calculated by using an Otsu method is used as the preset threshold.

7. The wafer edge defect detection method based on image recognition according to claim 1, wherein, The method further includes: calculating, based on the three-dimensional point cloud data, a three-dimensional geometric parameter of the target defect region by using the following method. The method further includes: registering the three-dimensional point cloud data with a theoretical geometric surface position in a wafer physical coordinate system that is mapped based on position information of the target defect region in the annular image data. The method further includes: performing denoising processing on the registered three-dimensional point cloud data to construct a triangular mesh surface model of the defect region. The method further includes: calculating, based on the triangular mesh surface model, a depth distribution, a volume, a surface area, and a projection aspect ratio of the defect region on a wafer tangent plane. The method further includes: extracting a three-dimensional coordinate of an edge contour of the defect region and calculating an average curvature and a curvature fluctuation of the contour line. 8.The wafer edge defect detection method based on image recognition of claim 7, wherein, The method further includes: extracting a boundary vertex of the defect region from the triangular mesh surface model and connecting the boundary vertex in a spatial order to form a closed edge contour line. The method further includes: calculating a discrete curvature of each vertex on the edge contour line. The method further includes: calculating an average curvature by calculating an arithmetic mean of the discrete curvatures of all the vertices on the edge contour line. The method further includes: calculating a standard deviation of all the discrete curvatures as a curvature fluctuation. The method further includes:

9. An image recognition-based wafer edge defect detection system, comprising: The data processing module is configured to acquire annular image data of a target wafer chamfer region, and process the annular image data to obtain target image data. The data prediction module is configured to input the target image data into an edge defect prediction model to obtain a defect probability distribution map representing a probability of a defect existing at each pixel position of the target wafer chamfer region. The defect identification module is configured to identify, based on the defect probability distribution map, a pixel region with a probability value greater than or equal to a preset threshold as a candidate defect region. The data extraction module is configured to perform feature extraction on the candidate defect region to obtain a target feature vector. The defect type module is configured to input the target feature vector into a defect classification model to determine a defect type of the candidate defect region and a corresponding classification confidence. ​ a defect determination module, configured to filter out a target defect region requiring three-dimensional measurement from the candidate defect region based on the defect type and / or the classification confidence; a geometry parameter module, configured to acquire three-dimensional point cloud data of the target defect region, and calculate three-dimensional geometry parameters of the target defect region based on the three-dimensional point cloud data; a detection result module, configured to determine a detection result of the target wafer based on the defect type and the three-dimensional geometry parameters.

10. A computer-readable storage medium storing a computer program, the computer-readable storage medium comprising: The computer program is executed by a processor to implement the steps of the method according to any one of claims 1 to 8. The computer program is executed by a processor to implement the steps of the method according to any one of claims 1 to 8.

Images