Wafer defect detection method and system based on edge features and gradient boosting tree
By combining edge features with gradient boosting trees, multi-scale features are extracted and morphological verification is performed, solving the problem of low-contrast slip line recognition in wafer defect detection and achieving efficient and accurate defect detection.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- 盖泽精密科技(苏州)有限公司
- Filing Date
- 2026-04-07
- Publication Date
- 2026-07-21
Smart Images
Figure CN122434835A_ABST
Abstract
Description
Technical Field
[0001] This application relates to semiconductor manufacturing, and in particular to a wafer defect detection method and system based on edge features and gradient boosting trees. Background Technology
[0002] In the wafer manufacturing process, surface defects are a key factor affecting chip yield. Typical defects include low-contrast, elongated slip line defects, as well as high-contrast, dot-shaped particle defects.
[0003] Among existing detection methods, traditional threshold segmentation is sensitive to illumination and has difficulty identifying low-contrast defects; deep learning models rely on massive labeled samples, are prone to overfitting in small sample scenarios, and have slow inference speeds that are difficult to adapt to online detection; single feature expression capabilities are insufficient, resulting in high false detection and false negative rates.
[0004] Therefore, there is an urgent need for a defect detection solution that is adaptable to small samples, has high inference efficiency, and can accurately distinguish between slip lines and particles. Summary of the Invention
[0005] To address the technical problems of existing wafer defect detection methods, such as difficulty in reliably identifying low-contrast slip lines, weak robustness of single features, high sample dependence of deep learning models, and high deployment difficulty, this invention provides a wafer defect detection method based on edge features and gradient boosting trees. The technical solution is as follows: On the one hand, a wafer defect detection method based on edge features and gradient boosting trees is provided, including: The acquired wafer surface images are preprocessed to suppress noise and normalize the grayscale distribution; Based on the multi-scale edge detection method, gradient response analysis and local texture analysis are performed on the neighborhood of each pixel in the wafer surface image to extract gradient features, orientation consistency features, texture features and edge density features, and construct pixel-level feature vectors. The feature vectors are used to characterize the difference between low-contrast linear defects and point defects. The feature vector is input into a pre-trained gradient boosting tree classification model to obtain the defect category prediction result for each pixel; Morphological processing and connected component geometric verification are performed on the defect category prediction results to output the final defect detection results.
[0006] Optionally, the gradient features include statistics on the gradient magnitude within the neighborhood; The directional consistency feature is obtained based on the gradient direction distribution within the neighborhood and is used to characterize the difference between linear structures with directional continuity and structures with anisotropic distributions. The texture features are obtained based on local texture coding; The edge density feature is obtained based on the proportion of pixels in the neighborhood that meet the edge response conditions.
[0007] Optionally, the gradient boosting tree classification model adopts a multi-classification structure and is configured to distinguish between a background category and at least two defect categories, wherein the at least two defect categories include slip line defects and particulate defects.
[0008] Optionally, class weights are introduced into the loss function of the gradient boosting tree classification model to balance the difference in the number of background samples and defective samples.
[0009] Optionally, the morphological processing includes performing at least one of opening, closing, erosion, and dilation operations on regions predicted as defects to filter out isolated noise or correct region boundaries.
[0010] Optionally, the connected component geometry verification includes a secondary confirmation of the defect category based on the aspect ratio and area threshold of the connected component.
[0011] Optionally, the multi-scale edge detection method includes extracting edge response information at at least two different scales.
[0012] Optionally, the gradient features, orientation consistency features, texture features, and edge density features are used together to describe the morphological features and spatial distribution characteristics of defects, forming a joint feature representation for judging defect types.
[0013] On the other hand, a wafer defect detection system based on edge features and gradient boosting trees is provided, including: The image preprocessing unit is configured to preprocess the acquired wafer surface image to suppress noise and normalize the grayscale distribution. The multi-scale feature extraction unit is configured to perform gradient response analysis and local texture analysis on the neighborhood of each pixel in the wafer surface image based on the multi-scale edge detection method, extract gradient features, orientation consistency features, texture features and edge density features, and construct pixel-level feature vectors. The feature vectors are used to characterize the difference between low-contrast linear defects and point defects. The classification reasoning unit is configured to input the feature vector into a pre-trained gradient boosting tree classification model to obtain the defect category prediction result corresponding to each pixel; The post-processing unit is configured to perform morphological processing and connected component geometric verification on the defect category prediction results and output the final defect detection results.
[0014] On the other hand, a computer-readable storage medium is provided on which a computer program is stored, which, when executed by a processor, implements the wafer defect detection method based on edge features and gradient boosting trees as described above.
[0015] This invention discloses a wafer defect detection method and system based on edge features and gradient boosting trees. First, the wafer image undergoes denoising and grayscale normalization preprocessing. Then, multi-scale edge detection extracts gradient, orientation consistency, texture, and edge density features to construct pixel-level vectors. A gradient boosting tree model is used to achieve pixel-level defect classification. Finally, morphological processing and connected component geometric verification are applied to output the results. This improves the representation ability of low-contrast slip lines and point-like particles through multi-scale edge and joint feature representation; enhances classification adaptability in small-sample scenarios and reduces deployment complexity through the gradient boosting tree model; and improves the stability of defect detection results through morphological processing and connected component geometric verification. Attached Figure Description
[0016] Figure 1 The diagram illustrates a process flow of a wafer defect detection method based on edge features and gradient boosting trees according to the present invention. Figure 2 The diagram illustrates the modular structure of a wafer defect detection system based on edge features and gradient boosting trees according to the present invention. Detailed Implementation
[0017] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings.
[0018] In this article, "multiple" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. The character " / " generally indicates that the preceding and following related objects have an "or" relationship. Example 1
[0019] like Figure 1 The diagram illustrates the overall flowchart of a wafer defect detection method based on edge features and gradient boosting trees. It provides a wafer defect detection method based on edge features and gradient boosting trees, including: Step 101: Preprocess the acquired wafer surface image to suppress noise and normalize the grayscale distribution.
[0020] The preprocessing includes filtering and denoising the grayscale image of the wafer surface and normalizing the grayscale range. Filtering is used to eliminate high-frequency noise introduced by image acquisition and transmission, and grayscale normalization is used to eliminate grayscale shift caused by uneven illumination, so as to provide a stable and consistent image basis for subsequent feature extraction.
[0021] The formula for grayscale normalization is: I norm =(II min ) / (Imax -I min ).
[0022] Where I is the original grayscale value, I min and I max These are the minimum and maximum gray values in the current image or the current processing area, respectively. norm The normalized gray value is represented by this formula, which maps the image gray level uniformly to the [0,1] interval, eliminating gray level differences caused by different acquisition environments.
[0023] Step 102: Based on the multi-scale edge detection method, gradient response analysis and local texture analysis are performed on the neighborhood of each pixel in the wafer surface image to extract gradient features, orientation consistency features, texture features and edge density features, and construct pixel-level feature vectors. The feature vectors are used to characterize the difference between low-contrast linear defects and point defects.
[0024] Among them, multi-scale edge detection calculates the edge response at different scales, taking into account the expression of both small and large defects; gradient response analysis is used to extract the intensity information of defect edges; local texture analysis is used to characterize the differences in microscopic surface structure; the four types of features describe defects from the dimensions of edge intensity, direction regularity, texture structure, and edge density, respectively, to achieve differentiated characterization of low-contrast slip lines and dot-like particles.
[0025] In this embodiment, the Sobel operator is used to calculate the gradient response of the image, and edge response information is extracted under multiple scale windows of 3×3, 5×5, and 7×7. The 3×3 window is used to characterize small-sized edge changes, the 5×5 window is used to balance local noise suppression and edge preservation, and the 7×7 window is used to enhance the overall response of continuous linear structures. Different scales are adapted to small-sized defects and large-scale continuous defects, respectively. The edge responses at each scale are fused by weighted averaging to form a stable multi-scale gradient response feature.
[0026] The multi-scale gradient fusion formula is: .
[0027] Where G3, G5, and G7 represent the gradient responses obtained at window scales of 3×3, 5×5, and 7×7, respectively, and ω1, ω2, and ω3 are the corresponding scale weights. These weights can be obtained through cross-validation, empirical setting, or training set performance optimization. In one implementation, the weights at each scale can be equal and satisfy the following conditions: .
[0028] Local texture analysis is achieved through the Local Binary Pattern (LBP) operator, which is used to extract the differences in microscopic texture structure in the region surrounding the pixel. Slip lines exhibit continuous directional texture, while particles exhibit irregular isotropic texture.
[0029] The directional consistency feature is calculated from the gradient direction distribution entropy: .
[0030] in, Binning is performed for the i-th gradient direction. The gradient direction represents the percentage of pixels within the window that fall into the bin of that direction. The gradient direction can be divided into several preset angle intervals. Summation is performed only for bins with a probability greater than 0. When =0, the corresponding item is treated as 0. The lower the entropy value, the more concentrated the direction, corresponding to slip line defects; the higher the entropy value, the more scattered the direction, corresponding to particles or background.
[0031] The formula for calculating edge density features is: D=N edge / N win .
[0032] Where, N edge The number of pixels N required to satisfy the edge response threshold condition. win This represents the total number of pixels within the current neighborhood window. The edge response threshold can be preset based on the statistical results of the training set.
[0033] Step 103: Input the feature vector into the pre-trained gradient boosting tree classification model to obtain the defect category prediction result for each pixel.
[0034] The gradient boosting tree model takes pixel-level feature vectors as input and outputs predictions of whether a pixel belongs to the background, slip line defect, or particle defect. It relies on nonlinear fitting capabilities to reliably distinguish defects from the background and between different types of defects. The gradient boosting tree model outputs the probability value of each pixel belonging to each candidate category, and uses the category with the highest probability as the initial classification result.
[0035] This embodiment uses the Gradient Boosting Tree (GBDT) model with a CART regression tree as the base learner, employs a multi-class log loss function, and balances the imbalance between background and defect samples through class weights. In one implementation, the defect class weight can be determined based on the total number of samples, the sample size of that class, and the number of defect classes, for example, it can be expressed as: w defect =N total / (N defect ×K).
[0036] Among them, w defect For the defect category weights, N total N represents the total number of samples. defect Let K be the number of samples in the current defect category, and K be the number of defect categories. The weight calculation method is not limited to the above expression. The model optimizes hyperparameters such as tree depth and learning rate through cross-validation.
[0037] In one implementation, the tree depth can be set to 3 to 7, and the learning rate can be set to 0.01 to 0.1. The model exhibits strong generalization ability under small sample conditions, is less prone to overfitting, and has a faster inference speed than deep learning models, making it more suitable for industrial online deployment.
[0038] Step 104: Perform morphological processing and connected component geometric verification on the defect category prediction results, and output the final defect detection results.
[0039] Among them, morphological processing is used to remove isolated noise points, and connected domain geometric verification further corrects the classification results based on the geometric shape of the defects, thereby improving the accuracy and stability of the detection results.
[0040] Morphological processing includes performing at least one of the following operations on the defect region: opening, closing, erosion, and dilation. Opening removes isolated noise points, closing fills in defect discontinuities, and erosion and dilation are used to smooth boundaries.
[0041] The fortune-enhancing algorithm is defined as follows: .
[0042] Where A represents the binary defect region to be processed, and B represents the structuring element. Indicates corrosion. It indicates expansion.
[0043] Connected component geometric verification achieves secondary confirmation by calculating aspect ratio and area. The aspect ratio is calculated as R=W / H, where W and H are the long and short sides of the smallest bounding rectangle of the connected component, respectively. The aspect ratio of the slip line is large, and the aspect ratio of the particles is close to 1. Combined with the area threshold to filter interference, the area threshold can be preset based on the statistical results of the training samples or the device resolution to further reduce misclassification.
[0044] Therefore, by stabilizing image quality through preprocessing, relying on multi-scale edge features to achieve reliable characterization of low-contrast defects, employing gradient boosting trees for classification and recognition under small sample sizes, and combining morphological and geometric verification to reduce false detections, the overall scheme exhibits fast inference speed, is adaptable to online detection scenarios, and helps distinguish between slip lines and particulate defects. The four types of features work together: gradient represents intensity, directional consistency represents linearity, texture represents microstructure, and edge density represents aggregation degree, achieving a complete expression of defect features and improving detection accuracy. Example 2
[0045] Based on Example 1, the gradient features include statistics on the gradient magnitude within the neighborhood; the orientation consistency feature is obtained based on the gradient orientation distribution within the neighborhood, used to characterize the difference between linear structures with directional continuity and anisotropic distribution structures; the texture feature is obtained based on local texture encoding; and the edge density feature is obtained based on the proportion of pixels in the neighborhood that satisfy the edge response condition.
[0046] Among them, gradient magnitude statistics reflect the intensity of defect edges; directional consistency features characterize the directionality of defects through gradient direction distribution, slip lines show a continuous and consistent directional distribution, and particles show an irregular anisotropic distribution; local texture encoding characterizes the differences in surface microstructure; and edge density ratio reflects the degree of defect edge aggregation.
[0047] The four types of features complement each other. Using any single feature alone is insufficient to fully distinguish defect types. Using them in combination allows for a comprehensive characterization of defect morphology and spatial distribution from multiple perspectives, including edge intensity, directional patterns, micro-texture, and density, thus improving defect separability. Single features are susceptible to noise and texture interference, while combined features possess stronger anti-interference and discrimination capabilities.
[0048] Thus, the four types of features correspond to different physical properties of defects, and the feature expression dimensions are complete, which helps to distinguish between directional linear defects and irregular point defects, and improves the robustness of defect characterization. Example 3
[0049] Based on Example 1, the gradient boosting tree classification model adopts a multi-classification structure and is configured to distinguish between background categories and at least two defect categories, including slip line defects and particle defects.
[0050] The model supports expanding the number of defect categories, primarily covering the two most typical defects on wafer surfaces: slip lines and microparticles, balancing detection versatility and specificity. Therefore, the classification architecture is flexible and scalable, focusing on core defect types to meet the actual needs of online wafer inspection. Example 4
[0051] Based on Example 1, class weights are introduced into the loss function of the gradient boosting tree classification model to balance the difference in the number of background samples and defective samples.
[0052] In wafer defect detection, the number of background samples is much larger than that of defect samples. By increasing the loss contribution of defect samples through class weighting, the problem of missed detection caused by sample imbalance can be alleviated, the classification bias caused by sample imbalance can be effectively improved, and the defect missed detection rate can be reduced. Example 5
[0053] Based on Example 1, the morphological processing includes performing at least one of opening, closing, erosion, and dilation operations on regions predicted as defects to filter out isolated noise or correct region boundaries.
[0054] Among them, opening operation is used to remove small isolated noise points, closing operation is used to fill the discontinuity in the defect area, and erosion and expansion are used to smooth the boundary and regularize the defect outline. Thus, noise is filtered out without destroying the real defect morphology, improving the cleanliness and accuracy of the detection results. Example 6
[0055] Based on Example 1, the connected domain geometry verification includes a secondary confirmation of the defect category based on the aspect ratio and area threshold of the connected domain.
[0056] Among them, the sliding line connected domains are slender with an aspect ratio greater than the set threshold; the particles are compact and blocky with an aspect ratio close to 1. Combined with the area threshold, invalid micro-interferences can be eliminated. Thus, the classification results are verified a second time through geometric rules, which helps to reduce the defect misclassification rate and improve the reliability of detection. Example 7
[0057] Based on Example 1, the multi-scale edge detection method includes extracting edge response information at at least two different scales.
[0058] Among them, multiple scales are adapted to both small-sized defects and large-scale continuous defects, avoiding the loss of defect information caused by a single scale. This takes into account the feature extraction of defects of different sizes and improves the adaptability to various types of defects. Example 8
[0059] Based on Example 1, gradient features, orientation consistency features, texture features, and edge density features are used together to describe the morphological features and spatial distribution characteristics of defects, so as to form a joint feature representation for judging defect types.
[0060] The four types of features complement each other. A single feature cannot fully distinguish defects, while the joint features can comprehensively characterize the morphology, orientation, texture, and density of defects. Through the complementary use of multi-dimensional information, the joint features eliminate the limitations of single features, making defect type identification more stable and reliable. Therefore, the feature combination forms a joint representation, which is more discriminative and robust than a single feature, thus improving classification accuracy.
[0061] Furthermore, this application also provides a wafer defect detection system based on edge feature enhancement and gradient boosting tree, including: The image preprocessing unit is configured to preprocess the acquired wafer surface image to suppress noise and normalize the grayscale distribution. The multi-scale feature extraction unit is configured to perform gradient response analysis and local texture analysis on the neighborhood of each pixel in the wafer surface image based on the multi-scale edge detection method, extract gradient features, orientation consistency features, texture features and edge density features, and construct pixel-level feature vectors. The feature vectors are used to characterize the difference between low-contrast linear defects and point defects. The classification reasoning unit is configured to input the feature vector into a pre-trained gradient boosting tree classification model to obtain the defect category prediction result corresponding to each pixel; The post-processing unit is configured to perform morphological processing and connected component geometric verification on the defect category prediction results and output the final defect detection results.
[0062] In addition, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described in any one of embodiments 1 to 8 above.
[0063] Among them, the storage medium can deploy the detection solution in embedded devices, industrial control computers and edge devices, expanding the applicable scenarios of the solution.
[0064] Those skilled in the art will understand that all or part of the steps of the above embodiments can be implemented by hardware or by a program instructing related hardware. The program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk.
[0065] The above description is merely an optional embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.
Claims
1. A wafer defect detection method based on edge features and gradient boosting trees, characterized in that, include: The acquired wafer surface images are preprocessed to suppress noise and normalize the grayscale distribution; Based on the multi-scale edge detection method, gradient response analysis and local texture analysis are performed on the neighborhood of each pixel in the wafer surface image to extract gradient features, orientation consistency features, texture features and edge density features, and construct pixel-level feature vectors. The feature vectors are used to characterize the difference between low-contrast linear defects and point defects. The feature vector is input into a pre-trained gradient boosting tree classification model to obtain the defect category prediction result for each pixel; Morphological processing and connected component geometric verification are performed on the defect category prediction results to output the final defect detection results.
2. The method according to claim 1, characterized in that: The gradient features include statistics on the gradient magnitude within the neighborhood; The directional consistency feature is obtained based on the gradient direction distribution within the neighborhood and is used to characterize the difference between linear structures with directional continuity and structures with anisotropic distributions. The texture features are obtained based on local texture coding; The edge density feature is obtained based on the proportion of pixels in the neighborhood that meet the edge response conditions.
3. The method according to claim 1, characterized in that, The gradient boosting tree classification model adopts a multi-class structure and is configured to distinguish between a background category and at least two defect categories, including slip line defects and particulate defects.
4. The method according to claim 1, characterized in that, The gradient boosting tree classification model incorporates class weights in its loss function to balance the difference in the number of background samples and defective samples.
5. The method according to claim 1, characterized in that, The morphological processing includes performing at least one of opening, closing, erosion, and dilation operations on regions predicted as defects to filter out isolated noise or correct region boundaries.
6. The method according to claim 1, characterized in that, The connected component geometry verification includes a secondary confirmation of the defect category based on the aspect ratio and area threshold of the connected component.
7. The method according to claim 1, characterized in that, The multi-scale edge detection method includes extracting edge response information at at least two different scales.
8. The method according to claim 1, characterized in that, The gradient features, directional consistency features, texture features, and edge density features are used together to describe the morphological features and spatial distribution characteristics of defects, forming a joint feature representation for distinguishing defect types.
9. A wafer defect detection system based on edge features and gradient boosting trees, characterized in that, include: The image preprocessing unit is configured to preprocess the acquired wafer surface image to suppress noise and normalize the grayscale distribution. The multi-scale feature extraction unit is configured to perform gradient response analysis and local texture analysis on the neighborhood of each pixel in the wafer surface image based on the multi-scale edge detection method, extract gradient features, orientation consistency features, texture features and edge density features, and construct pixel-level feature vectors. The feature vectors are used to characterize the difference between low-contrast linear defects and point defects. The classification reasoning unit is configured to input the feature vector into a pre-trained gradient boosting tree classification model to obtain the defect category prediction result corresponding to each pixel; The post-processing unit is configured to perform morphological processing and connected component geometric verification on the defect category prediction results and output the final defect detection results.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the wafer defect detection method based on edge features and gradient boosting trees as described in any one of claims 1 to 8.