Construction method of wafer defect data set, training method of recognition system and application

Abstract

The invention belongs to the technical field of wafer defect detection, and particularly relates to a wafer defect data set construction method, an identification system training method and application. The construction method of the wafer defect data set comprises the following steps that n ideal scratch images are generated through a Bezier curve, and one ideal scratch image comprises a plurality of ideal scratches; performing break mark processing and background noise adding processing on each ideal scratch image to form a corresponding simulation scratch image; one simulated scratch image is bound with the corresponding ideal scratch image to serve as one piece of data in the wafer defect data set, and the wafer defect data set comprises n pieces of data. According to the method, the wafer defect data set for scratch detection can be efficiently and accurately constructed.

Description

Technical Field

[0001] The present invention belongs to the technical field of wafer defect detection, and in particular relates to a method for constructing a wafer defect data set, a training method for a recognition system, and applications. Background Art

[0002] In the semiconductor manufacturing industry, wafer surface quality is crucial to the performance and reliability of the final chip. Due to various factors during the manufacturing process, a variety of surface defects often appear on the wafer surface, the most common of which are scratches and particles. These defects reduce production and yield. Different defects have different causes, so accurately identifying defect types and obtaining defect characteristic information (such as overall scratch length, scratch morphology, scratch direction, and particle size) is crucial for tracing defect causes and improving yield processes.

[0003] The commonly used wafer surface defect detection method is based on dark-field scattering technology, which analyzes scattered images to achieve defect detection and classification. For scratch detection, traditional scratch detection methods often require a lot of manpower and time. In addition, for a scratch with uneven thickness and shallow depth, several shallow parts of the scratch will not be displayed in the scattering image (for ease of description, they will be referred to as defect binary images) (these parts of the scratch are recorded as connection scratches). As a result, a scratch can easily be detected as multiple short scratches and particle defects. The absence of connection scratches in the scattering image not only seriously affects the accuracy of defect type identification, but also directly leads to incorrect defect feature information obtained.

[0004] To improve the accuracy and efficiency of defect detection, technicians are increasingly using deep learning models to detect scratches on wafer surfaces. These models require pre-training using datasets containing large amounts of training data. This training data is manually collected and labeled, which is not only costly and inefficient, but also varies widely among technicians, affecting the accuracy of the annotations.

[0005] Therefore, how to efficiently construct a wafer defect dataset for scratch detection has become a difficult problem that needs to be solved urgently in the field of wafer defect detection technology. Summary of the Invention

[0006] The purpose of the present invention is to overcome the deficiencies of the above-mentioned prior art and provide a method for constructing a wafer defect dataset, which can efficiently and accurately construct a wafer defect dataset for scratch detection.

[0007] To achieve the above object, the present invention adopts the following technical solutions:

[0008] A method for constructing a wafer defect dataset includes the following steps: generating n ideal scratch images using Bezier curves, where each ideal scratch image contains several ideal scratches; performing fracture processing and background noise processing on each ideal scratch image to form a corresponding simulated scratch image; and binding a simulated scratch image with the corresponding ideal scratch image to form a piece of data in the wafer defect dataset, where the wafer defect dataset contains n pieces of data.

[0009] Preferably, the method of generating n ideal scratch images by using Bezier curves further includes the following steps: in a background image, a plurality of ideal scratches are generated by using a third-order Bezier curve to obtain an ideal scratch image: Curve(t) = (1-t) 3 A+3t(1-t) 2 B+3(1-t)t 2 C+t 3 D; where Curve(t) represents the trajectory of any ideal scratch; A represents the starting coordinate of the current ideal scratch; D represents the ending coordinate of the current ideal scratch; B represents the first control point coordinate of the current ideal scratch; C represents the second control point coordinate of the current ideal scratch; and t represents the adjustment parameter of the current ideal scratch.

[0010] Preferably, the starting point coordinate A and the end point coordinate D are two non-overlapping points randomly selected in the current background image; the line segment connected by the starting point coordinate A and the end point coordinate D is recorded as line segment AD, the first control point coordinate B and the second control point coordinate C are located within a circular area with the line segment AD as the diameter, and 0≤distance(B, AD)≤d1 and 0≤distance(C, AD)≤d2, wherein distance(B, AD) represents the distance from the first control point coordinate B to the line segment AD, distance(C, AD) represents the distance from the second control point coordinate C to the line segment AD, d1 represents the first distance threshold, and d2 represents the second distance threshold.

[0011] Preferably, the scratch processing of the ideal scratch image further includes the following:

[0012] Step 11: After the technician determines the scratch width interval [α1, α2], the scratch width interval is divided into several width segment intervals, and the occurrence probability of each width segment interval is set so that the sum of the occurrence probabilities of all width segment intervals is 1; 0<α1<1<α2;

[0013] Step 12: After dividing the current ideal scratch into k segments, determine the width segmentation interval of each segment in the current ideal scratch according to the probability of occurrence of each width segmentation interval, and randomly select a value within the width segmentation interval as the width of the corresponding segment in the current ideal scratch;

[0014] Step 13: If the scratch width in the current segment is greater than 1, set 1 break in the current segment; if the scratch width in the current segment is equal to 1, set i breaks in the current segment; the break setting method is to erase all pixels within the set range.

[0015] Preferably, the background noise addition processing also includes the following contents: randomly obtaining a pixel point in the current ideal scratch image, recorded as the initial pixel point; the DFS algorithm and the BFS algorithm each have their own corresponding number of remaining traversed pixel points before the traversal begins, and each time a pixel point is traversed, the corresponding number of remaining traversed pixel points will be reduced by 1; using the DFS algorithm and / or the BFS algorithm to traverse the current ideal scratch image based on the initial pixel point, and changing the traversed pixel points to the scratch color; until the number of remaining traversed pixel points in the DFS algorithm and / or the BFS algorithm is 0, the background noise addition processing of the current ideal scratch image is completed.

[0016] The present invention also provides a training method for a wafer defect recognition system, wherein the wafer defect recognition system includes a scratch repair model and a defect recognition model, wherein the output of the scratch repair model serves as the input of the defect recognition model;

[0017] Training the scratch repair model also includes the following steps:

[0018] S1, extracting n1, n2, and n3 pieces of data from the wafer defect dataset to form a first training set, a first validation set, and a first test set, respectively; wherein the data in the first training set, the first validation set, and the first test set do not overlap, and n1+n2+n3=n, and n>3; the wafer defect dataset is obtained using the wafer defect dataset construction method described above;

[0019] S2, using the simulated scratch images and the corresponding ideal scratch images in the first training set to perform mask training on the U-Net image segmentation network model;

[0020] S3, using the first validation set to verify the scratch repair effect of the U-Net image segmentation network model on the simulated scratch image, and iteratively optimizing the U-Net image segmentation network model during the validation process;

[0021] S4, use the first test set to test the scratch repair accuracy ACC1 of the U-Net image segmentation network model: ACC1 = COR1 / n3; where COR1 represents the number of simulated scratch images repaired by the U-Net image segmentation network model that are identical to the corresponding ideal scratch images; if the scratch repair accuracy ACC1 is greater than the first accuracy threshold, the U-Net image segmentation network model at this time is recorded as the trained scratch repair model; if the scratch repair accuracy ACC1 is below the first accuracy threshold, return to S1 and perform a new round of training on the scratch repair model.

[0022] Preferably, the defect recognition model is trained, including S1 and S2′ to S4′;

[0023] S2′, the image obtained by performing an OR operation on the simulated scratch image and the corresponding ideal scratch image is recorded as a fused defect image; all wafer defects in the fused image corresponding to the first training set are scratches; the fused defect image corresponding to the first training set and several images of wafer defects of particle type are combined to form a second training set; the second training set is used to train the classification model to identify scratches and particles.

[0024] S3′, combining the fused defect images corresponding to the first validation set with several images of wafer defects with particles to form a second validation set; using the second validation set to verify the wafer defect recognition performance of the classification model, and iteratively optimizing the classification model during the verification process;

[0025] S4′, the fused defect image corresponding to the first test set and several images of wafer defect type particles are combined to form a second test set; the second test set is used to test the wafer defect recognition accuracy ACC2 of the classification model: ACC2 = COR2 / n4; where COR2 represents the number of images in which the classification model correctly recognizes all wafer defects in an image; n4 represents the number of images in the second test set; if the wafer defect recognition accuracy ACC2 is greater than the second accuracy threshold, the classification model at this time is recorded as a trained defect recognition model; if the wafer defect recognition accuracy ACC2 is below the first accuracy threshold, return to S1 and conduct a new round of training for the defect recognition model.

[0026] Preferably, the scratch repairing includes removing noise pixels and repairing broken marks.

[0027] The present invention also provides an application of a wafer defect recognition system: after obtaining a defect binary image, the image is sent to a scratch repair model in the wafer defect recognition system for scratch repair to obtain a repair image; the scratch repair model performs an OR operation on the repair image and the corresponding defect binary image, and outputs the corresponding fused defect image to the defect recognition model in the wafer defect recognition system; the defect recognition model identifies all wafer defects in the fused defect image and outputs the recognition result; before application, the wafer defect recognition system is trained using the training method of the wafer defect recognition system as described above.

[0028] The beneficial effects of the present invention are:

[0029] (1) In the method for constructing a wafer defect data set of the present invention, an ideal scratch image is first obtained, and then a corresponding simulated scratch image is obtained based on the ideal scratch image. Moreover, on the basis of the Bezier curve, the scratch shape, fracture degree, width and background noise are parameterized and controlled. Thus, ideal scratch images and simulated scratch images with various forms can be generated in batches. There is no need to spend a lot of manpower and economic costs to collect the defect binary images obtained during actual detection, and there is no need to manually label the defect binary images obtained during actual detection after feature analysis. This saves a lot of manpower costs and greatly improves the efficiency of constructing the wafer defect data set.

[0030] (2) Scratches with different degrees of curvature are generated by adjusting the control points, and random fractures and background noise are introduced to simulate the complex conditions in the real detection environment, making the simulated scratch images more realistic.

[0031] (3) In the prior art, technicians obtain a binary image of defects during actual detection, and then perform feature analysis on the binary image based on experience and then label the type of real wafer defects that the technicians believe. However, in the process of feature analysis, due to different standards of each technician, it will directly affect the accuracy of determining whether the scratches at the connection are missed in the binary image of defects, which will also directly affect the determination of the type of wafer defects, that is, it will affect the accuracy of the annotation of technicians in the process of constructing the wafer defect data set. The method for constructing the wafer defect data set of the present invention is completely opposite to the prior art. It first obtains an ideal scratch image, and the wafer defect type (that is, the wafer defect type is a scratch) and the parameters of the ideal scratch (that is, the parameters of the real scratch) are determined when the ideal scratch image is generated. The actual scratch part corresponding to the break in the subsequent simulated scratch image can also be determined without a doubt; that is, the data in the wafer defect data set constructed by the present invention does not have labeling errors, which directly improves the recognition accuracy of the defect recognition model trained using the wafer defect data set obtained in this embodiment.

[0032] (4) The wafer defect dataset constructed by the present invention contains a variety of real, reasonable, and random scratches and fractures, which improves the richness and diversity of the data.

[0033] (5) The training method of the wafer defect recognition system of the present invention can simultaneously train the scratch repair model and the defect recognition model based on the wafer defect data set, thereby improving the training efficiency, accuracy and robustness of the entire wafer defect recognition system.

[0034] (6) In the training method of the wafer defect recognition system of the present invention, the scratch repair model is trained based on the wafer defect data set, so that the repair result after the scratch repair model repairs the defect binary image detected by the prior art is closer to the corresponding real scratch.

[0035] (7) In the training method of the wafer defect recognition system of the present invention, the defect recognition model is trained based on the fused defect image obtained based on the wafer defect data set, thereby improving the accuracy of the defect recognition model in recognizing two types of wafer defects, scratches and particles.

[0036] (8) The application of the wafer defect recognition system of the present invention is to perform wafer defect recognition based on the fused defect image obtained by repairing the scratches on the defect binary image, thereby effectively improving the efficiency and accuracy of wafer defect recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 Flowchart of a method for constructing a wafer defect dataset according to the present invention;

[0038] Figure 2 is an ideal scratch image;

[0039] Figure 3 is a simulated scratch image;

[0040] Figure 4 This is a comparison chart of image fusion effects;

[0041] Figure 5 Flowchart of the training method of the wafer defect recognition system of the present invention. DETAILED DESCRIPTION

[0042] In order to make the technical solution of the present invention clearer and more specific, the present invention is clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments. Ordinary technicians in this field, without making any creative work, make equivalent substitutions for the technical features of the technical solution of the present invention and solutions derived from conventional reasoning all fall within the scope of protection of the present invention.

[0043] The wafer defect dataset required to be constructed in the present invention is for scratches.

[0044] Example 1

[0045] like Figure 1 FIG. 1 is a flowchart of a method for constructing a wafer defect dataset according to this embodiment, including the following contents:

[0046] N ideal scratch images are generated using Bezier curves, where n is a positive integer and one ideal scratch image contains several ideal scratches. Each ideal scratch image is then subjected to fracture processing and background noise addition to form a corresponding simulated scratch image. A simulated scratch image is bound to the corresponding ideal scratch image and used as a piece of data in the wafer defect dataset, which contains n pieces of data.

[0047] "Generating n ideal scratch images using Bezier curves" also includes the following:

[0048] In a black background image, after generating several ideal white scratches using a third-order Bezier curve, an ideal scratch image is obtained:

[0049] Curve(t)=(1-t) 3 A+3t(1-t) 2 B+3(1-t)t 2 C+t 3 D;

[0050] Among them, Curve(t) represents the trajectory of any ideal scratch; A represents the starting coordinate of the current ideal scratch; D represents the ending coordinate of the current ideal scratch; B represents the first control point coordinate of the current ideal scratch; C represents the second control point coordinate of the current ideal scratch; t represents the adjustment parameter of the current ideal scratch, which is set by technical personnel.

[0051] A, B, C, and D are four non-overlapping points randomly selected from the current black background image.

[0052] Optionally, the starting point coordinate A and the end point coordinate D are two non-overlapping points randomly selected in the current black background image; the line segment connected by the starting point coordinate A and the end point coordinate D is recorded as line segment AD, the first control point coordinate B and the second control point coordinate C are located within a circular area with the line segment AD as the diameter, and 0≤distance(B, AD)≤d1 and 0≤distance(C, AD)≤d2, where distance(B, AD) represents the distance from the first control point coordinate B to the line segment AD, distance(C, AD) represents the distance from the second control point coordinate C to the line segment AD, d1 represents the first distance threshold, and d2 represents the second distance threshold.

[0053] In this embodiment, the pixel value of an ideal scratch is 255, and the background pixel value is 0.

[0054] That is, the range of the first control point coordinate B and the second control point coordinate C is further limited, so that the trajectory of the ideal scratch is closer to the actual scratch condition.

[0055] According to the ideal scratch generation method described above, we can obtain a number of ideal scratches that are random, have different shapes, and fit the real shape, which greatly improves the realism and randomness of the ideal scratch image.

[0056] The number of ideal scratches contained in each ideal scratch image is also random.

[0057] like Figure 2 As shown, Figure 2 2a, 2b and 2c in FIG are three different ideal scratch images.

[0058] "Fracturing each ideal scratch image" also includes the following:

[0059] Step 11: After the technicians determine a reasonable scratch width interval [α1, α2] based on experience, they divide the scratch width interval into several width segmentation intervals and set the occurrence probability of each width segmentation interval so that the sum of the occurrence probabilities of all width segmentation intervals is 1; 0<α1<1<α2.

[0060] Step 12: After dividing the current ideal scratch into k segments, where k is a positive integer, determine the width segmentation interval of each segment in the current ideal scratch according to the probability of occurrence of each width segmentation interval, and randomly select the value within the width segmentation interval as the width of the corresponding segment in the current ideal scratch.

[0061] When k=1, the width of the entire ideal scratch is consistent; when k>1, some sections of an ideal scratch are wider, while some sections are narrower.

[0062] Long-term observation by technicians has shown that the widths of different sections of an actual scratch on the wafer surface vary. The method of determining the widths of each section of the ideal scratch in step 12 further increases the randomness of the scratch.

[0063] Step 13: If the scratch width in the current segment is greater than 1, set 1 break in the current segment; if the scratch width in the current segment is equal to 1, set i breaks in the current segment, where i is a positive integer; the break setting method is to erase all pixels within the set range.

[0064] The position and range of the fracture within the corresponding segment are random, that is, the range of pixels erased when setting the fracture is random.

[0065] A fracture mark is a scratch that includes a fracture surface.

[0066] In this embodiment, the number of pixels erased when setting the break is in the range of (0, 100).

[0067] According to the experience of technicians: For scratches with a width greater than 1, only normal disconnection will appear in the defect binary image (i.e., one fracture will appear); while for scratches with a width of 1, not only normal disconnection may appear in the defect binary image, but also dotted disconnection may appear (i.e., more than two fractures, making the scratch look like a dotted line). Therefore, in order to make the shape of the ideal scratch close to the actual scratch, we adopted the content described in step 13, ensuring the rationality of the fracture while increasing the randomness of the fracture.

[0068] "Background noise processing" also includes the following:

[0069] Randomly obtain a pixel in the current ideal scratch image and record it as the initial pixel. Use the DFS (depth-first search) algorithm and / or the BFS (breadth-first search) algorithm to traverse the initial pixel and change the traversed pixels in the current ideal scratch image to white. When the number of remaining traversed pixels in the DFS algorithm and / or the BFS algorithm reaches 0, the background noise processing of the current ideal scratch image is completed.

[0070] The DFS algorithm and / or BFS algorithm each have a corresponding number of remaining traversal pixels before traversal begins. Each time a pixel is traversed, the corresponding number of remaining traversal pixels decreases by 1. Technicians can set the number of remaining traversal pixels before the DFS algorithm and / or BFS algorithm traversal begins.

[0071] By adding randomly distributed noise pixels to the ideal scratch image, the resulting simulated scratch image more closely resembles the binary defect image obtained during actual inspection. The DFS algorithm explores adjacent pixels in a depth-first manner, while the BFS algorithm explores adjacent pixels in a breadth-first manner. The random application of these two algorithms further increases the randomness and realism of the generated noise pixels.

[0072] The ideal scratches in the ideal scratch image correspond to the actual scratches on the wafer surface, while the simulated scratch image corresponds to the defect binary image obtained when actually detecting the actual scratches on the wafer surface. Because this embodiment first obtains the ideal scratch image and then obtains the corresponding simulated scratch image based on the ideal scratch image, and based on the Bezier curve, this embodiment performs parameterized control of the scratch shape, degree of fracture, width, and background noise, it can batch generate ideal scratch images and simulated scratch images with various forms. This eliminates the need for consuming a large amount of manpower and economic costs to collect the defect binary images obtained during actual detection, and eliminates the need for manual labeling of the defect binary images obtained during actual detection after feature analysis. This saves a lot of manpower costs and greatly improves the efficiency of constructing wafer defect datasets.

[0073] By adjusting the control points, scratches with different degrees of curvature are generated. By introducing random fractures and background noise, the complex conditions in the real detection environment are simulated, making the simulated scratch images more realistic.

[0074] In the existing technology, technicians obtain a binary defect image during actual inspection, and then perform feature analysis on the binary defect image based on experience before labeling the type of real wafer defects that the technicians believe. However, in the process of feature analysis, due to different standards among technicians, it will directly affect the accuracy of judging whether the scratches at the connection are missed in the binary defect image, which will also directly affect the judgment of the wafer defect type, that is, it will affect the accuracy of the annotation of technicians in the process of constructing the wafer defect data set. For example, a real scratch on the wafer surface appears as a broken dotted line in the detected binary defect image, but the dotted scratch is extremely short. If the technicians believe that this is not a broken dotted line scratch, but rather two scratches, and there is a granular wafer defect in the middle of the two scratches. The method for constructing the wafer defect dataset of this embodiment is completely opposite to the prior art. The present invention first obtains an ideal scratch image, and the wafer defect type (that is, the wafer defect type is a scratch) and the parameters of the ideal scratch (that is, the parameters of the real scratch) are determined when the ideal scratch image is generated. The actual scratch part corresponding to the break in the subsequent simulated scratch image can also be determined without a doubt; that is, the data in the wafer defect dataset constructed by this embodiment does not have labeling errors, which directly improves the recognition accuracy of the defect recognition model subsequently trained using the wafer defect dataset obtained by this embodiment.

[0075] The wafer defect dataset constructed in this embodiment includes various real, reasonable, and randomly shaped scratches and fractures, which improves the richness and diversity of the data.

[0076] Will Figure 2 After the three ideal scratch images in the image are processed by fracture and background noise, the image is formed. Figure 3 The three corresponding simulated scratch images. Figure 3 3a in corresponds to Figure 2 2a in Figure 3 3b in corresponds to Figure 2 2b in Figure 3 3c in corresponds to Figure 2 2c in.

[0077] Example 2

[0078] The wafer defect recognition system includes a scratch repair model and a defect recognition model, and the output of the scratch repair model serves as the input of the defect recognition model.

[0079] This embodiment also provides a training method for a wafer defect recognition system, such as Figure 5 As shown, the following steps are included:

[0080] S1, extract n1, n2 and n3 data from the wafer defect data set to form a first training set, a first validation set and a first test set respectively; wherein the data in the first training set, the first validation set and the first test set do not overlap, and n1+n2+n3=n, and n>3.

[0081] The wafer defect dataset is constructed using the wafer defect dataset construction method described in Example 1.

[0082] After S1, the scratch repair model is trained, including S2 to S4:

[0083] S2, using the simulated scratch images and the corresponding ideal scratch images in the first training set, perform mask training on the U-Net image segmentation network model.

[0084] S3, using the first validation set to verify the scratch repair effect of the U-Net image segmentation network model on the simulated scratch image, and iteratively optimize the U-Net image segmentation network model during the verification process.

[0085] S4, use the first test set to test the scratch repair accuracy ACC1 of the U-Net image segmentation network model:

[0086] ACC1 = COR1 / n3; where COR1 represents the number of images whose simulated scratch images after restoration by the U-Net image segmentation network model are identical to the corresponding ideal scratch images;

[0087] If the scratch repair accuracy ACC1 is greater than the first accuracy threshold, the U-Net image segmentation network model at this time is recorded as the trained scratch repair model; if the scratch repair accuracy ACC1 is below the first accuracy threshold, return to S1 and perform a new round of training on the scratch repair model.

[0088] Scratch repair includes removing noisy pixels and repairing broken marks.

[0089] The scratch repair model is a trained U-Net image segmentation network model. The U-Net image segmentation network model has an encoder-decoder structure and skip connections, which can effectively capture features of different scales and retain spatial information.

[0090] After S1, the defect recognition model is trained, including S2' to S4':

[0091] S2′, the image obtained by performing an OR operation on the simulated scratch image and the corresponding ideal scratch image is recorded as a fused defect image; all wafer defect types in the fused image corresponding to the first training set are scratches; the fused defect image corresponding to the first training set and several images of wafer defect types of particles together constitute a second training set; the second training set is used to train the classification model to identify scratches and particles.

[0092] S3′, the fused defect images corresponding to the first verification set and several images of wafer defects with particles are combined to form a second verification set; the second verification set is used to verify the wafer defect recognition effect of the classification model, and the classification model is iteratively optimized during the verification process.

[0093] S4′, combining the fused defect images corresponding to the first test set with several images of wafer defects of particle type to form a second test set; using the second test set to test the classification model's wafer defect recognition accuracy ACC2: ACC2 = COR2 / n4; where COR2 represents the number of images in which the classification model correctly identifies all wafer defects in an image; n4 represents the number of images in the second test set, where n4>n3 and n4 is a positive integer;

[0094] If the wafer defect recognition accuracy ACC2 is greater than the second accuracy threshold, the classification model at this time is recorded as the trained defect recognition model; if the wafer defect recognition accuracy ACC2 is below the first accuracy threshold, return to S1 and perform a new round of training on the defect recognition model.

[0095] In S2′, the image of the wafer defect type being particles is the defect binary image actually collected.

[0096] like Figure 4 The following is a comparison chart of image fusion effects. Figure 4 4a in the figure is a simulated scratch image. Figure 4 4b in the figure is the ideal scratch image corresponding to 4a. Figure 4 4c is the fused defect image obtained by performing an OR operation on 4a and 4b.

[0097] For the scratch repair model / defect recognition model: If the input image exceeds the input size limit, the entire large image can be decomposed into smaller images that meet the format requirements. After scratch repair / defect recognition is performed on the smaller images, they can be reassembled. If there are smaller images smaller than the input size limit after the large image is decomposed, they can be padded with zero pixel values until the smaller images meet the input size limit.

[0098] This embodiment trains a scratch repair model based on a wafer defect dataset, so that the scratch repair model can produce a repair result that is closer to the corresponding real scratch after repairing the defect binary image detected by the prior art.

[0099] This embodiment trains the defect recognition model based on fused defect images obtained from a wafer defect dataset. The fused defect images ensure a complete distribution of scratches and particles, making it easier for the defect recognition model to extract features of scratches and particles, thereby improving the accuracy of the defect recognition model in identifying these two types of wafer defects. The defect recognition model performs a connected domain analysis on the fused defect images or images with particles as the wafer defect type, extracting the geometric features of each defect area, including area, perimeter, and shape parameters (such as aspect ratio). Based on the features extracted through the connected domain analysis, scratches and particles are classified based on the following rules: scratches are typically characterized by a large axial length ratio, a linear shape, and a small area ratio, while particles are characterized by a small area and a shape that is nearly circular or irregularly dotted.

[0100] As can be seen from the above description, based on the wafer defect dataset, the scratch repair model and defect recognition model can be trained simultaneously, which improves the training efficiency, accuracy, and robustness of the entire wafer defect recognition system.

[0101] Example 3

[0102] This embodiment also provides an application of a wafer defect recognition system, including: obtaining a defect binary image and sending it into a scratch repair model in the wafer defect recognition system for scratch repair to obtain a repair image; the scratch repair model performs an OR operation on the repair image and the corresponding defect binary image, and outputs the corresponding fused defect image to the defect recognition model in the wafer defect recognition system; the defect recognition model identifies all wafer defects in the fused defect image and outputs the identification results.

[0103] Before application, the wafer defect recognition system is trained using the training method for the wafer defect recognition system described in Example 2.

[0104] Some binary defect images contain not only wafer defects like scratches but also wafer defects like particles; some binary defect images contain only wafer defects like particles. The scratch repair model will also repair binary defect images containing particles, but this repair may only remove noise pixels. If there are no scratches in the current binary defect image, no break repair will be performed. Therefore, when a binary defect image containing particles is fed into the scratch repair model, the corresponding repaired image will be generated.

[0105] The application of the wafer defect recognition system in this embodiment performs scratch repair on the defect binary image and then performs wafer defect recognition based on the fused defect image obtained from the repair image and the defect binary image, effectively improving the efficiency and accuracy of wafer defect recognition.

[0106] The technologies, shapes, and structures not described in detail in this embodiment are all well-known technologies. It should also be pointed out that the above are only preferred embodiments of the present invention and are not intended to limit the present invention. The components or steps in this embodiment can be decomposed and / or recombined, and such decompositions and / or recombinations should be regarded as equivalent solutions of this application and should fall within the scope of protection of the present invention.

Claims

1. A method for constructing a wafer defect dataset, characterized in that: Includes the following: N ideal scratch images are generated using Bezier curves, where each ideal scratch image contains several ideal scratches. Each ideal scratch image is then subjected to fracture processing and background noise addition to form a corresponding simulated scratch image. A simulated scratch image is bound to the corresponding ideal scratch image and used as a piece of data in the wafer defect dataset, which contains n pieces of data.

2. The method for constructing a wafer defect dataset according to claim 1, wherein: Generate n ideal scratch images through Bezier curves, including the following: In a background image, after generating several ideal scratches using a third-order Bezier curve, an ideal scratch image is obtained: Curve(t)=(1-t) 3 A+3t(1-t) 2 B+3(1-t)t 2 C+t 3 D; Where Curve(t) represents the trajectory of any ideal scratch; A represents the starting coordinate of the current ideal scratch; D represents the ending coordinate of the current ideal scratch; B represents the coordinate of the first control point of the current ideal scratch; C represents the coordinate of the second control point of the current ideal scratch; and t represents the adjustment parameter of the current ideal scratch.

3. The method for constructing a wafer defect dataset according to claim 2, wherein: The starting point coordinate A and the end point coordinate D are two non-overlapping points randomly selected in the current background image; the line segment connected by the starting point coordinate A and the end point coordinate D is recorded as line segment AD, the first control point coordinate B and the second control point coordinate C are located within a circular area with the line segment AD as the diameter, and 0≤distance(B, AD)≤d1 and 0≤distance(C, AD)≤d2, where distance(B, AD) represents the distance from the first control point coordinate B to the line segment AD, distance(C, AD) represents the distance from the second control point coordinate C to the line segment AD, d1 represents the first distance threshold, and d2 represents the second distance threshold.

4. The method for constructing a wafer defect dataset according to claim 1, wherein: The ideal scratch image is processed for fracture, which also includes the following: Step 11: After the technician determines the scratch width interval [α1, α2], the scratch width interval is divided into several width segment intervals, and the occurrence probability of each width segment interval is set so that the sum of the occurrence probabilities of all width segment intervals is 1; 0<α1<1<α2; Step 12: After dividing the current ideal scratch into k segments, determine the width segmentation interval of each segment in the current ideal scratch according to the probability of occurrence of each width segmentation interval, and randomly select a value within the width segmentation interval as the width of the corresponding segment in the current ideal scratch; Step 13: If the scratch width in the current segment is greater than 1, set 1 break in the current segment; if the scratch width in the current segment is equal to 1, set i breaks in the current segment; the break setting method is to erase all pixels within the set range.

5. The method for constructing a wafer defect dataset according to claim 1, wherein: The background noise addition process also includes the following: randomly obtaining a pixel point in the current ideal scratch image, recorded as the initial pixel point; the DFS algorithm and the BFS algorithm each have their own corresponding number of remaining traversal pixels before the traversal begins, and each time a pixel point is traversed, the corresponding number of remaining traversal pixels will be reduced by 1; using the DFS algorithm and / or the BFS algorithm to traverse the current ideal scratch image based on the initial pixel point, and changing the traversed pixel points to the scratch color; until the number of remaining traversal pixels in the DFS algorithm and / or the BFS algorithm is 0, the background noise addition process of the current ideal scratch image is completed.

6. A training method for a wafer defect recognition system, characterized by: The wafer defect recognition system includes a scratch repair model and a defect recognition model. The output of the scratch repair model serves as the input of the defect recognition model. Training the scratch repair model also includes the following steps: S1, extracting n1, n2, and n3 pieces of data from a wafer defect dataset to form a first training set, a first validation set, and a first test set, respectively; wherein the data in the first training set, the first validation set, and the first test set do not overlap, and n1+n2+n3=n, and n>3; the wafer defect dataset is obtained using the wafer defect dataset construction method according to any one of claims 1 to 5; S2, using the simulated scratch images and the corresponding ideal scratch images in the first training set to perform mask training on the U-Net image segmentation network model; S3, using the first validation set to verify the scratch repair effect of the U-Net image segmentation network model on the simulated scratch image, and iteratively optimizing the U-Net image segmentation network model during the validation process; S4, use the first test set to test the scratch repair accuracy ACC1 of the U-Net image segmentation network model: ACC1 = COR1 / n3; where COR1 represents the number of simulated scratch images repaired by the U-Net image segmentation network model that are identical to the corresponding ideal scratch images; if the scratch repair accuracy ACC1 is greater than the first accuracy threshold, the U-Net image segmentation network model at this time is recorded as the trained scratch repair model; if the scratch repair accuracy ACC1 is below the first accuracy threshold, return to S1 and perform a new round of training on the scratch repair model.

7. The method for training a wafer defect recognition system according to claim 6, wherein: Training the defect recognition model, including S1 and S2′ to S4′; S2′, the image obtained by performing an OR operation on the simulated scratch image and the corresponding ideal scratch image is recorded as a fused defect image; all wafer defects in the fused image corresponding to the first training set are scratches; the fused defect image corresponding to the first training set and several images of wafer defects of particle type are combined to form a second training set; the second training set is used to train the classification model to identify scratches and particles. S3′, combining the fused defect images corresponding to the first validation set with several images of wafer defects with particles to form a second validation set; using the second validation set to verify the wafer defect recognition performance of the classification model, and iteratively optimizing the classification model during the verification process; S4′, the fused defect image corresponding to the first test set and several images of wafer defect type particles are combined to form a second test set; the second test set is used to test the wafer defect recognition accuracy ACC2 of the classification model: ACC2 = COR2 / n4; where COR2 represents the number of images in which the classification model correctly recognizes all wafer defects in an image; n4 represents the number of images in the second test set; if the wafer defect recognition accuracy ACC2 is greater than the second accuracy threshold, the classification model at this time is recorded as a trained defect recognition model; if the wafer defect recognition accuracy ACC2 is below the first accuracy threshold, return to S1 and conduct a new round of training for the defect recognition model.

8. The method for training a wafer defect recognition system according to claim 6, wherein: Scratch repair includes removing noisy pixels and repairing broken marks.

9. Application of wafer defect recognition system, characterized by: After obtaining the defect binary image, it is sent to the scratch repair model in the wafer defect recognition system for scratch repair to obtain a repair image; the scratch repair model performs an OR operation on the repair image and the corresponding defect binary image, and outputs the corresponding fused defect image to the defect recognition model in the wafer defect recognition system; the defect recognition model identifies all wafer defects in the fused defect image and outputs the recognition results; Before application, the wafer defect recognition system is trained using the wafer defect recognition system training method described in any one of claims 6 to 8.

Images