Wafer defect recognition method and system under multi-feature fusion

Abstract

The application discloses a wafer defect recognition method and system under multi-feature fusion, and relates to the related field of semiconductor manufacturing.The method comprises the following steps: an optical detector is activated to acquire an initial surface image of a target wafer, and the initial surface image is processed according to a preprocessing strategy to obtain a target surface image; a first image corresponding to a first structure is extracted from the target surface image, and a first target image is constructed according to a plurality of die images in the first image; the first target image is compared with a first standard image to obtain a first defect image, and a defect recognition result of the target wafer is obtained based on the first defect image.The technical problem of insufficient recognition accuracy and robustness of the existing wafer defect recognition is solved, and the technical effect of improving the accuracy and robustness of defect recognition is achieved.

Description

Technical Field

[0001] This application relates to the field of semiconductor manufacturing, and in particular to a method and system for identifying wafer defects using multi-feature fusion. Background Technology

[0002] In the semiconductor manufacturing industry, wafers are the core material for chip manufacturing, and their quality directly determines the performance and yield of the final chip. Even the smallest defects on the wafer surface can lead to chip malfunctions and significant economic losses. Therefore, efficient and accurate identification of wafer defects is a crucial step in the semiconductor manufacturing process. Currently, the main method for solving wafer defect identification is based on traditional image processing and machine learning methods using single features. This typically involves first acquiring images of the wafer surface using equipment such as optical microscopes, then optimizing the images using preprocessing techniques such as image enhancement and filtering. Next, single features such as texture and edges are extracted from the images, and finally, machine learning algorithms are used to classify and identify wafer defects based on these extracted features. However, current methods, relying primarily on single features for defect identification, suffer from several problems. On the one hand, wafer surface defects are complex and diverse, and different types of defects may exhibit similar characteristics in terms of a single feature. This makes it difficult to accurately distinguish various defects based on a single feature, easily leading to misjudgments and missed detections. On the other hand, single features are highly sensitive to interference factors such as image noise and changes in lighting. In complex industrial environments, these interferences can severely affect the accuracy of single feature extraction, thereby reducing the robustness and stability of defect identification.

[0003] Currently, wafer defect identification technologies suffer from insufficient accuracy and robustness. Summary of the Invention

[0004] This application provides a wafer defect identification method and system based on multi-feature fusion. The method employs an activated optical detector to acquire an initial surface image of the target wafer, preprocesses the initial image to obtain a target surface image, extracts first image features from the processed image, constructs a first target image for comparison based on these features, compares the first target image with a preset first standard image, generates a first defect image by analyzing the differences, and outputs the defect identification result of the target wafer based on the defect image. These techniques solve the technical problems of insufficient accuracy and robustness in existing wafer defect identification methods, achieving the technical effect of improving the accuracy and robustness of defect identification.

[0005] This application provides a wafer defect identification method under multi-feature fusion, comprising: activating an optical detector to acquire an initial surface image of a target wafer, and processing the initial surface image according to a preprocessing strategy to obtain a target surface image; extracting a first image corresponding to a first structure from the target surface image, and constructing a first target image based on multiple grain images in the first image; comparing the first target image with a first standard image to obtain a first defect image, and obtaining a defect identification result of the target wafer based on the first defect image.

[0006] In a possible implementation, an optical detector is activated to acquire an initial surface image of the target wafer, and the following processing is performed: an initial three-dimensional model of the target wafer is acquired through a 3D defect detection component in the optical detector; the initial three-dimensional model is compared with a predetermined three-dimensional model to obtain a three-dimensional comparison result; if the three-dimensional comparison result meets a predetermined deviation constraint, a 2D defect detection component in the optical detector is activated; the initial surface image is acquired through the 2D defect detection component; wherein, the three-dimensional comparison result includes pit height difference and protrusion height difference.

[0007] In a possible implementation, an initial three-dimensional model of the target wafer is obtained through a 3D defect detection component in the optical detector, and the following processing is performed: multi-angle wafer images of the target wafer are collected through the 3D defect detection component; a three-dimensional wafer structure is constructed based on the multi-angle wafer images, and dynamic surface mapping is performed on the three-dimensional wafer structure to obtain the initial three-dimensional model.

[0008] In a possible implementation, the following processing is performed: the preprocessing strategy includes an illumination correction scheme and a noise filtering scheme; wherein, the initial surface image is separated according to the illumination correction scheme to obtain a separation result, and the low-frequency components in the separation result are subjected to gamma compression; wherein, a dynamic filtering threshold is obtained according to the noise filtering scheme, and the initial surface image is filtered based on the dynamic filtering threshold, and the dynamic filtering threshold refers to a filtering threshold dynamically adjusted based on the mapping relationship between the spot diameter and the defect size.

[0009] In a possible implementation, a first image corresponding to the first structure is extracted from the target surface image, and a first target image is constructed based on multiple grain images in the first image. The following processing is performed: obtaining a first grayscale value of the first grain image among the multiple grain images; obtaining a second grayscale value of a second grain image among the multiple grain images, wherein the second grain image is adjacent to the first grain image; establishing a grayscale value linked list based on the first grayscale value and the second grayscale value; performing iterative filtering analysis on the grayscale value linked list according to an iterative filtering mechanism to obtain a first target pixel value; and obtaining the first target image based on the first target pixel value.

[0010] In a possible implementation, the grayscale value linked list is iteratively filtered and analyzed according to an iterative filtering mechanism to obtain a first target pixel value, and the following processing is performed: the mean and standard deviation of the grayscale linked list are calculated respectively; it is determined whether the standard deviation is within a predetermined limit, and a judgment result is obtained, wherein the predetermined limit is contained in the iterative filtering mechanism; the grayscale value linked list is filtered according to the judgment result to obtain candidate pixel values; if the standard deviation meets a predetermined termination threshold, the mean of the candidate pixel values ​​is taken as the first target pixel value.

[0011] In a possible implementation, the following processing is performed: the predetermined limit includes a first limit and a second limit; wherein, when the standard deviation is higher than the first limit, a defect point removal operation is performed; wherein, when the standard deviation is lower than the second limit, a noise point correction operation is performed; and the candidate pixel value is obtained based on the defect point removal operation and the noise point correction operation.

[0012] In a possible implementation, the first target image is compared with the first standard image to obtain a first defect image, and the following processing is performed: the first target image and the first standard image are sequentially subjected to inverse pyramid transformation to obtain a first target transformation image and a first standard transformation image, respectively; the first target transformation image and the first standard transformation image are analyzed through dual generation channels to generate a dual-channel defect image; the dual-channel defect images are fused to obtain the first defect image; wherein, the dual generation channels include a first channel and a second channel, including: the first channel obtains a first channel defect image based on the structural similarity weighting principle; the second channel obtains a second channel defect image based on the 1-cube operation principle of similarity value; the first channel defect image and the second channel defect image constitute the dual-channel defect image.

[0013] In a possible implementation, the following process is performed: before fusing the dual-channel defect maps to obtain the first defect map, the first channel defect map is further subjected to frequency domain calibration based on the first phase spectrum information of the first target map.

[0014] This application also provides a wafer defect identification system based on multi-feature fusion, comprising: a surface image preprocessing module, used to activate an optical detector to acquire an initial surface image of a target wafer, and process the initial surface image according to a preprocessing strategy to obtain a target surface image; a first target image construction module, used to extract a first image corresponding to a first structure from the target surface image, and construct a first target image based on multiple grain images in the first image; and a defect identification module, used to compare the first target image with a first standard image to obtain a first defect image, and obtain a defect identification result of the target wafer based on the first defect image.

[0015] This application proposes a wafer defect identification method and system based on multi-feature fusion. First, an optical detector is activated to acquire an initial surface image of the target wafer. This initial surface image is then processed according to a preprocessing strategy to obtain a target surface image. Next, a first image corresponding to a first structure is extracted from the target surface image. A first target image is constructed based on multiple grain images from the first image. Finally, the first target image is compared with a first standard image to obtain a first defect image. The defect identification result of the target wafer is then obtained based on the first defect image. This achieves the technical effect of improving the accuracy and robustness of defect identification. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings of the embodiments of the present invention will be briefly described below. Flowcharts are used in this application to illustrate the operations performed by the system according to the embodiments of the present application. It should be understood that the preceding or following operations are not necessarily performed precisely in sequence. Instead, various steps can be processed in reverse order or simultaneously as needed. Furthermore, other operations can be added to these processes, or one or more steps can be removed from these processes.

[0017] Figure 1 This is a flowchart illustrating a wafer defect identification method based on multi-feature fusion, as provided in an embodiment of this application.

[0018] Figure 2 This is a schematic diagram of a wafer defect identification system based on multi-feature fusion, provided in an embodiment of this application.

[0019] Figure labeling: Surface image preprocessing module 10, first target image construction module 20, defect identification module 30. Detailed Implementation

[0020] The above description is merely an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, specific embodiments of this application are given below.

[0021] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description of this application will be provided in conjunction with the accompanying drawings. The described embodiments should not be considered as limitations on this application. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0022] In the following description, references to "some embodiments" describe a subset of all possible embodiments. However, it is understood that "some embodiments" can be the same or different subsets of all possible embodiments and can be combined with each other without conflict. The terms "first" and "second" are used merely to distinguish similar objects and do not represent a specific ordering of objects. The terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or modules not explicitly listed or inherent to these processes, methods, products, or devices. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only.

[0023] This application provides a method for wafer defect identification based on multi-feature fusion, such as... Figure 1 As shown, the method includes:

[0024] Step S100: Activate the optical detector to acquire an initial surface image of the target wafer, and process the initial surface image according to the preprocessing strategy to obtain the target surface image.

[0025] Specifically, the optical detector can employ a high-resolution industrial camera, which possesses high sensitivity and rapid imaging capabilities. It is activated by controlling the detector's power supply and signal triggering circuitry. For example, a programmable logic controller (PLC) can send a start signal to the detector. Upon receiving the signal, the detector images the target wafer according to preset parameters such as exposure time and gain, converting the light signal into an electrical signal to generate an initial surface image, which is then stored in the detector's buffer or external storage device in a digital image format (such as BMP, JPEG, etc.).

[0026] The initial surface image is processed according to a preprocessing strategy to obtain the target surface image. This preprocessing strategy includes several steps, such as: using a median filtering algorithm to remove salt-and-pepper noise from the image. Specifically, the gray value of each pixel in the image is replaced by the median gray value of its neighboring pixels. For example, for a 3×3 neighborhood window, the gray values ​​of the nine pixels within the window are sorted, and the median value is taken as the new gray value of the center pixel. Another example is using histogram equalization to enhance image contrast. First, the gray-level histogram of the image is calculated, and then a gray-level transformation function is calculated based on the histogram to transform the gray value of each pixel in the image, making the gray-level distribution of the transformed image more uniform, thereby enhancing the image's detail information. Finally, if the image has geometric distortions, such as perspective distortion, affine transformation can be used for correction. By selecting multiple feature points in the image (such as the edge points of a wafer), the correspondence before and after the transformation is determined, the affine transformation matrix is ​​calculated, and then the coordinates of each pixel in the image are transformed to obtain the corrected image. After these processing steps, the target surface image is obtained.

[0027] In one possible implementation, an optical detector is activated to acquire an initial surface image of the target wafer. Step S100 further includes step S110, where an initial three-dimensional model of the target wafer is acquired using a 3D defect detection component in the optical detector. Specifically, the optical detector is equipped with a 3D defect detection component, which can employ technologies such as laser scanning or structured light projection. Taking laser scanning as an example, a laser emitter emits a laser beam that illuminates the surface of the target wafer. Due to the unevenness of the wafer surface, the laser beam will deform. A high-speed camera captures the image of the deformed laser beam, while simultaneously recording the angle and position information of the laser emission. Using the principle of triangulation, the height information of each point on the wafer surface relative to a reference plane is calculated based on parameters such as the degree of deformation of the laser beam, the emission angle, and the camera position. By integrating this height information according to certain spatial coordinates, an initial three-dimensional model of the target wafer can be constructed, which accurately reflects the three-dimensional morphological characteristics of the wafer surface.

[0028] Step S120 involves comparing the initial 3D model with a predetermined 3D model to obtain a 3D comparison result, which includes pit height difference and protrusion height difference. Specifically, the predetermined 3D model is a pre-stored wafer 3D model that meets quality standards, representing the ideal 3D shape of the wafer. The initial 3D model obtained in step S110 is registered with the predetermined 3D model to ensure consistency in spatial position and orientation between the two models. Then, the height difference between each point in the initial 3D model and the corresponding point in the predetermined 3D model is calculated using a point-by-point comparison method. For pitted areas on the wafer surface, the calculated height difference is the height difference between the bottom of the pit and the ideal surface, i.e., the pit height difference; for protrusion areas, the calculated height difference is the height difference between the top of the protrusion and the ideal surface, i.e., the protrusion height difference. These height difference data can intuitively reflect the defects on the wafer surface in three-dimensional space.

[0029] In step S130, if the three-dimensional comparison result meets the predetermined deviation constraint, the 2D defect detection component in the optical detector is activated. Specifically, the predetermined deviation constraint is a pre-set allowable range of three-dimensional defects based on the wafer's manufacturing process and quality requirements. For example, it is specified that the pit height difference should not exceed 0.1 mm, and the bump height difference should not exceed 0.05 mm. When the three-dimensional comparison results (pit height difference and bump height difference) calculated in step S120 are both within the predetermined deviation constraint range, it indicates that the three-dimensional defects on the wafer surface are within an acceptable range. At this time, the 2D defect detection component in the optical detector can be further activated to perform more detailed two-dimensional defect detection on the wafer surface. If the three-dimensional comparison result exceeds the predetermined deviation constraint, it indicates that there are relatively serious three-dimensional defects on the wafer surface. Further 2D detection is not required, and the wafer is directly determined to be a defective product.

[0030] Step S140: Acquire the initial surface image using the 2D defect detection component. Specifically, the 2D defect detection component employs a high-resolution industrial camera. Once activated, the component's power supply and signal trigger circuit are controlled to initiate operation. A programmable logic controller (PLC) sends a start signal to the detection component. Upon receiving the signal, the component images the target wafer according to preset parameters such as exposure time and gain. During the image capture, appropriate lighting ensures that the details of the wafer surface are clearly displayed. After converting the light signal into an electrical signal, an initial surface image is generated and stored in a digital image format (such as BMP, JPEG, etc.) in the detection component's cache or external storage device.

[0031] This implementation method acquires an initial 3D model through a 3D defect detection component and compares it, which can quickly detect 3D defects on the wafer surface. If the 3D comparison result meets the predetermined deviation constraints, the 2D defect detection component is then activated to acquire an initial surface image. 2D detection can further detect other defects on the wafer surface. The combination of 3D and 2D detection can comprehensively detect defects on the wafer surface from different dimensions, improving the accuracy of defect detection.

[0032] In one possible implementation, an initial three-dimensional model of the target wafer is acquired through a 3D defect detection component in the optical detector. Step S110 further includes step S111, whereby the 3D defect detection component collects multi-angle wafer images of the target wafer. Specifically, the 3D defect detection component can be equipped with multiple image acquisition devices at different angles, such as multiple high-resolution industrial cameras. These cameras are arranged around the target wafer to ensure that the wafer can be photographed from multiple different perspectives. In actual operation, the detection system controls the wafer to rotate or controls the cameras to move (depending on the specific design) to achieve multi-angle imaging. For example, in a typical setup, six cameras are evenly distributed around the wafer, with an angular interval of 60 degrees between each camera. During wafer rotation, each camera takes an image at a specific point in time, thereby acquiring images of the wafer at different rotation angles. These images contain detailed information about the wafer surface from different perspectives, including shape, texture, edge features, etc. Through multi-angle imaging, the manifestation of various defects that may exist on the wafer surface from different perspectives can be captured.

[0033] Step S112 involves constructing a three-dimensional wafer structure based on the multi-angle wafer images and performing dynamic surface mapping on the three-dimensional wafer structure to obtain the initial three-dimensional model. Specifically, the multi-angle wafer images obtained in step S111 are preprocessed, including denoising, image enhancement, and correction, to improve image quality. Then, representative feature points are extracted from each image using feature extraction algorithms (such as SIFT). These feature points reflect the local features of the wafer surface. Next, the same feature points in different images are matched using a feature matching algorithm to determine their correspondence in three-dimensional space. Based on these matched feature points, the coordinates of each feature point in three-dimensional space are calculated using methods such as triangulation or multi-view geometry. Finally, the coordinates of all calculated feature points are connected to construct a preliminary three-dimensional wafer structure framework.

[0034] After constructing the initial 3D wafer structure, dynamic surface mapping is needed to make the model more accurate and realistic, based on the physical and optical properties of the wafer surface. For example, the wafer surface exhibits optical phenomena such as reflection and refraction, which affect the appearance of the wafer surface in the image. Through dynamic surface mapping, combined with information such as color and texture from multi-angle images, the surface of the initially constructed 3D structure is corrected and optimized. That is, based on the color and texture changes at corresponding locations in images from different angles, the curvature and shape of the 3D model surface are adjusted to better match the characteristics of the actual wafer surface. After dynamic surface mapping, the resulting initial 3D model of the target wafer can more accurately reflect the true morphology and defects of the wafer surface.

[0035] This approach collects multi-angle images of the target wafer, enabling the acquisition of wafer surface information from different perspectives. Images from different angles can complement each other, reducing information loss and occlusion issues caused by a single viewpoint. Dynamic surface mapping can finely adjust the model surface based on color and texture information from the multi-angle images, making the model not only consistent with the actual wafer in shape but also more realistic in surface appearance, thus improving the accuracy of the initial 3D model construction.

[0036] In one possible implementation, step S100 further includes step S150, where the preprocessing strategy includes an illumination correction scheme and a noise filtering scheme. The initial surface image is separated according to the illumination correction scheme to obtain a separation result, and the low-frequency components in the separation result are subjected to gamma compression. Specifically, when acquiring the initial surface image of the target wafer, due to the non-uniformity of illumination conditions, the image may exhibit localized over-brightness or under-brightness. This situation can cause defect features to be masked or distorted by abnormal illumination, thus affecting subsequent detection and analysis of wafer surface defects. The illumination correction scheme is used to make the illumination distribution in the image more uniform, thereby enabling more accurate extraction of the true information of the wafer surface.

[0037] According to the illumination correction plan, a separation operation is performed on the initial surface image. The separation method can be based on frequency domain analysis of the image, such as using Fourier transform to convert the image from the spatial domain to the frequency domain. In the frequency domain, the low-frequency components of the image represent the overall brightness and slowly changing parts, while the high-frequency components contain detailed information such as edges and textures. By using a filter, the low-frequency and high-frequency components of the image can be separated, yielding a separation result. This separation result contains both low-frequency and high-frequency information from the image.

[0038] Gamma compression is a non-linear image enhancement technique that adjusts image contrast and brightness by non-linearly transforming pixel values. In illumination correction, applying gamma compression to the separated low-frequency components can effectively improve the brightness distribution of an image. Specifically, the gamma compression formula is I... out =I in γ , where I in It is the input pixel value, I out γ is the output pixel value, and γ is the gamma coefficient. When γ < 1, the pixel values ​​in low-brightness areas are amplified to make details in dark areas clearer; when γ > 1, the pixel values ​​in high-brightness areas are compressed to avoid information loss in overly bright areas. Through gamma compression, the overall brightness of the illumination-corrected image can be made more uniform while preserving the image's detail information.

[0039] Step S160: Obtain a dynamic filtering threshold according to the noise filtering scheme, and filter the initial surface image based on the dynamic filtering threshold. The dynamic filtering threshold refers to a filtering threshold dynamically adjusted based on the mapping relationship between the spot diameter and the defect size. Specifically, during image acquisition, various noises, such as electronic noise and thermal noise, are inevitably introduced. These noises can interfere with the detection of wafer surface defects, making the defect features unclear and potentially leading to misjudgments. The noise filtering scheme is used to remove noise from the image, improve image quality, and thus more accurately identify wafer surface defects.

[0040] Dynamic filtering thresholds are dynamically adjusted based on the mapping relationship between spot diameter and defect size. In wafer inspection, there is a certain relationship between spot diameter and defect size. For example, a larger spot diameter may mask smaller defects, while a smaller spot diameter may not accurately detect larger defects. By establishing a mapping model between spot diameter and defect size, a suitable filtering threshold can be dynamically determined based on the spot diameter currently used for inspection. This mapping relationship can be obtained through statistical analysis of experimental data or theoretical derivation based on optical principles. For example, through numerous experiments, the minimum defect size that can be accurately detected under different spot diameters can be recorded, and then a functional relationship between spot diameter and minimum defect size can be fitted. During the inspection process, the corresponding filtering threshold is calculated using this functional relationship based on the actual spot diameter used.

[0041] After obtaining the dynamic filtering threshold, a filtering algorithm is used to process the initial surface image. Filtering algorithms include median filtering, mean filtering, and Gaussian filtering. Taking median filtering as an example, median filtering is a non-linear filtering method. Its basic principle is to sort the gray values ​​of all pixels in the neighborhood of a given pixel in the image, and then take the median value as the new gray value of that pixel. Median filtering can effectively remove impulse noise and salt-and-pepper noise from the image. When filtering based on the dynamic filtering threshold, the parameters of the filtering algorithm can be adjusted according to the filtering threshold, such as adjusting the neighborhood size of the median filter, so that the filtering effect is more adapted to the current image features and detection requirements, thereby preserving as much image detail information as possible while removing noise.

[0042] This approach improves image quality through illumination correction and noise filtering, thereby enhancing the accuracy of subsequent defect detection.

[0043] Step S200: Extract the first image corresponding to the first structure from the target surface image, and construct the first target image based on the multiple grain images in the first image.

[0044] Specifically, the first structure is a specific region on the wafer, such as a chip region or a specific circuit structure region. Image segmentation algorithms are used to extract the image corresponding to this structure. For example, a threshold-based segmentation method first calculates the grayscale histogram of the target surface image. Based on the valleys of the histogram or a preset threshold, the image is divided into foreground (the first structure) and background, thus extracting the first image (the image portion corresponding to the first structure extracted from the target surface image; it is the image representation of the specific structural region in the target surface image). Alternatively, edge detection-based segmentation methods, such as the Canny edge detection algorithm, can be used. This first detects edges in the image, then determines the boundary of the first structure based on the edge information, and then extracts the first image.

[0045] Each grain in the first image can be located and segmented using a template matching method. A template image of a standard grain is pre-stored. This template is then slid across the first image, and the similarity (e.g., cross-correlation coefficient) between the template and a local region of the image is calculated. When the similarity exceeds a preset threshold, the location of a grain is considered found. Each located grain is segmented to obtain a single grain image. These single grain images are then stitched together according to their actual arrangement on the wafer to construct the first target image. For example, if the grains on the wafer are arranged in a matrix, the single grain images are stitched together sequentially according to the row and column order of the matrix to form a complete first target image.

[0046] In one possible implementation, a first image corresponding to the first structure is extracted from the target surface image, and a first target image is constructed based on multiple grain images in the first image. Step S200 further includes step S210, obtaining the first grayscale value of the first grain image among the multiple grain images. Specifically, among the multiple grain images contained in the first image, one grain image (the first grain image) is selected as the starting point for analysis, and its grayscale value is obtained. The grayscale value is a quantitative representation of the brightness of pixels in an image. In a grayscale image, it reflects the brightness of pixels, and its value ranges from 0 to 255. Image processing algorithms, such as traversing all pixels of the first grain image, can be used to calculate the average, median, and other statistical quantities of the grayscale values ​​of these pixels, and these statistical quantities can be used as the first grayscale value of the first grain image.

[0047] Step S220: Obtain the second grayscale value of the second grain image among the plurality of grain images, wherein the second grain image is adjacent to the first grain image. Specifically, select the second grain image adjacent to the first grain image and obtain its grayscale value. By comparing the grayscale values ​​of adjacent grain images, information such as boundary features and texture changes between grains can be analyzed. Similar to obtaining the first grayscale value, first determine the pixel matrix of the second grain image, and then calculate its grayscale value statistics as the second grayscale value.

[0048] Step S230: Establish a grayscale value linked list based on the first grayscale value and the second grayscale value. Specifically, associate the grayscale values ​​of the first and second grain images to form an ordered data structure (grayscale value linked list). The grayscale value linked list is used to record the grayscale value change relationship between grain images. The first and second grayscale values ​​can be stored as nodes in the linked list, while recording their positional relationship (such as adjacency relationship). For example, define a linked list node structure containing grayscale values ​​and pointers to adjacent nodes, and then connect multiple nodes sequentially through pointers to form a grayscale value linked list. If there are multiple grain images in the first image, the grayscale values ​​of adjacent grain images can be obtained sequentially in a certain order (such as from left to right, from top to bottom), and the grayscale value linked list can be continuously expanded.

[0049] Step S240: The grayscale value linked list is iteratively filtered and analyzed according to an iterative filtering mechanism to obtain the first target pixel value. Specifically, by iteratively filtering and analyzing the grayscale value linked list, noise interference and outliers are removed, and pixel values ​​(the first target pixel value) that accurately reflect the first structural features are extracted. The iterative filtering mechanism can be designed with different filtering rules according to specific needs, such as threshold filtering based on grayscale values, filtering based on grayscale value change trends, etc. Assuming a threshold-based iterative filtering mechanism is used, an initial threshold is first set, and then each node in the grayscale value linked list is traversed. For each node's grayscale value, if the difference between its grayscale value and that of its neighboring nodes exceeds the threshold, the node is considered an outlier node and is removed from the linked list or corrected. After one round of filtering, the threshold is readjusted based on the grayscale values ​​of the remaining nodes, and the next round of filtering is performed until the preset iteration termination condition is met (such as the number of iterations reaching the upper limit or the change in the grayscale values ​​of nodes in the linked list becoming stable). Finally, the grayscale values ​​in the filtered linked list are statistically analyzed (e.g., averaged) to obtain the first target pixel value.

[0050] Step S250: Obtain the first target image based on the first target pixel value. Specifically, process the first image using the first target pixel value to construct a first target image that highlights the first structural features. Each pixel in the first image can be reassigned or adjusted according to the first target pixel value. For example, pixels with grayscale values ​​similar to the first target pixel value can be retained or enhanced, while pixels with significant differences can be suppressed or removed. Alternatively, an image segmentation algorithm can be used to segment the first image into different regions with the first target pixel value as a reference, highlighting regions related to the first structure to obtain the first target image.

[0051] This implementation method constructs a first target image based on the pixel values ​​of the first target, which can highlight the key information of the first structure and make the structural features more obvious. Compared with the original image, the first target image can filter out irrelevant background information and noise interference, improving the accuracy and efficiency of image feature extraction.

[0052] In one possible implementation, the grayscale value linked list is iteratively filtered and analyzed according to an iterative filtering mechanism to obtain the first target pixel value. Step S240 further includes step S241, calculating the mean and standard deviation of the grayscale linked list. Specifically, the mean reflects the average level of all grayscale values ​​in the grayscale value linked list, and can reflect the central tendency of grayscale values ​​in the grain image. The standard deviation measures the dispersion of grayscale values ​​relative to the mean. Through the standard deviation, we can understand the fluctuation of grayscale values ​​in the linked list, determine whether the data is concentrated or dispersed, and thus identify possible outliers or noise interference. The mean is obtained by calculating the sum of all grayscale values ​​divided by the number of grayscale values. The standard deviation is obtained by first calculating the square of the difference between each grayscale value and the mean, then calculating the average of these squared differences, and finally taking the square root.

[0053] Step S242: Determine whether the standard deviation is within a predetermined limit, and obtain a determination result. The predetermined limit is contained within the iterative filtering mechanism. Specifically, the predetermined limit is a range pre-set in the iterative filtering mechanism, used to determine whether the dispersion of gray values ​​in the current gray value chain is within a reasonable range. If the standard deviation is within the predetermined limit, it indicates that the distribution of gray values ​​is relatively concentrated, the data is relatively stable, and there may be no obvious outliers or noise interference. If the standard deviation exceeds the predetermined limit, it indicates that the dispersion of gray values ​​is large, and there may be outliers or noise, requiring further filtering of the gray value chain.

[0054] The iterative filtering mechanism stores the lower and upper limits of the predetermined limit. The calculated standard deviation is compared with the lower and upper limits. If the standard deviation is greater than or equal to the lower limit and less than or equal to the upper limit, the result is that the standard deviation is within the predetermined limit; otherwise, the result is that the standard deviation exceeds the predetermined limit.

[0055] Step S243: Based on the judgment result, the grayscale value chain is filtered to obtain candidate pixel values. Specifically, when the judgment result is that the standard deviation exceeds a predetermined limit, it indicates that there are outliers or noise in the grayscale value chain that may affect the accuracy of the result. These interfering factors need to be removed through filtering to obtain more reliable candidate pixel values. Candidate pixel values ​​are the set of grayscale values ​​that are more likely to represent the true features of the first structure after preliminary filtering. Various filtering methods can be used, such as a threshold-based filtering method. A threshold is set, which can be determined based on the mean and standard deviation, such as threshold = mean ± k × standard deviation (where k is a constant that can be adjusted according to the actual situation). Grayscale values ​​within the threshold range in the grayscale value chain are retained as candidate pixel values; values ​​outside this range are considered outliers or noise and removed. Alternatively, a sorting-based filtering method can be used, where the grayscale values ​​in the grayscale value chain are sorted, and a certain proportion of extreme values ​​at both ends are removed, with the remaining grayscale values ​​serving as candidate pixel values.

[0056] Step S244: If the standard deviation meets a predetermined termination threshold, the mean of the candidate pixel values ​​is used as the first target pixel value. Specifically, the predetermined termination threshold is the condition for determining whether the iterative filtering process can end. When the standard deviation meets the predetermined termination threshold, it indicates that after multiple filterings, the gray values ​​in the gray value chain have become relatively stable, and further filtering has little impact on the results. At this time, the mean of the candidate pixel values ​​can be used as the first target pixel value, which can better represent the features of the first structure. The predetermined termination threshold is a small value. After each iterative filtering, the standard deviation of the new gray value chain (or candidate pixel value set) is calculated. If the new standard deviation is less than or equal to the predetermined termination threshold, it is considered that the standard deviation meets the predetermined termination threshold, the iterative filtering process is stopped, the mean of the candidate pixel values ​​is calculated, and this mean is used as the first target pixel value; if the new standard deviation is greater than the predetermined termination threshold, iterative filtering continues, and steps S241-S244 are repeated.

[0057] This implementation method calculates the standard deviation and determines whether it falls within a predetermined limit, enabling timely detection of outliers and noise in the grayscale value chain. Based on the judgment result, the grayscale value chain is filtered to remove these interfering factors, making the candidate pixel values ​​closer to the true grayscale values ​​of the first structure, thus improving the accuracy of target pixel value extraction.

[0058] In one possible implementation, step S243 further includes step S2431, where the predetermined limit includes a first limit and a second limit. When the standard deviation is higher than the first limit, a defect point removal operation is performed. Specifically, the first limit is a relatively high threshold used to determine whether there are obvious outliers in the grayscale value chain. When the standard deviation is higher than the first limit, it indicates that the grayscale values ​​are highly dispersed, with some grayscale values ​​deviating significantly from the overall distribution. These values ​​are likely defect points. The defect point removal operation is performed to remove these outliers, preventing them from interfering with subsequent analysis and allowing candidate pixel values ​​to better reflect the true characteristics of the first structure. Statistical methods can be used to determine defect points. For example, points whose grayscale values ​​satisfy |grayscale value − mean| > k × standard deviation are identified as defect points and removed from the grayscale value chain. Alternatively, a sorting-based method can be used, sorting the grayscale values ​​in the grayscale value chain and removing a certain proportion (determined based on the first limit and the overall distribution) of extreme values ​​at both ends; these extreme values ​​are the defect points.

[0059] Step S2432: When the standard deviation is lower than the second limit, a noise point correction operation is performed. Specifically, the second limit is a low threshold. When the standard deviation is lower than the second limit, it indicates that the grayscale value dispersion is small, but there may be some small fluctuations caused by noise during image acquisition (such as electronic noise, thermal noise, etc.). Although these noise points have a small impact on the overall grayscale value, if they are not processed, they may accumulate errors during multiple iterations of screening, affecting the accuracy of the final first target pixel value. The noise point correction operation is used to smooth these noise points, reduce the impact of noise on the grayscale value chain, and improve the quality of candidate pixel values. The noise point correction operation can use mean filtering, median filtering, etc. Mean filtering takes the average of the grayscale values ​​of the pixels in the neighborhood around each pixel as the corrected grayscale value of that pixel. For example, for a 3×3 neighborhood, the corrected grayscale value of the center pixel is the average of the grayscale values ​​of its eight surrounding pixels and its own grayscale value. Median filtering sorts the gray values ​​of the pixels in the neighborhood around each pixel and takes the median value as the corrected gray value of that pixel.

[0060] Step S2433: Based on the defect point removal operation and the noise point correction operation, the candidate pixel values ​​are obtained. Specifically, after the defect point removal operation and the noise point correction operation, the outliers and noise in the gray value chain are effectively processed, and the resulting gray value set is closer to the true gray value distribution of the first structure. These processed gray values ​​are used as candidate pixel values.

[0061] The degree of defects and noise points varies in images of different qualities. This implementation method, by setting different limits and corresponding operations, can adapt to images of various qualities, thus enhancing the adaptability of the iterative screening process.

[0062] Step S300: Compare the first target image with the first standard image to obtain a first defect image, and obtain the defect identification result of the target wafer based on the first defect image.

[0063] Specifically, the first standard image is a pre-stored, defect-free standard image with a structure corresponding to the first target image. It serves as a comparison benchmark to determine whether defects exist in the first target image. The comparison method can employ image differencing, subtracting the grayscale values ​​of corresponding pixels in the first target image and the first standard image to obtain a difference image. Then, a threshold is applied to the difference image. A threshold is set; if the absolute value of the difference between any pixel in the difference image exceeds the threshold, the region corresponding to that pixel is considered defective and marked as a defective pixel; otherwise, the region is considered defect-free. After processing, a first defect image is obtained, where the region composed of defective pixels is the defective region.

[0064] The defect regions in the first defect map are analyzed to statistically analyze their number, area, shape, and other characteristics. For example, the number of defects is counted using a connected component analysis algorithm, the number of pixels in each connected region (i.e., a defect) is calculated to obtain the defect area, and the shape parameters (such as aspect ratio, roundness, etc.) of the defects are calculated based on the coordinates of the boundary points of the defect regions. These characteristics are compared with preset defect standards to determine the type of defect (such as scratches, particles, voids, etc.), thereby obtaining the defect identification result of the target wafer.

[0065] In one possible implementation, the first target image is compared with the first standard image to obtain a first defect image. Step S300 further includes step S310, sequentially performing inverse pyramid transformation on the first target image and the first standard image to obtain a first target transformed image and a first standard transformed image, respectively. Specifically, pyramid transformation is a multi-scale analysis method used to extract features at different scales in image processing. Inverse pyramid transformation is its inverse process, used to restore an image from different scales to an approximate representation of the original image while preserving detailed information at different scales. Performing inverse pyramid transformation on the first target image and the first standard image is used to synthesize the features of the image at different scales, thereby detecting defects more comprehensively. Because some defects manifest as overall structural differences at large scales, while manifesting as local texture changes at small scales, inverse pyramid transformation can capture these defect features at different scales.

[0066] The inverse pyramid transform employs upsampling and filtering operations. Taking a Gaussian pyramid as an example, its inverse transform process is as follows: Starting from the top layer of the pyramid, the image is upsampled, i.e., the image resolution is increased. Upsampling can be achieved through interpolation methods, such as bilinear interpolation and cubic spline interpolation. These methods estimate the value of the newly inserted pixel based on the values ​​of surrounding pixels. A Gaussian filter is then applied to the upsampled image to smooth it and reduce the jagged edges introduced during upsampling. The kernel function of the Gaussian filter is a Gaussian function, and its magnitude and standard deviation can be adjusted according to specific needs. The filtered image is then fused with the image of the next pyramid layer to obtain the inverse transform image of the current layer. This process is repeated until the resolution of the original image is restored, resulting in the first target transform image and the first standard transform image.

[0067] Step S320: Analyze the first target transformation image and the first standard transformation image using dual generation channels to generate a dual-channel defect map. The dual generation channels include a first channel and a second channel. The first channel generates a first-channel defect map based on the structural similarity weighting principle, and the second channel generates a second-channel defect map based on the 1-cube operation principle of similarity values. The first-channel defect map and the second-channel defect map together constitute the dual-channel defect map. Specifically, structural similarity (SSIM) is an indicator that measures the similarity between two images, considering brightness, contrast, and structural information. Generating the first-channel defect map based on the structural similarity weighting principle can highlight the structural differences between the first target image and the first standard image. Structural differences are an important manifestation of defects, such as object shape deformation and edge loss. By using structural similarity weighting, these structural defects can be located more accurately, and different weights can be assigned to them, making severe structural defects more obvious in the defect map.

[0068] First, the structural similarity of corresponding local regions in the first target transformation image and the first standard transformation image is calculated. For each pixel in the image, a local window (such as 3×3, 5×5, etc.) is selected centered on it. The mean brightness, standard deviation of contrast, and structural correlation coefficient of the images within the two windows are calculated, and then substituted into the structural similarity calculation formula to obtain the structural similarity. Then, a weighted processing is performed based on the structural similarity value. The structural similarity can be inverted or transformed in some way so that regions with lower similarity (i.e., regions with a higher probability of defects) have larger values ​​in the defect image.

[0069] Besides structural similarity, image similarity can be measured from other perspectives. Generating a second-channel defect map based on the 1-similarity value cube operation principle can capture defect features that might be overlooked in structural similarity analysis. The cube operation can amplify differences in similarity values, allowing even minute defects to be reflected in the defect map. For example, subtle differences in color or minor variations in texture can be more clearly displayed after the cube operation.

[0070] First, the similarity values ​​of corresponding pixels in the first target transformation image and the first standard transformation image are calculated. Different calculation methods can be chosen for the similarity values, such as Euclidean distance or cosine similarity, depending on the specific requirements. Then, a subtraction operation is performed on the similarity values, followed by a cube operation, to obtain the second channel defect image. In this way, regions with smaller similarity values ​​(i.e., regions with a higher probability of defect) will have larger values ​​in the second channel defect image, and due to the cube operation, the differences will be further amplified.

[0071] Step S330: The dual-channel defect images are fused to obtain the first defect image. Specifically, the dual-channel defect images detect defects from two perspectives: structural similarity and the cube operation of similarity values, each with different advantages and focuses. By fusing the dual-channel defect images, the information from both channels can be combined to obtain a more comprehensive and accurate defect image. The fused first defect image can simultaneously reflect structural defects as well as defects in color, texture, etc., improving the accuracy and reliability of defect detection.

[0072] A weighted fusion method can be used to fuse the dual-channel defect maps. For example, define the weight of the first channel defect map as w1 and the weight of the second channel defect map as w2, and w1 + w2 = 1. Then the calculation formula for the first defect map is: First Defect Map = w1 × First Channel Defect Map + w2 × Second Channel Defect Map. The weights can be adjusted according to the actual application scenario and the importance of the two channels. If structural defects are more critical, the value of w1 can be increased; if defects such as color and texture are more important, the value of w2 can be increased. Other fusion methods can also be used, such as maximum value fusion, minimum value fusion, etc., and the appropriate fusion strategy should be selected according to specific needs.

[0073] This implementation uses inverse pyramid transformation to restore images to approximate representations at different scales, enabling the detection of defects at various scales. At large scales, it can detect overall structural differences, while at small scales, it can detect local texture variations, thus capturing defect features more comprehensively and reducing the possibility of missed and false detections. The dual-generation channels analyze the image from two perspectives: structural similarity and the cube operation of similarity values, enabling the detection of different types of defects. The first channel highlights structural defects, while the second channel amplifies subtle differences in color, texture, etc. By fusing the dual-channel defect maps, information from both channels can be combined to obtain a more accurate and comprehensive defect map, improving the accuracy of defect detection.

[0074] In one possible implementation, before fusing the dual-channel defect maps to obtain the first defect map, step S300 further includes step S340, which performs frequency domain calibration processing on the first channel defect map based on the first phase spectrum information of the first target map.

[0075] Specifically, during image acquisition, transmission, or processing, the first target image may be affected by various factors, such as noise interference and optical distortion, leading to distortion in the frequency domain. This frequency domain distortion may affect the accuracy of the first channel defect image, causing some defect features to be incorrectly represented or masked. By introducing the first phase spectrum information of the first target image to perform frequency domain calibration on the first channel defect image, this frequency domain distortion can be corrected, making the first channel defect image more realistically reflect the structural differences between the first target image and the first standard image, thereby improving the reliability of subsequent defect detection.

[0076] Specifically, the first step is to perform a Fourier transform on the first target image, converting the image from the spatial domain to the frequency domain. The Fourier transform represents an image as a superposition of different frequency components, where the amplitude spectrum reflects the intensity of each frequency component, and the phase spectrum reflects the phase information of each frequency component. The first phase spectrum information is extracted from the Fourier transform result. The phase spectrum determines the relative position and shape of different parts of the image. Even with the same amplitude spectrum, different phase spectra will result in completely different image content.

[0077] Similarly, a Fourier transform is performed on the first channel defect map to obtain its frequency domain representation, including the amplitude spectrum and phase spectrum. The first phase spectrum information of the first target map is compared with the phase spectrum of the first channel defect map, and the difference between them is analyzed. This difference may be due to frequency domain distortion caused by the various interference factors mentioned earlier. Based on the comparison results, the phase spectrum of the first channel defect map is corrected. Phase correlation techniques can be used to calculate the phase difference between the first target map and the first channel defect map, and a phase correction factor is constructed and applied to the phase spectrum of the first channel defect map to make its phase information closer to the true situation.

[0078] The frequency domain representation of the corrected first channel defect map (including the corrected phase spectrum and the original amplitude spectrum) is subjected to inverse Fourier transform and converted back to the spatial domain to obtain the first channel defect map after frequency domain calibration.

[0079] After frequency domain calibration, the accuracy of the first channel defect map is improved. Then, it is fused with the second channel defect map using the previously described method (step S330). Because the first channel defect map more accurately reflects the structural defect information, the fused first defect map can more comprehensively and accurately reflect the defect situation of the first target map relative to the first standard map, thus improving the accuracy of defect detection.

[0080] This application employs various technical means, including activating an optical detector to acquire an initial surface image of the target wafer, preprocessing the initial image to obtain a target surface image, extracting first image features from the processed image, constructing a first target image for comparison based on these features, comparing the first target image with a preset first standard image, generating a first defect image by analyzing the differences, and outputting the defect identification result of the target wafer based on the defect image. These methods solve the technical problems of insufficient identification accuracy and robustness in existing wafer defect identification, and achieve the technical effect of improving the accuracy and robustness of defect identification.

[0081] In the above text, refer to Figure 1 A wafer defect identification method based on multi-feature fusion according to an embodiment of the present invention is described in detail. Next, reference will be made to... Figure 2 A wafer defect identification system based on multi-feature fusion is described according to an embodiment of the present invention.

[0082] A wafer defect identification system based on multi-feature fusion according to an embodiment of the present invention addresses the technical problems of insufficient accuracy and robustness in existing wafer defect identification methods, thereby improving the accuracy and robustness of defect identification. The wafer defect identification system based on multi-feature fusion includes: a surface image preprocessing module 10, a first target image construction module 20, and a defect identification module 30.

[0083] The surface image preprocessing module 10 is used to activate the optical detector to acquire an initial surface image of the target wafer, and process the initial surface image according to the preprocessing strategy to obtain a target surface image; the first target image construction module 20 is used to extract a first image corresponding to a first structure from the target surface image, and construct a first target image based on multiple grain images in the first image; the defect identification module 30 is used to compare the first target image with a first standard image to obtain a first defect image, and obtain the defect identification result of the target wafer based on the first defect image.

[0084] The specific configuration of the surface image preprocessing module 10 is described in detail below: As mentioned above, the optical detector is activated to acquire an initial surface image of the target wafer. The surface image preprocessing module 10 may further include: an initial three-dimensional model acquisition unit for acquiring an initial three-dimensional model of the target wafer through a 3D defect detection component in the optical detector; a three-dimensional model comparison unit for comparing the initial three-dimensional model with a predetermined three-dimensional model to obtain a three-dimensional comparison result, wherein the three-dimensional comparison result includes pit height difference and protrusion height difference; a 2D defect detection component activation unit for activating the 2D defect detection component in the optical detector if the three-dimensional comparison result meets a predetermined deviation constraint; and an initial surface image acquisition unit for acquiring the initial surface image through the 2D defect detection component.

[0085] The initial three-dimensional model of the target wafer is obtained through the 3D defect detection component in the optical detector. The initial three-dimensional model acquisition unit may further include: a multi-angle wafer image collection subunit for collecting multi-angle wafer images of the target wafer through the 3D defect detection component; and a three-dimensional wafer structure construction subunit for constructing a three-dimensional wafer structure based on the multi-angle wafer images and performing dynamic surface mapping on the three-dimensional wafer structure to obtain the initial three-dimensional model.

[0086] The surface image preprocessing module 10 may further include: a preprocessing strategy construction unit for constructing a preprocessing strategy, the preprocessing strategy including an illumination correction scheme and a noise filtering scheme, wherein the initial surface image is separated according to the illumination correction scheme to obtain a separation result, and the low-frequency components in the separation result are subjected to gamma compression, wherein a dynamic filtering threshold is obtained according to the noise filtering scheme, and the initial surface image is filtered based on the dynamic filtering threshold, and the dynamic filtering threshold refers to a filtering threshold dynamically adjusted based on the mapping relationship between the spot diameter and the defect size.

[0087] The detailed description of the specific configuration of the first target image construction module 20 is explained as follows: As described above, a first image corresponding to the first structure is extracted from the target surface image, and a first target image is constructed based on multiple grain images in the first image. The first target image construction module 20 may further include: a first grayscale value acquisition unit for acquiring the first grayscale value of the first grain image among the multiple grain images; a second grayscale value acquisition unit for acquiring the second grayscale value of the second grain image among the multiple grain images, wherein the second grain image is adjacent to the first grain image; a grayscale value linked list establishment unit for establishing a grayscale value linked list based on the first grayscale value and the second grayscale value; an iterative filtering analysis unit for performing iterative filtering analysis on the grayscale value linked list according to an iterative filtering mechanism to obtain the first target pixel value; and a first target image acquisition unit for obtaining the first target image based on the first target pixel value.

[0088] The iterative filtering analysis unit further includes: a calculation subunit for calculating the mean and standard deviation of the grayscale list; a judgment subunit for judging whether the standard deviation is within a predetermined limit and obtaining a judgment result, wherein the predetermined limit is contained within the iterative filtering mechanism; a filtering subunit for filtering the grayscale list according to the judgment result to obtain candidate pixel values; and a first target pixel value acquisition subunit for using the mean of the candidate pixel values ​​as the first target pixel value if the standard deviation meets a predetermined termination threshold.

[0089] The screening subunit may further include: a defect point removal component for performing a defect point removal operation when the predetermined limit includes a first limit and a second limit; a noise point correction component for performing a noise point correction operation when the standard deviation is lower than the second limit; and a candidate pixel value acquisition component for obtaining the candidate pixel value based on the defect point removal operation and the noise point correction operation.

[0090] The specific configuration of the defect identification module 30 is described in detail below: As mentioned above, the first target image is compared with the first standard image to obtain the first defect image. The defect identification module 30 may further include: an inverse pyramid transformation unit for sequentially performing inverse pyramid transformation on the first target image and the first standard image to obtain the first target transformation image and the first standard transformation image, respectively; a dual generation channel analysis unit for analyzing the first target transformation image and the first standard transformation image through dual generation channels to generate a dual-channel defect image, wherein the dual generation channels include a first channel and a second channel, including: the first channel obtaining the first channel defect image based on the structural similarity weighting principle; the second channel obtaining the second channel defect image based on the 1-similarity value cube operation principle; the first channel defect image and the second channel defect image constitute the dual-channel defect image; and a fusion unit for fusing the dual-channel defect images to obtain the first defect image.

[0091] Before fusing the dual-channel defect maps to obtain the first defect map, the defect identification module 30 may further include: a frequency domain calibration unit for performing frequency domain calibration processing on the first channel defect map based on the first phase spectrum information of the first target map.

[0092] The wafer defect identification system based on multi-feature fusion provided in this embodiment of the invention can execute the wafer defect identification method based on multi-feature fusion provided in any embodiment of the invention, and has the corresponding functional modules and beneficial effects of the method.

[0093] Although this application makes various references to certain modules in the system according to the embodiments of this application, any number of different modules can be used and run on user terminals and / or servers. The various units and modules included are only divided according to functional logic, but are not limited to the above division, as long as the corresponding functions can be achieved; in addition, the specific names of each functional unit are only for easy distinction between each other and are not used to limit the scope of protection of this invention.

[0094] The specific embodiments described above do not constitute a limitation on the scope of protection of this application. Those skilled in the art should understand that various modifications, combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the scope of protection of this application. In some cases, the actions or steps described in this application can be performed in a different order than that shown in the embodiments and still achieve the desired results. Furthermore, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

Claims

1. A wafer defect recognition method under multi-feature fusion, characterized in that, The method comprises: activating an optical detector to acquire an initial surface image of a target wafer, and processing the initial surface image according to a preprocessing strategy to obtain a target surface image; extracting a first image corresponding to a first structure in the target surface image, and constructing a first target image according to a plurality of die images in the first image; comparing the first target image with a first standard image to obtain a first defect map, and obtaining a defect identification result of the target wafer based on the first defect map; extracting a first image corresponding to a first structure in the target surface image, and constructing a first target image according to a plurality of die images in the first image, comprising: obtaining a first gray value of a first die image in the plurality of die images; obtaining a second gray value of a second die image in the plurality of die images, wherein the second die image is adjacent to the first die image; establishing a gray value linked list based on the first gray value and the second gray value; iteratively screening and analyzing the gray value linked list according to an iterative screening mechanism to obtain a first target pixel value; obtaining the first target image based on the first target pixel value; comparing the first target image with a first standard image to obtain a first defect map, comprising: sequentially performing inverse pyramid transformation on the first target image and the first standard image to obtain a first target transformed image and a first standard transformed image, respectively; analyzing the first target transformed image and the first standard transformed image through a double-generation channel to generate a double-channel defect map; fusing the double-channel defect map to obtain the first defect map; wherein the double-generation channel comprises a first channel and a second channel, comprising: the first channel obtains a first channel defect map based on a structural similarity weighted principle; the second channel obtains a second channel defect map based on a cubic operation principle of a 1-similarity value; the first channel defect map and the second channel defect map constitute the double-channel defect map.

2. The wafer defect recognition method under multi-feature fusion of claim 1, wherein, activating an optical detector to acquire an initial surface image of a target wafer, comprising: acquiring an initial three-dimensional model of the target wafer through a 3D defect detection component in the optical detector; comparing the initial three-dimensional model with a predetermined three-dimensional model to obtain a three-dimensional comparison result; if the three-dimensional comparison result meets a predetermined deviation constraint, activating a 2D defect detection component in the optical detector; acquiring the initial surface image through the 2D defect detection component; wherein the three-dimensional comparison result comprises a pit height difference and a protrusion height difference. 3.The wafer defect recognition method under multi-feature fusion of claim 2, wherein, acquiring an initial three-dimensional model of the target wafer through a 3D defect detection component in the optical detector, comprising: collecting a plurality of angle wafer images of the target wafer through the 3D defect detection component; constructing a three-dimensional wafer structure based on the plurality of angle wafer images, and performing dynamic surface mapping on the three-dimensional wafer structure to obtain the initial three-dimensional model.

4. The wafer defect recognition method under multi-feature fusion of claim 2, wherein, The preprocessing strategy comprises an illumination correction plan and a noise filtering plan; wherein the initial surface image is separated according to the illumination correction plan to obtain a separation result, and the low-frequency component in the separation result is gamma compressed. The dynamic filtering threshold is obtained according to the noise filtering plan, and the initial surface image is filtered based on the dynamic filtering threshold, and the dynamic filtering threshold refers to a filtering threshold dynamically adjusted based on a mapping relationship between a spot diameter and a defect size.

5. The wafer defect recognition method under multi-feature fusion of claim 1, wherein, The first target pixel value is obtained by iteratively screening and analyzing the gray value linked list according to an iteration screening mechanism, including: The mean value and the standard deviation of the gray value linked list are calculated respectively; It is judged whether the standard deviation is within a predetermined limit value to obtain a judgment result, wherein the predetermined limit value is stored in the iteration screening mechanism; The gray value linked list is screened according to the judgment result to obtain a candidate pixel value; If the standard deviation meets a predetermined termination threshold, the mean value of the candidate pixel value is taken as the first target pixel value.

6. The wafer defect recognition method under multi-feature fusion of claim 5, wherein, The predetermined limit value includes a first limit value and a second limit value; When the standard deviation is higher than the first limit value, a defect point rejection operation is performed; When the standard deviation is lower than the second limit value, a noise point correction operation is performed; The candidate pixel value is obtained based on the defect point rejection operation and the noise point correction operation.

7. The wafer defect recognition method under multi-feature fusion of claim 1, wherein, Before the first defect map is obtained by fusing the two-channel defect maps, the first-channel defect map is processed in the frequency domain based on the first phase spectrum information of the first target map. 8.A wafer defect recognition system under multi-feature fusion, characterized in that, The system is used to implement the wafer defect recognition method with multi-feature fusion according to any one of claims 1-7, and the system includes: A surface image preprocessing module is configured to activate an optical detector to obtain an initial surface image of a target wafer, and to process the initial surface image according to a preprocessing strategy to obtain a target surface image. A first target map construction module is configured to extract a first image corresponding to a first structure from the target surface image, and to construct a first target map according to a plurality of die images in the first image. A defect recognition module is configured to compare the first target map with a first standard map to obtain a first defect map, and to obtain a defect recognition result of the target wafer based on the first defect map.

Images