Method and system for detecting defects on semiconductor wafers based on machine vision

Abstract

This invention relates to the field of image defect detection technology, specifically to a semiconductor wafer defect detection method and system based on machine vision. The method includes acquiring an original wafer image and preprocessing it; generating candidate regions for suspected defects using the Canny edge detection algorithm based on the preprocessed wafer image; and extracting wafer surface feature vectors based on the candidate regions by constructing a gray-level co-occurrence matrix (GLCM). A reflection compensation mechanism is introduced during the GLCM construction process to reduce the interference of light spots on the image through a reflection compensation module. This invention parameterizes the edge blurring characteristics of different process layers and constructs an adaptive gradient kernel and optimal pixel distance mapping mechanism accordingly, enabling Canny edge detection and GLCM texture analysis to dynamically adapt to surface morphology differences caused by various manufacturing processes.

Description

Technical Field

[0001] This invention relates to the field of image defect detection technology, and more specifically to a semiconductor wafer defect detection method and system based on machine vision. Background Technology

[0002] In semiconductor manufacturing, wafer surfaces are highly susceptible to various micron- and even nanometer-scale defects introduced by complex processes such as photolithography, etching, chemical mechanical polishing (CMP), and thin film deposition. These defects include particle contamination, scratches, bridging, pits, and pattern distortion. If these defects are not detected in time, they will directly affect the electrical performance, yield, and even reliability of the chip. Therefore, high-precision and high-reliability image inspection technology has become an indispensable quality control link in modern semiconductor manufacturing. Currently, mainstream wafer defect image inspection methods are mainly divided into two categories: traditional methods and emerging algorithms. Traditional methods, such as manual visual inspection or image comparison based on fixed rules, are essentially first-generation image inspection technologies, and their detection efficiency heavily relies on operator experience or preset templates. Faced with high-resolution, large-area, and multi-process layer stacked wafer images of advanced processes, these methods have significant bottlenecks in detection efficiency, false negative rate, and generalization ability, making it difficult to meet mass production requirements. As process nodes continue to shrink, defect morphologies become increasingly complex and highly similar to normal circuit structures, placing almost stringent requirements on the sensitivity, specificity, and anti-interference capabilities of image inspection algorithms. Especially in real-world industrial scenarios, wafer surfaces often exhibit strong reflections, non-uniform illumination, or multi-textured backgrounds, resulting in images containing a significant amount of interference. Under such complex imaging conditions, conventional image detection processing is highly susceptible to interference from light spots, process noise, or linear artifacts, leading to frequent false positives and false negatives, thus becoming a core obstacle to improving the reliability of image detection. Summary of the Invention

[0003] The purpose of this invention is to address the problems existing in the background technology by proposing a semiconductor wafer defect detection method and system based on machine vision.

[0004] The technical solution of this invention: A semiconductor wafer defect detection method based on machine vision, comprising: S1. Acquire the original wafer image and preprocess the original wafer image; S2. Based on the preprocessed wafer image, the Canny edge detection algorithm is used to generate candidate regions for suspected defects. S3. Extract wafer surface feature vectors by constructing a gray-level co-occurrence matrix based on suspected defect candidate regions; Among them, a reflection compensation mechanism is introduced in the process of constructing the gray-level co-occurrence matrix. The reflection compensation module in the reflection compensation mechanism reduces the interference of light spots on the image, and the directional mask matrix is ​​used to reduce the impact of linear scratches on texture features. S4. Based on the feature vectors of the wafer surface, a convolutional neural network model is used to classify and identify candidate regions of suspected defects, and the severity level of defects is evaluated according to the defect size, grayscale deviation and distribution location.

[0005] As a further improvement to this technical solution, in step S2, the suspected defect candidate region is generated based on the preprocessed wafer image using the Canny edge detection algorithm, including the following steps: S2.1 Set edge blur width parameters for different process layers in semiconductor manufacturing and construct an adaptive gradient kernel; S2.2. Convolve the preprocessed wafer image using an adaptive gradient kernel to calculate the gradient intensity and direction of each pixel in the preprocessed wafer image; S2.3. Perform non-maximum suppression on each pixel based on gradient intensity and direction to generate edge pixels; S2.4. Use two thresholds to classify edge pixels; S2.5 Based on the edge pixel classification results, the edge pixels are aggregated into regions by the connected component analysis method to represent suspected defect candidate regions. The suspected defect candidate regions are then filtered, and the final suspected defect candidate regions are output.

[0006] As a further improvement to this technical solution, in step S2.1, setting the edge blur width parameter for different process layers and constructing an adaptive gradient kernel includes the following steps: S2.11. Based on the wafer's process node and layer type, preset different edge blur width parameters. ; S2.12, Based on the edge blur width parameter of each layer Calculate the standard deviation of the Gaussian kernel; S2.13, Generate edge profile model based on the standard deviation of Gaussian kernel; S2.14. Calculate the first derivative of the edge profile model to obtain the final adaptive gradient kernel.

[0007] As a further improvement to this technical solution, in step S3, the wafer surface feature vector is extracted by constructing a gray-level co-occurrence matrix based on the suspected defect candidate region, including the following steps: S3.1 Based on the suspected defect candidate regions output in step S2.5, they are cropped into individual small patches, and homomorphic filtering preprocessing is performed on the small patches; S3.2. Construct a reflection compensation mechanism based on small tiles. Reduce the interference of light spots on the image through the reflection compensation module in the reflection compensation mechanism, and reduce the impact of linear scratches on texture features by using a directional mask matrix. S3.3, adjust the edge blur width parameter The optimal pixel distance is mapped to the gray-level co-occurrence matrix. Based on the optimal pixel distance, the gray-level co-occurrence matrix is ​​constructed, and each small patch is analyzed to extract the texture feature vector of semiconductor wafer defects. This includes reflectance-corrected contrast, reflectance-invariant correlation, and scale-perceived homogeneity. S3.4 Perform morphological analysis on each small patch to extract the shape feature vector of semiconductor wafer defects. It should include at least the area, perimeter, and circularity; S3.5. By statistically analyzing the grayscale values ​​of pixels within small image patches, extract the grayscale feature vector of semiconductor wafer defects. It should include at least the mean, variance, and skewness; S3.6, Transfer the texture feature vector Shape feature vector and grayscale feature vector Combine them to generate wafer surface feature vectors. .

[0008] As a further improvement to this technical solution, in step S3.2, the interference of light spots on the image is reduced by the reflection compensation module in the reflection compensation mechanism, and the influence of linear scratches on texture features is reduced by the directional mask matrix, including the following steps: S3.21. Use a Gaussian difference filter to process each small patch and extract the low-frequency components of the reflection part; S3.22 Calculate the reflection normalization parameters for each pixel based on the reflection component. Based on reflection normalization parameters For small tiles Perform normalization; S3.23, Set the threshold for the reflection area; S3.24. Use Gabor filter banks to detect the dominant texture direction in small tiles; S3.25. Based on the difference between the local gradient direction and the dominant texture direction in the small patch, construct a direction-sensitive masking matrix. It is used to weight and suppress pixel regions that deviate from the dominant texture direction.

[0009] As a further improvement to this technical solution, in step S3.3, a gray-level co-occurrence matrix is ​​constructed based on the optimal pixel distance, and each small image patch is analyzed to extract the texture features of semiconductor wafer defects, including the following steps: S3.31, adjust the edge blur width parameter Distance to the optimal pixel in each small tile Perform mapping; S3.32. For each small image patch, use the optimal pixel distance obtained by mapping through the edge blur width parameter. Calculate the gray-level co-occurrence matrix; S3.33. Based on the gray-level co-occurrence matrix, extract texture features from each small patch, including at least reflection-corrected contrast, reflection-invariant correlation, and scale-aware homogeneity.

[0010] As a further improvement to this technical solution, in S3.33, the reflection correction contrast is extracted using a weighted modified gray-level co-occurrence matrix contrast formula, and a dynamic weighting factor is introduced into the gray-level co-occurrence matrix contrast formula. Reflection-invariant correlation is extracted based on small image patches preprocessed by homomorphic filtering using the standard gray-level co-occurrence matrix correlation formula. Scale-aware homogeneity is extracted using multi-scale fusion gray-level co-occurrence matrix homogeneity features.

[0011] As a further improvement to this technical solution, the dynamic weighting factor The calculation process is as follows: based on the gray-level co-occurrence matrix, traverse the small image tiles. All directions satisfy and optimal pixel distance pixel pairs Based on each pixel pair Obtain its grayscale value Based on grayscale values The pixel pair is calculated using the normalized joint intensity component, gradient suppression component, and phase consistency component. Corresponding reflection confidence level Based on reflection confidence Calculate dynamic weighting factors .

[0012] As a further improvement to this technical solution, in step S4, the suspected defect candidate regions are classified and identified based on the wafer surface feature vector using a convolutional neural network model, and the defect severity level is evaluated according to the defect size, grayscale deviation, and distribution location, including the following steps: S4.1 Standardize the candidate regions of suspected defects based on the feature vectors of the wafer surface; S4.2 Input the standardized suspected defect candidate regions into the pre-trained convolutional neural network model for defect classification and identification, and generate wafer defect regions; S4.3 Calculate the pixel size of each wafer defect region by binarizing the wafer defect region; S4.4 Calculate the grayscale deviation between the defective region and the normal region of the wafer, including the difference in mean grayscale and the standard deviation of grayscale. S4.5. Based on the overall geometric information of the semiconductor wafer, calculate the relative distance between the wafer defect region and the wafer center; S4.6. Based on pixel size, grayscale deviation, and relative distance, a comprehensive severity assessment value is constructed using a weighted method. S4.7. Assess the severity level of the defect based on the comprehensive severity assessment value.

[0013] On the other hand, the present invention provides a semiconductor wafer defect detection system based on machine vision, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the semiconductor wafer defect detection method based on machine vision described above.

[0014] Compared with the prior art, the above-mentioned technical solution of the present invention has the following beneficial technical effects: By parameterizing the edge blurring characteristics of different process layers (such as metal layers, via layers, and isolation layers), and constructing an adaptive gradient kernel and optimal pixel distance mapping mechanism accordingly, the Canny edge detection and gray-level co-occurrence matrix (GLCM) texture analysis can dynamically adapt to the surface morphology differences caused by various manufacturing processes; at the same time, the introduction of a reflection compensation mechanism and a directional mask matrix in the GLCM construction process effectively suppresses the interference of light spot interference and linear scratches on texture statistics, significantly improving the accuracy and stability of defect texture feature extraction in the image detection process under complex lighting and multi-process scenarios, and making the overall image detection system more adaptable to processes and robust to interference. Attached Figure Description

[0015] Figure 1 This is a flowchart of the overall method of the present invention. Detailed Implementation

[0016] Example 1: Please refer to Figure 1 As shown, this embodiment provides a semiconductor wafer defect detection method based on machine vision, including the following steps: S1. Acquire raw wafer images (using a high-resolution industrial camera combined with coaxial light, dark field light, or multispectral illumination devices to scan and image the wafer surface area by area, forming raw wafer images; calibrating the imaging system using a calibration board or a reference pattern of known size to obtain image resolution parameters). (Unit: pixels / micrometer, i.e., the actual pixel size represented by each pixel), and preprocess the original wafer image (preprocessing steps include noise reduction, illumination equalization, contrast enhancement, etc.); S2. Based on the preprocessed wafer image, the Canny edge detection algorithm is used to generate candidate regions for suspected defects. In this embodiment, the Canny edge detection algorithm is used to generate candidate regions for suspected defects based on the preprocessed wafer image, including the following steps: S2.1 During the wafer defect detection process, edge blur width parameters are set for different process layers in semiconductor manufacturing (such as metal layers (strong CMP effect), via layers (etching process), isolation layers (STI filling), etc.) to construct an adaptive gradient kernel; The adaptive gradient kernel is specifically designed to address the issue of decreased sensitivity and accuracy of traditional fixed gradient operators when dealing with defect edges with varying fuzziness characteristics caused by different process layers (such as metal layers, via layers, and isolation layers) in semiconductor manufacturing. Since different process steps (such as chemical mechanical polishing (CMP), etching, and filling) predictably affect the microstructure of the wafer surface, resulting in systematic differences in the sharpness and transition width of defect edges, this method injects prior process knowledge (in the form of fuzziness width parameters) into the initial stage of edge detection, generating a gradient kernel that adapts to different edge contours. This ensures the subsequent algorithm's ability to capture real defects under various process backgrounds. This invention quantifies and models the inherent edge fuzziness effects of different process layers, dynamically generating a matching adaptive gradient kernel. This allows gradient calculation to more accurately respond to the intensity changes of real defect edges, while effectively suppressing non-defective gradient regions or noise caused by the process itself. This fundamentally improves the robustness and accuracy of the edge detection step across multiple process scenarios, providing more reliable and less noisy candidate regions for subsequent feature extraction and defect classification. To set edge blur width parameters for different process layers and construct an adaptive gradient kernel, the following steps are included: S2.11. Based on the wafer's process node and layer type (such as metal layer, via layer, isolation layer, etc.), and using process knowledge or historical data, determine the typical edge blur physical width for each process layer. (Unit: nm), preset different edge blur width parameters (using image resolution parameters) (pixels / nm), converting physical width to pixel values: These values ​​are used to determine the impact of different processes on the edges: In this embodiment, each process layer (such as metal layer (strong CMP effect), via layer (etching process), isolation layer (STI filling), etc.) has a different impact on the edges, therefore, different edge blur width parameters need to be set for each layer; edge blur width parameters This indicates the width affecting edge transition, in pixels. For example, metal layers (with a strong CMP effect) will have blurred edges due to chemical mechanical polishing (CMP), requiring a larger blur width. Pixels; via layer (etching process): Due to the influence of the etching process, the edges are relatively sharp, but some blurring still needs to be considered. Pixels; the filling process of the isolation layer (STI fill) causes significant edge blurring, thus requiring a larger blur width: Pixel; S2.12, Based on the edge blur width parameter of each layer Calculate the standard deviation of the Gaussian kernel ; S2.13, Standard deviation based on Gaussian kernel An edge profile model is generated (this model, through the properties of the error function, makes the transition from blurred to sharp edges smoother, avoiding the generation of hard boundaries, and can more realistically reflect the edge gradient effect caused by processing in the image, thus providing accurate edge information for subsequent edge detection and feature extraction), using the error function ( The gradient characteristics of the edge are described by ); the edge profile model is represented as: (In the formula, The pixel values ​​or image grayscale values ​​in the edge profile model vary with position. It changes with the changes. This represents the pixel position from the edge, which is the horizontal or vertical distance from the edge center to the edge. These are the initial values ​​(constants) for the edge model. (A constant used to control the scale of grayscale changes in the edge transition region). S2.14. Calculate the first derivative of the edge profile model to obtain the final adaptive gradient kernel; the gradient kernel will simulate the transition characteristics of different process edges, forming gradient kernels in the horizontal and vertical directions. Specifically: First, the model about To find the derivative, use the derivative formula of the error function: ; Therefore, the first derivative of the edge profile for: ; Then, the first derivative As the core of the adaptive gradient kernel, it is used to describe the rate of change of edges in an image; the shape of this adaptive gradient kernel is closely related to the transition characteristics of the edges, and is adjusted... and The value of can flexibly control the magnitude and width of the gradient kernel, making it adaptable to different process layers or edge features; ultimately, this adaptive gradient kernel can be applied to image convolution to extract edge information in the image, especially to describe the changing characteristics of gradient regions. S2.2. Convolve the preprocessed wafer image using an adaptive gradient kernel to calculate the gradient intensity and direction of each pixel in the preprocessed wafer image; specifically: select an adaptive gradient kernel (horizontal and vertical directions) generated according to different process layers, and convolve it with the preprocessed wafer image; the convolution process obtains the horizontal gradient of each pixel by weighting the adaptive gradient kernel in each pixel of the image. ) and vertical gradient ( Then, calculate the gradient intensity and direction for each pixel: gradient strength Calculated using the following formula: ; gradient direction Then it is calculated using the following formula: ; For each pixel, its gradient intensity (reflecting edge intensity) and direction (reflecting edge directionality) are obtained. S2.3. Non-maximum suppression is performed on each pixel based on gradient intensity and direction to generate edge pixels: First, for each pixel, it is assigned to one of four main directions (0°, 45°, 90°, 135°) using the gradient direction, and the neighboring pixels of the pixel in the gradient direction are determined; then, according to the gradient direction, two pixels adjacent to the current pixel are selected along the direction, and the gradient intensity of the current pixel is compared with the gradient intensity of these two neighboring pixels. If the gradient intensity of the current pixel is greater than the gradient intensity of the two neighboring pixels, the pixel is retained as an edge pixel; otherwise, the intensity of the pixel is suppressed to zero (i.e., the pixel is not an edge). This process ensures the continuity of edges by comparing each pixel, while removing noise and false edges in the image, making the final edge more accurate and detailed; S2.4. Edge pixels are classified into three categories using two thresholds (high threshold and low threshold): Strong edges: pixels with edge intensity greater than the high threshold are directly retained as edges; Weak edges: pixels with edge intensity between the high and low thresholds are weak edges; Non-edges: pixels with edge intensity lower than the low threshold are considered noise and directly removed. Weak edges are connected: if a weak edge pixel is adjacent to a strong edge pixel, it is considered an edge; otherwise, it is suppressed. This step ensures edge continuity and removes discontinuous noise edges. The high threshold value ranges from 0.1 to 0.3, determined experimentally; the low threshold value ranges from 0.05 to 0.15, determined experimentally. S2.5. Based on the edge pixel classification results, edge pixels are aggregated into regions using connected component analysis methods (such as 8-neighborhood connection). (First, each edge pixel in the image is treated as an independent point, and its 8-neighborhood (i.e., the 8 adjacent pixels around the current pixel) is checked. For each edge pixel, it is checked whether there are other pixels classified as edges in its neighborhood. If so, these adjacent edge pixels are grouped into the same connected region. By continuously traversing the entire image, all connected edge pixels are merged into one region, ultimately generating multiple connected regions (i.e., edge regions). In this process, connectivity rules (such as 8-neighborhood connection) are used to ensure that adjacent edge pixels in the horizontal, vertical, and diagonal directions are correctly aggregated into the same region.) Adjacent edge pixels are connected together to form larger regions, representing suspected defect candidates. Select regions and filter suspected defect candidate regions (first calculate the geometric features of each region, including area, aspect ratio, perimeter, and roundness, etc. Based on the set threshold, exclude regions with too small an area, too large an aspect ratio, or low roundness, as these are usually noise or do not meet the defect characteristics. For example, exclude regions with an area less than 20 pixels, an aspect ratio greater than 5:1, or a roundness less than 0.3; then, further filter each region using edge strength and grayscale features. For example, by calculating the mean and standard deviation of the grayscale of pixels within the region, filter out regions with small grayscale fluctuations. The grayscale standard deviation of the candidate region must be higher than 1.5 times the standard deviation of its surrounding normal region, or the absolute value must be greater than 10 (within the grayscale range of 0-255). These are usually non-defect regions), and output the final suspected defect candidate regions after filtering.

[0017] S3. Extract wafer surface feature vectors by constructing a gray-level co-occurrence matrix based on suspected defect candidate regions; Among them, a reflection compensation mechanism is introduced in the process of constructing the gray-level co-occurrence matrix. The reflection compensation module in the reflection compensation mechanism reduces the interference of light spots on the image, and the directional mask matrix is ​​used to reduce the impact of linear scratches on texture features. In this embodiment, the wafer surface feature vector is extracted by constructing a gray-level co-occurrence matrix based on the suspected defect candidate region, including the following steps: S3.1 Based on the suspected defect candidate regions output in step S2.5, extract the corresponding parts in the image, crop them into individual small patches, and perform homomorphic filtering preprocessing on the small patches; these small patches will be used as input for subsequent feature extraction to avoid interference from other irrelevant regions. In order to eliminate low-frequency interference caused by uneven reflection on the wafer surface or poor illumination, uniform standardization preprocessing is performed on the small patches, specifically including: applying homomorphic filtering to decompose the image into illumination components and reflectivity components, and obtaining standardized small patches with enhanced texture details by suppressing low-frequency illumination components and enhancing high-frequency reflectivity components; or selecting local zero-mean normalization processing under specific processes. This standardized small image patch will serve as the sole input source for all subsequent texture, shape, and grayscale feature extractions. The homomorphic filtering preprocessing, to ensure that correlated features are not interfered with by reflective regions, involves reflection correction of the calculation results. This means that before calculating the above formula, a homomorphic filtering preprocessing operation is performed on the small image patch. The principle of this operation is to treat the image as the product of an "illumination component" (caused by reflection from the wafer surface) and a "reflectivity component" (representing the inherent texture of the object). Homomorphic filtering separates these two components in the frequency domain, suppressing the slowly changing illumination (reflection) component while enhancing the rapidly changing texture component. S3.2. Construct a reflection compensation mechanism based on small tiles. Reduce the interference of light spots on the image through the reflection compensation module in the reflection compensation mechanism, and reduce the impact of linear scratches on texture features by using a directional mask matrix. In this embodiment, the reflection compensation mechanism and directional mask matrix are used to address the specific problem of strong reflective spots and linear scratches on the semiconductor wafer surface, caused by the material's reflective properties and contact with process machinery, severely interfering with and masking the true texture features of defects. These interferences distort or obscure the true representation of defect areas in statistical features such as the gray-level co-occurrence matrix (GLCM), leading to misclassification or underclassification by subsequent classifiers. The reflection compensation mechanism is specifically designed to address the interference from strong light spots (highlight areas) on the surface of semiconductor wafers caused by material reflection and ambient lighting. Its core idea is to separate and suppress the reflection component (brightness changes caused by illumination but unrelated to defects) in the image from the true texture information, thereby preventing the light spot region from distorting texture statistical features (such as the gray-level co-occurrence matrix) and improving the accuracy of defect identification. Directional masking matrices are designed to address linear scratch-like interference commonly found on wafer surfaces. By analyzing the dominant texture direction in local image regions and constructing a weight matrix, they automatically suppress the contribution of linear structures (i.e., scratches) that are inconsistent with the dominant direction to texture feature calculation, thereby preventing scratches from being misidentified as defect textures or interfering with the extraction of true defect features. This invention achieves "physical property decoupling" and "directional adaptive suppression" of interference sources. It models and extracts low-frequency reflection components related to illumination at the image signal level through Gaussian difference (DoG) and reflection gradient normalization, rather than crudely suppressing all bright areas. This reduces light spot interference while preserving the underlying true texture as much as possible. Simultaneously, through Gabor filtering and direction-sensitive masking, it intelligently identifies and suppresses linear scratch features that do not conform to the dominant texture direction, preventing them from being mixed into the statistical feature vector. This two-pronged refined processing significantly improves the purity and robustness of extracting essential texture features of defects in complex backgrounds. The process involves reducing the interference of light spots on the image through the reflection compensation module in the reflection compensation mechanism, and reducing the impact of linear scratches on texture features using a directional mask matrix. This includes the following steps: S3.21. Use a Difference of Gaussian (DoG) filter to process each small tile and extract the low-frequency components of the reflection to distinguish them from the actual texture: ; in, This is the standard deviation of the Gaussian filter, used to adjust the filter's bandwidth when processing low-frequency reflection information. To represent the low-frequency portion of a Gaussian filter, The grayscale value of the small image patch. This represents the reflection component of a small image patch. The reflection portion after filtering reflects low-frequency reflection information. S3.22. Since the reflection portion often leads to a high reflection gradient in small patches, which adversely affects the analysis of small patches, the reflection normalization parameter of each pixel is calculated based on the reflection portion. This parameter is used to perform weighted suppression on regions with large reflection gradients: ; Based on reflection normalization parameters For small tiles Normalize: ; In the formula, To adjust the reflection gradient with respect to the normalized parameters The parameter affecting the degree of influence, ranging from 0.1 to 1.0, adjusts the suppression strength of the reflection gradient on normalization, and is optimized through grid search. Reflective component The squared gradient measures the rate of change of the reflected portion, reflecting the intensity or gradient change of the reflection. For small tiles The mean, The standard deviation of the small plot; S3.23, Set the threshold of the reflection area (The aim is to extract highly reflective areas from the image by comparing them with the average grayscale value of the wafer image, and then perform reflection compensation.) If the area is a reflective region, then it is considered to be a reflective region (in this embodiment, (40% of the average gray level of the wafer). When performing the gray-level co-occurrence matrix (GLCM) calculation, the weights of these pixel pairs are halved to reduce the impact of reflective areas on texture features. S3.24. Detecting the dominant texture direction in small patches using a Gabor filter bank; specifically: First, select a set of Gabor filters with different directions and scales. These filters contain different frequency and directional information, capable of capturing texture features in the image; then, convolve each Gabor filter with the small patch to obtain response images at multiple directions and scales; these response images reflect the texture information of the image in different directions; next, calculate the response intensity in each direction, and select the direction with the largest response intensity as the dominant texture direction of the small patch; this dominant direction represents the most significant texture feature within the small patch, which is helpful for subsequent texture analysis and feature extraction; through this method, directional textures in images can be effectively identified and described, especially when processing wafer defect images, which helps improve the accuracy of texture feature extraction; S3.25. Based on the difference between the local gradient direction in the small patch (calculated using a gradient operator (such as the Sobel operator) and the dominant texture direction), construct a direction-sensitive masking matrix. (Direction-sensitive masking matrix) Used to characterize pixels The degree of consistency between the local texture direction and the dominant texture direction of the target region (used to weight and suppress pixel regions that deviate from the dominant texture direction); masking matrix. The value of the masking matrix is ​​adjusted based on the difference between the local gradient direction and the dominant texture direction. When the local gradient direction is very close to the dominant texture direction (i.e., the difference between the local gradient direction and the dominant texture direction is less than the difference threshold m (m is 15° in this embodiment)), the value of the masking matrix is ​​1 (i.e., no suppression is performed), indicating that the region belongs to the dominant texture of the image. When the difference between the two is large (i.e., the difference between the local gradient direction and the dominant texture direction is greater than the difference threshold m), the value of the masking matrix is ​​close to 0 (i.e., strong suppression is performed), indicating that the region is a noise or interference region that deviates from the dominant texture direction. As a weighted matrix, it applies weighted suppression to regions that do not conform to the dominant texture direction in subsequent texture calculations (such as GLCM). This ensures that texture analysis mainly focuses on regions that conform to the dominant texture direction, while suppressing regions that do not conform to the direction (such as linear scratches, reflections, and other noise), reducing the impact of these regions on texture statistical features. S3.3, blur the edge width parameter of different process layers. The optimal pixel distance is mapped to a gray-level co-occurrence matrix (GLCM). A GLCM is constructed based on this optimal pixel distance, and each small patch is analyzed to extract texture feature vectors of semiconductor wafer defects. This includes reflectance-corrected contrast, reflectance-invariant correlation, and scale-perceived homogeneity. Furthermore, a gray-level co-occurrence matrix is ​​constructed based on the optimal pixel distance, and each small patch is analyzed to extract the texture features of semiconductor wafer defects, including the following steps: S3.31, Extract the edge blur width parameter of each process layer. Distance to the optimal pixel in each small tile Perform mapping; specifically: apply a non-linear mapping function to the edge blur width parameter. Dynamically converted to the optimal pixel distance for texture analysis of each small tile. ; S3.32. For each small image patch, use the optimal pixel distance obtained by mapping through the edge blur width parameter. Calculate the Gray-Level Co-occurrence Matrix (GLCM); specifically, for each small patch, calculate the offsets that satisfy the following conditions in the horizontal (0°), vertical (90°), and diagonal (45°, 135°) directions. , , and All pixel pairs The gray-level co-occurrence frequency (where, Location in the image The pixel value (grayscale value) at that location. Location in the image The pixel value (grayscale value) at that location; and These are the pixel offsets in the horizontal and vertical directions, respectively. The statistical method is as follows: for each pair that satisfies the offset... , , and The pixel pairs, based on the consistency of the two pixels with the dominant texture direction, are determined according to the directional mask matrix. Define its weighted contribution value ( This represents the offset pixels that have a co-occurrence relationship with it. (The degree of consistency of the dominant texture direction of the pixel); this weighted contribution value. Used to modulate the co-occurrence statistical intensity of the corresponding gray-level pair in the gray-level co-occurrence matrix: when both pixels are highly consistent with the dominant texture direction (when the directional consistency of the two pixels in the pixel pair is in the high consistency range (e.g.) When a pixel pair is considered to be highly consistent with the dominant texture direction, its gray-level co-occurrence frequency is accumulated with a large weight; when any pixel deviates from the dominant direction, the accumulated contribution of its corresponding gray-level co-occurrence frequency will be continuously weakened; the weighted contribution values ​​of all pixel pairs are calculated as follows. The gray values ​​are accumulated to obtain weighted gray-level co-occurrence matrices in four directions. The matrices in each direction are then normalized (so that the sum of the matrix elements is 1) to obtain the final gray-level co-occurrence matrix used for texture feature extraction. S3.33. Based on the gray-level co-occurrence matrix (GLCM), extract texture features from each small patch, including at least reflection-corrected contrast, reflection-invariant correlation, and scale-aware homogeneity. The reflection-corrected contrast ratio is extracted using a weighted modified gray-level co-occurrence matrix (GLCM) contrast ratio formula, and a dynamic weighting factor is introduced into the GLCM contrast ratio formula. Reflection-invariant correlation is extracted based on small image patches preprocessed with homomorphic filtering using the standard gray-level co-occurrence matrix (GLCM) correlation formula (since the input image has already undergone homomorphic filtering in step S3.1 to remove slowly changing illumination (reflection) interference, the extracted correlation features are reflection-invariant). Scale-aware homogeneity is extracted using multi-scale fusion gray-level co-occurrence matrix (GLCM) homogeneity features (constructing a scale set). A weighted fusion strategy was adopted, which not only preserved the core scale information based on the optimal pixel distance, but also integrated the texture features of adjacent scales, thereby enhancing the detection sensitivity of defects of different sizes. Furthermore, the contrast ratio is calculated based on the gray-level co-occurrence matrix. This contrast ratio measures the drastic change in pixel grayscale; to reduce the interference of reflection on texture, a reflection correction method (i.e., introducing a dynamic weighting factor) is introduced. The effects of reflection can be eliminated by reducing the weight of the reflective region or by performing image normalization. ; In the formula, To correct contrast for reflection, the intensity of grayscale changes within small image tiles is measured. Elements of the Gray-Level Co-occurrence Matrix (GLCM) represent gray-level values. and The frequency of pixel pairs, that is, all pixels in an image patch that satisfy a specified spatial relationship (i.e., direction is...) The optimal pixel distance is In a pixel pair (i.e., pixel interval), the grayscale value of the previous pixel is... The grayscale value of the next pixel is The joint probability (or normalized frequency) of the pairings occurring, here It has already been processed by a directional mask matrix. The weighted probability distribution ensures that the contribution of pixel pairs deviating from the dominant texture direction (such as linear scratches) to contrast features is effectively suppressed. This represents the grayscale value of a pixel in the Gray-Level Co-occurrence Matrix (GLCM). This represents the grayscale value of another pixel in the Gray-Level Co-occurrence Matrix (GLCM). This is a dynamic weighting factor, representing an adjustment of the weights of each pixel pair in the gray-level co-occurrence matrix; Furthermore, dynamic weighting factors The calculation process is as follows: based on the Gray-Level Co-occurrence Matrix (GLCM), traverse the small image tiles. All directions satisfy and optimal pixel distance pixel pairs Based on each pixel pair Obtain its grayscale value Based on grayscale values The pixel pair is calculated using the normalized joint intensity component, gradient suppression component, and phase consistency component. Corresponding reflection confidence level Based on reflection confidence Calculate dynamic weighting factors (In the formula, (This is an adjustment parameter for the influence of reflection confidence on the weighting factor, with a value ranging from 0.2 to 0.8, determined through expert experience). Among them, the confidence level of reflection It is composed of a weighted fusion of three local feature components: normalized joint intensity component, gradient suppression component, and phase consistency component, which capture the characteristics of reflection from different perspectives. The confidence level of the fused reflections is expressed by the following formula: ; In the formula, To normalize the joint strength components, For gradient suppression components, For phase-consistent components, To normalize the weights of the joint strength components, The weights of the gradient suppression components, The weights of the phase consistency components, , , The values ​​of are all in the range of 0 to 1, determined through experimental calibration. ; Reflective areas are typically brighter, therefore the likelihood of reflection is closely related to the brightness of the area. To reduce the impact of different lighting conditions, the grayscale values ​​of pixels were normalized. The joint intensity component was normalized. Calculated using the following formula: ; In the formula, It is a location The grayscale values, after normalization or global contrast stretching, range from [0,1]. It is a location The grayscale values ​​are normalized or global contrast stretched, and the range is [0,1]. In this way, it can adapt to different lighting conditions more stably, rather than relying solely on the brightness of a single pixel. Reflective regions weaken texture features, resulting in reduced gradients (texture variations) within them; conversely, real texture regions have stronger gradients and exhibit distinct boundaries. Therefore, a gradient suppression component is constructed. To measure pixel pairs Gradient strength: ; in, and They are the locations and The gradient magnitude (calculated using the Sobel operator). The gradient threshold is calculated using the gradient magnitude histogram of the image (e.g., using the top 20% quantile as the threshold); if pixel pairs If the gradients are very weak, it indicates that the region is a uniform reflection area. The gradient approaches 1; if the gradient is strong, it indicates a texture edge. Approaching 0; The edges of natural textures have consistent phase characteristics, while the edges of "highlights" caused by reflections are often inconsistent or broken in phase with the texture boundaries; therefore, a phase consistency component is introduced. To measure pixel pairs Phase difference between them: ; in, and They are the locations and The phase consistency measurement value (calculated using the classic frequency domain phase consistency method). This is a sensitivity parameter used to adjust the sensitivity to phase differences, with a value ranging from 1.0 to 3.0, determined through expert experience; when the pixel pair When they are both within a uniform reflection region, their phase characteristics tend to be consistent. (small value) The value approaches 1, which is important for the confidence level of reflection. It produces a high positive contribution; when pixel pairs are located at the boundary between reflective and non-reflective regions, their phase consistency differs significantly. (High value) The value approaches 0, thus reducing This value helps to distinguish the region of phase continuity (the reflection zone) from the boundary of a phase abrupt change; After calculating the reflection-corrected contrast, the correlation in the gray-level co-occurrence matrix is ​​calculated. To ensure that the correlation features are not affected by the reflection region, reflection correction is applied to the calculation results. The reflection-invariant correlation emphasizes the actual correlation of the texture under the influence of reflection, and the calculation formula is as follows: ; In the formula, The reflection-invariant correlation (achieved through preprocessing (homomorphic filtering)) measures the correlation of small tile textures. grayscale value The mean, grayscale value The mean, grayscale value standard deviation grayscale value Standard deviation; To capture multi-scale texture features in small image patches, homogeneity (i.e., a multi-scale fusion strategy) is calculated using gray-level co-occurrence matrices (GLCMs) at different scales. Texture information at different scales is obtained by adjusting the window size or pixel distance of the small image patches. Scale-aware homogeneity can more accurately reflect the multi-scale structure of the image. The calculation formula is as follows: ; ; In the formula, To achieve scale-perceived homogeneity, measures the smoothness and consistency of textures in small tiles (here). Is Based on this, optimizations were performed after introducing a multi-scale fusion strategy. Indicated in scale The weights of homogeneous features under the given conditions. For scale sets, For scale Homogeneity; Homogeneity reflects the smoothness and consistency of local areas of small patches, and scale-aware homogeneity can enhance the capture of multi-scale texture features; S3.4 Perform morphological analysis on each small patch to extract the shape feature vector of semiconductor wafer defects. The criteria include at least area, perimeter, and circularity. Specifically, morphological operations (such as dilation and erosion) are applied to process defective regions in small tiles to remove noise and small isolated areas, and to fill any possible holes. Then, connected component analysis methods (such as 8-neighborhood connectivity) are used to identify and extract each independent defective region. For each independent defective region, its shape features are calculated: the area is the number of pixels in the region; the perimeter is calculated using a contour extraction algorithm to determine the boundary length of the defective region; and the circularity is evaluated by calculating the relationship between the area and perimeter of the defective region. S3.5. By statistically analyzing the grayscale values ​​of pixels within small image patches, extract the grayscale feature vector of semiconductor wafer defects. The calculation includes at least the mean, variance, and skewness. Specifically, the grayscale values ​​of pixels in each small patch are extracted, and the grayscale values ​​within the small patch are considered as a grayscale distribution. Then, the mean of the grayscale values ​​within the small patch is calculated, which is the arithmetic mean of all pixel grayscale values, used to describe the overall brightness level of the image. Next, the variance is calculated, which is the average of the squares of the differences between the grayscale values ​​and the mean, reflecting the dispersion or range of variation of the grayscale distribution of the image. Finally, the skewness is calculated, which is the degree of deviation of the grayscale value distribution from the mean, reflecting the symmetry or skewness of the grayscale distribution. S3.6, Transfer the texture feature vector Shape feature vector and grayscale feature vector Combine them to generate wafer surface feature vectors. .

[0018] S4. Based on the feature vectors of the wafer surface, a convolutional neural network model is used to classify and identify suspected defect candidate regions, and the severity level of defects is evaluated according to the defect size, grayscale deviation and distribution location. In this embodiment, a convolutional neural network model is used to classify and identify suspected defect candidate regions based on wafer surface feature vectors, and the severity level of defects is evaluated according to defect size, grayscale deviation, and distribution location. This includes the following steps: S4.1 Standardize the candidate regions for suspected defects based on the wafer surface feature vectors. Specifically, each candidate region is first cropped and its size is normalized to meet the fixed input size requirements of the convolutional neural network model. At the same time, the wafer surface feature vectors extracted in the previous steps are aligned and encapsulated with the image data of the corresponding candidate regions for suspected defects (each candidate region for suspected defects is combined with the wafer surface feature vector) to construct a unified network input sample, providing standardized input for subsequent deep feature learning and classification recognition. S4.2 Input the standardized suspected defect candidate regions into the pre-trained convolutional neural network model for defect classification and identification, and generate wafer defect regions; The pre-trained convolutional neural network model consists of an input layer, a feature extraction skeleton, a feature fusion layer, and a classification output layer. The input layer receives standardized images of suspected defect candidate regions; the feature extraction skeleton contains multiple alternating sets of convolutional layers (utilizing...). or The filter extracts spatial features, and a pooling layer is used (to downsample and reduce computation and improve translation invariance); a feature fusion mechanism is introduced before the fully connected layer to integrate the wafer surface feature vector extracted in the previous step S3. The deep semantic features extracted by the convolutional layer are concatenated to construct a multimodal fusion feature space. Finally, a fully connected layer maps the features to a one-dimensional space, and the Softmax activation function outputs probability scores for "non-defect" and various real defect types. Combined with a discrimination threshold, accurate classification and recognition are achieved. The model outputs the probability score of each candidate region belonging to each category (including the "non-defect" category and each defect type). Based on the probabilities output by the model, a discrimination threshold n is set (in this embodiment, n is 0.5, calibrated experimentally). If the probability of the "non-defect" category is the highest and exceeds the discrimination threshold n, the candidate region is determined to be a false defect and excluded. If the probability of a certain defect category is the highest and exceeds the discrimination threshold n, the candidate region is determined to be a real defect region, and its defect type is recorded. After this step, all regions confirmed as real defects are the finally generated "wafer defect regions." S4.3. By binarizing the wafer defect region (applying a threshold to convert the wafer defect region into a binary image, where the pixel value of the defect region is 1 and the background region is 0), calculate the pixel size of each wafer defect region, at least the pixel area (obtained by counting the number of pixels in the binarized region), the pixel aspect ratio (obtained by calculating the width and height of the bounding rectangle of the defect region), and the maximum pixel diameter (obtained by calculating the distance between the two farthest points in the region); use the image resolution parameters calibrated in step S1. (pixels / micrometer), converting pixel size to physical size; S4.4 Calculate the grayscale deviation between the wafer defect area and the normal area, including the difference in grayscale mean and grayscale standard deviation; specifically: divide the wafer image into wafer defect area and normal area. For the wafer defect area, extract its binarized pixel values, and calculate the grayscale mean and grayscale standard deviation of the area; the grayscale mean is obtained by calculating the average of the grayscale values ​​of all pixels in the defect area, while the grayscale standard deviation represents the degree of fluctuation of the grayscale values ​​in the area, and is calculated by taking the square root of the sum of the squares of the differences between the pixel grayscale values ​​and the mean; next, select the normal area (the part of the original wafer image without defects) and perform the same grayscale statistics, calculating the grayscale mean and standard deviation of the area; S4.5. Based on the overall geometric information of the semiconductor wafer, calculate the relative distance between the wafer defect region and the wafer center (for each wafer defect region, obtain the center coordinates of the region by calculating its centroid, and then use the Euclidean distance formula to calculate the distance between the centroid of the wafer defect region and the wafer center). S4.6. Based on pixel size, grayscale deviation, and relative distance, a comprehensive severity assessment value is constructed using a weighted method. (That is, the pixel size, grayscale deviation, and relative distance are dimensionless, and the maximum normalization or deviation normalization method is used to map each index to the [0,1] interval to obtain the standardized pixel size, grayscale deviation, and relative distance. By assigning weights to the standardized pixel size, grayscale deviation, and relative distance, the standardized pixel size, grayscale deviation, and relative distance are weighted and summed based on the assigned weights to obtain the comprehensive severity assessment index.) ); S4.7, Based on the comprehensive severity assessment value Assess the severity level of the defect (including minor, moderate, and severe levels); specifically, based on the comprehensive severity assessment index. Within the range, the following grading criteria are set: if the comprehensive severity assessment value Below the lower limit threshold of the evaluation (In this embodiment, If the score is 0.3 (based on historical defect data and process yield standards), the severity of the defect is considered minor, indicating that it has little impact on wafer functionality; if the overall severity assessment value is... Greater than or equal to the lower limit threshold of evaluation And less than the upper limit threshold of the evaluation (In this embodiment, If the score is 0.7 (based on historical defect data and process yield standards), the severity of the defect is judged as moderate, indicating that it has a certain impact on wafer functionality; if the overall severity assessment value is... Exceeding the upper limit threshold of the evaluation If the defect is deemed severe, it indicates that it has a significant impact on the wafer's functionality and performance.

[0019] Example 2: This example provides a semiconductor wafer defect detection system based on machine vision, including a memory, a processor, and a computer program stored in the memory and executable on the processor. The processor executes the computer program to implement the semiconductor wafer defect detection method based on machine vision described above.

[0020] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited thereto. Various changes can be made within the scope of knowledge possessed by those skilled in the art without departing from the spirit of the present invention.

Claims

1. A method for detecting defects in a semiconductor wafer based on machine vision, characterized by, include: S1. Acquire the original wafer image and preprocess the original wafer image; S2. Based on the preprocessed wafer image, the Canny edge detection algorithm is used to generate candidate regions for suspected defects. S3. Extract wafer surface feature vectors by constructing a gray-level co-occurrence matrix based on suspected defect candidate regions; Among them, a reflection compensation mechanism is introduced in the process of constructing the gray-level co-occurrence matrix. The reflection compensation module in the reflection compensation mechanism reduces the interference of light spots on the image, and the directional mask matrix is ​​used to reduce the impact of linear scratches on texture features. S4. Based on the feature vectors of the wafer surface, a convolutional neural network model is used to classify and identify candidate regions of suspected defects, and the severity level of defects is evaluated according to the defect size, grayscale deviation and distribution location.

2. The machine vision-based semiconductor wafer defect detection method of claim 1, wherein, In step S2, the suspected defect candidate region is generated based on the preprocessed wafer image using the Canny edge detection algorithm, including the following steps: S2.1 Set edge blur width parameters for different process layers in semiconductor manufacturing and construct an adaptive gradient kernel; S2.

2. Convolve the preprocessed wafer image using an adaptive gradient kernel to calculate the gradient intensity and direction of each pixel in the preprocessed wafer image; S2.

3. Perform non-maximum suppression on each pixel based on gradient intensity and direction to generate edge pixels; S2.

4. Use two thresholds to classify edge pixels; S2.5 Based on the edge pixel classification results, the edge pixels are aggregated into regions by the connected component analysis method to represent suspected defect candidate regions. The suspected defect candidate regions are then filtered, and the final suspected defect candidate regions are output.

3. The semiconductor wafer defect detection method based on machine vision according to claim 2, characterized in that, In step S2.1, setting edge blur width parameters for different process layers and constructing an adaptive gradient kernel includes the following steps: S2.

11. Based on the wafer's process node and layer type, preset different edge blur width parameters. ; S2.12, Based on the edge blur width parameter of each layer Calculate the standard deviation of the Gaussian kernel; S2.13, Generate edge profile model based on the standard deviation of Gaussian kernel; S2.

14. Calculate the first derivative of the edge profile model to obtain the final adaptive gradient kernel.

4. The semiconductor wafer defect detection method based on machine vision according to claim 2, characterized in that, In step S3, the wafer surface feature vector is extracted by constructing a gray-level co-occurrence matrix based on the suspected defect candidate region, including the following steps: S3.1 Based on the suspected defect candidate regions output in step S2.5, they are cropped into individual small patches, and homomorphic filtering preprocessing is performed on the small patches; S3.

2. Construct a reflection compensation mechanism based on small tiles. Reduce the interference of light spots on the image through the reflection compensation module in the reflection compensation mechanism, and reduce the impact of linear scratches on texture features by using a directional mask matrix. S3.3, adjust the edge blur width parameter The optimal pixel distance is mapped to the gray-level co-occurrence matrix. Based on the optimal pixel distance, the gray-level co-occurrence matrix is ​​constructed, and each small patch is analyzed to extract the texture feature vector of semiconductor wafer defects. This includes reflectance-corrected contrast, reflectance-invariant correlation, and scale-perceived homogeneity. S3.4 Perform morphological analysis on each small patch to extract the shape feature vector of semiconductor wafer defects. It should include at least the area, perimeter, and circularity; S3.

5. By statistically analyzing the grayscale values ​​of pixels within small image patches, extract the grayscale feature vector of semiconductor wafer defects. It should include at least the mean, variance, and skewness; S3.6, Transfer the texture feature vector Shape feature vector and grayscale feature vector Combine them to generate wafer surface feature vectors. .

5. The semiconductor wafer defect detection method based on machine vision according to claim 4, characterized in that, In step S3.2, the interference of light spots on the image is reduced by the reflection compensation module in the reflection compensation mechanism, and the influence of linear scratches on texture features is reduced by the directional mask matrix, including the following steps: S3.

21. Use a Gaussian difference filter to process each small patch and extract the low-frequency components of the reflection part; S3.22 Calculate the reflection normalization parameters for each pixel based on the reflection component. Based on reflection normalization parameters For small tiles Perform normalization; S3.23, Set the threshold for the reflection area; S3.

24. Use Gabor filter banks to detect the dominant texture direction in small tiles; S3.

25. Based on the difference between the local gradient direction and the dominant texture direction in the small patch, construct a direction-sensitive masking matrix. It is used to weight and suppress pixel regions that deviate from the dominant texture direction.

6. The semiconductor wafer defect detection method based on machine vision according to claim 4, characterized in that, In step S3.3, a gray-level co-occurrence matrix is ​​constructed based on the optimal pixel distance, and each small patch is analyzed to extract the texture features of semiconductor wafer defects, including the following steps: S3.31, adjust the edge blur width parameter Distance to the optimal pixel in each small patch Perform mapping; S3.

32. For each small image patch, use the optimal pixel distance obtained by mapping through the edge blur width parameter. Calculate the gray-level co-occurrence matrix; S3.

33. Based on the gray-level co-occurrence matrix, extract texture features from each small patch, including at least reflection-corrected contrast, reflection-invariant correlation, and scale-aware homogeneity.

7. The semiconductor wafer defect detection method based on machine vision according to claim 6, characterized in that, In step S3.33, the reflection-corrected contrast is extracted using a weighted modified gray-level co-occurrence matrix contrast formula, and a dynamic weighting factor is introduced into the gray-level co-occurrence matrix contrast formula. Reflection-invariant correlation is extracted based on small image patches preprocessed by homomorphic filtering using the standard gray-level co-occurrence matrix correlation formula. Scale-aware homogeneity is extracted using multi-scale fusion gray-level co-occurrence matrix homogeneity features.

8. The semiconductor wafer defect detection method based on machine vision according to claim 7, characterized in that, The dynamic weighting factor The calculation process is as follows: based on the gray-level co-occurrence matrix, traverse the small image tiles. All directions satisfy and optimal pixel distance pixel pairs Based on each pixel pair Obtain its grayscale value Based on grayscale values The pixel pair is calculated using the normalized joint intensity component, gradient suppression component, and phase consistency component. Corresponding reflection confidence level Based on reflection confidence Calculate dynamic weighting factors .

9. The semiconductor wafer defect detection method based on machine vision according to claim 1, characterized in that, In step S4, suspected defect candidate regions are classified and identified based on wafer surface feature vectors using a convolutional neural network model. The severity level of the defect is then assessed based on its size, grayscale deviation, and distribution location. This includes the following steps: S4.1 Standardize the candidate regions of suspected defects based on the feature vectors of the wafer surface; S4.2 Input the standardized suspected defect candidate regions into the pre-trained convolutional neural network model for defect classification and identification, and generate wafer defect regions; S4.3 Calculate the pixel size of each wafer defect region by binarizing the wafer defect region; S4.4 Calculate the grayscale deviation between the defective region and the normal region of the wafer, including the difference in mean grayscale and the standard deviation of grayscale. S4.

5. Based on the overall geometric information of the semiconductor wafer, calculate the relative distance between the wafer defect region and the wafer center; S4.

6. Based on pixel size, grayscale deviation, and relative distance, a comprehensive severity assessment value is constructed using a weighted method. S4.

7. Assess the severity level of the defect based on the comprehensive severity assessment value.

10. A semiconductor wafer defect detection system based on machine vision, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: The processor executes a computer program to implement the semiconductor wafer defect detection method based on machine vision as described in any one of claims 1-9.

Images