A semiconductor polished surface detection method and system based on industrial vision

Abstract

The application provides a kind of semiconductor polishing surface detection method and system based on industrial vision, it is related to industrial vision detection technical field, by collecting the image of semiconductor wafer polishing surface, the pixel connected domain with the gray value higher than background area is extracted as preliminary abnormal area;For preliminary abnormal area, pixel value change degree measure, intensity change direction uniformity measure and geometric regularity measure are calculated;Through preset discrimination criterion, preliminary abnormal area is classified, and it is divided into background abnormal area and defect area needing to report, and finally the position and shape information of defect area needing to report are output, while shielding the detection data of background abnormal area, improve the precision and reliability of semiconductor polishing surface detection, provide strong guarantee for yield improvement of semiconductor manufacturing process.

Description

Technical Field

[0001] This application relates to the field of industrial vision inspection technology, and more specifically, to a method and system for inspecting semiconductor polished surfaces based on industrial vision. Background Technology

[0002] In the semiconductor manufacturing industry, high-precision inspection of polished wafer surfaces is essential to ensure the yield of subsequent critical processes such as photolithography and etching. Industrial vision systems, as a non-contact and highly efficient inspection method, are widely used to identify micron-level scratches, particle contamination, or oxide layer defects that may exist on wafer surfaces. These systems typically consist of a high-resolution industrial camera, a precision optical lens, a dedicated illumination source, and a motion platform that supports the wafer. They use complex image processing logic to distinguish defects from the normal background.

[0003] In actual production, after chemical mechanical polishing (CMP), semiconductor wafers may exhibit two visually very similar anomalies: one is an acceptable large-area, low-contrast background haze, caused by process variations and not inherently a fatal defect; the other is a critical large-area, low-contrast chemical residue that must be detected, as it severely impacts subsequent processes. Both anomalies visually manifest as large-area, low-contrast morphology, posing a challenge to traditional detection methods.

[0004] In existing technologies, if a simple brightness threshold is used for judgment, the system may misclassify harmless background glow as a defect, leading to a large number of false alarms and severely impacting production line yield. To solve the problem of false alarms caused by glow, engineers introduced background separation technology. This involves smoothing the original image to obtain a background image, and then subtracting the background image from the original image, thereby effectively suppressing the brightness of the glow area. However, while eliminating the glow effect, this background separation technology may also mistakenly "subtract" visually similar chemical residues as part of the background, causing the system to ignore defects that have a serious impact on subsequent processes, resulting in a large number of missed detections and significant economic losses.

[0005] Therefore, in the visual inspection of semiconductor polished surfaces, when faced with two surface anomalies with extremely similar visual characteristics—acceptable large-area, low-contrast background glow and critical large-area, low-contrast chemical residue marks that must be detected—how to design an image analysis method that can accurately distinguish and identify real defects, thereby achieving high-precision detection with low false alarm and low false alarm rates, is a pressing technical problem that needs to be solved. This method addresses the dilemma that simple brightness threshold judgment leads to excessively high false alarm rates, while background subtraction techniques result in excessively high missed detection rates due to their inability to differentiate between the two. Summary of the Invention

[0006] This application discloses a semiconductor polished surface inspection method and system based on industrial vision. It aims to solve the technical problem in semiconductor polished surface inspection where, when faced with two surface anomalies with extremely similar visual characteristics—acceptable large-area, low-contrast background glow and critical large-area, low-contrast chemical residue marks that must be detected—simple brightness threshold judgment leads to an excessively high false alarm rate, while background subtraction technology leads to an excessively high missed detection rate of real defects due to its inability to distinguish between them.

[0007] The technical solution of this application is as follows:

[0008] In a first aspect, this application discloses a semiconductor polished surface inspection method based on industrial vision, the method comprising the following steps:

[0009] Acquire images of the polished surface of a semiconductor wafer under preset lighting conditions; extract connected regions of pixels in the wafer surface image whose grayscale values ​​are higher than those of the background region, and mark them as preliminary abnormal regions;

[0010] For each initial abnormal region, calculate the measure of pixel value variation, the measure of uniformity of intensity variation direction, and the measure of geometric regularity;

[0011] Based on the measures of pixel value variation degree, intensity variation direction uniformity, and geometric regularity, the preliminary abnormal areas are classified by the preset discrimination criteria, and the background abnormal areas and the defect areas to be reported are obtained.

[0012] The output should report the location and shape information of the defect area, and mask the detection data of abnormal background areas.

[0013] Secondly, this application also discloses a semiconductor polished surface inspection system based on industrial vision, the system comprising:

[0014] The image acquisition module acquires images of the polished surface of a semiconductor wafer under preset lighting conditions.

[0015] The preliminary anomaly region identification module extracts connected components of pixels with gray values ​​higher than those of the background region from the wafer surface image and marks them as preliminary anomaly regions.

[0016] The measurement determination module calculates the degree of pixel value change, the uniformity of intensity change direction, and the geometric regularity for each preliminary abnormal region.

[0017] The classification module, based on the degree of pixel value change, the uniformity of intensity change direction, and the geometric regularity, classifies the preliminary abnormal areas according to preset discrimination criteria, and divides them into background abnormal areas and areas with defects that need to be reported.

[0018] The information output module is used to output the location and shape information of the defective area to be reported, and to mask the detection data of abnormal background areas.

[0019] Beneficial effects

[0020] This application provides a semiconductor polished surface inspection method based on industrial vision. The method acquires images of the polished surface of a semiconductor wafer and extracts connected regions of pixels with grayscale values ​​higher than the background area as preliminary anomalous regions. For these preliminary anomalous regions, this application calculates measures of pixel value variation, uniformity of intensity variation direction, and geometric regularity. Based on this, the preliminary anomalous regions are classified according to preset discrimination criteria, dividing them into background anomalous regions and reportable defect regions. Finally, the location and morphological information of the reportable defect regions are output, while the detection data of background anomalous regions are masked. This application effectively solves the problem in existing technologies where, during semiconductor polished surface inspection, the high visual similarity between background glow and chemical residue marks leads to a high false alarm rate with traditional brightness threshold judgment methods, while background subtraction techniques easily cause missed detection of real defects. By introducing multi-dimensional measures (pixel value variation, uniformity of intensity variation direction, and geometric regularity), this application can more comprehensively and finely characterize the features of anomalous regions, thereby achieving accurate differentiation between background anomalous regions and reportable defect regions. This multidimensional feature analysis method overcomes the limitations of single feature judgment, significantly reduces false alarm rate and false negative rate, improves the accuracy and reliability of semiconductor polished surface detection, and provides a strong guarantee for improving the yield of semiconductor manufacturing process. Attached Figure Description

[0021] To illustrate this application more clearly, the accompanying drawings used in the embodiments will be briefly described below. Obviously, those skilled in the art can obtain other drawings based on these drawings without any creative effort.

[0022] Figure 1 The diagram above illustrates a flowchart of a semiconductor polished surface inspection method based on industrial vision.

[0023] Figure 2 The diagram above illustrates a schematic of a semiconductor polished surface inspection system based on industrial vision.

[0024] Figure reference numerals: 100, Semiconductor polished surface inspection system based on industrial vision; 10, Image acquisition module; 20, Preliminary abnormal area identification module; 30, Measurement determination module; 40, Classification module; 50, Information output module. Detailed Implementation

[0025] The technical solutions of this application will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments. The components of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0026] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this application, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0027] In the semiconductor manufacturing industry, high-precision inspection of polished wafer surfaces is essential to ensure the yield of subsequent critical processes such as photolithography and etching. However, traditional industrial vision inspection systems face significant challenges when dealing with background glow and critical chemical residues that may coexist on the polished surface of semiconductor wafers, both exhibiting extremely similar visual characteristics. If a simple brightness threshold is used, the system may misclassify harmless background glow as a defect, leading to numerous false alarms. Conversely, while background separation technology eliminates the glow's influence, it may also mistakenly include actual chemical residues as part of the background, resulting in missed detections. Both of these situations severely impact the accuracy of inspection and the yield of the production line.

[0028] like Figure 1 As shown, an exemplary flowchart of a semiconductor polished surface inspection method based on industrial vision is illustrated. This application proposes a semiconductor polished surface inspection method based on industrial vision, comprising:

[0029] S10: Acquire a wafer surface image of the polished semiconductor wafer under preset lighting conditions; extract connected regions of pixels in the wafer surface image whose grayscale values ​​are higher than those of the background area, and mark them as preliminary abnormal regions;

[0030] S20, for each preliminary abnormal region, calculate the measure of pixel value change degree, the measure of intensity change direction uniformity, and the measure of geometric regularity;

[0031] S30, based on the pixel value change degree measure, intensity change direction uniformity measure and geometric regularity measure, classifies the preliminary abnormal area through preset discrimination criteria, and divides the background abnormal area and the defect area to be reported.

[0032] S40 outputs the location and shape information of the defective area to be reported, and masks the detection data of abnormal background areas.

[0033] This application aims to overcome the limitations of traditional methods in distinguishing between background anomalies and real defects by introducing multi-dimensional measurement analysis and dynamic discrimination criteria, thereby achieving high-precision, low-miss-detection-rate, and low-false-detection-rate semiconductor polished surface inspection.

[0034] To better understand the technical solution of this application, some key terms involved will be explained first.

[0035] A wafer surface image refers to a digital image of the polished surface of a semiconductor wafer acquired by an industrial camera under specific lighting conditions. It contains all the visual information of the wafer surface and forms the basis for subsequent analysis. A pixel connected region refers to a region in an image composed of adjacent pixels with similar or identical grayscale values. These regions are often treated as a whole for analysis in image processing. Preliminary anomalous regions refer to image regions that, after initial screening, may contain defects or background anomalies; they require further analysis to determine their nature.

[0036] Pixel value variation measures the severity or uniformity of pixel grayscale value changes within an initial anomaly region, reflecting the texture characteristics within the region. Intensity variation direction uniformity measures assess the consistency of pixel intensity gradient directions within the initial anomaly region, helping to distinguish between directional defects (such as scratches) and non-directional background noise. Geometric regularity measures describe the geometric characteristics of the initial anomaly region, such as shape, size, and boundary smoothness, to differentiate between regular or irregular defects and background anomalies. Background anomaly regions refer to those areas that, while visually abnormal, have no substantial impact on semiconductor device performance, such as background glow.

[0037] The detection method of this application first requires acquiring images of the polished surface of a semiconductor wafer under preset illumination conditions. Image acquisition is the starting point of the entire detection process, and its quality directly affects the accuracy of subsequent analysis. For example, a high-resolution industrial camera can be used in conjunction with specialized lighting equipment such as a ring lamp or coaxial light source to ensure that the wafer surface image has sufficient contrast and clarity to capture minute surface features. In actual operation, appropriate illumination angles and intensities can be selected based on the wafer material, polishing process, and possible defect types to maximize the contrast between defects and the background.

[0038] After acquiring the wafer surface image, the next step is to extract the connected components of pixels with gray values ​​higher than the background area and mark them as preliminary anomalous regions. This step aims to initially screen out all potentially anomalous regions in the image, providing targets for subsequent detailed analysis. For example, a global thresholding method can be used, setting a fixed gray value threshold and identifying all pixels with gray values ​​higher than that threshold as potential anomalous pixels. Another approach is to use a local adaptive thresholding method, dynamically adjusting the threshold based on the local gray value distribution of different regions of the image to better adapt to the unevenness of image brightness. For example, the local average gray value and standard deviation of the image can be calculated, and then the threshold for each pixel can be determined based on these statistics. Through connected component analysis, these potential anomalous pixels can be clustered into regions, and noisy regions that are too small or too large can be eliminated, thus obtaining preliminary anomalous regions.

[0039] For each initial anomalous region, it is necessary to calculate measures of pixel value variation, uniformity of intensity variation direction, and geometric regularity. These measures are crucial for distinguishing background anomalies from genuine defects.

[0040] Measuring the degree of pixel value variation can reflect the texture detail within a region. For example, the standard deviation of the grayscale values ​​of all pixels within an initial anomalous region can be calculated; a larger standard deviation indicates a more drastic change in pixel values ​​and richer texture. Another approach is to calculate local contrast, quantifying the degree of variation by comparing the grayscale differences between pixels within the region and their surrounding pixels. For instance, edge detection operators such as the Laplacian or Sobel operators can be used to process the initial anomalous region, and then the intensity of the edge response can be statistically analyzed to reflect the drasticness of the pixel value variation.

[0041] The uniformity of intensity variation direction can reveal the directionality of texture within a region. For example, the gradient direction of each pixel within an initial anomalous region can be calculated, and then the distribution of these gradient directions can be statistically analyzed. If the gradient directions are highly concentrated in a certain direction, it indicates that the region has strong directional texture, such as scratches; if the gradient directions are uniformly distributed, it indicates that the texture in the region has no obvious directionality, such as diffuse background glow. For example, a Gabor filter bank can be used to filter the initial anomalous region at different directions and scales, and then the energy distribution of the filtered response can be analyzed to assess the uniformity of intensity variation direction.

[0042] Geometric regularity measures can describe the shape characteristics of a region. For example, the ratio of the perimeter to the area of ​​a preliminary anomalous region, or the aspect ratio of its smallest bounding rectangle, can be calculated to assess its shape regularity. Circular or elliptical regions generally have high geometric regularity, while irregular defects (such as chemical residues) may have lower geometric regularity. For example, shape descriptors, such as Hu invariant moments or Fourier descriptors, can be used to quantify the contours of preliminary anomalous regions, thus obtaining a geometric regularity measure.

[0043] Based on measures of pixel value variation, uniformity of intensity variation direction, and geometric regularity, preliminary anomalous regions are classified using preset discrimination criteria, resulting in background anomalous regions and reportable defect regions. This step is the core of this application, utilizing multi-dimensional measurement information for refined classification of preliminary anomalous regions. For example, machine learning models such as Support Vector Machines (SVM) or neural networks can be used as discrimination criteria. During the training phase, a large amount of labeled measurement data of background anomalous regions and reportable defect regions is used to train the model, enabling it to learn the optimal classification boundary to distinguish between the two. During the detection phase, the measurements of the preliminary anomalous regions are input into the trained model, which outputs the probability or category of whether the region belongs to the background anomalous region or the reportable defect region. Another approach is to use a rule-based expert system, setting a series of thresholds and logical rules based on experience. For example, if the degree of pixel value variation in a preliminary abnormal area is below a certain threshold, and the degree of geometric regularity is above a certain threshold, it may be identified as a background abnormal area; conversely, if the degree of pixel value variation is high, and the degree of uniformity of intensity variation shows obvious directionality, it may be identified as a defect area that needs to be reported.

[0044] Finally, the system outputs the location and shape information of the defective areas that need to be reported, and masks the detection data of background abnormal areas. This step involves the presentation and application of the detection results. For areas classified as defective that need to be reported, the system extracts detailed information such as their coordinates, size, and shape in the wafer surface image and outputs it in the form of a report for subsequent processing by process engineers. For example, a detection report containing information such as defect image, defect type, defect location coordinates, and defect area can be generated. At the same time, the system masks the detection data of areas classified as background abnormalities and does not report them, thereby effectively reducing the false alarm rate and avoiding unnecessary interference to the production line.

[0045] This application presents a semiconductor polished surface inspection method based on industrial vision. By introducing multi-dimensional measurement analysis and dynamic discrimination criteria, it effectively solves the dilemma of traditional methods in distinguishing surface anomalies with similar visual features, such as background glow and chemical residue marks. Traditional methods often suffer from high false alarm rates or high false negative rates because they cannot accurately distinguish between these two types of anomalies. For example, simple brightness threshold judgment may misclassify harmless background glow as a defect, while background subtraction techniques may eliminate real defects as well.

[0046] The core innovation of this application lies in its comprehensive consideration of three complementary dimensions: the degree of pixel value variation, the uniformity of intensity variation direction, and the geometric regularity. The degree of pixel value variation captures the texture details within a region, the uniformity of intensity variation direction reveals the directional characteristics of the texture, and the geometric regularity describes the anomalous region from a shape perspective. Through the comprehensive analysis of these multi-dimensional measures, this application can more comprehensively and accurately characterize the essential features of the initial anomalous region.

[0047] Compared with existing technologies, the advantages of this application lie in its refined feature extraction and intelligent classification. For example, when faced with a large area of ​​low-contrast background glow, the pixel value variation measure may be low, the intensity variation direction uniformity measure may be uniformly distributed, and the geometric regularity measure may be high. However, for a similarly large area of ​​low-contrast chemical residue, the pixel value variation measure may be high, the intensity variation direction uniformity measure may exhibit a specific directionality, and the geometric regularity measure may be low. By pre-setting discrimination criteria, the system can accurately distinguish between these two visually similar but fundamentally different anomalies based on the combined features of these measures. This multi-dimensional measurement analysis method significantly improves the accuracy of detection, effectively reduces the false alarm rate and false negative rate, thereby providing a more reliable quality control means for semiconductor manufacturing processes.

[0048] In some of the embodiments described above in this application, when calculating the degree of pixel value change, the uniformity of intensity change direction, and the geometric regularity for the initial abnormal region, if only a single-scale analysis method is used, it may be difficult to effectively capture the defect features of different sizes or shapes, resulting in insufficient differentiation between fine structures and diffuse structures, thereby affecting the accuracy of subsequent classification.

[0049] In response, this application further proposes an optimization scheme, which performs multi-scale morphological filtering on the image data corresponding to the initial abnormal region to separate fine structure and diffuse structure at different spatial scales, and calculates the above-mentioned measures simultaneously based on multiple sets of filtered images, thereby improving the robustness and accuracy of detection.

[0050] Specifically, the steps for calculating the pixel value variation measure, intensity variation direction uniformity measure, and geometric regularity measure for each initial anomaly region include:

[0051] Multi-scale morphological filtering is performed on the image data corresponding to the preliminary abnormal region to separate fine structure and diffuse structure at different spatial scales, resulting in multiple sets of filtered images.

[0052] Based on multiple sets of filtered images, the degree of pixel value change, the uniformity of intensity change direction, and the geometric regularity of the preliminary abnormal region are calculated simultaneously.

[0053] Multi-scale morphological filtering refers to the morphological processing of images using a series of structuring elements of different sizes and shapes, such as erosion, dilation, opening, and closing operations. By employing structuring elements of different scales, features of different sizes can be effectively extracted from the image. For example, smaller structuring elements are suitable for detecting fine, localized defect features, while larger structuring elements are better suited for identifying diffuse, large-scale anomalous regions. Fine structures typically refer to small, well-defined defects, such as scratches and particles; diffuse structures refer to larger, blurred-bordered or gently changing anomalies, such as polishing residue and water stains. Thus, multi-scale filtering yields multiple sets of filtered images, each representing the feature extraction results of the initial anomalous region at a specific scale. Based on these multiple sets of filtered images, measures of pixel value variation, uniformity of intensity variation direction, and geometric regularity can be calculated simultaneously. Simultaneous calculation means that these measures are calculated in parallel or collaboratively when processing images at different scales to fully utilize multi-scale information and ensure the comprehensiveness and consistency of the measurement calculations.

[0054] This application's solution effectively addresses the limitations of single-scale analysis in distinguishing defects of different sizes and shapes by introducing multi-scale morphological filtering. Specifically, multi-scale filtering can effectively separate fine and diffuse structures in the initial anomaly region. For example, for fine structures, small-scale filtering can highlight their local features; for diffuse structures, large-scale filtering can capture their overall trend. Thus, processing image data at different spatial scales allows for a more comprehensive and accurate extraction of defect-related feature information. Subsequently, based on these multiple filtered images, measures of pixel value variation, intensity variation direction uniformity, and geometric regularity are simultaneously calculated, ensuring that these measures comprehensively reflect the characteristics of the initial anomaly region at different scales, thereby providing richer and more reliable discrimination criteria for subsequent defect classification.

[0055] Through the above technical solution, this application can significantly improve the accuracy and robustness of defect detection on semiconductor polished surfaces. Compared with methods using only single-scale analysis, multi-scale morphological filtering enables the system to effectively distinguish and identify defects of different sizes and shapes. For example, it can more clearly distinguish between minute scratches and large areas of polishing residue. Furthermore, by simultaneously calculating various metrics based on multiple sets of filtered images, it ensures that these metrics comprehensively reflect the characteristics of the initial abnormal area at different scales, avoiding missed detections or misjudgments due to scale mismatch. This improves the accuracy of defect classification, reduces the risk of background abnormal areas being falsely reported as defect areas, and ultimately improves detection efficiency and product quality.

[0056] In some preferred embodiments, multi-scale morphological filtering operations can employ a series of circular or square structuring elements, with their radii or side lengths gradually increasing from 1 pixel to 10 pixels. For example, for image data of a preliminary anomalous region, opening and closing operations are first performed using a structuring element with a radius of 1 pixel to obtain a first set of filtered images, used to highlight small particles or minor scratches. Next, the same operation is performed using a structuring element with a radius of 5 pixels to obtain a second set of filtered images, used to identify medium-sized defects or texture variations. Finally, an operation is performed using a structuring element with a radius of 10 pixels to obtain a third set of filtered images, used to capture large-area diffuse anomalies. Based on these three sets of filtered images, measures of pixel value variation, uniformity of intensity variation direction, and geometric regularity can be calculated in parallel for each preliminary anomalous region. For example, the pixel value variation measure can be weighted and fused with local variance or entropy values ​​at different scales; the uniformity of intensity variation direction can be statistically analyzed based on gradient direction histograms at different scales; and the geometric regularity measure can integrate geometric features such as the perimeter and area ratio of connected regions at different scales. In this way, we can ensure a more comprehensive and accurate description of the characteristics of the initial abnormal area.

[0057] Specifically, in some of the above embodiments, the calculation of the degree of pixel value change in the preliminary abnormal region based on multiple sets of filtered images can be further refined into the following steps:

[0058] The initial abnormal region is divided into multiple detection sub-regions;

[0059] For each detection sub-region, calculate the local dispersion coefficient of its pixel grayscale value;

[0060] By fusing the local discrete coefficients of the multiple detection sub-regions, a measure of the degree of pixel value change in the preliminary abnormal region is determined.

[0061] Specifically, to more precisely capture the changing characteristics of local pixel values ​​and avoid the averaging of the overall region obscuring local details, the initially identified anomalous regions are spatially subdivided into smaller, independently analyzable local units. These detection sub-regions can be regular grid-like divisions, such as square or rectangular areas, or irregular regions adaptively divided based on image content.

[0062] For each detection sub-region, the local coefficient of variation (LCV) of the pixel grayscale values ​​is calculated. The LCV can be understood as a statistical measure of the degree of fluctuation in pixel grayscale values ​​within that sub-region. For example, statistical methods such as the ratio of standard deviation to mean or the coefficient of variation can be used to quantify the degree of local dispersion. This step aims to accurately assess the uniformity or volatility of pixel grayscale values ​​within each local region, thereby reflecting the texture features or anomalies of that region.

[0063] The method of fusing local discrete coefficients from multiple detection sub-regions to determine the degree of pixel value variation in the initial anomaly region involves comprehensively processing the local discrete coefficients calculated from each detection sub-region to obtain a measure of the overall pixel value variation in the entire initial anomaly region. Fusion methods may include, but are not limited to, averaging, weighted averaging, maximizing, medianing, or performing statistical distribution analysis. This fusion yields a measure that comprehensively reflects the overall pixel value variation characteristics of the initial anomaly region, providing crucial information for subsequent defect classification.

[0064] The proposed solution subdivides the initial anomaly region into multiple detection sub-regions and calculates local discrete coefficients for each sub-region. This effectively avoids the loss of local detail information caused by excessively large regions or complex internal structures when performing a single measurement on the entire initial anomaly region. By fusing the local discrete coefficients, the degree of variation in pixel grayscale values ​​within the initial anomaly region can be characterized more accurately and precisely, thereby improving the ability to distinguish between different types of anomaly regions (such as background anomalies and reportable defects).

[0065] The above technical solution enables refined quantification of pixel value changes in the initial anomaly region, effectively improving the recognition accuracy of minute defects or complex background textures on semiconductor polished surfaces. Compared to directly performing coarse calculations on the entire initial anomaly region, this solution, through the calculation and fusion of local discrete coefficients, can more sensitively capture subtle pixel grayscale changes, thus providing more reliable and discriminative feature measures for subsequent defect classification. This helps reduce false alarms and false negatives, improving the overall performance of the detection system.

[0066] In some embodiments described above in this application, a measure of the uniformity of intensity change direction in preliminary anomalous regions is proposed. Specifically, calculating the measure of the uniformity of intensity change direction in preliminary anomalous regions based on multiple sets of filtered images may include the following steps:

[0067] The initial abnormal region is divided into multiple gradient detection regions;

[0068] For each gradient detection region, calculate the local concentration of intensity change direction;

[0069] Statistical analysis of the local concentration in all gradient detection regions yields a measure of the uniformity of intensity change direction in the preliminary anomaly region.

[0070] Dividing the initial anomaly region into multiple gradient detection regions refers to spatially subdividing the initial anomaly region to be analyzed into several smaller, independent sub-regions. These sub-regions can be regular, such as using a grid-like division, or they can be adaptively divided based on image features. The purpose is to capture and analyze the local texture and structural changes within the initial anomaly region in a more refined manner.

[0071] Furthermore, for each gradient detection region, the local concentration of intensity change directions is calculated. This can be understood as assessing whether the direction of pixel grayscale value changes tends to be consistent within that local region. For example, the concentration can be quantified by calculating the gradient direction of each pixel and then statistically analyzing the distribution of gradient directions across all pixels within the gradient detection region, such as by calculating the peak value, variance, or entropy of the orientation histogram. A higher local concentration indicates a more consistent intensity change direction in that region, potentially corresponding to regular textures or boundaries; conversely, a lower concentration may correspond to a diffuse, disordered structure.

[0072] In practical applications, statistical analysis of the local concentration levels across all gradient detection regions yields a measure of the uniformity of intensity variation direction in the initial anomaly region. This involves a comprehensive evaluation of the local concentration levels across all sub-regions. For example, the mean, median, variance, or specific percentile of all local concentration levels can be calculated to reflect the overall uniformity of intensity variation direction within the entire initial anomaly region. This statistical analysis effectively distinguishes defects with directional textures, such as scratches, from non-directional background noise or diffuse contamination.

[0073] The proposed solution subdivides the initial anomaly region into multiple gradient detection regions and calculates the local concentration of intensity change direction in each region, enabling precise capture of the directional features of image texture at the local level. This fine quantification of local features allows for accurate reflection of the overall uniformity of intensity change direction within the entire initial anomaly region when subsequent statistical analysis and fusion of this local information are performed. This divide-and-conquer strategy avoids the loss of detail that might result from a coarse analysis of the entire region, thereby improving the ability to distinguish between different types of anomaly regions, especially defects with specific directional textures.

[0074] The above technical solutions enable a more precise and accurate quantification of the uniformity of intensity variation in the initial anomaly region. This refined measurement helps distinguish defects with specific directional characteristics, such as scratches and crystal orientation textures, from non-directional or randomly distributed background noise or diffuse contamination. Statistical analysis of local concentration levels can effectively improve the accuracy and robustness of the detection method for identifying various defects, especially when dealing with complex and varied defects on semiconductor polished surfaces, providing a more reliable basis for judgment.

[0075] In some embodiments of this application, it is necessary to calculate the geometric regularity measure for the initial abnormal region.

[0076] In response, this application further proposes a method for calculating the geometric regularity measure of preliminary abnormal regions based on multiple sets of filtered images, specifically including:

[0077] The initial abnormal region is divided into multiple contour detection units;

[0078] For each contour detection unit, calculate the local smoothness and local compactness of its boundary;

[0079] By analyzing the distribution characteristics of the local smoothness and local compactness of the multiple contour detection units, a geometric regularity measure of the preliminary abnormal region is obtained.

[0080] Specifically, a contour detection unit refers to a sub-region within the initial anomaly region used for local geometric feature analysis. These units can be rectangular or circular regions of a preset size, or irregular regions adaptively divided based on image content. The purpose is to decompose the complex initial anomaly region into smaller, more easily analyzed parts for refined local geometric feature evaluation. Local smoothness of the boundary can be understood as the smoothness of the contour detection unit's boundary. For example, it can be quantified by calculating the curvature of boundary pixels, the rate of gradient change, or using polynomial fitting. Higher smoothness indicates that the boundary is more likely to be a straight line or a regular curve. Local compactness refers to the density of pixel distribution or the regularity of shape within the contour detection unit. For example, it can be measured by calculating the ratio of the unit's area to its perimeter (e.g., roundness, rectangularity) or the concentration of grayscale values ​​among its internal pixels. Higher compactness generally means that the region has a more regular shape and a denser internal structure. In practical applications, the distribution characteristics of local smoothness and compactness of multiple contour detection units can be analyzed using statistical methods. These methods include calculating statistical parameters such as the mean, variance, skewness, and kurtosis of these local measures, or constructing histograms and probability density functions. These distribution characteristics allow for a comprehensive assessment of the geometric regularity of the entire initial anomaly region. For example, if most contour detection units exhibit high smoothness and high compactness, the initial anomaly region is considered to have high geometric regularity.

[0081] The proposed solution subdivides the initial anomaly region into multiple contour detection units and quantifies the local smoothness and compactness of each unit's boundaries, thereby capturing the geometric features of the initial anomaly region at the local level. Subsequently, by analyzing the distribution characteristics of these local measurements, the geometric regularity of the entire initial anomaly region can be comprehensively evaluated. This divide-and-conquer strategy makes the evaluation of the geometric regularity of complex-shaped initial anomaly regions more accurate and detailed, avoiding information loss or misjudgment that may result from a single global measurement.

[0082] The above technical solutions enable more refined and accurate quantification of the geometric regularity of preliminary anomaly regions. By extracting local features and analyzing their distribution characteristics, it is possible to effectively distinguish anomaly regions with different geometric features, such as differentiating regularly shaped defects from irregularly shaped background noise, thereby improving the accuracy and robustness of defect detection.

[0083] In some embodiments described above, preset discrimination criteria are typically used to classify preliminary abnormal regions. However, in actual semiconductor polished surface inspection, the characteristic distribution of background abnormal regions and reportable defect regions on the wafer surface may drift with changes in time, environment, or process conditions. This makes it difficult for fixed discrimination criteria to maintain optimal classification results, potentially leading to false alarms or missed alarms, thus affecting the accuracy and stability of the inspection. Therefore, this application further proposes a method for dynamically adjusting classification discrimination criteria to adapt to changes in the actual production environment and improve the robustness of the inspection.

[0084] The steps described above, based on measures of pixel value variation, uniformity of intensity variation direction, and geometric regularity, to classify preliminary abnormal regions using preset discrimination criteria and divide them into background abnormal regions and regions requiring reporting defects, include:

[0085] Continuously monitor the real-time distribution of measures of pixel value change degree, intensity change direction uniformity, and geometric regularity.

[0086] The real-time distribution is compared with the preset benchmark distribution to identify the distribution deviation characteristics between the real-time distribution and the preset benchmark distribution;

[0087] Based on the degree and direction of the distribution deviation from the characteristics, the classification boundary in the preset discrimination criterion is modified;

[0088] When the measurement distributions of the background anomaly area and the defect area to be reported overlap, the classification threshold of the overlapping area is adjusted based on historical detection statistics to separate the background anomaly area and the defect area to be reported.

[0089] Specifically, during the detection process, the system continuously collects and statistically analyzes the real-time distributions of the measures of pixel value variation, intensity variation direction uniformity, and geometric regularity to obtain their statistical characteristics within the current time period, such as mean, variance, kurtosis, skewness, or multidimensional probability density function. These real-time distributions reflect the actual situation of the wafer surface features under the current production conditions.

[0090] The currently monitored real-time distribution of measurements is then compared with a reference distribution established by the system under stable or ideal conditions. This comparison aims to identify differences between the two, such as shifts in the distribution center, changes in the distribution width, or distortions in the distribution shape. These differences are known as distribution deviation characteristics, indicating possible changes in the detection environment or the characteristics of the object being measured.

[0091] In practical applications, the classification boundary in the preset discrimination criterion is modified according to the degree and direction of the distribution deviation features. Specifically, this involves dynamically adjusting the classification boundary used to distinguish between background anomaly areas and areas requiring reporting defects based on the identified distribution deviation features. For example, if the real-time distribution shows a general increase in background noise intensity, the classification boundary may be shifted upwards to avoid misclassifying enhanced background noise as a defect. The degree and direction of the modification correspond to the degree and direction of the deviation features, ensuring that the classification criterion can adapt to new data distributions.

[0092] Furthermore, in some cases, the measurement values ​​for background anomalies and actual defects may have ambiguous overlap. In such cases, the system utilizes historical detection data accumulated over a long period, including statistical information on confirmed defects and background anomalies, to finely adjust the classification threshold within the overlapping area. For example, by analyzing the ratio of false positives to false negatives in historical data within this overlapping area, the threshold can be optimized, thereby maximizing the defect detection rate while maintaining a low false positive rate.

[0093] This application's solution effectively addresses the performance degradation of traditional fixed discrimination criteria when facing changes in the production environment by introducing a dynamic adjustment mechanism. Specifically, continuous monitoring of the real-time distribution of measurements allows the system to promptly detect subtle changes in wafer surface features or background noise. By comparing the real-time distribution with a preset benchmark distribution, the system can quantify these changes and identify distribution deviations. Based on these deviations, the classification boundaries are adjusted accordingly, ensuring that the discrimination criteria always match the current actual data distribution. Especially when the measurement distributions of background anomaly areas and reportable defect areas overlap, fine-tuning the classification threshold for overlapping areas using historical detection statistics enables more accurate differentiation between these two types of areas, thus avoiding misclassification or missed classification due to fixed thresholds and significantly improving classification accuracy and robustness.

[0094] Through the above technical solution, this application enables adaptive adjustment of the discrimination criteria for defect detection on semiconductor polished surfaces, significantly improving the accuracy and stability of the detection system during long-term operation and in the face of process fluctuations. This solution effectively reduces the false alarm rate and false negative rate caused by environmental or process changes, ensuring the reliability of defect detection results. Furthermore, by utilizing historical detection statistics to optimize the classification of overlapping areas, the system's discrimination capability under complex and ambiguous conditions is further enhanced, enabling the detection method to better adapt to the dynamics and uncertainties of industrial production, thereby providing more reliable quality control for semiconductor manufacturing processes.

[0095] In some preferred embodiments, it is assumed that on a semiconductor wafer polishing production line, slight fluctuations in the composition of the polishing slurry cause a novel background texture with low contrast and irregular geometry to appear on the wafer surface. This texture may initially be misjudged as a defect, or in some cases, the true defect may be masked by this texture. The system continuously monitors the real-time distribution of pixel value variation measures, intensity variation direction uniformity measures, and geometric regularity measures. It detects a slight shift in the distribution of geometric regularity measures towards greater irregularity, and a slight decrease in the mean value of the pixel value variation measure. The system compares these real-time distributions with a preset baseline distribution, identifying deviations in the geometric regularity measure distribution, indicating a decrease in the regularity of the background texture. Based on the degree and direction of these deviations, the system dynamically adjusts the classification boundaries in the classification criteria; for example, it adjusts a classification threshold for the geometric regularity measure towards a more lenient direction to correctly classify the new background texture as an abnormal background region. Simultaneously, if an overlap is found between this new background texture and the metric distribution of certain minor defects, the system will fine-tune the classification threshold for that overlapping area based on historical detection statistics (e.g., the proportion of similar features previously manually identified as background or defects). For example, if historical data shows that the overlapping area is mostly a background anomaly rather than a real defect, the threshold will be adjusted accordingly to reduce false alarms. Through this dynamic adjustment, even with subtle changes in production conditions, the system can accurately distinguish between new background textures and genuine defects, thus maintaining high-precision detection performance.

[0096] In some embodiments described above, this application proposes a scheme that classifies preliminary abnormal regions based on measures of pixel value variation, uniformity of intensity variation direction, and geometric regularity, using preset discrimination criteria, and continuously monitors the real-time distribution of these measures to correct the classification boundaries. However, relying solely on real-time distribution for monitoring may have limitations. For example, instantaneous data fluctuations may be misjudged as process changes, leading to frequent and unnecessary adjustments to the classification boundaries; or, slow but continuous process drift may fail to be identified in time due to a lack of trend analysis, thus delaying the correction of classification boundaries and ultimately affecting the accuracy and stability of defect detection. If these problems are not addressed, the system may experience delayed or overly sensitive responses to process changes, thereby affecting the reliability of semiconductor polished surface inspection.

[0097] In response, this application further proposes a step for continuously monitoring the real-time distribution of the aforementioned pixel value change measure, intensity change direction uniformity measure, and geometric regularity measure, including:

[0098] A fixed-duration sliding time window is set, and within this sliding time window, the degree of change of the pixel value, the uniformity of the direction of intensity change, and the geometric regularity are periodically monitored.

[0099] The pixel value change degree measure, intensity change direction uniformity measure, and geometric regularity measure collected within the sliding time window are statistically fitted to generate the measure distribution characteristics of the current sliding time window.

[0100] The measurement distribution characteristics of the current sliding time window are compared with those of the previous multiple sliding time windows to identify the continuous changing trend of the measurement distribution characteristics.

[0101] When the continuous change trend exceeds the preset fluctuation threshold, an early warning signal is issued, and adjustment parameters indicating the classification boundary are generated.

[0102] Specifically, setting a fixed-duration sliding time window means determining a continuous time period as the unit for data accumulation and analysis based on the actual production line cycle time and the expected speed of process changes. For example, this sliding time window can be set to several minutes, several hours, or several batch production cycles. Within this sliding time window, the system periodically collects data on the degree of pixel value change, the uniformity of intensity change direction, and geometric regularity at a preset frequency or when triggered by specific events. Statistical fitting is performed on the measurements collected within the sliding time window to extract a representative statistical model or parameter set from the large amount of raw data within the window. For example, statistical moments such as the mean, variance, skewness, and kurtosis of these measurements can be calculated, or their distribution can be characterized using methods such as histograms and kernel density estimation, thereby generating the measurement distribution characteristics of the current sliding time window. The purpose of comparing the measurement distribution characteristics of the current sliding time window with those of multiple previous sliding time windows is to identify whether the measurement distribution characteristics exhibit a continuous and directional evolution, such as whether the mean is continuously rising or the variance is continuously expanding, rather than simply focusing on deviations at a single time point. This comparison can be achieved by calculating the distance between the distribution characteristics of different windows (such as KL divergence or Wasserstein distance) or the rate of change of parameters. When the continuous trend of change exceeds a preset fluctuation threshold, which is set based on historical data and process requirements, it is used to determine whether the change in the measurement distribution characteristics has reached a level requiring intervention. At this time, the system will issue an early warning signal, notify the operator or the automation system, and generate adjustment parameters indicating the classification boundary, providing a quantitative basis for subsequent correction of the classification boundary, such as indicating the direction and magnitude of the shift in the classification boundary.

[0103] This application's solution addresses the lag and misjudgment issues that may arise from relying solely on real-time distribution by introducing a sliding time window and trend analysis. Specifically, a fixed-duration sliding time window allows the system to accumulate data over a representative time span, smoothing out instantaneous noise and obtaining more stable measurement distribution characteristics. Statistical fitting of the data within the window accurately captures the measurement distribution characteristics under the current process state. More importantly, comparing the distribution characteristics of the current window with those of previous windows allows the system to identify continuous trends in the measurement distribution characteristics, rather than isolated deviations. This trend analysis effectively distinguishes random fluctuations from actual process drift or equipment degradation, providing early warning signals before problems become severe. When the continuous trend exceeds a preset fluctuation threshold, the system promptly generates adjustment parameters indicating the classification boundary, providing a forward-looking and quantitative basis for the dynamic correction of subsequent judgment criteria, avoiding missed or false alarms due to delayed adjustments.

[0104] Through the above technical solution, this application enables more refined and forward-looking monitoring of the distribution of key measurements during the inspection of semiconductor polished surfaces. Compared to judgments based solely on real-time distribution, this solution significantly improves the system's sensitivity and robustness to minute process changes and the evolution of potential defects by introducing a sliding time window and trend analysis. This effectively avoids misjudgments caused by instantaneous fluctuations and allows for earlier identification of continuous problems such as process drift or equipment degradation, thus providing a more timely and accurate basis for the dynamic correction of classification boundaries. This not only helps improve the accuracy and stability of defect detection but also effectively reduces the scrap rate in the production process, improving the yield and production efficiency of semiconductor products.

[0105] For example, suppose that during the polishing process of a semiconductor wafer, it is necessary to monitor the degree of pixel value variation, the uniformity of intensity variation direction, and the geometric regularity. The system can set a sliding time window with a fixed duration of 1 hour. Within this window, data for these three measurements is collected every 5 minutes. At the end of each 1-hour window, the system performs statistical fitting on all the data collected within the window, such as calculating the mean and standard deviation of each measurement to form the measurement distribution characteristics of the current window. Subsequently, the mean and standard deviation of the current window are compared with the mean and standard deviation of the past 10 windows (i.e., the past 10 hours). If it is found that the mean of a certain measurement has been rising continuously for 3 consecutive windows, and its rise exceeds a preset fluctuation threshold of 0.5%, the system will immediately issue an early warning signal and generate an adjustment parameter, such as indicating that the classification boundary should be moved 0.2 units towards higher values ​​to accommodate slight drifts in process parameters. This proactive monitoring and adjustment mechanism enables the system to adjust the defect identification criteria in a timely manner when the polishing process parameters change slowly but continuously, thus avoiding missed defects due to outdated criteria.

[0106] In some embodiments described above, when the continuous trend exceeds a preset fluctuation threshold, a warning signal is issued and adjustment parameters indicating the classification boundary are generated. However, if these adjustment parameters fail to fully consider the complex correlations between different measures, the identification of defect patterns may be inaccurate, thus affecting the dynamic correction effect of the classification boundary. Especially in semiconductor polished surface inspection, defect features are often the result of the combined effect of multiple dimensions of measures. Single or simple measure adjustments may not be able to effectively cope with complex process drift or defect evolution, which may lead to inaccurate division between background abnormal areas and areas requiring reporting defects.

[0107] In response, this application further proposes a step of issuing a warning signal and generating adjustment parameters indicating the classification boundary when the continuous change trend exceeds a preset fluctuation threshold, including:

[0108] Cross-analysis is performed on the variation trends of pixel value variation degree measure, intensity variation direction uniformity measure and geometric regularity measure to extract the correlation offset pattern between multiple measures;

[0109] Based on the correlation offset pattern, a multidimensional adjustment vector is generated, which includes adjustment components for each single metric and correlation weights indicating the mutual influence between the metrics.

[0110] The multidimensional adjustment vector is used as a classification boundary adjustment parameter for dynamic correction of the discrimination criterion.

[0111] Specifically, when the continuous change trend of the measured distribution characteristics exceeds the preset fluctuation threshold, the change of each measure is no longer viewed in isolation. Instead, statistical methods (such as correlation analysis, principal component analysis, or cluster analysis) are used to conduct cross-analysis of the change trends of the pixel value change degree measure, intensity change direction uniformity measure, and geometric regularity measure to explore whether there are correlations such as synchronous change, reverse change, or lag change among these measures, so as to identify the extent to which these measures influence each other and what type of process drift or defect mode they all point to.

[0112] Extracting the correlational shift patterns between multiple measures can be understood as identifying specific variation patterns in combinations of measures based on cross-analysis. For example, a significant increase in the degree of pixel value variation while a slight decrease in the geometric regularity measure might indicate a specific type of surface roughness defect. These patterns can be predefined or learned from historical data using machine learning algorithms. The aim is to simplify complex measure variations into recognizable, physically meaningful patterns.

[0113] In practical applications, a vector containing multiple components is constructed based on the identified correlation offset patterns. This multidimensional adjustment vector not only includes adjustment components for each individual metric (such as the degree of pixel value variation, the uniformity of intensity variation direction, and the geometric regularity metric), but also includes correlation weights indicating the mutual influence between these metrics. For example, if a certain offset pattern indicates that the degree of pixel value variation has a greater influence on defect classification, its corresponding adjustment component and correlation weight will be increased accordingly. The aim is to provide a comprehensive and fine-grained set of adjustment parameters that can simultaneously consider the independent changes of each metric and their interactions.

[0114] Furthermore, using the multidimensional adjustment vector as a classification boundary adjustment parameter for dynamic correction of the discrimination criterion means directly applying the generated multidimensional adjustment vector to the classification boundary in the preset discrimination criterion. For example, the classification boundary can be defined by a multidimensional hyperplane, and the multidimensional adjustment vector can be used to translate, rotate, or deform this hyperplane to more accurately adapt to the current defect feature distribution on the wafer surface. The aim is to achieve refined and adaptive adjustment of the classification boundary, thereby improving the accuracy and robustness of defect classification.

[0115] This application's scheme, through cross-analysis of the changing trends of pixel value variation measures, intensity variation direction uniformity measures, and geometric regularity measures, enables a more comprehensive understanding of the evolution patterns of defect features. It is precisely because of the extraction of the correlation offset patterns between multiple measures that the system can identify complex combinations of defect features that a single measure cannot reveal. By generating a multi-dimensional adjustment vector containing single-measure adjustment components and correlation weights, this application's scheme can more accurately quantify the impact of different measures on the classification boundary and consider their interactions. This refined adjustment parameter ensures that the dynamic correction of the discrimination criterion is no longer a simple linear adjustment, but rather an adaptive, multi-dimensional optimization based on actual process drift and defect evolution patterns, thus effectively solving the problem of insufficient classification accuracy of traditional methods under complex defect patterns.

[0116] Through the above technical solution, this application can achieve refined generation of classification boundary adjustment parameters, overcoming the limitations of traditional methods that adjust based on only a single metric or a simple combination of metrics. By considering the correlation and offset patterns between multiple metrics, the generated adjustment parameters can more accurately reflect actual defect evolution and process drift, thereby significantly improving the accuracy and adaptability of dynamic correction of the discrimination criteria. Therefore, in semiconductor polished surface inspection, it can more effectively distinguish between background anomaly areas and areas requiring reporting defects, reducing false alarms and false negatives, and improving the overall performance and reliability of the inspection system.

[0117] As a specific implementation method, it is assumed that over a period of time, the system monitors a continuous upward trend in the pixel value variation measure of the polished surface of a semiconductor wafer, while the intensity variation direction uniformity measure slightly decreases, and the geometric regularity measure remains relatively stable. In traditional adjustment methods, the classification boundary may be adjusted solely based on the change in the pixel value variation measure. However, in the scheme of this application, the changing trends of these three measures are first cross-analyzed. The analysis reveals that this combination of "increasing pixel value variation measure + decreasing intensity variation direction uniformity measure" is usually associated with a specific microscopic scratch defect, which may not significantly affect geometric regularity in the initial stage.

[0118] Based on this correlational offset pattern, the system generates a multidimensional adjustment vector. This vector may contain a large positive component to adjust the classification threshold for the pixel value variation measure, a small negative component to adjust the classification threshold for the intensity variation direction uniformity measure, and assign a specific correlation weight between the pixel value variation measure and the intensity variation direction uniformity measure to reflect their synergistic effect in identifying this type of scratch defect. For example, this multidimensional adjustment vector may indicate that when the pixel value variation measure reaches a certain threshold and the intensity variation direction uniformity measure is below a certain threshold, even if the geometric regularity measure has not changed significantly, it should be more likely to be classified as a reportable defect.

[0119] Ultimately, this multidimensional adjustment vector is used to dynamically correct the classification boundary in the discrimination criterion. For example, if the classification boundary is a hyperplane in three-dimensional space, this vector will guide the hyperplane to translate and / or rotate, enabling it to more accurately classify regions with the aforementioned characteristics of "increased pixel value variation + decreased intensity variation direction uniformity" as reported defect regions, while avoiding misclassification as background anomalies. In this way, even when faced with complex, multi-feature associated defect patterns, the system can achieve high-precision adaptive classification.

[0120] Specifically, the step of extracting connected regions of pixels with gray values ​​higher than the background area in the wafer surface image and marking them as preliminary anomalous regions can be further refined into the following operations:

[0121] The wafer surface image is subjected to grayscale normalization and background denoising.

[0122] Based on an adaptive grayscale threshold, a set of pixels with grayscale values ​​higher than the background area is selected.

[0123] Connectivity analysis is performed on the pixel set to remove tiny regions with an area smaller than a preset size, generating preliminary abnormal regions.

[0124] The process involves grayscale normalization of the wafer surface image to eliminate inconsistencies in image brightness caused by uneven illumination or differences in sensor response. This ensures that the pixel grayscale values ​​in the image are within a standardized range, providing a consistent basis for subsequent thresholding. Background denoising removes irrelevant noise points or textures from the image, such as interference caused by dust, optical scattering, or electronic noise, ensuring that any abnormal areas identified later are real and not noise artifacts.

[0125] Furthermore, an adaptive grayscale threshold is used to filter out a set of pixels with grayscale values ​​higher than the background area. Unlike a globally fixed threshold, the adaptive grayscale threshold dynamically adjusts the threshold based on the grayscale distribution characteristics of local image regions, thus better adapting to potential local brightness variations or background complexity on the wafer surface. In this way, potentially anomalous pixels can be more accurately separated from the background, forming a preliminary pixel set.

[0126] Furthermore, connected component analysis is performed on the pixel set to identify and aggregate spatially connected pixels, forming independent connected regions. During this process, tiny regions smaller than a preset size are removed. These tiny regions are typically considered noise or minor defects with no practical detection significance. Removing them effectively reduces false alarms and focuses on anomalous regions of a certain size and importance, ultimately generating preliminary anomalous regions.

[0127] The proposed solution first preprocesses the wafer surface image, including grayscale normalization and background denoising, to provide high-quality image data for subsequent anomaly region identification. Grayscale normalization ensures uniform image brightness, while background denoising effectively suppresses interference. Based on this, an adaptive grayscale threshold is used to flexibly adapt to local image characteristics, accurately filtering out pixels with grayscale values ​​higher than the background, avoiding under-segmentation or over-segmentation problems that may occur with a fixed threshold. Finally, connected component analysis is used to aggregate these filtered pixels into meaningful regions and remove minor noise regions, ensuring that the initially identified anomaly regions have sufficient size and connectivity, thereby improving the accuracy and reliability of anomaly region identification.

[0128] Through the above technical solutions, this application can significantly improve the accuracy and robustness of preliminary anomaly region identification. Gray-level normalization and background denoising effectively eliminate the influence of environmental and equipment factors on image quality, making subsequent analysis more reliable. The application of adaptive gray-level thresholds enables the system to flexibly respond to changes in local brightness and texture on the wafer surface, improving the detection sensitivity for different types of anomalies. Simultaneously, connected component analysis combined with micro-region removal effectively filters out noise and unimportant small artifacts, reducing the false alarm rate and ensuring that the identified preliminary anomaly regions more accurately reflect the potential defects on the wafer surface, providing more accurate and reliable input for subsequent defect classification.

[0129] like Figure 2 As shown, an exemplary semiconductor polished surface inspection system based on industrial vision is illustrated. Specific embodiments of this application also disclose a semiconductor polished surface inspection system 100 based on industrial vision, comprising:

[0130] Image acquisition module 10 acquires images of the polished surface of a semiconductor wafer under preset lighting conditions;

[0131] The preliminary anomaly region identification module 20 extracts the connected components of pixels with gray values ​​higher than those of the background region in the wafer surface image and marks them as preliminary anomaly regions.

[0132] The measurement determination module 30 calculates the pixel value change degree measure, intensity change direction uniformity measure, and geometric regularity measure for each preliminary abnormal region.

[0133] The classification module 40 classifies the preliminary abnormal region based on the pixel value change degree measure, the intensity change direction uniformity measure, and the geometric regularity measure, and divides the background abnormal region and the defect region to be reported by using a preset discrimination criterion.

[0134] The information output module 50 is used to output the location and shape information of the defect area to be reported, and to mask the detection data of the background abnormal area.

[0135] This application presents a semiconductor polished surface inspection system based on industrial vision. Through its unique modular design and multi-dimensional feature analysis capabilities, it effectively solves the dilemma of traditional inspection methods in distinguishing visually similar surface anomalies such as background glow and chemical residue marks. Traditional methods often suffer from high false alarm rates or high false negative rates because they cannot accurately distinguish between these two types of anomalies. For example, simple brightness threshold judgment may misjudge harmless background glow as a defect, while background subtraction techniques may eliminate real defects as well.

[0136] This application's system achieves comprehensive and in-depth analysis of wafer surface anomalies through the collaborative work of an image acquisition module, a preliminary anomaly region identification module, a measurement determination module, a classification module, and an information output module. In particular, the measurement determination module can extract three complementary dimensions of information: the degree of pixel value variation, the uniformity of intensity variation direction, and geometric regularity. The classification module then performs intelligent discrimination based on these multi-dimensional measurements. This system architecture allows detection to move beyond relying solely on single grayscale or brightness information, comprehensively considering the texture details, directional features, and shape characteristics within the region.

[0137] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A semiconductor polished surface inspection method based on industrial vision, characterized in that, include: Acquire images of the polished surface of a semiconductor wafer under preset lighting conditions; Extract connected regions of pixels with grayscale values ​​higher than the background area from the wafer surface image and mark them as preliminary abnormal regions; For each initial abnormal region, calculate the measure of pixel value variation, the measure of uniformity of intensity variation direction, and the measure of geometric regularity; Based on the pixel value change degree measure, the intensity change direction uniformity measure, and the geometric regularity measure, the preliminary abnormal area is classified by a preset discrimination criterion to divide the background abnormal area and the defect area to be reported. The output should report the location and shape information of the defective area, and mask the detection data of the background abnormal area.

2. The semiconductor polished surface inspection method according to claim 1, characterized in that, The steps for calculating the pixel value variation measure, intensity variation direction uniformity measure, and geometric regularity measure for each initial abnormal region include: Multi-scale morphological filtering is performed on the image data corresponding to the preliminary abnormal region to separate fine structure and diffuse structure at different spatial scales, resulting in multiple sets of filtered images. Based on multiple sets of filtered images, the degree of pixel value change, the uniformity of intensity change direction, and the geometric regularity of the preliminary abnormal region are calculated simultaneously.

3. The semiconductor polished surface inspection method according to claim 2, characterized in that, The measure of pixel value change in the initial abnormal region is calculated based on multiple sets of filtered images, specifically including: The initial abnormal region is divided into multiple detection sub-regions; For each detection sub-region, calculate the local dispersion coefficient of its pixel grayscale value; By fusing the local discrete coefficients of the multiple detection sub-regions, a measure of the degree of pixel value change in the preliminary abnormal region is determined.

4. The semiconductor polished surface inspection method according to claim 2, characterized in that, The uniformity of intensity variation direction in the preliminary anomalous region is calculated based on multiple sets of filtered images, specifically including: The initial abnormal region is divided into multiple gradient detection regions; For each gradient detection region, calculate the local concentration of intensity change direction; Statistical analysis of the local concentration in all gradient detection regions yields a measure of the uniformity of intensity change direction in the preliminary anomaly region.

5. The semiconductor polished surface inspection method according to claim 2, characterized in that, The initial geometric regularity measure of the abnormal region is calculated based on multiple sets of filtered images, specifically including: The initial abnormal region is divided into multiple contour detection units; For each contour detection unit, calculate the local smoothness and local compactness of its boundary; By analyzing the distribution characteristics of the local smoothness and local compactness of the multiple contour detection units, a geometric regularity measure of the preliminary abnormal region is obtained.

6. The semiconductor polished surface inspection method according to claim 1, characterized in that, The step of classifying the preliminary abnormal region based on the pixel value change degree measure, the intensity change direction uniformity measure, and the geometric regularity measure, and dividing it into background abnormal region and reportable defect region by means of a preset discrimination criterion, includes: Continuously monitor the real-time distribution of the pixel value change measure, the intensity change direction uniformity measure, and the geometric regularity measure; The real-time distribution is compared with a preset benchmark distribution to identify the distribution deviation characteristics between the real-time distribution and the preset benchmark distribution; Based on the degree and direction of the distribution deviation features, the classification boundary in the preset discrimination criterion is corrected; When the measurement distributions of the background anomaly area and the defect area to be reported overlap, the classification threshold of the overlapping area is adjusted based on historical detection statistics to separate the background anomaly area and the defect area to be reported.

7. The semiconductor polished surface inspection method according to claim 6, characterized in that, The steps for continuously monitoring the real-time distribution of the pixel value change measure, the intensity change direction uniformity measure, and the geometric regularity measure include: A fixed-duration sliding time window is set, and within this sliding time window, the degree of change of the pixel value, the uniformity of the direction of intensity change, and the geometric regularity are periodically monitored. The pixel value change degree measure, intensity change direction uniformity measure, and geometric regularity measure collected within the sliding time window are statistically fitted to generate the measure distribution characteristics of the current sliding time window. The measurement distribution characteristics of the current sliding time window are compared with those of the previous multiple sliding time windows to identify the continuous changing trend of the measurement distribution characteristics. When the continuous change trend exceeds the preset fluctuation threshold, an early warning signal is issued, and adjustment parameters indicating the classification boundary are generated.

8. The semiconductor polished surface inspection method according to claim 7, characterized in that, The steps of issuing a warning signal and generating adjustment parameters indicating the classification boundary when the continuous change trend exceeds a preset fluctuation threshold include: Cross-analysis is performed on the changing trends of the pixel value change degree measure, the intensity change direction uniformity measure, and the geometric regularity measure to extract the correlation offset patterns between multiple measures; Based on the correlation offset pattern, a multidimensional adjustment vector is generated, which includes adjustment components for a single metric and correlation weights indicating the mutual influence between the metrics. The multidimensional adjustment vector is used as a classification boundary adjustment parameter for dynamic correction of the discrimination criterion.

9. The semiconductor polished surface inspection method according to claim 1, characterized in that, The step of extracting connected regions of pixels with gray values ​​higher than those of the background region from the wafer surface image and marking them as preliminary abnormal regions specifically includes: The wafer surface image is subjected to grayscale normalization and background denoising. Based on an adaptive grayscale threshold, a set of pixels with grayscale values ​​higher than the background area is selected. Connectivity analysis is performed on the pixel set to remove tiny regions with an area smaller than a preset size, generating preliminary abnormal regions.

10. A semiconductor polished surface inspection system based on industrial vision, characterized in that, The system includes: The image acquisition module acquires images of the polished surface of a semiconductor wafer under preset lighting conditions. The preliminary anomaly region identification module extracts connected components of pixels with gray values ​​higher than those of the background region from the wafer surface image and marks them as preliminary anomaly regions. The measurement determination module calculates the degree of pixel value change, the uniformity of intensity change direction, and the geometric regularity for each preliminary abnormal region. The classification module classifies the preliminary abnormal region based on the pixel value change degree measure, the intensity change direction uniformity measure, and the geometric regularity measure, and divides it into background abnormal region and defect region to be reported. The information output module is used to output the location and shape information of the defective area to be reported, and to mask the detection data of the background abnormal area.

Images