A semiconductor wafer visual inspection method based on interactive feature learning

Abstract

The application provides a kind of semiconductor wafer visual inspection method based on interactive feature learning, comprising: including model training phase and online detection phase;The model training phase includes the following steps: inputting the digital image of semiconductor wafer;Target area is manually framed from the digital image through human-computer interaction interface;The target area image is pretreated to obtain the optimized image;Contour edge detection and feature vector extraction are carried out on the optimized image;Based on the extracted feature parameters, a standard parameter set is constructed and stored in the database;Train AI model to identify imaging abnormal categories, and establish an abnormal category database;The online detection phase includes the following steps: new image samples are analyzed using the trained AI model, and if an anomaly is detected, human intervention is required;If no anomaly is detected, the standard parameter set is used to identify the features of the new image sample.

Description

Technical Field

[0001] This invention relates to the fields of industrial visual inspection and semiconductor manufacturing technology, specifically to a semiconductor wafer visual inspection method based on interactive feature learning. Background Technology

[0002] In semiconductor manufacturing, feature inspection of the wafer surface is a critical step in ensuring product quality. As the feature size of integrated circuits continues to shrink, higher demands are placed on the accuracy and efficiency of inspection technologies. Traditional automated optical inspection systems typically employ fixed image processing algorithms, which suffer from poor adaptability and complex parameter adjustments.

[0003] In existing technologies, most detection schemes employ fully automated region localization methods, such as template matching and edge detection algorithms. While these methods can achieve automated detection under specific conditions, they often exhibit instability when faced with complex wafer surface features, varying lighting conditions, and different product types. Automated localization algorithms are susceptible to noise interference, changes in illumination, and other factors, which in turn affect the subsequent feature recognition performance.

[0004] Furthermore, fully automated inspection systems lack the flexibility of human-machine interaction and cannot fully utilize the operator's domain knowledge. In practical applications, operators can often quickly and accurately identify key feature areas based on their experience, but existing systems fail to effectively leverage this advantage.

[0005] Therefore, improving the accuracy of semiconductor wafer feature recognition is a problem that needs to be solved. Summary of the Invention

[0006] The purpose of this invention is to propose a semiconductor wafer visual inspection method based on interactive feature learning, which can improve the accuracy of semiconductor wafer feature recognition.

[0007] To achieve the above objectives, the present invention provides a semiconductor wafer visual inspection method based on interactive feature learning, comprising:

[0008] It includes a model training phase and an online detection phase; the model training phase includes the following steps:

[0009] Input a digital image of a semiconductor wafer;

[0010] The target area can be manually selected from the digital image through a human-computer interaction interface;

[0011] The target region image is preprocessed to obtain an optimized image;

[0012] The optimized image is then subjected to contour edge detection and feature vector extraction.

[0013] A standard parameter set is constructed based on the extracted feature parameters and stored in the database;

[0014] Train an AI model to identify imaging anomaly categories and build an anomaly category database;

[0015] The online detection phase includes the following steps: using a trained AI model to analyze new image samples; if an anomaly is detected, manual intervention is performed; if no anomaly is detected, the standard parameter set is used to perform feature recognition on the new image samples.

[0016] In an optional embodiment, the preprocessing includes a binarization step, wherein the binarization includes:

[0017] Peak-valley detection based on gray-level histogram automatically identifies the number of peaks and valleys K in the gray-level distribution as the threshold quantity;

[0018] The grayscale range of 0-255 is divided into K+1 intervals according to the peak position. An optimal threshold is output for each interval, forming a set of multiple thresholds to achieve simultaneous segmentation of multiple targets.

[0019] The inter-class variance is calculated within ±n gray values ​​near the peak, where n is a natural number, in order to compress the search space.

[0020] For each interval, the candidate threshold is calculated using the following formula:

[0021]

[0022] Where Score(t) is the candidate threshold, =1, determined through sample training. , The value, For inter-class variance, The mean of local gray-level entropy. The global grayscale entropy;

[0023] Compare the candidate thresholds for each interval, and select the one with the largest candidate threshold as the optimal segmentation threshold;

[0024] Binarization is performed according to the THRESH_BINARY formula.

[0025] Among the optional solutions, and The calculation formula is as follows:

[0026] =

[0027] in, Grayscale levels, ranging from 0 to 255; grayscale The probability of appearing in the entire image.

[0028]

[0029] Where R is a gray-level interval divided by the candidate threshold. This represents the number of gray levels contained within the interval R; This represents the subinterval formed after discretizing the interval R;

[0030] Represents grayscale level The conditional probability of occurrence within the interval R. .

[0031] In an optional implementation, the contour edge detection is performed using the Canny operator, including the following steps:

[0032] Image convolution denoising is performed using a two-dimensional Gaussian function;

[0033] The magnitude and direction of the image gradient are calculated using the Sobel operator.

[0034] Preserve local gradient maxima along the gradient direction to refine edges;

[0035] Set high and low thresholds to classify and link edge points to generate continuous contours.

[0036] In an optional scheme, the feature vector extraction includes:

[0037] Based on the contour output by contour edge detection, contour features are extracted, including contour perimeter, contour area and number of contour vertices.

[0038] By debugging the software, the threshold ranges of the contour perimeter, contour area, and number of contour vertices are limited to filter out the target contour.

[0039] Feature parameters are extracted from the target contour to obtain the feature vector.

[0040] In an optional embodiment, the number of contour vertices is the number of vertices obtained by approximating the contour with polygons.

[0041] In an optional approach, the step of performing feature recognition on new image samples using the standard parameter set includes:

[0042] The similarity distance between the contour features extracted from the image to be tested and the standard parameter set in the database in the multidimensional feature space is calculated, and the distance is compared with a tolerance threshold to determine whether they match.

[0043] In the optional solution, the categories of the anomalies include overexposure, underexposure, out-of-focus blur, motion blur, and foreign object obstruction.

[0044] In an optional embodiment, after the feature recognition step, there is also a step of outputting recognition results, which include one or more of the following: the existence of target features, location coordinates, type identifier, or confidence level.

[0045] In an optional embodiment, the AI ​​model is a discriminative model, which is trained using the geometric feature parameters of historical wafer images and their corresponding artificial quality rating labels to identify and classify known defect patterns.

[0046] The beneficial effect of this invention is that it improves the accuracy of semiconductor wafer feature identification. Attached Figure Description

[0047] The above and other objects, features and advantages of the present invention will become more apparent from the accompanying drawings, in which like reference numerals generally denote like parts.

[0048] Figure 1 This is a flowchart of a semiconductor wafer visual inspection method based on interactive feature learning in one embodiment of the present invention. Detailed Implementation

[0049] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. The advantages and features of the present invention will become clearer from the following description and drawings. However, it should be noted that the concept of the technical solution of the present invention can be implemented in many different forms and is not limited to the specific embodiments described herein. The accompanying drawings are all in a very simplified form and use non-precise proportions, and are only used to facilitate and clarify the illustration of the embodiments of the present invention.

[0050] It should be understood that when an element or layer is referred to as "on," "adjacent to," "connected to," or "coupled to" other elements or layers, it may be directly on, adjacent to, connected to, or coupled to other elements or layers, or there may be intervening elements or layers. Conversely, when an element is referred to as "directly on," "directly adjacent to," "directly connected to," or "directly coupled to" other elements or layers, there are no intervening elements or layers. It should be understood that although the terms first, second, third, etc., may be used to describe various elements, components, areas, layers, and / or portions, these elements, components, areas, layers, and / or portions should not be limited by these terms. These terms are only used to distinguish one element, component, area, layer, or portion from another element, component, area, layer, or portion. Therefore, without departing from the teachings of this invention, the first element, component, area, layer, or portion discussed below may be referred to as the second element, component, area, layer, or portion.

[0051] Spatial relation terms such as “below,” “under,” “below,” “under,” “above,” “above,” etc., are used herein for convenience of description to describe the relationship between one element or feature shown in the figure and other elements or features. It should be understood that, in addition to the orientation shown in the figure, spatial relation terms are intended to also include different orientations of the device in use and operation. For example, if the device in the figure is flipped, then the element or feature described as “below” or “under” the other element or feature will be oriented “above” the other element or feature. Therefore, the exemplary terms “below” and “under” can include both upper and lower orientations. The device may be otherwise oriented (rotated 90 degrees or otherwise) and the spatial descriptive terms used herein will be interpreted accordingly.

[0052] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the invention. When used herein, the singular forms “a,” “an,” and “the” are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the terms “comprising” and / or “including,” when used in this specification, identify the presence of the stated features, integers, steps, operations, elements, and / or components, but do not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups. When used herein, the term “and / or” includes any and all combinations of the associated listed items.

[0053] Example

[0054] Reference Figure 1 This embodiment provides a semiconductor wafer visual inspection method based on interactive feature learning, including a model training stage and an online detection stage; the model training stage includes the following steps:

[0055] Input a digital image of a semiconductor wafer;

[0056] The target area can be manually selected from the digital image through a human-computer interaction interface;

[0057] The target region image is preprocessed to obtain an optimized image;

[0058] The optimized image is then subjected to contour edge detection and feature vector extraction.

[0059] A standard parameter set is constructed based on the extracted feature parameters and stored in the database;

[0060] Train an AI model to identify imaging anomaly categories and build an anomaly category database;

[0061] The online detection phase includes the following steps: using a trained AI model to analyze new image samples; if an anomaly is detected, manual intervention is performed; if no anomaly is detected, the standard parameter set is used to perform feature recognition on the new image samples.

[0062] Specifically, the model training phase includes the following steps:

[0063] S1: Image acquisition and input;

[0064] S2: Interactive target region determination;

[0065] S3: Image preprocessing;

[0066] S4: Multi-dimensional feature extraction;

[0067] S5: Construction of standard parameter set;

[0068] S6: AI model learning and the establishment of an anomaly category database.

[0069] The specific steps in S1 are as follows: an industrial camera is used to acquire digital images of the semiconductor wafer, and a coaxial light source system is used to eliminate surface reflection interference to ensure that the image quality meets the requirements of subsequent processing.

[0070] The specific steps of S2 are as follows: The target area is manually selected through a graphical user interface (GUI). This manual selection of the ROI (Region of Interest) in the captured image eliminates irrelevant background interference. The selection operation can be a rectangular selection, an irregular polygonal selection, or an outline drawing operation.

[0071] The S3 steps consist of three sub-steps:

[0072] S301: Adaptive binarization is performed using the improved Otsu's Method formula.

[0073] S302: Perform a dilation operation using a rectangular structuring element, with the number of iterations defaulting to 1 time.

[0074] S303: Perform an erosion operation using the same structuring element, with the number of iterations defaulting to 1 time. This preprocessing sequence constitutes a morphological closing operation, which can effectively connect feature breaks and smooth the boundaries.

[0075] The specific operation of S301 is as follows: First, display the ROI of the target area selected by the user, and then perform binarization on the ROI. This binarization process is based on the THRESH_BINARY formula. Among them, in OpenCV, the THRESH_BINARY is the most basic and commonly used thresholding method, and its core formula and logic are very clear. The purpose is to convert a grayscale image (pixel value range 0 - 255) into a binary image containing only two values (usually 0 and 255). The mathematical definition of the THRESH_BINARY formula is as follows. For any pixel point in a grayscale image, its grayscale value is (range 0 - 255). Set a threshold and a maximum value , then the pixel value after binarization is:

[0076]

[0077] Adaptive binarization is based on an improved Otsu algorithm. The Otsu algorithm first calculates the grayscale histogram and counts the number of occurrences of the grayscale value (0 - 225) of each pixel; then traverses all possible thresholds T, from T = 0 to T = 255, and calculates the between-class variance one by one; finds the T corresponding to the maximum cumulative variance, and this T is the optimal threshold; finally, applies it to threshold segmentation, sets the pixels with grayscale value T to white and the pixels with grayscale value < T to black to obtain a binarized image.

[0078] The modified between-class variance formula of the traditional Otsu algorithm is .

[0079] We first select a candidate threshold T, which divides all pixels of the image into two categories:

[0080] Class C0: Pixel grayscale value (usually regarded as foreground or target)

[0081] Class C1: Pixel grayscale value > T (usually regarded as background)

[0082] : The total number of pixels in class C0

[0083] Total number of pixels in C1

[0084] The total number of pixels in the image, i.e. = +

[0085] Average grayscale value of all pixels in class C0

[0086] Average grayscale value of all pixels in class C1

[0087] The total average gray value of all pixels in the entire image.

[0088] The traditional Otsu algorithm has several drawbacks: it only outputs a single fixed threshold, which cannot segment multiple targets in a grayscale image (such as multiple parts, defects, and background); and it calculates the variance by traversing all grayscale values ​​(0-255), which is inefficient. In the case of multiple thresholds, it requires exhaustive search in a huge space, which is computationally expensive and makes it difficult to meet the needs of real-time monitoring.

[0089] In this embodiment, the Otsu algorithm is modified by dynamically determining the threshold number: a peak-valley detection method based on the grayscale histogram automatically identifies the number K of several peaks and valleys in the grayscale distribution, replacing the traditional manually preset K value. Specifically, a maximum suppression algorithm is used to perform modal analysis on the grayscale image histogram. Noise is suppressed by smoothing the histogram, and then local maxima are detected. The number of these maxima is used as the dynamically determined threshold number K. This enables the algorithm to self-configure based on the image.

[0090] Refined threshold calculation within intervals: Divide the grayscale range (0-255) into K+1 intervals according to the peak position (e.g., if the peak is at 50 or 150, divide it into [0-50], [50-150], [150-255]). Output a candidate threshold for each interval, forming a multi-threshold set (e.g., T1=30). T2=180 (T3=200) enables simultaneous segmentation of multiple targets. This ensures that the selected threshold not only separates grayscale differences but also captures defect areas with unique texture patterns, thereby improving segmentation accuracy at the fundamental level.

[0091] Threshold search pruning: only near the peak The variance is calculated within 5 grayscale values ​​(e.g., peak value is 50, search range is 45-55), compressing the search space from 255 to 20-50. This does not change the optimization objective of maximizing inter-class variance, but rather greatly improves the efficiency of achieving this objective by intelligently limiting its search space, thus meeting the requirements of industrial real-time performance.

[0092] For each interval, the inter-class variance formula is modified (incorporating grayscale structure information within the interval): Improved formula: ( =1, determined through sample training, for example , (This section emphasizes the differences in local features within the enhanced interval); synergy with traditional principles: the calculation of the inter-class variance formula is the root cause of time consumption. This improvement is based on the prior knowledge of image statistics that "the optimal segmentation threshold must be located between peaks and valleys," transforming global exhaustive search into local precise search. Among other things... Global gray-level entropy is used to measure the uncertainty and information content of the gray-level distribution of the entire image.

[0093] =

[0094] in, Gray levels, ranging from 0 to 255. Gray level The probability of it appearing in the entire image, its value is , It is the local gray-level entropy mean, used to measure the average texture complexity within a certain gray-level range defined at a specific candidate threshold t.

[0095]

[0096] R: A gray-level interval divided by the candidate threshold t : Represents the number of gray levels contained within the interval R. : Represents the subinterval formed after discretizing the interval R. Gray level The conditional probability of an event occurring within the interval R is denoted as . At this point, the optimal candidate segmentation threshold within the interval is obtained. :

[0097]

[0098] The most crucial improvement is the synergy between the modified inter-class variance and the traditional principle. We recognize that pure inter-class variance may fail in regions with complex textures; therefore, we do not abandon its principles but rather optimize and enhance them.

[0099] Complete the candidate thresholds within all intervals After calculation, a comparison is performed, and the one that makes the comparison is selected. The candidate threshold with the largest value is the optimal threshold. Determine the local optimal segmentation threshold for this ROI, and segment it according to the THRESH_BINARY formula.

[0100] The specific operation of S302 is as follows:

[0101] The binarized Region of Interest (ROI) is then dilated. Dilation refers to "sliding" the image matrix through a structuring element matrix to expand the foreground region. As long as a structuring element overlaps with at least one foreground (255, white) pixel, the anchor point will be "expanded" into the foreground. The following uses a pixel value matrix of 0 (black, background) and 255 (white, foreground) to visually demonstrate the principle and effect of dilation.

[0102] The core logic of the dilation is based on the judgment of pixel values ​​0 and 255. For example, consider a pixel matrix before dilation (5... 5) After dilation, it will become this pixel matrix (5 5).

[0103]

[0104] Therefore, the mathematical essence of inflation is the maximum value operation of the submatrix:

[0105]

[0106] The specific operation of S303 is as follows:

[0107] After the region of interest (ROI) has been dilated, an erosion operation is performed. The core of erosion is to perform a "sliding match" on the image matrix using a structuring element matrix, shrinking the foreground region and aligning the anchor points of the structuring elements with each pixel of the image matrix. If all the positions of "1" in the structuring element correspond to a pixel value of 255 in the image matrix (i.e., the structuring element is completely embedded in the white foreground), then the pixel is retained as 255 after erosion; otherwise, it becomes 0 (black background).

[0108] Therefore, the mathematical essence of corrosion is the operation of maximizing the submatrix:

[0109]

[0110] The specific operation of S4 is as follows: edge detection is implemented using OpenCV's Canny operator. The formula and principle of edge extraction by the Canny operator are divided into four core stages, each with a clear mathematical logic: S401: Gaussian smoothing (noise reduction stage); S402: gradient calculation (edge ​​strength and direction); S403: non-maximum suppression (edge ​​thinning); S404: double threshold detection and edge linking.

[0111] In stage S401, the Gaussian smoothing formula used is a two-dimensional Gaussian function, and its mathematical expression is:

[0112]

[0113] in, It is the standard deviation of the Gaussian kernel (which determines the smoothness). The larger the value, the better the smoothing effect. These are the relative coordinates within the Gaussian kernel. Convolving the Gaussian kernel with the original image achieves noise suppression; the formula is:

[0114]

[0115] Where * indicates convolution calculation. These are the pixel values ​​of the original image.

[0116] In stage S402, gradient calculation requires using the Sobel operator to calculate the gradient magnitude and direction of each pixel. The formula for the gradient magnitude is:

[0117]

[0118] ( yes directional gradient, yes (directional gradient)

[0119] The gradient direction formula is:

[0120]

[0121] Quantize the gradient direction as Four directions facilitate subsequent processing.

[0122] and Obtained by convolving the Sobel kernel with the smoothed image:

[0123] ,

[0124] In the S403 stage, nonmaximum suppression refers to suppression along the gradient direction. It determines whether the current pixel is at a local gradient maximum; if not, it suppresses the gradient. The logical formula is: for each pixel... Find two adjacent pixels along the gradient direction and ,like and Then retain Otherwise set .

[0125] In the S404 stage, dual threshold detection and edge linking refer to setting a high threshold. and low threshold (generally =2 ), classifying edge points: if > Marked as a strong edge, it can be preserved if... < < Marked as a weak edge, it needs to be checked whether it is connected to a strong edge. If connected, it is preserved; otherwise, it is suppressed. < Direct inhibition.

[0126] After the above operations, a certain number of feature contours can be extracted. Combined with the general information of the contours (contour perimeter, contour area, etc.), the approximate information of the target area features can be roughly determined.

[0127] S5: Standard parameter set construction, the specific steps are as follows:

[0128] S501: Image Acquisition and Input: Acquire a series of sample product images with known states (qualified or clearly labeled with abnormal types) to form a training set, and input them into the system. S502: Interactive Target Region Determination: For each sample image, specify the component or contour region to be detected through the human-computer interaction interface. The system records the coordinate information of this region for subsequent feature extraction. S503: Multi-dimensional Feature Extraction: For the target region determined in S502, automatically extract multiple features of its contour to form a high-dimensional feature vector. S504: Intelligent Feature Filtering and Standard Parameter Set (Template) Construction:

[0129] Feature selection: Using feature evaluation algorithms (such as feature importance ranking based on ANOVA or random forest), the most effective feature subset for distinguishing between qualified and abnormal states is selected from the multi-dimensional features extracted by S503.

[0130] Template construction: Based on the filtered feature subset, calculate the statistical center of the feature vectors of all qualified samples, and define the center vector as the standard parameter set or standard template of the test item.

[0131] S6: AI model learning and anomaly category database establishment, the specific operations are as follows:

[0132] Model learning: A classification model is trained using the feature vectors of samples and their labels (qualified / abnormal categories) as training data. This model learns the mapping relationship from the feature space to the state category.

[0133] Database establishment: The trained AI model parameters, standard parameter sets, and feature vector sets of various abnormal samples are stored together to form an abnormal feature database, which is used for comparison and inference in the online detection stage.

[0134] Training the AI ​​model involves training a dedicated deep learning model to achieve sub-pixel-level precise localization and geometric analysis of lithographic alignment marks (crosshairs, squares, etc.), specifically including:

[0135] S601, Data Annotation: Collect wafer images containing various alignment marks, use pixel-level instance segmentation technology to annotate the precise outline of each mark, and associate it with its design coordinates.

[0136] S602. Network Architecture: Construct a two-branch neural network. The backbone network adopts HRNet to maintain high-resolution features. The first branch outputs a pixel-level segmentation mask with labels, and the second branch outputs a heatmap with labels for key points (center and corners).

[0137] S603, Training Optimization:

[0138] 1. Loss function combination: Segmentation loss (Dice Loss) + Keypoint detection loss (Focal Loss variant) + Geometric constraint loss (enforcing conditions such as symmetry of the labels).

[0139] 2. Data Augmentation: Specifically simulates overlay errors (minor translation / rotation), process variations (edge ​​erosion, defocusing), and sensor noise to enhance model robustness.

[0140] 3. Two-stage training: First, use synthetic data to learn the ideal geometric structure, and then use real data to fine-tune it to adapt to process fluctuations.

[0141] 4. Training Cycle: Set the number of training epochs to 100 and the batch size to 32. In each training epoch, calculate the loss and compare the output with the true label; thereby optimizing the parameters and updating the weight function.

[0142] 5. Training monitoring: Record the loss and accuracy curves of the training set and parameter set. Terminate training when the accuracy of the validation set no longer improves for ten consecutive rounds.

[0143] 6. Model saving: Save the trained network weights as a .h5 or .pth format file for subsequent auxiliary detection.

[0144] S604, Performance Standards: The model must meet the requirements of localization repeatability (3σ) < 3 nm and edge contour error < 2 nm on the validation set.

[0145] Due to the diverse shapes (including lighting conditions, shooting angles, and focusing effects) and random distribution of captured wafer images, traditional thresholding or template matching methods are inadequate and struggle to handle wafer images of different shapes. The multi-scale feature fusion network approach, utilizing convolutional neural networks (CNNs), has become an important research direction in industrial vision due to its excellent feature extraction performance. Among numerous CNN models, the YOLO algorithm is widely used in various fields due to its high accuracy, speed, and adaptability.

[0146] YOLOv5 is the most widely used in industry. This example uses the lightweight YOLOv5s model because it has a fast detection speed and low resource consumption, making it very suitable as an auxiliary module that does not affect the main process.

[0147] The training objective of this model is unrelated to wafer defects; instead, it focuses on identifying problems within the imaging system itself. The detected categories are defined as follows: Overexposure: loss of detail in feature areas due to excessive lighting, either locally or globally; Underexposure: blurred areas of detail in the image due to insufficient lighting; Defocus blur: decreased overall image sharpness due to lens inaccuracy; Motion blur: ghosting caused by wafer or camera shake; Foreign object occlusion: fixed artifacts formed on the image by dust, fibers, etc., on the lens or light source.

[0148] The AI ​​model is a discriminative model. It is trained using the geometric feature parameters of historical wafer images and their corresponding artificial quality rating labels to identify and classify known defect patterns.

[0149] The YOLOv5 model was trained and labeled using a dataset containing images of the above-mentioned imaging anomalies.

[0150] "Purity" judgment: The model did not detect any predefined imaging categories in the image. This indicates that the image quality is reliable and requires no further manual review.

[0151] "Contamination" judgment: The model identifies one or more imaging aberration areas (such as an overexposed area) in the input image and outputs its category and location.

[0152] By debugging the software and limiting features such as contour perimeter and area, the selected contour is guaranteed to be the target contour. Contour features are extracted from the preprocessed image to obtain feature vectors. A contour is only identified as a target feature when its perimeter, area, and number of vertices all fall within their respective preset threshold ranges. The number of contour vertices is the number of vertices obtained after approximating the contour using polygons.

[0153] After extracting the target contour information, these parameters are then stored in the database as standard templates for this type of alignment mark.

[0154] When importing or acquiring new image samples, the AI ​​model (YOLOv5) first analyzes the image's imaging effect, then binarizes the entire image, systematically extracts all contours within the full image search area, and then automatically processes the image based on parameters stored in the database, finally filtering out target features. Images classified as B by the YOLOv5 model, after the above filtering and target feature identification, still require manual verification to ensure the correct selection of target features.

[0155] The steps for feature recognition using a standard parameter set include: calculating the similarity distance between the contour features extracted from the image to be tested and the standard parameter set in the multidimensional feature space, and comparing the distance with a tolerance threshold to determine whether a match is made.

[0156] The purpose of feature recognition using a standard parameter set is to quantitatively compare the features of the product image under test with a pre-established standard, thereby making an objective judgment. This step is achieved through the following specific and interconnected technical steps:

[0157] S1. Extraction of the feature vector to be tested

[0158] After acquiring an image of the wafer under test online, the system automatically locates the predefined region of interest (i.e., the region where the alignment mark is located). Then, using the same feature extraction algorithm and parameters as in the model training phase, a multidimensional feature vector representing the alignment mark is generated. =[f1,f2,...,fn], where n is the feature dimension.

[0159] S2. Calculation of similarity distance:

[0160] Calculate the feature vector to be measured The difference between the standard parameter set S and the standard parameter set S in the multidimensional feature space is quantified as "similarity distance". This embodiment preferably uses Euclidean distance as the metric, and its calculation formula is as follows:

[0161]

[0162] in, It is the value of the standard parameter set S on the i-th feature. The smaller the Euclidean distance D, the more similar the feature to be tested is to the standard template.

[0163] S3. Obtaining and Meaning of Tolerance Threshold

[0164] The "tolerance threshold" T is not arbitrarily set, but a dynamic quality threshold scientifically determined based on the statistical distribution of the qualified sample population. Its acquisition method is completed during the model training phase, specifically as follows:

[0165] 1. After constructing the standard parameter set S, calculate the feature vectors of all M qualified samples.

[0166] The Euclidean distance to S is obtained, resulting in the distance set. .

[0167] 2. Perform statistical analysis on this distance set and determine the tolerance threshold T using one of the following two methods:

[0168] (1) Maximum value method: This method is conservative, ensuring that all qualified samples pass.

[0169] (2) Statistical interval method: Calculate the mean μ and standard deviation σ of the distance, and take T=μ+k, where k is a constant (usually 2 or 3). This method can more effectively screen out samples with abnormal dispersion, while allowing normal fluctuations in qualified samples.

[0170] The final determined threshold T is used as a core process parameter and stored in the system configuration file or database.

[0171] S4. Matching Logic

[0172] The decision-making process is a deterministic, programmable logical comparison:

[0173] .

[0174] In this embodiment, after the feature recognition step, there is also a step of outputting the recognition result, which includes one or more of the following: the existence of the target feature, location coordinates, type identifier, or confidence level.

[0175] This embodiment improves the accuracy and efficiency of semiconductor wafer feature identification, making it suitable for industrial quality control.

[0176] The above description is merely a description of preferred embodiments of the present invention and is not intended to limit the scope of the present invention in any way. Any changes or modifications made by those skilled in the art based on the above disclosure shall fall within the protection scope of the claims.

Claims

1. A semiconductor wafer visual inspection method based on interactive feature learning, characterized in that, It includes a model training phase and an online detection phase; the model training phase includes the following steps: Input a digital image of a semiconductor wafer; The target area can be manually selected from the digital image through a human-computer interaction interface; The target region image is preprocessed to obtain an optimized image; The optimized image is then subjected to contour edge detection and feature vector extraction. A standard parameter set is constructed based on the extracted feature parameters and stored in the database; Train an AI model to identify imaging anomaly categories and build an anomaly category database; The online detection phase includes the following steps: using a trained AI model to analyze new image samples; if an anomaly is detected, manual intervention is performed; if no anomaly is detected, the standard parameter set is used to perform feature recognition on the new image samples. The preprocessing includes a binarization step, which includes: Peak-valley detection based on gray-level histogram automatically identifies the number of peaks and valleys K in the gray-level distribution as the threshold quantity; The grayscale range of 0-255 is divided into K+1 intervals according to the peak position. An optimal threshold is output for each interval, forming a set of multiple thresholds to achieve simultaneous segmentation of multiple targets. The inter-class variance is calculated within ±n gray values ​​near the peak, where n is a natural number, in order to compress the search space. For each interval, the candidate threshold is calculated using the following formula: Where Score(t) is the candidate threshold. =1, determined through sample training. , The value, For inter-class variance, The mean of local gray-level entropy. The global grayscale entropy; Compare the candidate thresholds for each interval, and select the one with the largest candidate threshold as the optimal segmentation threshold; Binarization is performed according to the THRESH_BINARY formula.

2. The semiconductor wafer visual inspection method based on interactive feature learning as described in claim 1, characterized in that, and The calculation formula is as follows: = in, Grayscale levels, ranging from 0 to 255; grayscale The probability of appearing in the entire image. Where R is a gray-level interval divided by the candidate threshold. This represents the number of gray levels contained within the interval R; This represents the subinterval formed after discretizing the interval R; Represents grayscale level The conditional probability of occurrence within the interval R. .

3. The semiconductor wafer visual inspection method based on interactive feature learning as described in claim 1, characterized in that, The contour edge detection is implemented using the Canny operator and includes the following steps: Image convolution denoising is performed using a two-dimensional Gaussian function; The magnitude and direction of the image gradient are calculated using the Sobel operator. Preserve local gradient maxima along the gradient direction to refine edges; Set high and low thresholds to classify and link edge points to generate continuous contours.

4. The semiconductor wafer visual inspection method based on interactive feature learning as described in claim 1, characterized in that, The feature vector extraction includes: Based on the contour output by contour edge detection, contour features are extracted, including contour perimeter, contour area and number of contour vertices. By debugging the software, the threshold ranges of the contour perimeter, contour area, and number of contour vertices are limited to filter out the target contour. Feature parameters are extracted from the target contour to obtain the feature vector.

5. The semiconductor wafer visual inspection method based on interactive feature learning as described in claim 4, characterized in that, The number of contour vertices is the number of vertices obtained by approximating the contour with polygons.

6. The semiconductor wafer visual inspection method based on interactive feature learning as described in claim 1, characterized in that, The steps for feature recognition of new image samples using the aforementioned standard parameter set include: The similarity distance between the contour features extracted from the image to be tested and the standard parameter set in the database in the multidimensional feature space is calculated, and the distance is compared with a tolerance threshold to determine whether they match.

7. The semiconductor wafer visual inspection method based on interactive feature learning as described in claim 1, characterized in that, The categories of anomalies include overexposure, underexposure, out-of-focus blur, motion blur, and foreign object obstruction.

8. The semiconductor wafer visual inspection method based on interactive feature learning as described in claim 1, characterized in that, Following the feature recognition step, there is also a step of outputting recognition results, which include one or more of the following: the existence of target features, location coordinates, type identifier, or confidence level.

9. The semiconductor wafer visual inspection method based on interactive feature learning as described in claim 1, characterized in that, The AI ​​model is a discriminative model, which is trained using the geometric feature parameters of historical wafer images and their corresponding artificial quality rating labels to identify and classify known defect patterns.

Images