A method for detecting, recognizing, and aligning silicon wafer overlay patterns

Abstract

This invention proposes a method for detecting, recognizing, and aligning silicon wafer overlay patterns. First, image segmentation is used to automatically find regions of interest (ROIs) and generate an SVM dataset. Second, traditional feature extraction methods are used to extract salient features from the image for training the SVM, and the ROIs are obtained for SVM model prediction. Then, shape-invariant moments (HUs) are used for feature point similarity matching to verify the SVM recognition results and eliminate mispredictions. Finally, the angle and position offset are determined for alignment. This method can automatically detect silicon wafer overlay patterns and achieve overlay alignment.

Description

Technical Field

[0001] This invention relates to a method for detecting, recognizing, and aligning silicon wafer overlay patterns, belonging to the field of image processing, and involving machine learning and image recognition classification methods. It is used for adaptive selection of templates before overlay alignment in lithography equipment and to achieve overlay alignment. Background Technology

[0002] As a key piece of equipment in microelectronics manufacturing, the lithography machine plays a vital role in modern technology. Lithography technology is one of the fundamental processes for manufacturing integrated circuits (ICs). It uses an optical projection system to transfer patterns onto silicon wafers, forming intricate electronic components and circuit structures. Its development is of great significance to national scientific and technological progress and economic development. Factors affecting the lithography process include photoresist materials, light sources, masks, and alignment accuracy. Alignment technology, as one of the three core technologies of a lithography system, is a crucial component. Based on the alignment marks on the mask and the sample to be exposed, it ensures the positional alignment between them, guaranteeing the precision and accuracy of the lithographic pattern. Continuous optimization and innovation are key to improving the performance and pattern quality of the lithography system.

[0003] Patent document CN202010834499.1, published on December 11, 2020, discloses an automatic alignment method. First, it captures an image of the mask's crosshair markings and an aligned image. Second, it extracts and processes the crosshair markings on the silicon wafer using a computer. Finally, it uses image analysis to obtain the positional difference between the mask and the silicon wafer to achieve alignment. However, this method requires the design of crosshair alignment markings. The conference presentation by Zhang Shaoyu et al., "Lithography alignment method based on image rotation matching," uses a grayscale-based template matching method to achieve image alignment with different markings. However, its implementation is limited by the initial generation and selection of the marking template. When the alignment markings change, or when the user designs different alignment markings for each overlay, the alignment template needs to be redesigned and generated, which is not conducive to the automated production of lithography machines.

[0004] To address the above issues, this invention proposes a method for detecting, recognizing, and aligning silicon wafer overlay patterns. This method automatically acquires a training set without requiring manual labeling or knowledge of the label content. An SVM model is trained, and HU invariant moments are used to verify the SVM recognition results. When changing equipment or masks, the alignment template does not need to be manually replaced; it adaptively detects and selects the template, achieving automatic overlay alignment. Summary of the Invention

[0005] This invention proposes a method for detecting, recognizing, and aligning silicon wafer overlay patterns. First, the image is preprocessed to enhance its features. Image segmentation is used to automatically obtain a labeled template image, which is then used as a training file. Second, traditional feature extraction methods are used to extract HOG features from the image, generating feature vectors. Machine learning methods are then used to train a detection and recognition model, and the region of interest (ROI) of the image to be predicted is obtained for SVM model detection and recognition, returning label values. Then, to eliminate the influence of other patterns on the prediction, based on the label values ​​returned by SVM recognition, HU shape invariant moment feature point matching is used for secondary verification and screening to ensure the accuracy of image detection and recognition. Finally, template matching and alignment are completed, and angle and position offsets are obtained for equipment alignment debugging. This method mainly aims to complete the detection and recognition of silicon wafer alignment templates, laying the foundation for the next step of selecting the corresponding template for image alignment, and obtaining angle and position offset values ​​to achieve equipment overlay alignment debugging.

[0006] A method for detecting, identifying, and aligning silicon wafer overlay patterns, characterized by comprising the following steps:

[0007] Step S1: Automatically generate a silicon wafer template dataset. Divide the region of interest based on the exposed image on the silicon wafer, select the region that meets the requirements, and expand it into a dataset.

[0008] Step S2: Extract HOG features from the dataset, generate feature vectors, and train the corresponding SVM model using the SVM method;

[0009] Step S3: Perform image preprocessing on the image to be predicted to remove background noise and obtain the region of interest;

[0010] Step S4: Extract the HOG features of the region of interest, input them into the SVM model for prediction, and return the predicted label;

[0011] Step S5: When the shapes of the graphics are different, their HOG gradient features may be similar with a small probability. In this case, SVM may misdetect. Based on the SVM prediction result label, select a suitable image and perform shape invariant moment similarity matching based on HU to verify the SVM recognition result. Combine the prediction results of the similarity matching between SVM and HU, return the final detection and recognition label value, and select the image with the same label value as the alignment template. If there is no region that matches the shape features of SVM and HU after full image detection, randomly extract another region for subsequent silicon wafer template matching and alignment.

[0012] Step S6: Perform template matching and alignment. Using a single graphic, obtain the angle and position offset. The angle is the rotational offset of the detected shape relative to the template, and the position is the offset of the shape's center point relative to the center point of the entire image. Adjust the device based on the angle and position offsets. The selected alignment template is an unknown graphic in an unknown region of the image; that is, the correct position of the final alignment result in the image cannot be determined. Therefore, the angle of the template itself and the center point of the overall image are selected as evaluation criteria.

[0013] Furthermore, in step S1, the generation of the silicon wafer template dataset includes the following steps:

[0014] Step S11: Perform preprocessing on the image, such as denoising and filtering;

[0015] Step S12: Use image segmentation to obtain regions where graphics may exist. If the conditions are met, save the region and obtain its upper left corner coordinates. If the conditions are not met, discard the region and continue to judge other regions.

[0016] Step S13: Based on the obtained coordinates of the top left corner of marker 1, crop an image of size m*n;

[0017] Step S14: Expand and save the dataset using rotation and translation methods, and save the label as 1.

[0018] Furthermore, in step S12, the judgment condition is that the number of pixels Q in the region should be within a certain range. If it exceeds the range, the region is abandoned and other graphic regions are judged. If the condition is met, the region is saved. According to the characteristics of the exposed image, there must be two or more independent graphics in the captured image.

[0019] Furthermore, in step S13, the size of each sample data is consistent as m*n, and the specific values ​​of m and n are determined according to the sample image. The m*n image is cropped only by region cropping compared with the original image.

[0020] Furthermore, in step S14, the number of sample categories is greater than 2, meaning that the SVM training includes not only positive and negative samples, but also multiple samples in a multi-class classification problem. The generation of other datasets such as label 2 and 3 is the same as the generation method of the label 1 dataset.

[0021] Furthermore, in step S3, the image to be predicted is preprocessed to obtain the region of interest, which is convenient for subsequent SVM model prediction. The method for cropping the region of interest is the same as the method for cropping the generated dataset, and the size is also the same.

[0022] Furthermore, in step S5, returning the final detected and identified label value includes the following steps:

[0023] Step S51: Based on the label value results returned by SVM detection and recognition, select the corresponding label image from the database;

[0024] Step S52: Perform image preprocessing operations on the corresponding label image and the corresponding region of interest;

[0025] Step S53: Extract the HU shape invariant moment features from the processed image;

[0026] Step S54: Perform HU invariant moment similarity matching and recognition verification. The score is 1 when the HU invariant moment similarity matching is completely consistent. That is, when the score is greater than or equal to 0.9, the matching result is considered to be consistent with the SVM prediction result. The label result is retained, and the image with the same label value is selected as the alignment template. When the prediction results are inconsistent, return to step S1 and proceed to step S13, using the selected region as the alignment template.

[0027] This invention provides a method for detecting, identifying, and aligning silicon wafer overlay patterns, which mainly has the following beneficial effects:

[0028] 1. The training set used for SVM training can be generated automatically by the software, without the need for manual labeling or knowledge of the label content;

[0029] 2. By using SVM (Support Vector Machine) and HU (Human Shape Invariant Moment) feature point matching, template types can be accurately detected and identified, preparing for the next step of lithography alignment. This method is applicable to lithography alignment equipment where alignment marks change, paving the way for the automation of lithography equipment. Attached Figure Description

[0030] Figure 1 This is a flowchart of the overall detection and identification method of the present invention.

[0031] Figure 2 This is a flowchart illustrating the process of generating the dataset independently in this invention.

[0032] Figure 3 This is a flowchart of the identification and verification process for this invention.

[0033] Figure 4 This is a schematic diagram of the identification results of the present invention. Detailed Implementation

[0034] The present invention will now be described in further detail with reference to the accompanying drawings.

[0035] like Figure 1 The diagram shows the overall flowchart of a method for detecting, recognizing, and aligning silicon wafer overlay patterns, which includes the following steps.

[0036] Step S1: Automatically generate the silicon wafer template dataset. Using image segmentation, regions of interest are divided based on the exposed images on the silicon wafer. Images with possible shapes are extracted, and regions meeting the requirements are selected. Data augmentation is performed using methods such as rotation and translation to automatically generate the silicon wafer template dataset. The ratio of the training set to the test set is 7:3.

[0037] Step S2: Extract HOG features from the dataset and generate feature vectors. Each dataset image yields a 1*3780 feature vector. Integrate the feature vectors of images with the same label into an n*3780 training file, where n is the number of dataset images under the same label. The number of images with different labels should be consistent. Select the SVM linear kernel function and train the corresponding SVM model. Test the model's recognition accuracy on the test set. If the detection accuracy on the test set is not 100%, return to step S1; otherwise, continue to step S3. The test set and training set have the same content, consisting of the images captured in step S1, including labels 1, 2, 3, etc. The number is less than that of the training set. The formula for calculating the accuracy is shown in formula (1):

[0038] (1)

[0039] Where P represents accuracy, a1 is the total number of data points labeled 1 in the test set, and a2 and a3 are calculated similarly. r1 is the number of times the SVM correctly predicts label 1 in the test set, and r2 and r3 are calculated similarly. The formula for calculating P changes when the number of data points between classes is increased or decreased.

[0040] The HOG algorithm first calculates the gradient of the image to capture edge and texture information. The image is divided into a series of identically sized cells, each containing a set of pixels. For each pixel within a cell, its gradient direction is calculated. A common gradient calculation method is to convolve the image using the Sobel operator and project it into a histogram. Each bar of the histogram represents a gradient direction within a certain range, and the gradient intensity in each direction within that cell is statistically analyzed. Adjacent cells are grouped into a block, and the histogram within each block is normalized to increase robustness to changes in lighting. Finally, the normalized histograms from all blocks are concatenated to form the final HOG feature vector.

[0041] Step S3: First, perform region cropping on the overall image. The cropping size is m1*n1, where m1 is smaller than the length m of the original image, and n1 is smaller than the width n of the original image. The starting coordinates for cropping in the original image to be tested are (x, y). The cropped region includes the shape the client expects to use for alignment. Perform image preprocessing operations such as filtering and dilation / erosion on the cropped image to remove background noise, which is more conducive to image segmentation and extraction of the region of interest. The method for extracting the region of interest is the same as the method for generating the dataset, but no data augmentation or storage is required. The resulting image is then sent to the SVM model for prediction.

[0042] Step S4: SVM model prediction is similar to SVM model training. It is necessary to extract the HOG features of the region of interest, input them into the SVM model for prediction, and return the prediction result labels. When there are multiple images in the prediction map, multiple prediction label values ​​are returned.

[0043] Step S5: When the shapes of the graphics are different, their HOG gradient features may be similar with a small probability. In this case, SVM may make false detections. Based on the SVM prediction result label, select a suitable image and perform shape invariant moment similarity matching based on HU to verify the SVM recognition result. Combine the prediction results of SVM and HU shape invariant moment similarity matching to return the final detection and recognition label value. Select the image with the same label value as the alignment template. If there is no region that matches the shape features of SVM and HU after full image detection, randomly extract another region for subsequent silicon wafer template matching and alignment.

[0044] Step S6: Complete template matching and alignment. Obtain the angle and position offset using a single graphic. The angle is the rotational offset of the shape relative to the template, and the position is the offset of the shape's center point relative to the center point of the entire image. Since we are selecting an unknown graphic in an unknown region of the image, we cannot determine the correct position of the graphic in the image. Therefore, we select the angle of the template itself and the center point of the entire image as the evaluation criteria. Based on the calculated angle and position offset, we then perform actual device adjustments.

[0045] SVM model training requires a dataset, such as... Figure 2 The diagram shown is a flowchart of the process for generating a dataset independently according to the present invention, which includes the following steps.

[0046] Step S11: First, perform a rough area cropping on the overall image, and then perform preprocessing such as denoising and filtering on the image to remove background noise and small objects that may affect subsequent operations.

[0047] Step S12: Obtain the coordinates of the upper left corner of possible marker 1 using image segmentation. In the cropped area, perform image segmentation using connected component analysis to find possible shape regions. Eliminate shapes with fewer than Q pixels. The Q value can be determined based on the number of pixels in each overlaid pattern in the overall image; for example, "Q=20". If the number of pixels is less than Q, the location is considered background noise and is excluded. If the background pixel count Q>20, the location is considered to have a shape and is retained. However, the Q value cannot be too large, as it is necessary to exclude cases where detection fails or the detected image is too large overall (e.g., the chip itself is too large), in which case a clear shape cannot be obtained for that region.

[0048] Step S13: Obtain the centroid information and the length and width information of the minimum bounding rectangle of the shape. Based on this information, obtain the coordinates of the top-left corner of the shape. Using the obtained top-left corner coordinates (x, y) of marker 1, crop an image of size m*n. Shift the x-axis coordinate of the top-left corner to the left, i.e., move it 1 position in the direction of decreasing value. Based on the current top-left corner coordinates (x-1, y), crop an image of size m*n. Shift the y-axis coordinate to the up, i.e., move it 1 position in the direction of decreasing value. Based on the current top-left corner coordinates (x, y-1), crop an image of size m*n. Similarly, let the top-left corner coordinates be (x-1, y-1), etc., and crop images. Each cropped image must completely include the shape. The amount of shift in the top-left corner coordinates depends on the actual situation.

[0049] Step S14: Expand and save the dataset using rotation and translation methods, and save the label as 1.

[0050] Each sample dataset has a consistent size of m*n, where m and n are determined based on the image resolution. The extracted m*n images are cropped only from the original images, with only region cropping performed. There are 20 samples within each class, which can be increased or decreased by 5. The number of samples between classes is greater than 2, meaning that SVM training involves not only positive and negative samples but also a multi-class classification problem with multiple samples. The generation of other datasets, such as Label 2 and Label 3, follows the same method as Label 1. The dataset is divided into training and test sets in a 7:3 ratio.

[0051] like Figure 3 The diagram shown is a flowchart of the secondary determination process of the present invention, which includes the following steps.

[0052] Step S51: Based on the label value results returned by SVM detection and recognition, select the overlay graphic template image with the same label from the database as the corresponding label image.

[0053] Step S52: Perform image preprocessing operations on the corresponding label image and the corresponding region of interest, including image processing such as noise reduction and filtering. Find connected regions, separate the found shape regions from the background regions, and fill them with white and black respectively. Only one shape region will exist in a single image.

[0054] Step S53: Extract the HU shape invariant moment features of the processed image.

[0055] Based on Hu's shape-invariant moment method, seven invariant moment features can be extracted:

[0056] (2)

[0057] (3)

[0058] (4)

[0059] (5)

[0060] (6)

[0061] (7)

[0062] (8)

[0063] The normalized result of the central moments of each order is calculated using the following formula:

[0064] (9)

[0065] In equation (9): p and q represent the p-th derivative with respect to x and the q-th derivative with respect to y of the coordinates (x, y), respectively, and are the central moments of each order, which are defined as:

[0066] (10)

[0067] In equation (10): x0 and y0 are obtained by calculating the moments at the origin, and the calculation formula is:

[0068] (11)

[0069] (12)

[0070] In the above formula, m is the p+q order raw moment, which is defined as:

[0071] (13)

[0072] In equation (13): M and N represent discrete images Img(x, y) of size M*N.

[0073] Define the invariant moments of the region of interest HU as SH[i], and the invariant moments of the template graph HU as TH[i].

[0074] Step S54: Perform HU invariant moment similarity matching. A score of 1 is given when the HU invariant moment similarity matching is completely consistent. When the score is greater than or equal to 0.9, the matching result is considered to be consistent with the SVM prediction result, and the image corresponding to the current label result is selected as the alignment template. If the prediction results are inconsistent, return to step S1 and proceed to step S13, using the selected region as the alignment template.

[0075] The extracted HU invariant moment information is extremely large, making it inconvenient for image matching calculations. Therefore, the value is converted to a two-digit integer for easier calculation using the logarithm function. The calculation formula is as follows:

[0076] M[i]={log}_{10}\left ( \left | {H[i]} \right | \right ) (14)

[0077] The sign of M[i] is opposite to that of H[i].

[0078] The formula for similarity matching is:

[0079] (15)

[0080] Where dSigmaST is the sum of the products of the template image and the absolute values ​​of the invariant moments of each HU in the region of interest, and its calculation formula is:

[0081] dSigmaST=\sum ^{7}_{i=1} \left | {SH[i]} \right |\times \left | {TH[i]} \right | (16)

[0082] dSigmaS is the sum of squares of the numerical invariant moments of each HU in the region of interest, and its calculation formula is:

[0083] dSigmaS=\sum ^{7}_{i=1} {SH{[i]}^{2}} (17)

[0084] dSigmaT is the sum of squares of the invariant moments of each HU in the template diagram, and its calculation formula is:

[0085] dSigmaT=\sum ^{7}_{i=1} {TH{[i]}^{2}} (18).

Claims

1. A method for detecting, identifying, and aligning overlay patterns on silicon wafers, characterized in that, Includes the following steps: Step S1: Automatically generate a silicon wafer template dataset. Divide the region of interest based on the exposed image on the silicon wafer, select the region that meets the requirements, and expand it into a dataset. Step S2: Extract HOG features from the dataset, generate feature vectors, and train the corresponding SVM model using the SVM method; Step S3: Perform image preprocessing on the image to be predicted to remove background noise and obtain the region of interest; Step S4: Extract the HOG features of the region of interest, input them into the SVM model for prediction, and return the predicted label; Step S5: When the shapes of the graphics are different, their HOG gradient features may be similar with a small probability. In this case, SVM may misdetect. Based on the SVM prediction result label, select a suitable image and perform shape invariant moment similarity matching based on HU to verify the SVM recognition result. Combine the prediction results of SVM and HU shape invariant moment similarity matching to return the final detection and recognition label value. Select the image with the same label value as the alignment template. If there is no region that matches the shape features of SVM and HU after full image detection, randomly extract another region for subsequent silicon wafer template matching and alignment. Step S6: Perform template matching and alignment. Obtain the angle and position offset through a single graphic. The angle is the rotational offset of the detected shape relative to the template, and the position is the offset of the center point of the detected shape relative to the center point of the entire image. Perform actual debugging of the device based on the angle and position offset.

2. The method for detecting, identifying, and aligning silicon wafer overlay patterns according to claim 1, characterized in that: In step S1, the silicon wafer template dataset is automatically generated, including the following steps: Step S11: Perform preprocessing operations such as denoising and filtering on the image; Step S12: Use image segmentation to obtain regions where graphics may exist. If the conditions are met, save the region and obtain its upper left corner coordinates. If the conditions are not met, discard the region and continue to judge other regions. Step S13: Based on the obtained coordinates of the top left corner of marker 1, crop an image of size m*n; Step S14: Expand and save the dataset using rotation and translation methods, and save the label as 1.

3. The method for detecting, identifying, and aligning silicon wafer overlay patterns according to claim 2, characterized in that: In step S12, the judgment condition is that the number of pixels Q in the region should be within a certain range. If it exceeds the range, the region is abandoned and other graphic regions are judged. If the condition is met, the region is saved. According to the characteristics of the exposed image, there must be two or more independent graphics in the captured image.

4. The method for detecting, identifying, and aligning silicon wafer overlay patterns according to claim 2, characterized in that: In step S13, the size of each sample data is the same as m*n, and the specific values ​​of m and n are determined according to the sample image. The m*n image is cropped only by region cropping compared with the original image.

5. The method for detecting, identifying, and aligning silicon wafer overlay patterns according to claim 2, characterized in that: In step S14, the number of sample categories is greater than 2, meaning that the SVM training includes not only positive and negative samples, but also multiple samples in a multi-class classification problem. The generation of other datasets such as label 2 and 3 is the same as the generation method of label 1 dataset.

6. The method for detecting, identifying, and aligning silicon wafer overlay patterns according to claim 1, characterized in that: In step S3, the image to be predicted is preprocessed to obtain the region of interest, which is convenient for subsequent SVM model prediction. The method for cropping the region of interest is the same as the cropping method for the generated dataset, and the cropping size is also the same.

7. The method for detecting, identifying, and aligning silicon wafer overlay patterns according to claim 1, characterized in that: In step S5, similarity matching and filtering are performed using the shape invariant moments of HU, including the following steps: Step S51: Based on the label value results returned by SVM detection and recognition, select the corresponding label image from the database; Step S52: Perform image preprocessing operations on the corresponding label image and the corresponding region of interest; Step S53: Extract the HU shape invariant moment features from the processed image; Step S54: Perform HU invariant moment similarity matching and recognition verification. When the matching is consistent with the SVM prediction result, retain the label result and select the image with the same label value as the alignment template. When the prediction results are inconsistent, return to step S1 and proceed to step S13 to use the selected region as the alignment template.

Images