An integrated circuit scanning electron microscope image noise reduction method

Abstract

The present invention discloses a method for reducing noise in integrated circuit scanning electron microscope images. Compared with the traditional multi-frame superposition technology in the industry, it can avoid photoresist shrinkage and make the image more clear. Specifically, the present invention obtains a wafer with consistent critical feature sizes on a semiconductor manufacturing production line; obtains SEM images of target points in different dielets; finds one image as a reference image among the above SEM images, and the remaining images as images to be compared; obtains all overlapping regions between the reference image and the images to be compared; calculates the maximum intersection region for all overlapping regions, and performs pixel-level averaging on the maximum intersection region to obtain a noise-reduced SEM image. The present invention can reduce the noise in SEM images, improve the image quality, and provide a reliable data source for product quality inspection and lithography model driving in the semiconductor manufacturing process.

Description

Technical Field

[0001] The present invention relates to the technical fields of semiconductor manufacturing and image processing, and relates to a method for reducing noise in integrated circuit scanning electron microscope images, particularly to CD-SEM image data, which can provide convenient operations for calculating lithography synthesis CD-SEM measurement data and the yield analysis department to collect clear SEM images. Background Art

[0002] In the field of integrated circuit manufacturing, with the development of semiconductor processes towards miniaturization and complexity, the importance of scanning electron microscopes (SEM) in measurement and detection has become increasingly prominent. After each process step in the semiconductor production line is completed, it is necessary to use critical dimension scanning electron microscopes (CD-SEM) to sample and inspect wafers to evaluate whether the integrated circuits after processes such as exposure, development, etching, and polishing meet the standards, thereby ensuring the smooth progress of subsequent process flows. If abnormalities are found in the early stage, measures can be taken in a timely manner to reduce losses. During the semiconductor process R & D process, the computational lithography and yield improvement teams frequently use CD-SEM to collect data. The computational lithography department constructs lithography models, etching models, and conducts model verification and other work by collecting CD-SEM images of a large number of test patterns. These images provide a solid data basis for lithography simulation verification. The yield improvement department also widely collects SEM images of hotspots and proposes targeted yield improvement strategies based on them.

[0003] SEM images play a crucial role in semiconductor manufacturing. However, due to the imaging principle of the machine, there will be a lot of noise. To effectively reduce the noise generated by CD-SEM equipment and obtain high-quality images, the semiconductor industry generally adopts a technology called multi-frame superposition. This technology involves scanning the electron beam at the same position multiple times to obtain multiple CD-SEM images, and then superimposing and averaging these images to generate a synthetic image with higher quality. Since the noise distribution in each image is random, through superposition, the noise in different directions can cancel each other out, while the common signal components in the images are retained, thereby forming a clearer image. Although the multi-frame superposition technology can reduce noise to a certain extent, each electron beam scan may cause slight shrinkage of the photoresist, resulting in a non-linear deviation between the measurement result and the actual process condition. In addition, since the acquisition of each frame of image requires a certain amount of time, collecting multiple frames of images will significantly increase the total image acquisition time. At the same time, a large number of electron beam bombards may also cause a decline in the performance of the material being detected.

[0004] Both semiconductor process R & D and actual production require collecting a large amount of SEM data, and this data requires the CD - SEM to work overtime. Due to the high cost of the machine, and when the imaging quality requirements are relatively high (low noise), the collection of a large amount of data is overwhelming, which greatly hinders the production and R & D cycle. At present, the application of deep learning in the field of noise reduction also requires a large number of paired image labels of noise - free and noisy images, which is extremely costly for researchers who have not delved deeply into the semiconductor manufacturing industry. Summary of the Invention

[0005] Based on the above problems, the present invention proposes a method for reducing noise in integrated circuit scanning electron microscope images to solve the problem of photoresist shrinkage caused by the multi - frame image superposition technology currently used in the industry, and to provide clearer and more actual - process - close SEM image data for the semiconductor manufacturing industry.

[0006] The object of the present invention is achieved by the following technical solutions:

[0007] In the first aspect, the present invention provides a method for reducing noise in integrated circuit scanning electron microscope images, and the method includes:

[0008] Obtain a wafer with consistent critical feature sizes on the semiconductor manufacturing production line;

[0009] Obtain SEM images of the target points in different dielets;

[0010] Find an image as the reference image from the above - mentioned SEM images, and the remaining images as the images to be compared;

[0011] Obtain all overlapping regions between the reference image and the images to be compared;

[0012] Calculate the maximum intersection region for all overlapping regions, and perform pixel - level averaging on the maximum intersection region to obtain the noise - reduced SEM image.

[0013] Preferably, the exposure conditions of each dielet in the wafer are the same.

[0014] Preferably, the process of finding the reference image is as follows:

[0015] Use the Fourier analysis method to calculate the amplitude spectrum information of each SEM image;

[0016] In the amplitude spectrum, intercept the low - frequency region with the origin of the frequency domain as the center;

[0017] Calculate the amplitude average value of the low - frequency region of each SEM image, and use the SEM image corresponding to the maximum amplitude average value as the reference image.

[0018] Preferably, the process of obtaining all overlapping regions between the reference image and the images to be compared is as follows:

[0019] Search for reference images and the images to be compared , the feature points between i = [1, k], where k represents the number of images to be compared;

[0020] According to the feature points of the reference image, match them one by one with the feature points of the images to be compared to obtain matching pairs, and then obtain the homography matrix;

[0021] Obtain the overlapping regions between each image to be compared and the reference image.

[0022] Preferably, the process of finding feature points is as follows:

[0023] For each image , q = [0, k], convolve with the image through a multi-scale Gaussian filter to form a multi-scale Gaussian pyramid .

[0024] For the same original image, subtract adjacent-sized images of the Gaussian pyramid to obtain the scale space of Gaussian difference .

[0025] Find the extreme points in the scale space of Gaussian difference as the feature points of the current image;

[0026] Use sub-pixel interpolation to offset each feature point to the correct position to obtain the corrected feature points;

[0027] Calculate the direction and amplitude of the corrected feature points in the scale space of Gaussian difference, and use position and eigenvector to describe each feature point.

[0028] Preferably, the process of calculating the eigenvector of the feature points is as follows:

[0029] Take a neighborhood around the position of the corrected feature point and calculate the gradient magnitude and direction of this neighborhood;

[0030] Create an orientation histogram containing 8 orientation bins, and count the gradient magnitudes into the histogram to obtain an orientation vector containing different orientation intensity values.

[0031] Preferably, the implementation process of matching the feature points of the reference image with the feature points of the images to be compared one by one to obtain matching pairs and then obtaining the homography matrix is as follows:

[0032] 1) Construct a KD tree with spatial geometric features according to the eigenvectors of all feature points of the reference image;

[0033] 2) For each feature point of the comparison image, find the feature point with the nearest neighbor in the KD tree as the reference feature point. The reference feature point and the current feature point of the comparison image form a pair of matching points, thereby obtaining the corresponding relationship of feature points between the comparison image and the reference image;

[0034] 3) Calculate the Euclidean distance between the feature point of the comparison image and the feature point of the reference image for each pair of matching points, and eliminate the feature points of the comparison image whose distance exceeds the threshold;

[0035] 4) Solve the non - homogeneous linear equations of the following formula to obtain the homography matrix H;

[0036]

[0037] where represents the feature vector of the feature point of the reference image, represents the feature vector of the feature point of the comparison image.

[0038] Preferably, the implementation process of obtaining the overlapping area between each comparison image and the reference image is as follows:

[0039] 1) Map the four edge points in the reference image to the comparison image through the homography matrix H to obtain the projection coordinates , and eliminate the projection coordinates that do not fall inside the current comparison image.

[0040] 2) Map the four edge points in the comparison image to the reference image through the homography matrix H to obtain the projection coordinates ;

[0041] 3) Take , as the vertex coordinates to determine an overlapping area in the comparison image;

[0042] 4) Repeat the above steps for all comparison images to form an overlapping area pool.

[0043] In a second aspect, the present invention provides a computer - readable storage medium, on which a computer program is stored. When the computer program is executed in a computer, the computer is made to execute the method.

[0044] In a third aspect, the present invention provides a computing device, including a memory and a processor. An executable code is stored in the memory, and when the processor executes the executable code, the method is implemented.

[0045] The beneficial effects of the present invention are:

[0046] The present invention obtains wafers with consistent critical feature sizes in a semiconductor manufacturing production line, acquires SEM images of target points from different dielets, uses the images of different dielets as the data basis to replace the superposition of multiple frames of images, avoids the problem of photoresist shrinkage from the source, ensures that the measurement results truly reflect the process conditions, improves the measurement accuracy, which is particularly important for semiconductor manufacturing at advanced nodes. The present invention obtains images from different dielets, can collect images of multiple dielets in a single scan, reduces the number of electron beam scans, reduces the impact on the performance of the material to be detected, improves production efficiency, and alleviates the problems of high data acquisition cost and long production and R & D cycle in the semiconductor manufacturing industry.

[0047] The present invention forms a Gaussian pyramid by convolving a multi-scale Gaussian filter with an image, subtracts adjacent-sized images to obtain a scale space of Gaussian difference, finds extreme points as feature points and corrects them, and calculates feature vectors to describe the feature points. Constructs a KD tree to match feature points, eliminates feature points with distances exceeding the threshold, solves a non-homogeneous linear equation system to obtain a homography matrix H, thereby obtains the overlapping region and calculates the maximum intersection region, and performs pixel-level averaging to obtain a denoised image, effectively removing noise, improving image clarity, and ensuring accurate measurement results. Performing pixel-level averaging on the maximum intersection region effectively removes the noise in the SEM image, can more accurately reveal the process information of the actual semiconductor process, and will not cause severe shrinkage of the photoresist. Description of the Drawings

[0048] Figure 1 It is an SEM image formed by superposing multiple frames of traditional electron beam scanning. Among them, (a) has a line width of 146.04 nm for 1 frame, (b) has a line width of 140.38 nm for 4 frames, (c) has a line width of 137.34 nm for 16 frames, (d) has a line width of 131.64 nm for 64 frames, (e) has a line width of 127.86 nm for 128 frames, and (f) has a line width of 122.75 nm for 256 frames.

[0049] Figure 2 It is a schematic diagram of the method flow of the present invention.

[0050] Figure 3 It is a conceptual diagram of the method of the present invention.

[0051] Figure 4 It is a schematic diagram of a CDU wafer.

[0052] Figure 5 They are SEM single-frame images formed by scanning in different directions. Among them, (a) is an image of the horizontal scan of the storage circuit area, (b) is an image of the vertical scan of the memory area; (c) is an image of the horizontal scan of the logic circuit area, and (d) is an image of the vertical scan of the logic circuit area.

[0053] Figure 6The Fourier transform amplitude spectra of SEM images for different dies, where (a) is the amplitude spectrum distribution of the reference image, and (b), (c), (d), (e), (f) are the amplitude spectra of 5 different images to be compared respectively.

[0054] Figure 7 Schematic diagram of the process for obtaining all overlapping regions between the reference image and the images to be compared.

[0055] Figure 8 Positions of feature points extracted from an example SEM image.

[0056] Figure 9 Spatial index matching results of the KD tree adopted by the present invention.

[0057] Figure 10 Feature point matching correspondence between the image to be compared (left figure) and the reference image (right figure).

[0058] Figure 11 SEM images of different dies collected and the noise-reduced SEM images obtained by the present invention, where (a) is a single-frame SEM image scanned in the horizontal direction (x-axis), (b) is a single-frame SEM image scanned in the vertical direction (y-axis), (c) is a double-die superposition SEM image of horizontal and vertical scans, (d) is a 4-die superposition SEM image, (e) is a 6-die superposition SEM image, (f) is a 10-die superposition SEM image, (g) is a 20-die superposition SEM image, and (h) is a 40-die superposition SEM image. Detailed implementation manners

[0059] The specific implementation manners of the present invention will be described in more detail below with reference to the schematic diagrams. According to the following description, the advantages and features of the present invention will be clearer. It should be noted that the drawings are in a very simplified form and use non-precise scales, only for conveniently and clearly assisting in explaining the purpose of the embodiments of the present invention.

[0060] CDSEM metrology plays an important role in both the field of computational lithography and yield improvement. The computational lithography department collects a large number of CDSEM images of test patterns to build lithography models, etching models, model verification, etc. The yield improvement department also collects a large number of SEM images of hotspots for analysis. Both semiconductor process R & D and actual production require the collection of a large amount of SEM data, and this data requires the CD-SEM to work overtime. Due to the high cost of the machine, and when high imaging quality (low noise) is required, the collection of a large amount of data is overwhelmed, which greatly hinders the production and R & D cycle. Currently, the application of deep learning in the field of noise reduction also requires a large number of noiseless and noisy paired image labels, which is extremely costly for researchers who have not delved deeply into the semiconductor manufacturing industry. The semiconductor manufacturing industry usually uses the image multi-frame superposition technology to suppress noise and obtain higher-quality SEM images, that is, the same position is scanned by the electron beam multiple times to obtain multiple CD-SEM images, and then these pictures are superimposed and averaged to obtain a relatively good multi-frame image.

[0061] Refer to Figure 1 in (a) to Figure 1 in (f) of the figure, which shows the SEM images of the same line pattern with different scanning frame numbers in different dielets, where 'Mean’ represents the line width value measured by the mean value. Due to the randomness of the noise distribution in each picture, the noise in different directions will cancel each other out, and the common signal components will be retained. It can be seen that as the scanning frame number increases, the image gradually becomes clear. However, as the scanning frame number increases, the line width also becomes smaller and smaller. This is due to the shrinkage of the photoresist on the wafer caused by multiple reciprocating electron beam scans. The measured line width of the photoresist line has shrunk from 146.04 nm to 122.75 nm, and the SEM image cannot effectively reflect the current actual process information. This defect becomes more obvious in more advanced nodes.

[0062] Inspired by the multi-frame image superposition technology, the present invention proposes a SEM image multi-dielet superposition technology. Refer to Figures 2 - 3 , a method for reducing noise in an integrated circuit scanning electron microscope image according to this embodiment includes the following steps:

[0063] Step S1, obtain a post-exposure critical dimension uniformity wafer (CDU wafer) on a semiconductor production line, and the exposure conditions of each dielet in the wafer are the same.

[0064] Refer to Figure 4 , the exposure energy and focal length of the CDU wafer are fixed at 46 mJ / cm² and 5 , and the exposure energy and exposure focal length of each dielet on the wafer remain unchanged. It is mainly used to evaluate the critical dimension uniformity during the lithography process to ensure that uniform pattern sizes can be obtained across the entire wafer.

[0065] Step S2, obtaining SEM images of target points in different core particles:

[0066] According to the coordinate information of the target point, SEM images of the target point are collected at different core positions of the wafer to obtain multiple SEM images. Parts with significant defects in the image are removed from the above SEM images, such as overexposure of the image, inconsistency between the developed image and the target image, photoresist collapse, photoresist stickiness, etc., and SEM images with normal images and no defects are screened out. If the number of defect-free images is too small, additional images should be collected. The present invention recommends not less than 10 defect-free images. In particular, in order to ensure the clarity of the final imaging, the present invention recommends scanning and acquiring SEM images from multiple angles. See Figure 5 In (a) to (d) in FIG5 , the present invention performs horizontal and vertical scanning respectively in the actual process. The first column is a horizontal scan, and it can be seen that the longitudinal lines are relatively clear; the second column is a vertical scan, and it can be seen that the horizontal lines are relatively clear. It can be observed that the vertical contour of the horizontal scan image is relatively clear, while the horizontal contour of the longitudinal scan image is relatively clear. Therefore, by superimposing the horizontal and longitudinal SEM images, an image with a clearer contour can be obtained.

[0067] Step S3, finding a reference image:

[0068] S3-1 Since there is a certain degree of displacement between each SEM image, the traditional spatial distance cannot effectively measure the similarity between images, so as to find the optimal reference image. Therefore, the present invention adopts the Fourier analysis method to select the optimal reference image. SEM images often have a lot of noise pollution. In the spectrum analysis, the low frequency represents the general outline of the image, and the high frequency is the image details and noise part.

[0069] Specifically, the amplitude spectrum information of each SEM image screened in step S2 is calculated using the Fourier analysis method;

[0070]

[0071] in is the frequency domain coordinate The amplitude value at , Re is the amplitude function extracted after Fourier transform, Coordinates on the image The pixel value of the image, H and W are the image height and width respectively.

[0072] See also Figure 6 Middle (a) to Figure 6 In (f), the amplitude spectrum of different SEM images is shown. The middle area and the horizontal and vertical center axis represent low-frequency information. Since each image has different noise pollution, the signal intensity in the low-frequency part will also be different.

[0073] In the amplitude spectrum of S3-2, the low-frequency region is intercepted with the origin of the frequency domain as the center; preferably, the size of the low-frequency region is one-tenth of the image pixels.

[0074] In S3-3, calculate the amplitude average value of the low-frequency region of each SEM image, and take the SEM image corresponding to the maximum amplitude average value as the reference image, and the remaining SEM images as the images to be compared.

[0075] In this embodiment, the SEM image corresponding to the first spectral information is selected as the reference image. It can be seen from the figure that its low-frequency information is more obvious, so the degree of noise pollution is lower.

[0076] Step S4: Obtain all overlapping regions between the reference image and the images to be compared. See the appendix Figure 7 :

[0077] S41 Search for the reference image and the images to be compared , i = [1, k]. The main steps are as follows:

[0078] 1) For each image , q = [0, k], convolve with the image through a multi-scale Gaussian filter to form a multi-scale Gaussian pyramid ; where G is the multi-scale Gaussian filter function, and its variance can be selected according to the actual situation. For example, taking values 3, 4, 7, and 9 respectively can generate 4 Gaussian convolution kernels with different blurring scales, and then complete the convolution to generate a layer of scale space.

[0079] 2) For the same original image, subtract the adjacent-size images of the Gaussian pyramid to obtain the Difference of Gaussian scale space (DOG), .

[0080] 3) Search for the extreme points in the Difference of Gaussian scale space as the feature points of the current image, and repeat the above steps to obtain the set T of feature points of all images.

[0081] 4) Use sub-pixel interpolation to offset each feature point to the correct position, and the offset is to obtain the corrected feature points.

[0082] 5) Calculate the direction and amplitude of the corrected feature points in the Difference of Gaussian scale space, and use the position and feature vector to describe each feature point;

[0083] The calculation process of the feature vector of the feature point:

[0084] Take a neighborhood around the position of the corrected feature point and calculate the gradient magnitude and direction of this neighborhood.

[0085] Create an orientation histogram containing 8 orientation bins, each bin representing 45 degrees, and count the gradient magnitudes into the histogram to obtain an orientation vector containing intensity values in different directions.

[0086] Refer to Figure 8 , through the above steps, the feature points in an image can be obtained. The positions of the respective feature points are marked by the circled dots of different colors in the figure. When storing, a set of mathematical vectors (position, scale, direction, etc.) are used to describe the feature points to ensure the complete information of the feature points.

[0087] S42 Match the feature points of the reference image with those of the image to be compared one by one to obtain matching pairs, and then obtain the homography matrix. Homography refers to the transformation from a point in one image to the corresponding point in another image. Specifically:

[0088] 1) Construct a KD tree with spatial geometric features based on the feature vectors describing all the feature points of the reference image.

[0089] 2) For each feature point of the image to be compared, find the feature point closest to it in the KD tree as the reference feature point. The reference feature point and the current feature point of the image to be compared form a set of matching pairs, thereby obtaining the correspondence of feature points between the image to be compared and the reference image.

[0090] 3) Calculate the Euclidean distance between the feature points of the image to be compared and the reference feature points in each set of matching pairs, and eliminate the feature points of the image to be compared whose distances exceed the threshold;

[0091] The Euclidean distance is calculated as follows:

[0092]

[0093] where n represents the spatial dimension of the feature vector describing the feature points, represents the feature vector of the reference feature point in the i-th set of matching pairs, represents the feature vector of the feature point of the image to be compared in the i-th set of matching pairs; represents the feature vector of the reference feature point, represents the feature vector of the feature point of the image to be compared;

[0094] Refer to Figure 9, a KD tree constructed using the feature point information of the reference image is adopted. This tree has spatial geometric information. When matching any feature point of the image to be compared, the nearest neighbor matching point can be quickly found according to the spatial tree structure, which speeds up the matching process. The blue dots in the figure represent the position distribution of the feature points in the multi-dimensional space, the horizontal and vertical lines constitute different spatial dimensions, and the intersection points are the intersections between the spatial dimensions. Refer to Figure 10 , which shows the matching situation of two different die SEM images. The feature points can be effectively paired by adopting the process described in the present invention.

[0095] 4) Solve the non-homogeneous linear equations of the following formula to obtain the homography matrix H;

[0096]

[0097] S43 Obtain the overlapping regions between each image to be compared and the reference image, and the specific method is as follows.

[0098] 1) Map the four edge points in the reference image to the image to be compared through the homography matrix H to obtain the projection coordinates , where , r = 1, 2, 3, 4, which are the image edge points (0, 0), (0, H), (W, 0), (H W) respectively. is the intersection point between the reference image and the current image to be compared . Retain the projection coordinates inside the current image to be compared, that is, .

[0099] 2) Map the four edge points in the image to be compared to the reference image through the homography matrix H to obtain the projection coordinates ;

[0100] 3) Use as the vertex coordinates to determine an overlapping region in the image to be compared.

[0101] 4) Repeat the above steps for all images to be compared to form an overlapping region pool.

[0102] Step S5: Calculate the maximum intersection region for all overlapping regions in the overlapping region pool, and perform pixel-level averaging on the maximum intersection region. The output can obtain a clear and low-noise SEM image. The final effect formed by the present invention is shown in Figure 11 from (a) to Figure 11 (h) in. In this embodiment, the number of SEM images collected in the vertical scanning direction and the horizontal scanning direction each accounts for half. The yellow frames in the figure represent the overlapping regions of these two original single-frame SEM images.

[0103] The calculation formula for pixel-level averaging of the largest intersection area is as follows:

[0104]

[0105] This operation cleverly utilizes the information of pixel points at the same position in multiple images. Through averaging processing, it effectively reduces the noise interference in a single image while retaining the common signal components in the image. Finally, an image after pixel-level averaging is output, obtaining a clear and low-noise SEM image. This image is not only clearer visually but also has a significant improvement in data accuracy, effectively removing the noise in the SEM image, more accurately revealing the process information of the actual semiconductor process, and not causing severe shrinkage of the photoresist.

[0106] Through the above series of implementation steps, the present invention has successfully realized an effective method for reducing noise in integrated circuit scanning electron microscope images. From the careful preparation of the wafer and image acquisition, to the scientific selection of reference images and the precise matching of feature points, and then to the accurate determination of the overlapping area and image superposition processing, this method not only solves the problem of photoresist shrinkage caused by traditional multi-frame superposition technology but also significantly improves the clarity of the image and the accuracy of measurement results, providing a reliable source of image data for product quality inspection and lithography model driving in the semiconductor manufacturing process.

Claims

1. An integrated circuit scanning electron microscope image noise reduction method, characterized in that The method includes: Obtaining a key feature size consistency wafer on a semiconductor manufacturing production line; Obtaining SEM images of a target point in different dielets; Finding one image as a reference image among the above SEM images, and the remaining images as images to be compared; Obtaining all overlapping regions between the reference image and the images to be compared; Calculating the maximum intersection region for all overlapping regions, and performing pixel-level averaging on the maximum intersection region to obtain a noise-reduced SEM image; Among them, the process of obtaining all overlapping regions between the reference image and the images to be compared is as follows: (1)Search for a reference image and the image to be compared , the feature points between i = [1, k], where k represents the number of images to be compared; (2) According to the feature points of the reference image, match them one by one with the feature points of the images to be compared to obtain matching pairs, and then obtain a homography matrix; the implementation process is as follows: (2.1) Construct a KD tree with spatial geometric features according to the feature vectors describing all feature points of the reference image; (2.2) For each feature point of the image to be compared, find the nearest neighbor feature point in the KD tree as the reference feature point. The reference feature point and the current feature point of the image to be compared form a group of matching pairs, so as to obtain the correspondence of feature points between the image to be compared and the reference image; (2.3) Calculate the Euclidean distance between the feature points of the image to be compared and the feature points of the reference image in each group of matching pairs, and eliminate the feature points of the image to be compared whose distance exceeds the threshold; (2.4) Solve the non-homogeneous linear equations of the following formula to obtain the homography matrix H; ； Among them represents the feature vector of the feature points of the reference image, represents the feature vector of the feature points of the image to be compared; (3) Obtain the overlapping regions between each image to be compared and the reference image.

2. The method according to claim 1, wherein The exposure conditions of each dielet in the wafer are the same.

3. The method according to claim 1, characterized in that, The process of finding the reference image is as follows: Calculate the amplitude spectrum information of each SEM image using the Fourier analysis method; In the amplitude spectrum, intercept the low-frequency region centered on the origin of the frequency domain; Calculate the amplitude average value of the low-frequency region of each SEM image, and take the SEM image corresponding to the maximum amplitude average value as the reference image.

4. The method according to claim 1, wherein The process of finding the feature points is as follows: For each image , q = [0, k] is convolved with the image through a multi-scale Gaussian filter to form a multi-scale Gaussian pyramid ; For the same original image, the Gaussian pyramid is subtracted from adjacent-sized images to obtain the scale space of difference of Gaussian ; Finding the Difference of Gaussian Scale Space The extreme points are used as the feature points of the current image; Use sub-pixel interpolation to offset each feature point to the correct position to obtain corrected feature points; Calculate the direction and amplitude of the corrected feature points in the difference of Gaussian scale space, and use the position and feature vector to describe each feature point.

5. The method according to claim 4, wherein The process of calculating the feature vector of the feature points is as follows: Take a neighborhood around the position of the corrected feature point and calculate the gradient magnitude and direction of the neighborhood; Create a direction histogram containing 8 direction bins, and count the gradient magnitudes into the histogram to obtain a direction vector containing different direction intensity values.

6. The method according to claim 1, wherein The implementation process of obtaining the overlapping regions between each image to be compared and the reference image is as follows: (3.1) Map the four edge points in the reference image to the image to be compared through the homography matrix H to obtain the projected coordinates , and eliminate the projected coordinates that do not fall within the current image to be compared; (3.2) Map the four edge points in the image to be compared to the reference image through the homography matrix H to obtain the projected coordinates ; (3.3) Using , as vertex coordinates, determine an overlapping region in the image to be compared ; (3.4) Repeat the above steps for all images to be compared to form an overlapping region pool.

7. A computer-readable storage medium, on which a computer program is stored. When the computer program is executed on a computer, the computer is made to execute the method according to any one of claims 1-6.

8. A computing device, including a memory and a processor. An executable code is stored in the memory. When the processor executes the executable code, the method according to any one of claims 1-6 is implemented.

Images