An image matching method

Abstract

The application discloses an image matching method, comprising: preprocessing, preprocessing an initial template image and an initial search image respectively to obtain a template image and a search image; coarse positioning, matching the template image and the search image based on at least two different deep learning image matching methods to obtain at least two matching regions with the same size as the template image, and determining a coarse matching region in the search image based on the at least two matching regions; fine positioning, matching the coarse matching region and the template image based on a phase correlation image matching method to determine a fine matching region in the search image. In the image matching method, coarse positioning is realized based on the deep learning image matching method first, and then fine positioning is realized based on the phase correlation image matching method, so that the high robustness and accuracy of the deep learning image matching method and the high accuracy of the phase correlation image matching method are considered, and the robustness, accuracy and accuracy of the image matching method are effectively improved.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to an image matching method. Background Technology

[0002] Image matching methods are used to match objects or scenes in different images. Traditional image matching methods typically perform matching by calculating the similarity of gray levels in various regions of a template image and a search image. These similarity scoring methods include the sum of squared errors algorithm, the sum of squared average differences algorithm, the normalized cross-correlation algorithm, the phase correlation algorithm, and the sequential similarity detection algorithm. When the template image and the search image are very clear, with little noise, relatively simple textures, and the transformation relationship between them is not complex (e.g., without rotation or scaling), these methods can accomplish this task well. However, fluctuations in the illumination (beam or electron beam) of image acquisition equipment (optical microscope or scanning electron microscope) can cause noise in the image, making traditional image matching methods very fragile, with low robustness and accuracy.

[0003] Deep learning-based image matching methods are characterized by efficiency, accuracy, and robustness, such as deep learning-based keypoint matching methods, block feature value similarity matching methods, and feature descriptor generation methods. Although deep learning-based image matching methods have achieved good results in many applications, some shortcomings still exist, such as insufficient accuracy.

[0004] Therefore, an image matching method is needed to effectively improve the robustness, accuracy, and precision of image matching methods. Summary of the Invention

[0005] The purpose of this invention is to overcome the aforementioned deficiencies in the prior art and provide an image matching method, comprising:

[0006] Preprocessing: The initial template image and the initial search image are preprocessed separately to obtain the template image and the search image;

[0007] Coarse localization involves matching the template image and the search image using at least two different deep learning image matching methods to obtain at least two matching regions of the same size as the template image, and determining a coarse matching region in the search image based on the at least two matching regions.

[0008] Fine-tuning involves matching the coarse matching region with the template image using a phase-correlation image matching method to determine the fine matching region in the search image.

[0009] Preferably, determining a coarse matching region in the search image based on the at least two matching regions includes:

[0010] Obtain the center point of each of the matching regions, and obtain the average coordinate position of all the center points or the geometric center of the polygon formed by all the center points;

[0011] The coarse matching region is obtained in the search image by taking the average coordinate position or geometric center as the center point of the coarse matching region and the size of the template image as the size of the coarse matching region.

[0012] Preferably, the at least two different deep learning image matching methods are at least two of the following: key point matching method, block feature value similarity matching method, and feature descriptor generation method.

[0013] Preferably, the coarse positioning includes:

[0014] Based on the key point matching method, the template image and the search image are matched to determine a first matching region in the search image that has the same size as the template image;

[0015] The template image and the search image are matched based on the block feature value similarity matching method to determine a second matching region in the search image that has the same size as the template image;

[0016] The coarse matching region is determined based on the first matching region and the second matching region.

[0017] Preferably, determining the coarse matching region based on the first matching region and the second matching region includes: using the average coordinate position of the center points of the first matching region and the second matching region as the center point of the coarse matching region, using the size of the template image as the size of the coarse matching region, and obtaining the coarse matching region in the search image.

[0018] Preferably, the keypoint matching method matches the template image and the search image to determine a first matching region in the search image that has the same size as the template image, including:

[0019] S101. Extract the descriptors di and location information pi of key points from the template image and the search image respectively;

[0020] S102. Based on the key point descriptor di, location information pi, and a multilayer perceptron, a vector is obtained. This vector information is then fed into a deep neural network to obtain the feature value fj of the template image. Tk and the feature value fi of the search image Sk ;

[0021] S103, Obtain the eigenvalue fj Tk and fi Sk Similarity matrix between them;

[0022] S104. By setting a similarity threshold, obtain at least two pairs of key points in the similarity matrix whose similarity satisfies the similarity threshold, and obtain a first matching region in the search image with the same size as the template image based on the at least two pairs of key points.

[0023] Preferably, in S104, the position offsets of each pair of key points in the template image and the search image are first obtained, then several pairs of key points that do not meet the offset threshold conditions are filtered out according to the preset offset threshold conditions, and finally, a first matching region with the same size as the template image is obtained in the search image based on the remaining pair of key points.

[0024] Preferably, the block-based feature value similarity matching method matches the template image and the search image, and determines a second matching region in the search image that has the same size as the template image, including:

[0025] S201. Use a convolutional neural network to cut the template image and the search image into blocks to obtain the first sub-block in the template image, the image block in the search image with the same size as the template image, and the second sub-block of the image block, and extract the features ft of the first sub-block and the features fs of the second sub-block;

[0026] S202. Calculate the cosine similarity between feature fs and ft, and the cosine similarity between feature ft and fs, respectively. Then, use the normalized exponential function to obtain the similarity ranking Soft(t|s) between feature fs and ft, and the similarity ranking Soft(s|t) between ft and fs, respectively. Calculate the matching score Score. (s,t) =Soft(t|s)*Soft(s|t);

[0027] S203. Obtain the image block with the highest matching score as the second matching region.

[0028] Preferably, the precise positioning includes:

[0029] S301. Construct formula (7) for the coarse matching region f1 and the template image f2.

[0030] f2(x,y)=f i (x-x0, y-y0) (7)

[0031] Where x0, y0 are the displacements between the coarse matching region f1 and the template image f2, and x and y are the coordinates of the template image f2;

[0032] S302. Perform a Fourier transform on the above formula (7) to obtain formula (8).

[0033]

[0034] Where u and v are the coordinates in the frequency domain, and F2(u,v) and F1(u,v) are the functions that transform the template image f2 and the coarse matching region f1 into the frequency domain, respectively.

[0035] S303. Obtain the pulse function shown in formula (9) according to formula (8), and obtain the position of the frequency where the maximum value of the pulse function is located, as the offset of the corresponding position of the template image f2 and the coarse matching region f1. Then, based on the position of the coarse matching region f1, combine the offset to obtain the position of the fine matching region, and use the size of the template image as the size of the fine matching region to determine the fine matching region in the search image.

[0036]

[0037] Where F1*(u,v) is the conjugate of F1(u,v), It is an impulse function.

[0038] Preferably, both the initial template image and the initial search image are images of a wafer.

[0039] The beneficial effects of this invention are as follows: First, coarse localization is achieved based on a deep learning image matching method, and then fine localization is achieved based on a phase correlation image matching method. This balances the high robustness and accuracy of deep learning image matching with the high precision of phase correlation image matching, effectively improving the robustness, accuracy, and precision of the image matching method. Furthermore, by determining the coarse matching region based on the matching regions obtained from at least two deep learning image matching methods, the problem of relatively low robustness in image matching methods caused by using a single deep learning image matching method can be avoided, further improving robustness. Attached Figure Description

[0040] Figure 1A A flowchart of an image matching method provided in a preferred embodiment of the present invention;

[0041] Figure 1B Another flowchart of an image matching method provided in a preferred embodiment of the present invention;

[0042] Figure 2 This is a schematic diagram illustrating the matching of the template image and the search image using a deep learning keypoint matching method in the coarse localization step of a preferred embodiment of the present invention.

[0043] Figure 3 This is a schematic diagram of the second matching region determined after matching by the deep learning block feature value similarity matching method in the coarse localization step of a preferred embodiment of the present invention;

[0044] Figure 4 This is a schematic diagram illustrating the determination of the coarse matching region in the coarse positioning step of a preferred embodiment of the present invention;

[0045] Figure 5 This is a template image used in the fine positioning step of a preferred embodiment of the present invention;

[0046] Figure 6 This is the coarse matching region used in the fine positioning step of a preferred embodiment of the present invention;

[0047] Figure 7 As a preferred embodiment of the present invention Figure 5 and Figure 6 Images obtained by direct overlay;

[0048] Figure 8 This is an image of the pulse function in the fine positioning step of a preferred embodiment of the present invention;

[0049] Figure 9 For the implementation of a preferred embodiment of the present invention, after precise positioning, Figure 5 and Figure 6 The image obtained by overlay. Detailed Implementation

[0050] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Unless otherwise defined, the technical or scientific terms used herein should have the ordinary meaning understood by those skilled in the art. The terms "comprising" and similar expressions used herein mean that the element or object preceding the word covers the element or object listed after the word and its equivalents, but does not exclude other elements or objects.

[0051] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0052] like Figure 1A As shown, an embodiment of the present invention provides an image matching method, including:

[0053] S10. Preprocessing, including preprocessing the initial template image and the initial search image to obtain the template image and the search image respectively;

[0054] In this embodiment, taking the initial template image (and similarly the initial search image) as an example, the preprocessing includes one or more of filtering and denoising the image, histogram stretching, and enhancing edge contrast. The purpose of filtering and denoising is to reduce noise in the image, such as noise caused by fluctuations in illumination.

[0055] In this embodiment, both the initial template image and the initial search image are images of the wafer; for example, the initial template image and the initial search image are obtained by acquiring images of the markings on the wafer based on an image acquisition device (optical microscope or electron microscope), but are not limited thereto.

[0056] S20. Coarse localization, including matching the template image and the search image based on at least two different deep learning image matching methods to obtain at least two matching regions of the same size as the template image, and determining a coarse matching region in the search image based on the at least two matching regions.

[0057] S30. Fine localization, including matching the coarse matching region and the template image based on the phase correlation image matching method to determine the fine matching region in the search image.

[0058] In this context, "localization" refers to determining the location. Therefore, determining the coarse matching region in the search image means determining the location of the coarse matching region in the search image, and determining the fine matching region in the search image means determining the location of the fine matching region in the search image. Furthermore, the dimensions of both the coarse and fine matching regions are the same as the dimensions of the template image.

[0059] The beneficial effects of this invention are as follows: First, coarse localization is achieved based on a deep learning image matching method, and then fine localization is achieved based on a phase correlation image matching method. This balances the high robustness and accuracy of deep learning image matching with the high precision of phase correlation image matching, effectively improving the robustness, accuracy, and precision of the image matching method. Furthermore, by determining the coarse matching region based on the matching regions obtained from at least two deep learning image matching methods, the problem of relatively low robustness in image matching methods caused by using a single deep learning image matching method can be avoided, further improving robustness.

[0060] In some embodiments, determining a coarse matching region in the search image based on at least two matching regions includes:

[0061] Obtain the center point of each matching region, and obtain the average coordinate position of all center points (i.e., the average position of the coordinates of all center points) or the geometric center of the polygon formed by all center points;

[0062] The coarse matching region is obtained in the search image by using the average coordinate position or geometric center as the center point and the size of the template image as the size of the coarse matching region.

[0063] In this embodiment, a matching region can be obtained based on any deep learning image matching method. It should be noted that regardless of whether the number of matching regions is two (e.g., the first matching region and the second matching region) or at least three, the average coordinate position of all center points can be obtained. However, when obtaining the geometric center of the polygon formed by all center points, the number of matching regions is required to be at least three. Therefore, in one embodiment, at least three matching regions are obtained based on at least three deep learning image matching methods.

[0064] In some embodiments, at least two different deep learning image matching methods are at least two of the following: (deep learning-based) keypoint matching method, (deep learning-based) block feature value similarity matching method, and (deep learning-based) feature descriptor generation method.

[0065] refer to Figure 1A-Figure 1B In this embodiment, coarse positioning includes:

[0066] The template image and the search image are matched based on the key point matching method to determine the first matching region in the search image that has the same size as the template image;

[0067] The template image and the search image are matched based on the block feature value similarity matching method to determine the second matching region in the search image that has the same size as the template image;

[0068] The coarse matching region is determined based on the first and second matching regions.

[0069] For example, such as Figure 2 As shown, image matching based on keypoint matching is performed in the search image S( Figure 2 The search image S in the figure only illustrates a portion of the area; the complete search image S is as follows: Figure 3 As shown), the first matching region 100 is obtained; Figure 3 As shown, the block feature value similarity matching method obtains the second matching region 200 in the search image S.

[0070] In some embodiments, determining the coarse matching region based on the first matching region and the second matching region includes: using the average coordinate position of the center points of the first matching region and the second matching region as the center point of the region of interest (ROI), and using the size of the template image as the size of the ROI, obtaining the ROI in the search image S. The ROI is thus the coarse matching region.

[0071] In this embodiment, as Figure 4 As shown, the center point of the line segment connecting the center points of the first matching region 100 and the second matching region 200 is taken as the center point of the region of interest (ROI). The ROI can also be called the coarse matching region S300. The size of the template image is taken as the size of the ROI. In the search image S( Figure 4 The search image S in the figure only illustrates a portion of the area; the complete search image S is as follows: Figure 3 The region of interest (ROI) is obtained as shown in the figure.

[0072] In this embodiment, matching the template image and the search image based on the key point matching method, and determining the first matching region in the search image that has the same size as the template image includes:

[0073] S101. Extract the descriptors di and location information pi of key points from the template image T and the search image S respectively;

[0074] In this embodiment, a deep learning method such as the SuperPoint method is used to extract the descriptors di and location information pi of key points from the template image T and the search image S, respectively. In other embodiments, traditional image processing methods such as SIFT can also be used.

[0075] S102. Based on the key point descriptor di, location information pi, and a multilayer perceptron, a vector is obtained. This vector information is then fed into a deep neural network to obtain the feature value fj of the template image T. Tk and the feature values ​​fi of the search image S Sk ;

[0076] Specifically, key points are encoded according to formula (1).

[0077] (0) x i =d i +MLP(p i (1)

[0078] The first layer of a deep neural network is... (0) x. For each keypoint, the keypoint descriptor di and location information pi are combined and embedded into a high-dimensional vector via a multilayer perceptron (MLP). Then, according to formula (2), the feature value fi S Deep neural networks can be used to learn to focus on subsets of keypoints based on specific attributes. This includes focusing on nearby keypoints and retrieving the relative positions of similar or prominent keypoints. W represents the weights of the connections between neurons, b is the bias term, Sk is the number of keypoints in the search image S, i is the i-th keypoint in the search image S, and L is the L-th layer of the deep neural network. Similarly, fj can be obtained. TkThe formula.

[0079] f i Sk =w· (L) x i Sk +b,i∈Sk (2)

[0080] In one embodiment, a neural network with two or more hidden layers is called a deep neural network.

[0081] S103, Obtain the eigenvalue fj Tk and fi Sk Similarity matrix between them;

[0082] Specifically, the eigenvalue fj is calculated according to formula (3). Tk and fi Sk The score for pairwise similarity comparison between them.

[0083] Score (i，j) = <f i Sk ,f j Tk >,(i,))∈Sk×Tk (3)

[0084] Among them, <fi Sk fj Tk > indicates fi Sk and fj Tk The inner product between them, j is the j-th key point in the template image T, Tk is the number of key points in the template image T, and Score(i,j) represents the score, which is the similarity matrix.

[0085] S104. By setting a similarity threshold, obtain at least two pairs of key points in the similarity matrix whose similarity satisfies the threshold. Based on these at least two pairs of key points, obtain a matching region in the search image S that is the same size as the template image. Figure 2 The first matching region in the [theory / information].

[0086] For example, in S104, the positional offsets of each pair of key points in the template image T and the search image S are first obtained. Then, several pairs of key points that do not meet the offset threshold conditions are filtered out according to a preset offset threshold condition. Finally, a first matching region with the same size as the template image is obtained in the search image based on the remaining pair of key points. The offset threshold condition includes an upper threshold. When the positional offset is greater than or equal to the upper threshold, the corresponding pair of key points is considered an anomaly and should be filtered out, i.e., discarded.

[0087] In this embodiment, the second matching region in the search image S is determined based on the block feature value similarity matching method. For the block feature value similarity matching method, it is implemented using formula (4), where R is a candidate window of fixed size (the same size as the template image T), which is the image block. The goal is to find the parameters (x, y, x) of a set of image blocks R. m and y m The optimal matching region, also known as the second matching region, is obtained by selecting the image block R with the highest overall matching score among all image blocks in the template image T and the search image S.

[0088]

[0089] Where, x m and y m Let w be the coordinates of image block R (such as the coordinates of its upper left corner), w and h be the width and height of image block R, and have the same size as the template image, s be a sub-block in image block R, called the second sub-block, and t be a sub-block in template image T, called the first sub-block.

[0090] In one embodiment, matching the template image and the search image based on the block feature value similarity matching method, and determining a second matching region in the search image that has the same size as the template image includes:

[0091] S201. Use a convolutional neural network to divide the template image T and the search image S into blocks to obtain the first sub-block t in the template image T, the image block R in the search image S with the same size as the template image T, and the second sub-block s in the image block R, and extract the features ft of the first sub-block t and the features fs of the second sub-block s.

[0092] S202. Calculate the cosine similarity between features fs and ft, and the cosine similarity between features ft and fs, respectively. Then, use a softmax activation layer (i.e., a normalized exponential function) to obtain the similarity ranking (also called similarity) Soft(t|s) between features fs and ft, and the similarity ranking Soft(s|t) between features ft and fs, respectively. Calculate the matching score Score. (s,t) ;

[0093] Specifically, the similarity ranking Soft(t|s) is calculated according to formula (5).

[0094]

[0095] Where * represents multiplication, and P is a preset value, such as a learnable parameter, that is, to allow the convolutional neural network to find a suitable P such that Soft(t|s) is 1 when matching and Soft(t|s) is 0 when not matching.

[0096] Calculate the matching score according to formula (6). (s,t) ,

[0097] Score (s，t) =Soft (t|s) ·Softe (s|t) (6)

[0098] Among them, Score (s,t) Soft(t|s) is the product of the similarity ranking Soft(t|t) of the second sub-block s in the search image S matching the first sub-block t in the template image T and the similarity ranking Soft(s|t) of the first sub-block t in the template image T matching the second sub-block s in the search image S.

[0099] S203. Obtain the image block R with the maximum matching score as the second matching region.

[0100] As shown in formulas (4) and (6), obtain the parameter x mentioned above in formula (6) that corresponds to the maximum similarity score. m and y m These parameters correspond to image patches R, which are the matching regions, such as... Figure 3 The second matching region is 200.

[0101] In this embodiment, precise positioning includes:

[0102] S301. Construct formula (7) for the coarse matching region f1 and template image f2 (i.e., template image T).

[0103] f2(x,y)=f1(x-x0,y-y0) (7)

[0104] Where x0,y0 is the displacement between the coarse matching region f1 and the template image f2, and x and y are the coordinates of the template image f2;

[0105] S302. Perform a Fourier transform on formula (7) to obtain formula (8).

[0106]

[0107] Where * represents multiplication, u and v are the coordinates in the frequency domain, and F2(u,v) and F1(u,v) are the functions that transform the template image f2 and the coarse matching region f1 into the frequency domain, respectively.

[0108] S303. Obtain the pulse function shown in formula (9) according to formula (8), and obtain the position of the frequency where the maximum value of the pulse function is located, as the offset of the corresponding position of the template image f2 and the coarse matching region f1. Then, based on the position of the coarse matching region f1, combine the offset to obtain the position of the fine matching region, and use the size of the template image as the size of the fine matching region to determine the fine matching region in the search image.

[0109]

[0110] Where F1*(u,v) is the conjugate of F1(u,v), It is an impulse function.

[0111] Specifically, in this embodiment, using Figure 5 The template image shown is used to determine the first and second matching regions using keypoint matching and block feature value similarity matching methods. The average coordinates of the center points of the first and second matching regions are used as the center point of the region of interest (ROI), and the size of the template image is used as the size of the ROI. The ROI is then obtained in the search image, i.e., as shown... Figure 6 The coarse matching area is shown. Figure 7 for Figure 5 and Figure 6 The images are directly superimposed, and a clear displacement can be seen between them. Figure 8 The graph shows the impulse function, with a distinct white dot in the lower right corner. This dot can be used to calculate... Figure 5 and Figure 6 The displacement. Figure 9 for Figure 5 and Figure 6 After successful registration (i.e., image matching followed by overlay), it's easy to see that areas with the same texture overlap. The term "overlay" refers to the overlapping of successfully matched keypoint pairs after template matching, placing the matching regions in both the template and search images on a single image. This facilitates observation of the positional relationship between the template and the matching regions.

[0112] This embodiment also provides a wafer positioning method, including: obtaining a fine matching region and the relative positional relationship between the template image and the fine matching region based on the image matching method described above, storing the relative positional relationship or correcting the wafer position based on the relative positional relationship.

[0113] For example, the relative positional relationship includes translation and / or rotation angle, and the relative positional relationship can be represented in matrix form. Furthermore, those skilled in the art will understand that whether storing the relative positional relationship or directly correcting the wafer position based on it, wafer positioning is achieved.

[0114] While embodiments of the present invention have been described in detail above, it will be apparent to those skilled in the art that various modifications and variations can be made to these embodiments. However, it should be understood that such modifications and variations fall within the scope and spirit of the invention as set forth in the claims. Furthermore, the invention described herein may have other embodiments and can be implemented or carried out in various ways.

Claims

1. An image matching method, characterized in that, include: Preprocessing: The initial template image and the initial search image are preprocessed separately to obtain the template image and the search image; Coarse localization involves matching the template image and the search image using at least two different deep learning image matching methods to obtain at least two matching regions of the same size as the template image, and determining a coarse matching region in the search image based on the at least two matching regions. The coarse positioning includes: Based on the key point matching method, the template image and the search image are matched to determine a first matching region in the search image that has the same size as the template image; The template image and the search image are matched based on the block feature value similarity matching method to determine a second matching region in the search image that has the same size as the template image; The coarse matching region is determined based on the first matching region and the second matching region; The keypoint matching method matches the template image and the search image, and determines a first matching region in the search image that has the same size as the template image, including: S101. Extract the descriptors di and location information pi of key points from the template image and the search image respectively; S102. Based on the key point descriptor di, location information pi, and a multilayer perceptron, a vector is obtained. This vector information is then fed into a deep neural network to obtain the feature value fj of the template image. Tk and the feature value fi of the search image Sk ; S103, Obtain the feature value fj Tk and fi Sk Similarity matrix between them; S104. By setting a similarity threshold, obtain at least two pairs of key points in the similarity matrix whose similarity satisfies the similarity threshold, and obtain a first matching region in the search image with the same size as the template image based on the at least two pairs of key points. Fine-tuning involves matching the coarse matching region with the template image using a phase-correlation image matching method to determine the fine matching region in the search image.

2. The image matching method according to claim 1, characterized in that, Determining a coarse matching region in the search image based on the at least two matching regions includes: Obtain the center point of each of the matching regions, and obtain the average coordinate position of all the center points or the geometric center of the polygon formed by all the center points; The coarse matching region is obtained in the search image by taking the average coordinate position or geometric center as the center point of the coarse matching region and the size of the template image as the size of the coarse matching region.

3. The image matching method according to claim 1, characterized in that, Determining the coarse matching region based on the first matching region and the second matching region includes: using the average coordinate position of the center points of the first matching region and the second matching region as the center point of the coarse matching region, using the size of the template image as the size of the coarse matching region, and obtaining the coarse matching region in the search image.

4. The image matching method according to claim 1, characterized in that, In S104, the position offset of each pair of key points in the template image and the search image is first obtained. Then, according to the preset offset threshold condition, several pairs of key points that do not meet the offset threshold condition are filtered out. Finally, the remaining pair of key points are used to obtain a first matching region in the search image with the same size as the template image.

5. The image matching method according to claim 1, characterized in that, The block-based feature value similarity matching method matches the template image and the search image, and determines a second matching region in the search image that has the same size as the template image, including: S201. Use a convolutional neural network to cut the template image and the search image into blocks to obtain the first sub-block in the template image, the image block in the search image with the same size as the template image, and the second sub-block of the image block, and extract the features ft of the first sub-block and the features fs of the second sub-block; S202. Calculate the cosine similarity between feature fs and ft, and the cosine similarity between feature ft and fs, respectively, and obtain the similarity ranking between feature fs and ft using the normalized exponential function. ( | ), and obtain the similarity ranking of the features ft and fs. ( | ), calculate matching score core (s,t) = ( | )* ( | ); S203. Obtain the image block with the highest matching score as the second matching region.

6. The image matching method according to claim 1, characterized in that, The precise positioning includes: S301. Construct the formula (7) for the coarse matching region f1 and the template image f2. （7） Where x0 and y0 are the displacements between the coarse matching region f1 and the template image f2, and x and y are the coordinates of the template image f2; S302. Perform a Fourier transform on the above formula (7) to obtain formula (8). （8） Where u and v are coordinates in the frequency domain. and These are functions that transform the template image f2 and the coarse matching region f1 into the frequency domain, respectively. S303. Obtain the pulse function shown in formula (9) according to formula (8), and obtain the position of the frequency where the maximum value of the pulse function is located, as the offset of the corresponding position of the template image f2 and the coarse matching region f1. Then, based on the position of the coarse matching region f1, combine the offset to obtain the position of the fine matching region, and use the size of the template image as the size of the fine matching region to determine the fine matching region in the search image. （9） Where F1*(u,v) is the conjugate of F1(u,v), It is an impulse function.

7. The image matching method according to claim 1, characterized in that, Both the initial template image and the initial search image are images of wafers.

Images