A high-precision alignment method for detecting moiré fringes based on convolutional neural networks

Abstract

The present invention discloses a high-precision alignment method based on convolutional neural network detection of moiré fringes, including a calibration stage and an alignment stage; in the design of alignment marks, only two pairs of circular grating marks are used, and the alignment mark pattern is simple. The circular grating can serve as a rough alignment and a fine alignment mark at the same time, and belongs to an alignment method combining geometric images with phase information. Compared with the traditional light intensity information alignment method, it does not require the design of a complex optical structure, and has better process adaptability and lower implementation cost; compared with the simple geometric image alignment method, the present invention combines the characteristics of the moiré fringes generated by the superposition of circular gratings to amplify tiny displacements, and can achieve deviation measurement higher than the pixel level; compared with the simple phase information alignment method, the present invention combines the use of grating circle outer contour positioning as a rough alignment deviation measurement, and can achieve a wider range of deviation measurement.

Description

Technical Field

[0001] The invention relates to an alignment method, and in particular to a high-precision alignment method based on convolutional neural network detection of moiré fringes. Background Art

[0002] Alignment technology is a technology that accurately adjusts objects to the desired position or state. It is widely used in the manufacturing industry, such as photolithography machines, vapor deposition machines, and placement machines. The mainstream alignment methods can be divided into three categories: geometric image alignment, light intensity information alignment, and phase information alignment. The geometric image alignment method uses the image of the alignment mark to achieve alignment. It is simple and has a wide range of applications, but the accuracy is low. Among them, the method based on machine vision and Mark points is the most common. The light intensity information alignment method achieves alignment by detecting the intensity of a specific diffraction order or the maximum critical intensity. It has a high signal-to-noise ratio, high accuracy, and good real-time registration performance, but the process adaptability is poor. The phase information alignment method uses the moiré fringe signal generated between the substrate and the mask to achieve alignment. Combining fringe processing and high-precision interference technology can achieve higher accuracy. Common grating marks used in the alignment methods based on phase information include linear gratings, circular gratings, and some specially designed combined linear gratings. These grating marks can all measure the deviations in the alignment process very well. Among them, for linear gratings, although the generated Moire fringe pattern is easy to detect, it can only measure the position deviation in one direction at the same time, and there is a problem of phase periodicity in the measurement process; for circular gratings, the position deviation in two directions can be measured simultaneously, but the Moire fringes generated by the superposition of its gratings are complex and difficult to detect accurately; for specially designed combined linear gratings, although the unidirectional measurement problem existing in linear gratings can be solved, the marking will become more complicated, and the use of traditional signal analysis methods based on phase information will inevitably lead to spectrum leakage problems, and the accuracy of its extraction and detection and the robustness of the method will also be affected. Summary of the invention

[0003] In order to solve the deficiencies of the prior art, the object of the present invention is to provide a high-precision alignment method for detecting circular grating moiré fringes based on a convolutional neural network. The present invention is achieved through the following technical solutions:

[0004] The invention discloses a high-precision alignment method based on convolutional neural network to detect moiré fringes, comprising: a calibration stage and an alignment stage;

[0005] Calibration phase:

[0006] Design auxiliary point circular grating alignment marks: circular grating A and circular grating B;

[0007] Place the designed circular grating alignment marks, circular grating A and circular grating B, at two corner points on the rigid substrate and the rigid mask;

[0008] An image including a rigid substrate is obtained, and the angle between the horizontal direction of the camera field of view and the horizontal direction of the rigid substrate is calibrated for the image so that the angles overlap in the horizontal direction;

[0009] Obtain a checkerboard image, calibrate the internal and external parameter matrix of the industrial camera on the checkerboard image, and obtain the actual size of the image pixels;

[0010] Counterpoint Phase:

[0011] Obtaining an image including circular grating A and circular grating B at corresponding corner positions during the alignment process of the rigid substrate and the rigid mask;

[0012] For the image containing circular grating A and circular grating B, use coarse alignment to preliminarily locate the center of the circular grating and obtain the distance between circular grating A and circular grating B;

[0013] When the distance between circular grating A and circular grating B is large, coarse alignment is used to obtain the pixel-level eccentricity and eccentricity angle at the corner point position;

[0014] When the distance between circular grating A and circular grating B is small, the circular grating moiré fringe detection is used to perform precise alignment to obtain the circular grating moiré fringe image. The circular grating moiré fringe image amplifies the tiny displacement of the circular grating, and the tiny eccentricity and eccentricity angle beyond the pixel level at the corner position are obtained by detecting the circular grating moiré fringe image.

[0015] According to the eccentricity and eccentricity angle at the two diagonal points, the overall alignment deviation is calculated to obtain the position deviation of the rigid substrate and the rigid mask center;

[0016] The position deviation of the center of the rigid substrate and the rigid mask is input into the UVW displacement platform to complete the alignment of the rigid substrate and the rigid mask.

[0017] As a further improvement, the design auxiliary point circular grating alignment mark is specifically as follows: circular grating indicator marks are placed at the two corner points on the rigid substrate and the rigid mask respectively, and for each pair of detection marks, it includes a part located on the rigid substrate and a part located on the rigid mask; there is a circular grating B with a grating constant of p1, a number of turns of n1, and an outermost radius of R1 on the rigid substrate; there is a circular grating A with a grating constant of p2, a number of turns of n2, and an outermost diameter of R2 on the rigid mask; the background of the mark is transparent by default.

[0018] As a further improvement, the angle between the horizontal direction of the image calibration camera field of view and the horizontal direction of the rigid substrate is overlapped in the horizontal direction. Specifically, the image captured by the industrial camera is first binarized, and then the ROI area at the edge of the calibration plate is cropped, and the Canny operator is used to perform edge detection on it. Then, the Hough line transform is used to detect the slope of the frame edge line to obtain the angle β between the camera field of view and the horizontal direction of the substrate, and then the position of the rigid substrate is adjusted according to the angle between the camera field of view and the horizontal direction of the substrate, and the camera field of view is repeatedly overlapped with the rigid substrate in the horizontal direction.

[0019] As a further improvement, the rough alignment is specifically as follows: when the distance between the centers of the circular grating B on the rigid substrate and the circular grating A on the rigid mask at the corner point is greater than R1, rough alignment is performed; after the rough alignment performs binarization processing on the image collected by the industrial camera, the image ROI containing the two gratings is cropped, Fourier transform is performed to the frequency domain, Gaussian low-pass filtering is performed, and Fourier inverse transform is performed to obtain the image after the frequency domain low-frequency filtering, the high-frequency periodic structure on the grating is eliminated, the outermost circular contour is retained, the outermost circular contour is subjected to Hough circle transform, and the center coordinates of the two circular gratings are respectively obtained, and the eccentricity d and the eccentricity angle θ are converted.

[0020] As a further improvement, the circular grating moiré fringes are specifically: the superposition of two circular gratings with similar periods will produce moiré fringes, the moiré fringes have the characteristic of amplifying small displacements, and the amplitude distribution of the moiré fringes satisfies

[0021] I m (r,θ)=cos{2π[r(f1-f2)+εf2 cos(θ-α)]};

[0022] Where, f1 and f2 are the grating frequencies of circular grating A and circular grating B, ε and α are the eccentricity and eccentricity angle of circular grating A and circular grating B; when the eccentricity is constant, the change of different eccentricity angles will cause the direction of the entire moiré fringe pattern to change significantly, and the opposite number of the direction of the moiré fringe pattern is the eccentricity angle; when the eccentricity angle is constant, the change of different eccentricity will cause the number and shape of the dark fringes of the entire moiré fringe pattern to change significantly; the magnification of the circular grating moiré fringe to small displacement is

[0023]

[0024] When the pixel size of the industrial camera is u, the minimum displacement d that can be detected theoretically is min for

[0025]

[0026] As a further improvement, the circular grating moiré fringe detection is precisely aligned, specifically: when the center distance between the circular grating B on the rigid substrate and the circular grating A on the rigid mask at the corner point position is less than R1, precise alignment is performed; the corner point deviation is obtained by detecting the complex moiré fringes generated by the periodic superposition of circular gratings; images of moiré fringes generated by the superposition of circular gratings A and B at various eccentricities and eccentricity angles are collected in advance, and input into a convolutional neural network for training to learn the mapping of complex moiré fringes to eccentricities and eccentricity angles; in the actual measurement process, the moiré fringe image generated by the superposition of circular gratings A and B collected by the industrial camera is input into the convolutional neural network model for prediction to obtain the eccentricity and eccentricity angle.

[0027] As a further improvement, the input is trained in the convolutional neural network, specifically: a classification network model is adopted; a network model is adopted for each eccentricity and eccentricity; N p2s are classified in the category of eccentricity classification, and the number of categories is c1; 360° is classified in the category of eccentricity classification, and the number of categories is c2; by using a displacement platform, when the circular grating A and the circular grating B are completely overlapped, the step length of each eccentricity is Or the step length of the eccentricity angle Collect moiré fringe images and input them into the classification network model for model training; in the selection of specific parameters for eccentricity detection, it is required Slightly larger than d min , the minimum displacement resolution that can actually be detected is In the selection of specific parameters for eccentric angle detection, the minimum angle resolution that can actually be detected is

[0028] As a further improvement, the overall alignment deviation calculation is specifically as follows: after obtaining the eccentricity d1, d2 and eccentricity angles θ1, θ2 of two corner points, the posture deviation between the rigid substrate and the rigid mask is obtained.

[0029]

[0030] Among them, e1 and e2 are the installation position parameters of corner points 1 and 2; finally, the actual position deviation of the rigid substrate and the rigid mask is obtained by conversion according to the actual size u of each pixel on the image.

[0031]

[0032] Compared with the prior art, the present invention has the following beneficial effects:

[0033] 1. In the design of alignment marks, only two pairs of circular grating marks are used, and the alignment mark pattern is simple. The circular grating can serve as both a rough alignment and a fine alignment mark, and is an alignment method that combines geometric images with phase information. Compared with the traditional light intensity information alignment method, it does not require the design of a complex optical structure, and has better process adaptability and lower implementation costs; compared with the simple geometric image alignment method, the present invention combines the characteristics of the moiré fringes generated by the superposition of circular gratings to amplify tiny displacements, and can achieve deviation measurement higher than the pixel level; compared with the simple phase information alignment method, the present invention combines the use of grating circle outer contour positioning as a rough alignment deviation measurement, and can achieve a wider range of deviation measurement.

[0034] 2. In terms of grating selection, the present invention selects a circular grating. Compared with a linear grating, it can simultaneously measure the position deviation in two directions in the plane, and there is no problem of narrow measurement range caused by the phase periodicity of the linear grating. No additional methods (such as symmetrical grating, grating encoding, etc.) are required to overcome the measurement periodicity problem.

[0035] 3. In terms of the detection of circular grating moiré fringe phase information, the present invention adopts a convolutional neural network model to learn the mapping of complex moiré fringe patterns to eccentricity and eccentricity angle. Compared with the traditional Fourier transform method of extracting the phase of moiré fringe patterns, there is no spectrum leakage problem, and the influence of factors such as grating manufacturing defects and image noise on deviation measurement can be reduced, and it has higher accuracy and robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 It is a schematic diagram of alignment in the present invention;

[0037] In the figure, 1 is a rigid substrate 1, which remains stationary during the alignment process; 2 is a rigid mask, which is rigidly fixed on the displacement platform. During the alignment process, the position of the rigid mask is adjusted to align the rigid substrate 1 with the rigid mask; 3 is a circular grating A, which is located on the rigid mask, and has a grating constant of p2 = 10 μm, a number of turns of 25, and an outermost diameter of R2 = 252.5 μm; 4 is a circular grating B, which is located on the rigid substrate 1, and has a grating constant of p1 = 9.5 μm, a number of turns of 25, and an outermost radius of R1 = 248.875 μm; 5 is a moiré fringe, which is generated by the superposition of circular gratings A and B;

[0038] Figure 2 It is a flowchart of the alignment of the method of the present invention;

[0039] Figure 3 It is a flow chart of deviation measurement of rough alignment of the method of the present invention;

[0040] Figure 4It is the moiré fringe pattern generated by the superposition of the circular grating A and the circular grating B in the present invention at different eccentricity distances d and eccentricity angles α. It can be seen that the moiré fringe pattern generated by the superposition of the two circular gratings at different eccentricity distances and eccentricity angles is complex, but the difference is obvious;

[0041] Figure 5 It is a flow chart of deviation measurement of precise alignment of the method of the present invention. DETAILED DESCRIPTION

[0042] The implementation steps of the present invention are as follows:

[0043] Step 1: Design auxiliary point circular grating alignment marks, and place the designed circular grating alignment marks at the two corner points on the rigid substrate 1 and the rigid mask 2. Figure 1 There are two pairs on the rigid substrate 1 and the rigid mask 2, distributed at two corner points; for each pair of detection marks, including the part located on the rigid substrate 1 and the part located on the rigid mask 2, there is a circular grating B4 mark 1 on the rigid substrate 1 with a grating constant of p1=9.5μm, a number of turns of 25, and an outermost radius of R1=248.875μm, and there is a circular grating A3 mark on the rigid mask 2 with a grating constant of p2=10μm, a number of turns of 25, and an outermost diameter of R2=252.5μm; the background of the mark is transparent by default; the superposition of circular grating A3 and circular grating B4 will produce moiré fringes 5.

[0044] Step 2: System calibration. Calibrate the internal and external parameter matrix of the industrial camera by shooting a checkerboard image to obtain the actual size u of the image pixel; calibrate the angle between the camera field of view and the horizontal direction of the rigid substrate 1, binarize the image captured by the industrial camera, and then crop the ROI area at the edge of the calibration plate, use the Canny operator to perform edge detection, and then use the Hough line transform to detect the slope of the border edge line, so as to obtain the angle β between the camera field of view and the horizontal direction of the rigid substrate 1, and then adjust the position of the substrate according to the angle between the camera field of view and the horizontal direction of the rigid substrate 1, and repeat it continuously to make the camera field of view overlap with the rigid substrate 1 in the horizontal direction.

[0045] Step 3: Train the convolutional neural network for moiré fringe 5 detection. The network model uses the ResNet101 classification model; for the prediction tasks of eccentricity and eccentricity angle, a classification model is used to train and predict the eccentricity and eccentricity angle respectively; 25 p2 are classified in the eccentricity classification category, the classification step is 0.025p2, a total of 1000 categories, the measurement range is 250μm, and the measurement resolution can reach 0.25μm; 360° is classified in the eccentricity classification category, the classification step is 0.36°, a total of 1000 categories, and the measurement resolution can reach 0.36°; in the data set, refer to Figure 4 , the moiré fringe 5 images generated by circular grating A3 and circular grating B4 at different eccentricities and eccentricity angles are taken as the training data set. Specifically, by using a displacement platform, when the circular grating A3 and circular grating B4 overlap, the moiré fringe 5 images generated by the overlap of circular grating A3 and circular grating B4 are collected at every eccentricity or eccentricity angle step, totaling 1 million images, and these 1 million images are input into the classification network model for model training; in model training, the training strategy of transfer learning is adopted, and the ResNet101 model parameters used for other classification tasks are transferred to train this classification task, thereby reducing the model learning time and improving the model learning efficiency.

[0046] Step 4: When the distance between the circular grating A3 and the circular grating B4 is large, use rough alignment to obtain the eccentricity and eccentricity angle at the corner point. Figure 3 When the posture deviation of the rigid substrate 1 and the rigid mask 2 is large, that is, when the center distance between the circular grating B4 on the rigid substrate 1 and the circular grating A3 on the rigid mask 2 at any corner point is greater than R1, a rough alignment is performed; the rough alignment performs binarization processing on the image collected by the industrial camera, cuts the image ROI containing the two gratings, performs Fourier transform to the frequency domain, performs Gaussian low-pass filtering, and performs inverse Fourier transform to obtain the image after low-frequency filtering in the frequency domain, eliminates the high-frequency periodic structure on the grating, retains the outermost circular contour, performs Hough circle transform on the outermost circular contour, and obtains the center coordinates of the two circular gratings respectively, which are converted into eccentricity d and eccentricity angle θ; this method can obtain pixel-level measurement accuracy.

[0047] Step 5: When the distance between the circular grating A3 and the circular grating B4 is small, the circular grating moiré fringe 5 is detected for precise alignment to obtain the eccentricity and eccentricity angle at the corner point. Figure 5 , when the posture deviation of the rigid substrate 1 and the rigid mask 2 is small, that is, when the center distance between the circular grating B4 on the rigid substrate 1 and the circular grating A3 on the rigid mask 2 at any corner point is less than R1, precise alignment is performed; since the distance between the two circular gratings is relatively close at this time, it is difficult to correctly detect the corner deviation through conventional image detection algorithms. At this time, the corner deviation is obtained by detecting the moiré fringes 5 generated by the superposition of the periodic circular gratings A3 and B4; after the image collected by the industrial camera is binarized, the image ROI containing the two gratings is cropped, and the images are input into the eccentricity and eccentricity angle classification network model trained in step 2, and the eccentricity and eccentricity angle are predicted, and the corresponding label category is obtained and multiplied by the corresponding classification step to obtain the eccentricity and eccentricity angle; this method can obtain measurement accuracy beyond the pixel level.

[0048] Step 6: According to the eccentric angle and the eccentric angle at the two diagonal points, the overall alignment deviation is calculated to obtain the position deviation of the center of the rigid substrate 1 and the rigid mask 2. Figure 2 , Steps 4 and 5 are aimed at measuring the position deviation of a single corner point. In the process of aligning two rigid planes, it is necessary to convert the position deviation measurement at the corner point into the posture deviation of the center points of the two rigid planes, and then input this posture deviation as an input parameter into the displacement control platform (such as the UVW displacement platform) to complete the final alignment; specifically, the offset images of the circular grating A3 and the circular grating B4 at the two corner point positions are continuously collected, and the eccentricity and eccentricity angle between the circular grating A3 and the circular grating B4 at the corresponding corner point positions are extracted using step 4 or step 5, and then the center point posture deviation between the rigid substrate 1 and the rigid mask 2 is calculated.

[0049]

[0050] Among them, the eccentricity d1, d2 and the eccentricity angle θ1, θ2 are the eccentricity and eccentricity of the circular grating A3 and the circular grating B4 at the positions of the corner points 1 and 2; e1 and e2 are the installation position parameters of the corner points 1 and 2; finally, the actual position deviation of the rigid substrate 1 and the rigid mask 2 is obtained by conversion according to the actual size of each pixel on the image

[0051]

[0052] Step 7: Input the posture deviation of the center of the rigid substrate 1 and the rigid mask 2 into the UVW displacement platform, and control the displacement platform to move to complete the alignment of the rigid substrate 1 and the rigid mask 2.

[0053] Finally, it should be noted that the above examples are only specific embodiments of the present invention. Obviously, the present invention is not limited to the above examples, and there are many variations. All variations that can be directly derived or associated with the content disclosed by a person skilled in the art should be considered as the protection scope of the present invention.

Claims

1. A high-precision alignment method for detecting moiré fringes based on a convolutional neural network, characterized in that: include: The calibration phase and the alignment phase; The calibration phase described: Design auxiliary point circular grating alignment marks: circular grating A and circular grating B; Place the designed circular grating alignment marks, circular grating A and circular grating B, at two corner points on the rigid substrate and the rigid mask; An image including a rigid substrate is obtained, and the angle between the horizontal direction of the camera field of view and the horizontal direction of the rigid substrate is calibrated for the image so that the angles overlap in the horizontal direction; Obtain a checkerboard image, calibrate the internal and external parameter matrix of the industrial camera on the checkerboard image, and obtain the actual size of the image pixels; The alignment phase: Obtaining an image including circular grating A and circular grating B at corresponding corner positions during the alignment process of the rigid substrate and the rigid mask; For the image containing circular grating A and circular grating B, use coarse alignment to preliminarily locate the center of the circular grating and obtain the distance between circular grating A and circular grating B; When the distance between circular grating A and circular grating B is greater than R1, coarse alignment is used to obtain the pixel-level eccentricity and eccentricity angle at the corner point position; When the distance between circular grating A and circular grating B is less than R1, the circular grating moiré fringe detection is used to perform precise alignment to obtain a circular grating moiré fringe image, the circular grating moiré fringe image amplifies the tiny displacement of the circular grating, and the tiny eccentricity and eccentricity angle beyond the pixel level at the corner position are obtained by detecting the circular grating moiré fringe image; According to the eccentricity and eccentricity angle at the two diagonal points, the overall alignment deviation is calculated to obtain the position deviation of the rigid substrate and the rigid mask center; Input the position deviation of the center of the rigid substrate and the rigid mask into the UVW displacement platform to complete the alignment of the rigid substrate and the rigid mask; R1 is the outermost radius of the ring B.

2. The high-precision alignment method for detecting moiré fringes based on a convolutional neural network according to claim 1, characterized in that: The design auxiliary point circular grating alignment mark is specifically as follows: circular grating indicator marks are placed at two corner points on the rigid substrate and the rigid mask respectively, and for each pair of detection marks, it includes a part located on the rigid substrate and a part located on the rigid mask; there is a circular grating B with a grating constant of p1, a number of turns of n1, and an outermost radius of R1 on the rigid substrate; there is a circular grating A with a grating constant of p2, a number of turns of n2, and an outermost diameter of R2 on the rigid mask; the background of the mark is transparent by default.

3. The high-precision alignment method for detecting moiré fringes based on a convolutional neural network according to claim 1, characterized in that: The method of calibrating the angle between the horizontal direction of the camera field of view and the horizontal direction of the rigid substrate so that they overlap in the horizontal direction is as follows: first, the image captured by the industrial camera is binarized, and then the ROI area at the edge of the calibration plate is cropped, and the Canny operator is used to perform edge detection on it, and then the Hough line transform is used to detect the slope of the frame edge line, so as to obtain the angle β between the camera field of view and the horizontal direction of the substrate, and then the position of the rigid substrate is adjusted according to the angle between the camera field of view and the horizontal direction of the substrate, and the camera field of view and the rigid substrate are repeatedly overlapped in the horizontal direction.

4. The high-precision alignment method based on convolutional neural network detection of moiré fringes according to claim 1, 2 or 3, characterized in that: The rough alignment is specifically as follows: when the distance between the centers of the circular grating B on the rigid substrate and the circular grating A on the rigid mask at the corner point is greater than R1, rough alignment is performed; after the image collected by the industrial camera is binarized, the image ROI containing the two gratings is cropped, Fourier transform is performed to the frequency domain, Gaussian low-pass filtering is performed, and Fourier inverse transform is performed to obtain a picture after low-frequency filtering in the frequency domain, the high-frequency periodic structure on the grating is eliminated, the outermost circular contour is retained, the outermost circular contour is subjected to Hough circle transform, the center coordinates of the two circular gratings are respectively obtained, and the eccentricity d and the eccentricity angle θ are converted.

5. The high-precision alignment method for detecting moiré fringes based on a convolutional neural network according to claim 4, characterized in that: The circular grating moiré fringes are specifically: the superposition of two circular gratings with similar periods will produce moiré fringes, the moiré fringes have the characteristic of amplifying tiny displacements, and the amplitude distribution of the moiré fringes satisfies I m (r,θ)=cos{2π[r(f1-f2)+εf2cos(θ-α)]}; Where, f1 and f2 are the grating frequencies of circular grating A and circular grating B, ε and α are the eccentricity and eccentricity angle of circular grating A and circular grating B; when the eccentricity is constant, the change of different eccentricity angles will cause the direction of the entire moiré fringe pattern to change significantly, and the opposite number of the direction of the moiré fringe pattern is the eccentricity angle; when the eccentricity angle is constant, the change of different eccentricity will cause the number and shape of the dark fringes of the entire moiré fringe pattern to change significantly; the magnification of the circular grating moiré fringe to small displacement is When the pixel size of the industrial camera is u, the minimum displacement d that can be detected theoretically is min for 6. The high-precision alignment method for detecting moiré fringes based on a convolutional neural network according to claim 4, characterized in that: The precise alignment of the circular grating moiré fringe detection is specifically as follows: when the distance between the centers of the circular grating B on the rigid substrate and the circular grating A on the rigid mask at the corner position is less than R1, precise alignment is performed; the corner point deviation is obtained by detecting the complex moiré fringe generated by the periodic superposition of circular gratings; and images of the moiré fringe generated by the superposition of circular gratings A and circular gratings B under various eccentricities and eccentricity angles are collected in advance, and input into the convolutional neural network for training to learn the mapping of complex moiré fringe to eccentricity and eccentricity angle; In the actual measurement process, the circular grating A and circular grating B captured by the industrial camera are superimposed to generate a moiré fringe image, which is input into the convolutional neural network model for prediction to obtain the eccentricity and eccentricity angle.

7. The high-precision alignment method for detecting moiré fringes based on a convolutional neural network according to claim 6, characterized in that: The input is trained in the convolutional neural network, specifically: adopting a classification network model; adopting a network model for each of the eccentricity and the eccentricity angle; classifying N p2s in the eccentricity classification category, the number of categories is c1; classifying 360° in the eccentricity classification category, the number of categories is c2; By using a translation platform, when the ring grating A and the ring grating B are completely overlapped, the step length of each eccentricity is Or the step length of the eccentricity angle Collect moiré fringe images and input them into the classification network model for model training; in the selection of specific parameters for eccentricity detection, it is required Slightly larger than d min , the minimum displacement resolution that can actually be detected is In the selection of specific parameters for eccentric angle detection, the minimum angle resolution that can actually be detected is 8. The high-precision alignment method based on convolutional neural network detection of moiré fringes according to claim 1 or 2 or 3 or 5 or 6 or 7, characterized in that: The overall alignment deviation calculation is specifically as follows: after obtaining the eccentricity d1, d2 and eccentricity angles θ1, θ2 of two corner points, the posture deviation between the rigid substrate and the rigid mask is obtained. Among them, e1 and e2 are the installation position parameters of corner points 1 and 2; finally, the actual position deviation of the rigid substrate and the rigid mask is obtained by conversion according to the actual size u of each pixel on the image.

Images