A method for extracting full-band errors of optical component surface shape based on neural network

Abstract

The present invention relates to the field of surface shape detection technology, and in particular to a method for extracting full-band errors of optical element surface shape based on a neural network, including S1: collecting surface shape detection results of different optical elements, and extracting corresponding full-band errors from each surface shape detection result, and packaging each surface shape detection result and the corresponding full-band error as a training data set to train three error extraction neural networks; S2: sorting and packaging the surface shape detection results of the optical element to be measured and inputting them into the three error extraction neural networks that have completed the training, and extracting the errors of each frequency band of the surface shape of the optical element to be measured through the three error extraction neural networks. The present invention can realize the automatic extraction of the full-band error of any optical surface shape without manual operation; the error neural network can process a large amount of surface shape data in batches, greatly improving the extraction efficiency of the full-band error, and facilitating subsequent analysis.

Description

Technical Field

[0001] The invention belongs to the technical field of surface shape detection, and in particular relates to a method for extracting full-band errors of optical element surface shape based on a neural network. Background Art

[0002] With the continuous development of optical component testing technology and the increasing demand for optical system applications, people have proposed a way to describe the surface error of optical components in the field of optical testing. Due to certain surface defects in the surface of optical components, surface manufacturing residuals introduced in the actual processing process, and slight changes introduced by the testing environment, there are different degrees of deviation between the physical distribution of surface vector height and the theoretical surface vector height. This deviation is reflected in space as different spatial frequencies. According to the different distributions of spatial frequencies, the surface error of optical components can be divided into different frequency band errors, namely low frequency (Figure, LSF), mid-frequency (Mid-Spatial Frequency, MSF), and high frequency (High Spatial Frequency, HSF), such as Figure 1 As shown in the figure, the three are collectively referred to as the surface frequency band error of the optical element, that is, the full frequency band error.

[0003] Among them, low-frequency errors usually refer to errors with longer spatial period lengths, which generally correspond to surface errors of optical elements, that is, traditional aberrations, which can be described by Zernike aberrations or Seidel aberrations. They represent the deviation between the surface morphology of optical elements and their theoretical design morphology, and will directly affect the imaging quality, wavefront aberrations and system stray light of the optical system; medium-frequency errors usually refer to errors with medium spatial period lengths, which generally correspond to waviness errors such as tool marks introduced during the processing process. The impact on the optical system is mainly reflected in small-angle scattering, which may produce flare and affect the system contrast; high-frequency errors usually refer to errors with shorter spatial period lengths, which generally correspond to microscopic features or roughness of the surface, such as scratches and pockmarks. The impact on the optical system is mainly reflected in large-angle scattering, which affects the clarity and sharpness of the image. The impact of errors in different frequency bands on the imaging of the optical system is as follows: Figure 2 Therefore, when evaluating the quality of the optical component surface or analyzing the performance of the optical system, it is very important to accurately obtain the full-band error of the optical component surface, which helps to analyze the errors in different frequency bands separately, so as to explore the impact of errors in different frequency bands on the optical system separately, and facilitate targeted convergence in the subsequent processing process.

[0004] To extract full-band errors from the surface of an optical component, the main principle is that the low-frequency error, mid-frequency error, and high-frequency error of the same optical component surface correspond to different frequency ranges in the frequency domain, so the full-band errors can be separated by frequency domain filtering. Currently, the commonly used method for extracting full-band errors is to use optical measurement and analysis software such as MetroPro, Mx, or mathematical software such as MATLAB for filtering. The main implementation methods of the two are as follows:

[0005] Method 1: Full-band error extraction based on optical measurement and analysis software such as MetroPro and Mx

[0006] MetroPro, Mx and other optical measurement and analysis software are supporting analysis software for using interferometer to detect surface shape. This type of software can directly read and display the test results of interferometer, roughness meter, atomic force microscope and other testing instruments, and integrate the function of processing and analyzing the test results in the software. When using MetroPro and Mx to extract full-band errors from surface shape detection results, taking MetroPro as an example, you need to first import the .dat file of the test results to be processed in the "GPI.app" of the software, and then select the appropriate filter and filtering method in the simulation control (Analyze Cntrl) option of the software, and set the frequency range corresponding to the different frequency band errors of the current surface shape for the filter, such as Figure 3 After the settings are completed, click Filter to obtain the frequency band error or full frequency band error of the current surface shape.

[0007] However, when using MetroPro to extract full-band errors, only one frequency band error in a face shape can be extracted each time, which is inefficient and cannot process large quantities of face shape data; in addition, frequency band errors of different shapes and sizes correspond to different frequency ranges, and there is no unified standard for frequency band division. Sometimes it is necessary to make empirical judgments based on the application range of optical components or the morphological characteristics of errors in different frequency bands to fine-tune the frequency range of the filter. Therefore, each filtering process must input different frequency ranges according to the actual situation of different face shapes to ensure that the most accurate frequency band error is obtained in the end. Therefore, even the Mx software that can extract full-band errors in batches by writing scripts cannot provide the correct filter frequency range in the process of extracting full-band errors for different face shapes, and cannot ensure the extraction accuracy of full-band errors when processing large quantities of face shape data.

[0008] Method 2: Full-band error extraction based on mathematical software such as MATLAB

[0009] Mathematical software such as MATLAB is often used for numerical analysis and matrix calculation of data. Taking MATLAB as an example, when extracting full-band errors of optical component surfaces, unlike optical measurement and analysis software such as MetroPro, which has built-in filtering functions, MATLAB requires writing scripts to perform filtering operations. First, the surface detection result data to be processed is imported into MATLAB, and a two-dimensional Fourier transform is performed on it to convert it from the spatial domain to the frequency domain. Subsequently, a filter corresponding to the frequency range is constructed according to the frequency band to be extracted, and the filter is multiplied with the frequency domain image. Finally, the filtered result is converted from the frequency domain back to the spatial domain through an inverse Fourier transform, and the frequency band error or full-band error of the current surface can be obtained. Although MATLAB can realize batch extraction of full-band errors by writing scripts, the full-band error extraction of MATLAB has the same problem as the full-band error extraction of MetroPro. It cannot give the corresponding accurate frequency band range for the actual situation of different optical surfaces, and cannot ensure the extraction accuracy of the full-band error when processing large quantities of surface data. Summary of the invention

[0010] In view of this, the present invention aims to provide a method for extracting full-band errors of optical element surface shape based on a neural network, so as to solve the technical problem that the prior art cannot ensure the accuracy of full-band error extraction when processing large quantities of surface shape data.

[0011] To achieve the above object, the technical solution created by the present invention is implemented as follows:

[0012] A method for extracting full-band errors of optical element surface shape based on a neural network comprises the following steps:

[0013] S1: Collect the surface detection results of different optical elements to obtain a surface detection result data set, and extract the corresponding full-band error from each surface detection result to obtain a full-band error data set, and package the surface detection result data set and the full-band error data set into a training data set to train three error extraction neural networks; wherein, the three error extraction neural networks are respectively a low-frequency error extraction neural network, a medium-frequency error extraction neural network and a high-frequency error extraction neural network; the full-band error includes a low-frequency error, a medium-frequency error and a high-frequency error, and the surface detection results are used as the training input data of the three error extraction neural networks, and the low-frequency error, medium-frequency error and high-frequency error corresponding to the surface detection results are used as the training targets of the three error extraction neural networks;

[0014] S2: The surface shape detection results of the optical element to be tested are sorted and packaged and input into three error extraction neural networks that have completed training. The low-frequency error, medium-frequency error and high-frequency error of the surface shape of the optical element to be tested are respectively extracted by the three error extraction neural networks and used as extraction results.

[0015] Furthermore, before the face detection result data set is put into training, the face detection result data set is preprocessed, specifically including the steps of:

[0016] Step 1: Eliminate unusable face detection results;

[0017] Step 2: Eliminate the face detection results that do not contain the actual size or add the actual size to the face detection results that do not contain the actual size;

[0018] Step 3: Remove the displacement and aberration in the surface detection results in MetroPro software. The displacement includes translation and tilt, and the aberration includes spherical aberration and astigmatism.

[0019] Furthermore, the corresponding full-band error is extracted from the face shape detection result, specifically including the steps of:

[0020] Step 1: Import the face detection results into MetroPro software;

[0021] Step 2: Determine the filtering range based on the actual size and morphological characteristics of the surface detection result, and select the corresponding filter according to the filtering range to separate at least one of the low-frequency error, medium-frequency error and high-frequency error corresponding to the surface detection result.

[0022] Furthermore, the face shape detection result data set and the full-band error data set are packaged into a training data set using MATLAB software, which specifically includes the following steps:

[0023] Step 1: Scale the pixels of each data in the face detection result dataset and the full-band error dataset to a predetermined size according to the scaling ratio;

[0024] Step 2: Find the peak value and trough value of the wavefront height of each data in the surface detection result data set and the full-band error data set, calculate the median based on the peak value and trough value, subtract the corresponding median from the wavefront height of each data, normalize the wavefront height to the range of [-1,1], and record the normalized ratio;

[0025] Step 3: Unify the numbers of the normalized face detection result dataset and the full-band error dataset and package them into a training dataset.

[0026] Furthermore, the network structures of the three error extraction neural networks are the same, which are composed of the first convolution layer, the first pooling layer, the second convolution layer, the second pooling layer, the third convolution layer, the third pooling layer, the fourth convolution layer, the first transposed convolution layer, the second transposed convolution layer, the third transposed convolution layer, the output convolution layer, and the regression layer stacked in sequence.

[0027] Furthermore, the prediction accuracy of the three error extraction neural networks is evaluated using the loss function. for:

[0028] ;

[0029] Where m is the number of samples, is the actual value of the i-th sample, is the predicted value of the i-th sample.

[0030] Further, different measuring instruments are used to detect the surface shapes of different optical elements respectively to obtain surface shape detection result data sets; wherein the measuring instruments include an interferometer, a roughness meter and an atomic force microscope;

[0031] Before the measuring instrument detects the surface shape, different resolutions and the pixel ratios corresponding to the different resolutions are set in the optical measurement and analysis software.

[0032] Furthermore, the interferometers include a 6-inch interferometer, an HDX interferometer and an MST interferometer from Zygo Corporation; wherein the interferometers include a 6-inch interferometer, an HDX interferometer and an MST interferometer from Zygo Corporation; wherein the resolution of the 6-inch interferometer is set to 640 480, the pixel ratio of the 6-inch interferometer is set to 0.3624; the resolution of the HDX interferometer is set to 1696 1696, the pixel ratio of the HDX interferometer is set to 0.06123, or the resolution of the HDX interferometer is set to 3396 3396, the pixel ratio of the HDX interferometer is set to 0.03062; the resolution of the MST interferometer is set to 1472 1472, the magnification of the MST interferometer is set to 1x, the pixel ratio of the MST interferometer is set to 0.07600, or the resolution of the MST interferometer is set to 1472 1472, the magnification of the MST interferometer is set to 1.7x, the pixel ratio of the MST interferometer is set to 0.04418, or the resolution of the MST interferometer is set to 736 736, the magnification of the MST interferometer is set to 1x, the pixel ratio of the MST interferometer is set to 0.1519, or the resolution of the MST interferometer is set to 736 736, the magnification of the MST interferometer was set to 1.7x, and the pixel ratio of the MST interferometer was set to 0.08828.

[0033] Furthermore, before the training data set is put into training, the types and quantity of the training data set are expanded using data enhancement methods, including data enhancement methods based on image processing and data enhancement methods based on generative adversarial networks.

[0034] Furthermore, after step S2, the following steps are also included:

[0035] S3: Restore the extracted results to the original resolution and wavefront height ratio according to the scaling ratio and normalization ratio, and finally output the full-band error .dat file.

[0036] Compared with the prior art, the invention can achieve the following beneficial effects:

[0037] 1. The present invention trains three error extraction neural networks respectively. The three error extraction neural networks learn the characteristic relationship between the actual surface shape and the low-frequency error, the medium-frequency error, and the high-frequency error, so as to realize the automatic extraction of the full-band error of any optical surface shape without manual filtering operation.

[0038] 2. Neural networks can process large amounts of surface data in batches, greatly improving the efficiency of extracting full-band errors and facilitating subsequent analysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] The drawings constituting part of the present invention are used to provide a further understanding of the present invention. The exemplary embodiments and descriptions of the present invention are used to explain the present invention and do not constitute an improper limitation on the present invention. In the drawings:

[0040] Figure 1 It is a schematic diagram of frequency band error division based on frequency and power spectral density (PSD) in the frequency domain;

[0041] Figure 2 It is a schematic diagram of the effect of frequency band error on imaging;

[0042] Figure 3 This is a schematic diagram of the control options for filtering operations in the MetroPro software;

[0043] Figure 4 It is a flow chart of a method for extracting full-band errors of optical element surface shape based on a neural network according to an embodiment of the invention;

[0044] Figure 5 is a schematic diagram of surface shape detection results of different detection instruments according to an embodiment of the invention;

[0045] Figure 6 It is a schematic diagram of a process of normalizing the wavefront height of the surface detection result according to an embodiment of the invention;

[0046] Figure 7 Schematic diagram of the network structure of the error extraction neural network according to the embodiment of the present invention. DETAILED DESCRIPTION

[0047] In order to make the purpose, technical solution and advantages of the invention more clear, the invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described here are only used to explain the invention and do not constitute a limitation of the invention.

[0048] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments may be combined with each other.

[0049] In the description of the present invention, it should be understood that the terms "center", "longitudinal", "lateral", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside" and the like indicate positions or positional relationships based on the positions or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on the present invention. In addition, the terms "first", "second", etc. are only used for descriptive purposes, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Thus, features defined as "first", "second", etc. may explicitly or implicitly include one or more of the features. In the description of the present invention, unless otherwise specified, "multiple" means two or more.

[0050] In the description of the invention, it should be noted that, unless otherwise clearly specified and limited, the terms "installation", "connection" and "connection" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, or it can be the internal communication of two components. For ordinary technicians in this field, the specific meanings of the above terms in the invention can be understood according to specific circumstances.

[0051] The present invention will be described in detail below with reference to the drawings and in combination with embodiments.

[0052] like Figure 4 As shown, the method for extracting full-band errors of optical element surface shape based on neural network provided by the present invention comprises the following steps:

[0053] S1: Collect the surface detection results of different optical elements to obtain a surface detection result data set, and extract the corresponding full-band error from each surface detection result to obtain a full-band error data set, and package the surface detection result data set and the full-band error data set into a training data set to train three error extraction neural networks; wherein, the three error extraction neural networks are respectively a low-frequency error extraction neural network, a medium-frequency error extraction neural network and a high-frequency error extraction neural network; the full-band error includes low-frequency error, medium-frequency error and high-frequency error, and the surface detection results are used as the training input data of the three error extraction neural networks, and the low-frequency error, medium-frequency error and high-frequency error corresponding to the surface detection results are used as the training targets of the three error extraction neural networks.

[0054] The low-frequency error extraction neural network is used to extract the low-frequency error of the optical element surface shape, the medium-frequency error extraction neural network is used to extract the medium-frequency error of the optical element surface shape, and the high-frequency error extraction neural network is used to extract the high-frequency error of the optical element surface shape. The present invention performs full-band error extraction on the surface shapes of optical elements with different morphologies by training these three error neural networks.

[0055] Before training the three error neural networks, it is necessary to collect a training data set for training. The training data set includes two parts: the face detection result and the full-band error separated from the face detection result.

[0056] The method for collecting surface shape detection results is: use different detection instruments to detect the surface shapes of optical components with different morphologies and scales. The detection instruments include interferometers (such as Zygo's 6-inch, HDX, MST, etc.), roughness meters (such as Zygo's 9000 white light interferometer, 7200 roughness meter, etc.) and atomic force microscopes. Before the detection instrument performs surface shape detection, it is necessary to set the corresponding pixel ratios at different resolutions in the optical measurement and analysis software. The purpose is to ensure that the detection results contain the actual size and can be filtered later. Taking the 6-inch interferometer, HDX interferometer, and MST interferometer as examples, the pixel ratios of these three detection instruments at different resolutions are shown in Tables 1, 2, and 3. After completing the surface shape detection, save the .dat file of the current surface shape detection result. In this way, a large number of surface shape detection results from different detection instruments are collected as a surface shape detection result data set, such as Figure 5 shown.

[0057] Table 1 Pixel ratio of 6-inch interferometer

[0058]

[0059] Table 2 Pixel ratio of HDX interferometer

[0060]

[0061] Table 3 Pixel ratio of MST interferometer

[0062]

[0063] The method for collecting full-band errors is to separate the corresponding low-frequency errors, medium-frequency errors and high-frequency errors from a large number of face detection results collected from different detection instruments as a full-band error data set. Specifically, the face detection results are imported into the MetroPro software, and the filtering range required for filtering processing is input into the MetroPro software in combination with the actual size and morphological characteristics of the face detection results. Based on the filtering range, a suitable filter is selected to separate the low-frequency error, medium-frequency error or high-frequency error corresponding to the face detection result through filtering. Low-frequency errors are mainly large-scale, relatively smooth overall morphology, while medium- and high-frequency errors are mainly waviness errors and roughness errors. Therefore, it should be ensured that the error of a certain frequency band obtained after filtering does not include the characteristics of the errors of the other two frequency bands as much as possible, and this should be used as a basis for fine-tuning the filtering frequency range.

[0064] Before putting the full-band error data set and the face detection result data set into training, the face detection result data set needs to be input into the MetroPro software for preprocessing, and then the preprocessed face detection result data set and the full-band error data set are sent to the MATLAB software for normalization processing and uniformly packaged as training data sets that can be used by the error extraction neural network.

[0065] In the process of face detection by the detection instrument, inaccurate face position placement, irregular operation process or differences between different detection instruments may introduce errors in the face detection results, causing the detection results to deviate from the actual face shape. Therefore, it is necessary to preprocess the face detection result data set. The preprocessing process of the face detection result data set is: first, remove some incomplete and obviously unusable face detection results; second, remove the face detection results that do not contain actual dimensions or add actual dimensions to these face detection results, otherwise these face detection results cannot be filtered later; finally, remove the translation (PST) and tilt (TLT) in the face detection results in the MetroPro software as appropriate. Some faces need to remove spherical aberration / spherical degree (PWR), astigmatism (AST), trim or other necessary operations based on actual conditions.

[0066] The specific process of normalization and packaging of the face detection result dataset and the full-band error dataset is as follows: first, the pixels of each data in the face detection result dataset and the full-band error dataset are uniformly scaled to a fixed pixel size according to the pixel ratio; then, the face detection result dataset and the full-band error dataset are normalized: the peak value and trough value of the wavefront height of each data in the face detection result dataset and the full-band error dataset are found, the median is calculated based on the peak value and trough value, the overall wavefront height of each data (that is, all positions of the wavefront height) is subtracted from the median, the wavefront height of each data is uniformly normalized to the range of [-1,1], and the normalization ratio is recorded to facilitate the subsequent restoration of the wavefront height of the full-band error extraction result, such as Figure 6 As shown; finally, check whether the face detection result dataset and the full-band error dataset correspond one-to-one and are correctly normalized, unify the numbering of the face detection result dataset and the full-band error dataset, and package them into a training dataset that can be used by the error extraction neural network.

[0067] In addition, data enhancement can be used to expand the types and quantity of training data sets, thereby further improving the extraction accuracy of the error extraction neural network, such as data enhancement based on image processing and data enhancement based on generative adversarial networks (GAN).

[0068] The training data set is divided into training set, validation set and test set according to the ratio of 7:1.5:1.5, and three error extraction neural networks are trained.

[0069] The network structures of the three error extraction neural networks obtained by training are the same, which are composed of the first convolution layer, the first pooling layer, the second convolution layer, the second pooling layer, the third convolution layer, the third pooling layer, the fourth convolution layer, the first transposed convolution layer, the second transposed convolution layer, the third transposed convolution layer, the output convolution layer, and the regression layer stacked in sequence, as shown in Figure 2. Figure 7 shown.

[0070] The convolution kernels of the first convolution layer, the first pooling layer, the second convolution layer, the second pooling layer, the third convolution layer, the third pooling layer, and the fourth convolution layer are all 3*3; the convolution kernels of the first transposed convolution layer, the second transposed convolution layer, the third transposed convolution layer, and the output convolution layer are all 4*4. The number of input channels of the first convolutional layer is 1, and the number of output channels of the first convolutional layer is 16. The number of input channels of the second convolutional layer is 16, and the number of output channels of the second convolutional layer is 32. The number of input channels of the third convolutional layer is 32, and the number of output channels of the third convolutional layer is 64. The number of input channels of the fourth convolutional layer is 64, and the number of output channels of the fourth convolutional layer is 128. The number of input channels of the first transposed convolutional layer is 128, and the number of output channels of the first transposed convolutional layer is 64. The number of input channels of the second transposed convolutional layer is 64, and the number of output channels of the second transposed convolutional layer is 32. The number of input channels of the third transposed convolutional layer is 32, and the number of output channels of the third transposed convolutional layer is 16. The number of output channels of the output convolutional layer is consistent with the number of channels of the input image (i.e. channels).

[0071] Because the training data set and the model prediction results are both the surface shapes of optical components, they can be understood as two-dimensional single-channel images, where the xy axis represents the size range of the surface shape and the z axis represents the wavefront height at the current point.

[0072] The regression layer uses the loss function during training to calculate the error between the predicted result and the actual value and updates the weights of the network.

[0073] In the training process of error extraction neural network, a common learning rate is 1e -3 , the optimizer is Adam, the training cycle epoch is 150, and the batch-size is 48.

[0074] Using loss function Evaluating the prediction accuracy of three error extraction neural networks, loss functions Specifically, it is the root mean square error RMSE, and the calculation method of the root mean square error RMSE is:

[0075] ;

[0076] S2: The surface shape detection results of the optical element to be tested are sorted and packaged and input into three error extraction neural networks that have completed training. The low-frequency error, medium-frequency error and high-frequency error of the surface shape of the optical element to be tested are respectively extracted by the three error extraction neural networks and used as extraction results.

[0077] For different frequency ranges, the corresponding detection instruments are used to obtain the surface detection results of the optical components to be tested. The surface detection results are sorted and packaged and input into three error extraction neural networks. The three error extraction neural networks automatically scale and normalize the surface detection results, and record the scaling ratio and normalization ratio at the same time. After the full-band error is extracted by the three error extraction neural networks, the extracted results are restored to the original resolution and wavefront height ratio according to the scaling ratio and normalization ratio, and finally the required full-band error .dat file is output.

[0078] It should be understood that the various forms of processes shown above can be used to reorder, add or delete steps. For example, the steps described in the disclosure of the present invention can be performed in parallel, sequentially or in different orders, as long as the desired results of the technical solution disclosed in the present invention can be achieved, and this document does not limit this.

[0079] The above specific implementations do not constitute a limitation on the protection scope of the present invention. It should be understood by those skilled in the art that various modifications, combinations, sub-combinations and substitutions can be made according to design requirements and other factors. Any modification, equivalent substitution and improvement made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for extracting full-band errors of optical element surface shape based on neural network, characterized in that: The steps include: S1: Collect the surface detection results of different optical elements to obtain a surface detection result data set, and extract the corresponding full-band error from each surface detection result to obtain a full-band error data set, and package the surface detection result data set and the full-band error data set into a training data set to train three error extraction neural networks; wherein, the three error extraction neural networks are respectively a low-frequency error extraction neural network, a medium-frequency error extraction neural network and a high-frequency error extraction neural network; the full-band error includes a low-frequency error, a medium-frequency error and a high-frequency error, and the surface detection results are used as the training input data of the three error extraction neural networks, and the low-frequency error, medium-frequency error and high-frequency error corresponding to the surface detection results are used as the training targets of the three error extraction neural networks; S2: The surface shape detection results of the optical element to be tested are sorted and packaged and input into three error extraction neural networks that have completed training. The low-frequency error, medium-frequency error and high-frequency error of the surface shape of the optical element to be tested are respectively extracted by the three error extraction neural networks and used as extraction results.

2. The method for extracting full-band errors of optical element surface shape based on neural network according to claim 1, characterized in that: Before putting the face detection result data set into training, the face detection result data set is preprocessed, which specifically includes the following steps: Step 1: Eliminate unusable face detection results; Step 2: Eliminate the face detection results that do not contain the actual size or add the actual size to the face detection results that do not contain the actual size; Step 3: Remove the displacement and aberration in the surface detection results in MetroPro software. The displacement includes translation and tilt, and the aberration includes spherical aberration and astigmatism.

3. The method for extracting full-band errors of optical element surface shape based on neural network according to claim 1, characterized in that: Extracting the corresponding full-band error from the face shape detection result includes the following steps: Step 1: Import the face detection results into MetroPro software; Step 2: Determine the filtering range based on the actual size and morphological characteristics of the surface detection result, and select the corresponding filter according to the filtering range to separate at least one of the low-frequency error, medium-frequency error and high-frequency error corresponding to the surface detection result.

4. The method for extracting full-band errors of optical element surface shape based on neural network according to claim 1, characterized in that: Use MATLAB software to package the face detection result dataset and the full-band error dataset into a training dataset, which specifically includes the following steps: Step 1: Scale the pixels of each data in the face detection result dataset and the full-band error dataset to a predetermined size according to the scaling ratio; Step 2: Find the peak value and trough value of the wavefront height of each data in the surface detection result data set and the full-band error data set, calculate the median based on the peak value and trough value, subtract the corresponding median from the wavefront height of each data, normalize the wavefront height to the range of [-1,1], and record the normalized ratio; Step 3: Unify the numbers of the normalized face detection result dataset and the full-band error dataset and package them into a training dataset.

5. The method for extracting full-band errors of optical element surface shape based on neural network according to claim 1, characterized in that: The network structures of the three error extraction neural networks are the same, which are composed of the first convolution layer, the first pooling layer, the second convolution layer, the second pooling layer, the third convolution layer, the third pooling layer, the fourth convolution layer, the first transposed convolution layer, the second transposed convolution layer, the third transposed convolution layer, the output convolution layer, and the regression layer stacked in sequence.

6. The method for extracting full-band errors of optical element surface shape based on neural network according to claim 1, characterized in that: The prediction accuracy of the three error extraction neural networks is evaluated using the loss function. for: ； Where m is the number of samples, is the actual value of the i-th sample, is the predicted value of the i-th sample.

7. The method for extracting full-band errors of optical element surface shape based on neural network according to claim 1, characterized in that: Using different measuring instruments to detect the surface shapes of different optical elements respectively, and obtaining a surface shape detection result data set; wherein the measuring instruments include an interferometer, a roughness meter and an atomic force microscope; Before the measuring instrument detects the surface shape, different resolutions and the pixel ratios corresponding to the different resolutions are set in the optical measurement and analysis software.

8. The method for extracting full-band errors of optical element surface shape based on neural network according to claim 7, characterized in that: The interferometers include Zygo's 6-inch interferometer, HDX interferometer, and MST interferometer. The resolution of the 6-inch interferometer is set to 640 480, the pixel ratio of the 6-inch interferometer is set to 0.3624; the resolution of the HDX interferometer is set to 1696 1696, the pixel ratio of the HDX interferometer is set to 0.06123, or the resolution of the HDX interferometer is set to 3396 3396, the pixel ratio of the HDX interferometer is set to 0.03062; the resolution of the MST interferometer is set to 1472 1472, the magnification of the MST interferometer is set to 1x, the pixel ratio of the MST interferometer is set to 0.07600, or the resolution of the MST interferometer is set to 1472 1472, the magnification of the MST interferometer is set to 1.7x, the pixel ratio of the MST interferometer is set to 0.04418, or the resolution of the MST interferometer is set to 736 736, the magnification of the MST interferometer is set to 1x, the pixel ratio of the MST interferometer is set to 0.1519, or the resolution of the MST interferometer is set to 736 736, the magnification of the MST interferometer was set to 1.7x, and the pixel ratio of the MST interferometer was set to 0.08828.

9. The method for extracting full-band errors of optical element surface shape based on neural network according to claim 1, characterized in that: Before putting the training data set into training, the types and quantity of the training data set are expanded using data augmentation methods. The data augmentation methods include data augmentation methods based on image processing and data augmentation methods based on generative adversarial networks.

10. The method for extracting full-band errors of optical element surface shape based on neural network according to claim 4, characterized in that: After step S2, the method further includes the following steps: S3: Restore the extracted results to the original resolution and wavefront height ratio according to the scaling ratio and normalization ratio, and finally output the full-band error .dat file.

Images