Mosaic mirror common error detection method based on wavelet transform multi-channel input

Abstract

The application discloses a splicing mirror common phase error detection method based on wavelet transform multi-channel input, comprising the following steps: designing an optical system based on far field imaging according to the principle of the optical system; introducing 10000 groups of common phase errors within one wavelength and 20dB Gaussian white noise; collecting 10000 far field images and recording corresponding common phase error values, respectively constructing an image dataset and a label dataset, and selecting a training set from the image dataset; configuring a deep learning environment and building a neural network model; preprocessing the image data in the training set as the input of the neural network model, obtaining a predicted common phase error value through the neural network model; and calculating the loss value between the predicted common phase error value and a given label, optimizing and adjusting the parameters of the neural network. The method has good noise resistance and higher precision, and has practical application significance for common phase error detection based on far field images.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of photoelectric measurement, image processing and deep learning, and in particular to a co-phase error detection method for a mosaic mirror based on wavelet transform multi-channel input. BACKGROUND

[0002] According to the Rayleigh criterion, for a given wavelength λ, the resolution of an optical system is inversely proportional to the pupil diameter D, that is, the larger the pupil diameter, the higher the system resolution; at the same time, the light collecting ability of the system is proportional to the square of the pupil diameter This means that a larger mirror can collect more light to achieve clearer imaging. To meet the needs of deeper space and higher resolution observation, scientists have begun to develop large-aperture primary mirrors for giant telescopes. However, the use of a single large-aperture mirror has many technical limitations, including difficulty in maintaining the accuracy of mirror manufacturing and testing, high difficulty in processing and transportation, and in addition, the size of the fairing of the launch vehicle limits the practical application of large-aperture single mirrors. The mosaic mirror or synthetic aperture technology is an important technical means for developing the next generation of super telescopes, which realizes the high-resolution imaging capability of a single main mirror with the same aperture through co-phase splicing of sub-mirrors or co-phase beam combining of subsystems. Therefore, the detection of co-phase error (piston error) is particularly important in the mosaic / synthetic aperture telescope system.

[0003] In recent years, with the development of artificial intelligence and deep learning, co-phase error detection based on convolutional neural networks (CNN) has made significant progress. As early as 2018, Guerra-Ramos et al. in Spain used CNN to realize the piston term detection of the co-phase error of a 36-mosaic mirror; later, Li Dequan et al. of the Changchun Institute of Optics, Fine Mechanics and Physics, Chinese Academy of Sciences studied the co-phase detection performance of CNN on a face target; Ma Xiafei et al. of the Institute of Optoelectronic Technology, Chinese Academy of Sciences constructed a deep CNN, and through simulation and experiment, the application of this method in the co-phase error detection of the mosaic mirror was confirmed.

[0004] The above-mentioned co-phase error detection method based on convolutional neural networks improves the detection accuracy, but still faces some problems. In actual application, the far-field image of the telescope system often contains noise, which can cause the image features to be blurred or distorted, thereby affecting the prediction accuracy of the neural network. This greatly reduces the performance of the CNN-based co-phase error detection method under noisy conditions. SUMMARY

[0005] The technical problem solved by the present application is that the traditional neural network has poor performance in the presence of noise in the co-phase error detection of the mosaic mirror.

[0006] The technical solution adopted by the present application to solve the above technical problem is a mosaic mirror co-phase error detection method based on wavelet transform multi-channel input, comprising the following steps:

[0007] Step one: design an optical system based on far-field imaging according to the principle of optical system, which includes a main mirror formed by two sub-mirrors, a mask, a focusing lens and a detector, wherein the incident parallel light beam is reflected by the sub-mirror, passes through the mask, is converged by the focusing lens, and is imaged on the image plane of the detector, and there is a co-phase error between the sub-mirrors;

[0008] Step two: take a circular hole as a reference, randomly introduce 10000 sets of co-phase error within one wavelength to another circular hole, and introduce 20dB of Gaussian white noise to simulate the readout noise of the camera in actual use;

[0009] Step three: collect 10000 far-field images and record the corresponding co-phase error values, and construct image data set and label data set respectively, and select training set from the image data set;

[0010] Step four: configure the deep learning environment and build the neural network model;

[0011] Step five: preprocess the image data in the training set, decompose the single-channel image data into 4-channel components through one Haar wavelet decomposition, and input the neural network model, and obtain the predicted co-phase error value through the neural network model;

[0012] Step six: calculate the loss value between the predicted co-phase error value and the given label, and optimize and adjust the parameters of the neural network.

[0013] Compared with the prior art, the present application has the following beneficial effects:

[0014] (1) The present application decomposes the far-field image into multi-scale features through wavelet transform, and inputs the CNN with multi-channel, so that the model can more accurately capture the details of the co-phase error and improve the prediction accuracy.

[0015] (2) The wavelet multi-channel input has the advantage of noise resistance, enhances the robustness of the model in the presence of noise, and enables the system to maintain high-precision detection in the presence of noise.

[0016] (3) The application adopts the ResNet18 structure, improves the feature extraction efficiency through the residual module, and is light and fast in network optimization, and is suitable for the actual application of the mosaic mirror system. BRIEF DESCRIPTION OF DRAWINGS

[0017] Fig. 1(a) is a Michelson type mosaic telescope system composed of a Michelson type primary mirror 1-1, a Michelson type secondary mirror 1-2 and a Michelson type image plane 1-3; Fig. 1(b) is a Pease type synthetic aperture telescope system composed of an afocal subsystem 2-1, a phase retarder 2-2, a beam combiner 2-3 and a Pease type image plane 2-4.

[0018] Figure 2 The corresponding optical imaging system schematic diagram of the application is taken as an example of two hexagonal sub-mirrors, which is composed of a sub-mirror 3-1, a mask 3-2, a focusing lens 3-3 and a detector 3-4.

[0019] Figure 3 The wavelet transform multi-channel decomposition schematic diagram in the application. The original noisy image (224x224 pixels) is decomposed into (a) LL (low-frequency approximate component, 112x112 pixels), (b) LH (horizontal high-frequency component, 112x112 pixels), (c) HL (vertical high-frequency component, 112x112 pixels), and (d) HH (diagonal high-frequency component, 112x112 pixels) by one Haar wavelet decomposition.

[0020] Figure 4 The schematic diagram of the mosaic mirror co-phase detection method based on wavelet transform multi-channel input proposed in the application.

[0021] Figure 5 The specific model used by the convolutional neural network in the application is ResNet18. The network contains 17 convolutional layers, 2 pooling layers and 1 fully connected layer, and the input is directly transmitted to the output through residual connection in ResNet18.

[0022] Figure 6 The prediction residual error distribution diagram of the method proposed in the application compared with the traditional transfer learning. DETAILED DESCRIPTION

[0023] In order to make the process, technical scheme and advantages of the application more clear, the following will be further described with the embodiment of the application of the mosaic double hexagonal sub-mirror system as the embodiment of the application.

[0024] Figure 1(a) is a Michelson type mosaic telescope system. The incident parallel light beam is reflected by the mosaic Michelson primary mirror 1-1 and then reflected by the mosaic Michelson secondary mirror 1-2, and then imaged on the mosaic Michelson image plane 1-3. Figure 1(b) is a Pechiney type synthetic aperture telescope system. The parallel light beam is reflected by the afocal subsystem 2-1, and then the phase retarder 2-2 adds different phases, and then the light beam synthesized by the beam combiner 2-3 is imaged on the Pechiney image plane 2-4.

[0025] Figure 2 Figure 3 is a schematic diagram of the optical imaging system corresponding to the present application. The primary mirror can be the Michelson type mosaic telescope system shown in Figure 1(a) or the Pechiney type synthetic aperture telescope system shown in Figure 1(b), which is spliced by two sub-mirrors 3-1, and there is a common phase error between the sub-mirrors 3-1. The incident parallel light beam is reflected by the spliced sub-mirror 3-1, passes through the mask 3-2, is focused by the focusing lens 3-3, and is imaged on the imaging plane of the detector 3-4. Figure 3 Figure 4 is a schematic diagram of the wavelet transform multi-channel decomposition in the present application. The original noisy image (224x224 pixels) is decomposed by one Haar wavelet to generate four sub-images: (a) LL component (low frequency approximation component, 112x112 pixels), containing the main low frequency information of the image; (b) LH component (horizontal high frequency component, 112x112 pixels), reflecting the horizontal direction high frequency details; (c) HL component (vertical high frequency component, 112x112 pixels), representing the vertical direction high frequency details; and (d) HH component (diagonal high frequency component, 112x112 pixels), containing the diagonal line direction high frequency information.

[0026] Figure 4 Figure 5 is the flow of the mosaic mirror common phase error detection method based on wavelet transform multi-channel input proposed in the present application. The four channel images obtained by wavelet transform of the original single channel image are used as the input of the convolutional neural network (CNN), and the CNN model (ResNet18 structure is used in the present application) is used to extract features and regression mapping of the input image, and a predicted value .

[0027] The mosaic mirror common phase error detection method based on wavelet transform multi-channel input includes the following steps:

[0028] Step 1: Design an optical system based on far-field imaging according to the principles of optical systems. This optical system includes a primary mirror formed by splicing two sub-mirrors 3-1, a mask 3-2, a focusing lens 3-3, and a detector 3-4. An incident parallel beam of light is reflected by the spliced ​​sub-mirror 3-1, passes through the mask 3-2, and is converged by the focusing lens 3-3 before being imaged onto the image plane of the detector 3-4. There is a common-phase error between the sub-mirrors 3-1. Different phase differences are applied to the spliced ​​sub-mirrors 3-1 to simulate the actual common-phase error. The design process includes designing the incident light wavelength, wavenumber, focal length of the focusing lens 3-3, and the circular aperture, etc. In one example, the incident light wavelength... , wave number Lens focal length round hole The circular aperture is the white circular area in mask 3-2. The design of the circular aperture includes the diameter of the aperture and the spacing between the two apertures.

[0029] Step 2: Using one circular aperture as a reference, randomly introduce 10,000 sets of one wavelength into the other circular aperture. The common phase error is within a certain range, and 20dB of Gaussian white noise is introduced to simulate the readout noise of the camera in reality.

[0030] Step 3: Acquire 10,000 far-field images and record the corresponding phase error values. Construct an image dataset and a label dataset, respectively, and select a training set from the image dataset. The acquired diffraction images contain phase errors between the stitching sub-mirrors. The 10,000 images were obtained by applying different phase differences to stitching sub-mirrors 3-1.

[0031] Figure 2 In the splicing mirror optical system shown, sub-mirror 3-1 suffers from co-phase error. In the simulation, this co-phase error can be simulated by applying a phase difference (label) through the circular hole in mask 3-2. After convergence by focusing lens 3-3, the image is formed on detector 3-4. When the applied phase difference (label) changes, the final image changes accordingly; that is, the label and image are in one-to-one correspondence. Constructing the image dataset and label dataset can include: creating one folder to store all named images (image dataset); and creating another folder to store all label data containing image indices (label dataset). The two are in one-to-one correspondence, with one image corresponding to one label.

[0032] In one embodiment, 80% of the image dataset is selected as the training set for the neural network model to learn the nonlinear mapping relationship between the image after Haar wavelet decomposition and the co-phase error value, and 20% of the image dataset is selected as the validation set to measure the accuracy and real-time performance of the neural network model.

[0033] Step four: configure the deep learning environment and build the neural network model;

[0034] Step five: preprocess the image data in the training set, decompose the single-channel image data into four-channel components through one Haar wavelet, and input the neural network model to obtain the predicted common phase error value through the neural network model.

[0035] Step six: calculate the loss value between the predicted common phase error value and the given label, and optimize and adjust the parameters of the neural network. The calculation of the loss value includes: calculating the difference between the predicted common phase error value and the given label through the mean square error loss function (MSE), thereby obtaining the loss value. The given label is in the label data set. The determination of the given label is completed by the path index method. For example, the name of an image in the image data set is: "0001.png", and the network finds the corresponding label value in the label according to this path, for example, the path of the corresponding label is: "0001.png_0.025", and the algorithm extracts the index value of the label under the image path, that is, "0.025" (given label). In this way, the corresponding extraction relationship between the image and the given label is completed.

[0036] After the training network converges, only a single-frame far-field diffraction image needs to be given to the network, and the network can output the predicted common phase error value. Compared with the traditional common phase error detection method based on transfer learning, the method proposed in the present application has stronger robustness to noisy images, and the prediction accuracy is greatly improved. Simulation shows that the time required for the present application to complete one common phase error prediction is about 10 milliseconds.

[0037] In step three, the mathematical expression of the image plane light intensity is:

[0038] ,

[0039] wherein, represents the ideal distribution of the point source on the image plane, represents the convolution operation, is the camera noise, are the horizontal and vertical coordinates of the image plane, respectively, and the intensity point spread function PSF can be expressed as:

[0040] ,

[0041] wherein, is the wavelength of the incident light, is the focal length of the lens, represents the generalized pupil function of the optical system, is the Fourier transform.

[0042] Figure 5For the specific model used in the convolutional neural network in the present application, ResNet18. ResNet18 contains 17 convolutional layers, 2 pooling layers (1 max pooling layer and 1 global average pooling layer), and 1 fully connected layer. In the network, the input is directly transmitted to the output through residual connection (Residual Connection), which alleviates the problem of gradient disappearance. In Figure 5 The construction of the residual module is shown in the lower right corner, and each residual module contains two layers of convolution. The input is directly transmitted to the output through residual connection (i.e. jump connection), which alleviates the problem of gradient disappearance in deep network training, improves the stability and training efficiency of the model.

[0043] In summary, a splicing mirror co-phase detection method based on wavelet transform multi-channel input, first, the wavelet decomposition of the splicing mirror far-field diffraction noisy image containing piston error is carried out, and the multi-scale feature components are obtained as multi-channel input provided to the convolutional neural network (CNN). The network uses ResNet18 model structure, contains multiple residual modules, to improve the robustness and noise resistance of feature extraction. The network output is a single regression value, which is used to predict the co-phase error (piston error) of the splicing mirror. In the training process, the mean square error loss function (MSE) is used, and the prediction error is minimized by optimizing the network parameters.

[0044] Figure 6 is the prediction residual error distribution map of the method proposed in the present application compared with the traditional transfer learning.

[0045] The contents not described in detail in the specification of the present application belong to the prior art known to those skilled in the art.

Claims

1. A method for detecting phase error in splicing mirrors based on wavelet transform multi-channel input, characterized in that, Includes the following steps: Step 1: Design an optical system based on far-field imaging according to the principle of optical systems. The optical system includes a primary mirror formed by splicing two sub-mirrors (3-1), a mask (3-2), a focusing lens (3-3), and a detector (3-4). The incident parallel beam is reflected by the sub-mirror (3-1), passes through the mask (3-2), and is converged by the focusing lens (3-3) and imaged on the image plane of the detector (3-4). There is a common phase error between the sub-mirrors (3-1). Step 2: Using one circular hole as a reference, randomly introduce 10,000 sets of co-phase errors within one wavelength to the other circular hole, and introduce 20dB of Gaussian white noise to simulate the readout noise of the camera in reality. Step 3: Collect 10,000 far-field images and record the corresponding common phase error values. Construct an image dataset and a label dataset respectively, and select a training set from the image dataset. Step 4: Configure the deep learning environment and build the neural network model; Step 5: Preprocess the image data in the training set by decomposing the single-channel image data into 4-channel components using a Haar wavelet decomposition, which are then used as input to the neural network model. The neural network model is then used to obtain the prediction co-phase error value. Step 6: Calculate the loss value between the predicted common phase error value and the given label, and optimize and adjust the parameters of the neural network.

2. The method for detecting phase co-phase error of splicing mirrors based on wavelet transform multi-channel input according to claim 1, characterized in that, The far-field image acquired in step three includes the co-phase error between sub-mirrors.

3. The method for detecting phase co-phase error of splicing mirrors based on wavelet transform multi-channel input according to claim 1, characterized in that, In step three, 80% of the image dataset is selected as the training set for the neural network model to learn the nonlinear mapping relationship between the image after Haar wavelet decomposition and the co-phase error value, and 20% of the image dataset is selected as the validation set to measure the accuracy and real-time performance of the neural network model.

4. The method for detecting phase co-phase error of splicing mirrors based on wavelet transform multi-channel input according to claim 1, characterized in that: The neural network model in step four uses a convolutional neural network ResNet18, which contains 17 convolutional layers, 2 pooling layers and 1 fully connected layer. In ResNet18, the input is directly passed to the output through residual connections.

5. The method for detecting phase co-phase error of splicing mirrors based on wavelet transform multi-channel input according to claim 1, characterized in that, In step six, the difference between the predicted co-phase error value and the given label is calculated using the mean squared error loss function (MSE) to obtain the loss value.

6. The method for detecting phase co-phase error of splicing mirrors based on wavelet transform multi-channel input according to claim 1, characterized in that, The design of an optical system based on far-field imaging according to the principles of optical systems includes: design of incident light wavelength, wavenumber, focal length of focusing lens (3-3), and circular aperture.

7. The method for detecting phase co-phase error of splicing mirrors based on wavelet transform multi-channel input according to claim 6, characterized in that, Incident light wavelength , wave number The focal length of the focusing lens (3-3) round hole .

8. The method for detecting phase co-phase error of splicing mirrors based on wavelet transform multi-channel input according to claim 1, characterized in that, Step three includes: Phase difference is applied in the circular hole of the mask (3-2) to simulate the common phase error. The phase difference is the tag. After being converged by the focusing lens (3-3), it is imaged in the detector (3-4). When the applied phase difference changes, the final far-field image changes accordingly, thus obtaining a one-to-one corresponding tag and far-field image.

9. The method for detecting phase co-phase error of splicing mirrors based on wavelet transform multi-channel input according to claim 8, characterized in that, Building the image dataset and label dataset involves creating one folder to store all named far-field images and another folder to store all labels containing image indices.

10. The method for detecting phase co-phase error of splicing mirrors based on wavelet transform multi-channel input according to claim 1, characterized in that, In step three, the mathematical expression for the image surface light intensity is: ， in, This represents the ideal distribution of point sources on the image plane. Represents convolution operation. It's camera noise. These are the x and y coordinates of the image plane, respectively. The intensity point spread function (PSF) is expressed as: ， in, It is the wavelength of the incident light. It is the focal length of the lens. The generalized pupil function of an optical system. It is a Fourier transform.

Images