Method for correcting non-common-path aberration of adaptive optical system based on neural network

Abstract

The present invention discloses a non-common path aberration correction method for an adaptive optical system based on a neural network. The steps include: constructing a lightweight neural network CP-Net; simulating and generating a perfect point spread function image; generating a deformable mirror control voltage matrix based on this image; using this voltage matrix as the input of CP-Net for training, and ending the training when the loss function value is less than a preset minimum value; connecting the trained CP-Net to an actual control loop, receiving the real-time image obtained by a scientific camera, generating a perfect reference voltage, and directly using it to control and change the mirror surface shape of the deformable mirror. The method of the present invention can directly generate deformable mirror control commands without measuring aberrations, and can achieve real-time automatic adjustment. Compared with the traditional phase diversity method and the stochastic gradient descent algorithm, it can greatly reduce the operation difficulty and correction time, and there is no need for an operator to manually adjust relevant parameters during the correction process.

Description

Technical Field

[0001] The present invention relates to an adaptive control method for a deformable mirror, and particularly to a non-common path aberration correction method for an adaptive optical system based on a neural network. Background Art

[0002] With the in-depth exploration of the universe, direct imaging of targets such as exoplanets outside the solar system is an important goal in the field of astronomy. However, it is very difficult to observe high-value targets such as exoplanets at night, and the light signals emitted by most observation targets are relatively weak. Moreover, when the light signal passes through the Earth's atmosphere, it is affected by adverse factors such as air flow disturbance and temperature change in the Earth's atmosphere, and will be distorted, resulting in the observed target being blurred and the original physical characteristics being unable to be analyzed. Therefore, an adaptive optical system came into being to correct the distortion of the light signal and enable the telescope to obtain an image with diffraction-limited ability.

[0003] An adaptive optical system mainly consists of a deformable mirror, a wavefront sensor, a controller, and an optical imaging component. The wavefront sensor detects the wavefront slope, thereby calculating the distortion information of the wavefront, and then sends the distortion information to the controller. The controller calculates the control signal according to the distortion information and sends it to the deformable mirror actuator to compensate for the wavefront distortion caused by factors such as atmospheric turbulence by changing the mirror surface shape of the deformable mirror. In order to improve the correction ability of the entire adaptive system, the detection accuracy of the wavefront sensor is crucial. However, there is a difference between the optical path where the wavefront sensor is located and the scientific imaging optical path where the final imaging camera is set, resulting in some aberrations that cannot be detected by the wavefront sensor, thus forming non-common path aberration (Non-Common Path Error, NCPE).

[0004] At present, there are two conventional methods for correcting non-common-path aberrations: the first is the phase diversity method (phase diversity algorithm), which has cumbersome steps, requires a large number of focal plane and defocus images to be collected on-site, and has a low degree of intelligence; the second is the Stochastic Parallel Gradient Descent algorithm (SPGD), and the performance of this algorithm is extremely sensitive to the setting of the evaluation function. The correction results vary greatly when different evaluation functions are used, and the correction process takes a long time, requiring thousands or even tens of thousands of iterations. Moreover, both of these methods rely on operators to manually adjust relevant parameters during the correction process and are difficult to adjust automatically in real time. In addition, the existing wavefront neural network method is mainly used to generate the Zernike polynomial coefficients of the wavefront signal and then calculate the control command rate, and still has not got rid of the two-step strategy of detection first and then correction in the traditional method. At the same time, it requires high-cost customized hardware such as DSPs and is difficult to directly integrate the algorithm into the existing hardware system. Therefore, the existing methods for correcting non-common-path aberrations are difficult to meet the intelligent and high-performance requirements of future large telescope adaptive optical systems. Summary of the Invention

[0005] In view of the above problems existing in the prior art, the present invention proposes a method based on a neural network that can directly control a deformable mirror to autonomously correct non-common-path aberrations in an adaptive optical system. By using a simulated software to generate a perfect point spread function (PSF) image as a benchmark for training, the neural network algorithm directly generates the control voltage of the deformable mirror to compensate for the non-common-path aberrations in the optical system, thereby improving the correction ability of the adaptive optical system.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] A method for correcting non-common-path aberrations in an adaptive optical system based on a neural network, comprising the following steps:

[0008] Step 1: Construct a lightweight neural network CP-Net;

[0009] Step 2: Simulate and generate a perfect point spread function image;

[0010] Step 3: Based on the perfect point spread function image used as the training set, generate a control voltage matrix of the deformable mirror as the reference voltage;

[0011] Step 4: Use the control voltage matrix of the deformable mirror as the input of the neural network CP-Net for training. When the loss function value is less than the preset minimum value, the output voltage is the perfect reference voltage, and the training ends at this time;

[0012] Step 5: Connect the trained neural network CP-Net to the actual control loop, receive the real-time images obtained by the scientific camera, generate the perfect reference voltage, and directly use it to control the change of the mirror surface shape of the deformable mirror.

[0013] Further, in Step 1, the neural network CP-Net includes three input channels, namely the upper, middle, and lower channels, and one output channel; where:

[0014] The upper and lower input channels are the same, and each includes a number of repeated comprehensive layers. Each comprehensive layer sequentially includes a convolutional layer, a non-linear activation layer, a normalization layer, and a max-pooling layer from the input side backward;

[0015] The middle input channel sequentially includes a fusion layer, a convolutional layer, a non-linear activation layer, a normalization layer, and a max-pooling layer from the input side backward;

[0016] The output channel sequentially includes a fusion layer, a normalization layer, two consecutive convolutional layers, a non-linear activation layer, two consecutive fully connected layers, and a regression layer from the input side backward.

[0017] Further, in the upper and lower input channels, the convolutional kernel of the convolutional layer is 4*4, the stride is 1, the non-linear activation layer uses the ReLu function, the data size processed by the normalization layer is 8, the number of channels is 2, the mean is normalized to 0.5, the variance is normalized to 1, and the max-pooling layer uses 2*2 max-pooling processing; in the middle input channel, the fusion layer unifies the data stream into 4*4*2, the convolutional kernel of the convolutional layer is 1*1, the stride is 1, the non-linear activation layer uses the ReLu function, the data size processed by the normalization layer is 4, the number of channels is 2, the mean is normalized to 0.5, the variance is normalized to 1, and the max-pooling layer uses 2*2 max-pooling processing; in the output channel, the fusion layer unifies the data stream into 4*4*2, the data size processed by the normalization layer is 4, the number of channels is 2, the mean is normalized to 0.5, the variance is normalized to 1, the convolutional kernel of the convolutional layer is 2*2, the stride is 1, the non-linear activation layer uses the ReLu function, the fully connected layer uses 1*1 full connection, and the regression layer outputs a control signal of 1*1*88.

[0018] Further, in Step 2, the generation formula of the perfect point spread function image is:

[0019] ;

[0020] where A is the light intensity of the input light source, Besselj( ) represents the Bessel function, (x, y) are the horizontal and vertical coordinates of each point on the PSF image, k is the wave number of the input light source, and Na represents the numerical aperture.

[0021] Further, in Step 3, partially cover the perfect point spread function images used as the training set to simulate the signal loss under weak light sources.

[0022] Further, in step 3, the random gradient descent method is used to generate the control voltage of the deformable mirror, forming a deformable mirror control voltage matrix, which is used as the label for generating the voltage by training the CP-Net network.

[0023] The formula for generating the control voltage by the random gradient descent method is as follows:

[0024] ;

[0025] where k is the number of iterations, u is the control voltage of the deformable mirror actuator, is the gain coefficient, which is positive when the evaluation function is optimized towards the minimum value, and negative otherwise, is the change amount of the evaluation function, is the applied random perturbation voltage signal and follows the Bernoulli distribution.

[0026] Further, the following evaluation function is used to evaluate the random gradient descent method:

[0027] ;

[0028] where (x, y) are the horizontal and vertical coordinates of each point on the generated perfect point spread function image, and (x i , y i ) are the horizontal and vertical coordinates of each point on the actual captured image of the scientific camera in the adaptive optical system; when the value of the evaluation function reaches the preset threshold, the control voltage v ir output by the CP-Net network is the perfect reference voltage, which is used as the training label of the CP-Net network.

[0029] Further, the loss function adopted in step 4 is:

[0030] ;

[0031] where v i represents the actual control voltage of the deformable mirror.

[0032] Further, the calculation stage of the non-common path aberration is skipped, and the control command of the deformable mirror is directly generated to automatically correct the non-common path aberration in the adaptive optical system and make automatic adjustments during the correction process.

[0033] Compared with the prior art, the beneficial effects of the present invention are:

[0034] (1) The CP-Net network structure adopted by the non-common path aberration correction method of the adaptive optical system based on neural network in the present invention is simple and can be directly run on a commercial computer without professional equipment such as DSP.

[0035] (2) The method of the present invention can directly generate deformable mirror control commands without measuring the aberration, and can achieve real-time automatic adjustment. Compared with the traditional phase diversity method and the stochastic gradient descent algorithm, it can greatly reduce the operation difficulty and the correction time, and there is no need for the operator to manually adjust the relevant parameters during the correction process.

[0036] (3) It can be simply incorporated into the existing adaptive optical system. Description of the Drawings

[0037] Figure 1 It is a schematic diagram of the CP-Net structure.

[0038] Figure 2 It is a sample of the perfect PSF image in the CP-Net dataset.

[0039] Figure 3 It is a flowchart of the overall training and application of CP-Net.

[0040] Figure 4 They are the effect diagrams of the adaptive optical system before and after adopting the method of the present invention; among them, (a) is the effect diagram of the adaptive optical system before adopting the method of the present invention; (b) is the effect diagram of the adaptive optical system after adopting the method of the present invention. Detailed Embodiment

[0041] The present invention will be further described in detail below with reference to the drawings.

[0042] The non-common path aberration correction method of the adaptive optical system based on neural network of the present invention can enable the adaptive optical system to be free from the influence of non-common path aberration during scientific observation and perform real-time automatic adjustment of the surface shape of the deformable mirror. The method includes the following steps:

[0043] Step 1: Construct a lightweight neural network CP-Net;

[0044] Step 2: Simulate and generate a perfect point spread function image;

[0045] Step 3: Based on the perfect point spread function image used as the training set, generate a deformable mirror control voltage matrix as the reference voltage;

[0046] Step 4: Use the deformable mirror control voltage matrix as the input of the neural network CP-Net for training. When the loss function value is less than the preset minimum value, the output voltage is the perfect reference voltage, and the training ends at this time;

[0047] Step 5: Connect the trained neural network CP-Net to the actual control loop, receive the real-time image obtained by the scientific camera, generate the perfect reference voltage, and directly use it to control and change the surface shape of the deformable mirror.

[0048] Figure 1 Schematic diagram of the specific structure of the neural network CP-Net of this embodiment. The neural network CP-Net includes three input channels, namely upper, middle and lower, and one output channel. Among them:

[0049] The upper and lower input channels are the same and both include three repeated comprehensive layers. Each comprehensive layer performs the following operations sequentially from the input side to the back: first, perform a convolution operation with a convolution kernel of 4*4 and a stride of 1, then perform non-linear activation processing using the ReLu function, and then perform a normalization operation. The data size for normalization is 8 and the number of channels is 2. The mean is normalized to 0.5 and the variance is normalized to 1. Then, perform a 2*2 max pooling operation. After repeating the above operations three times, enter the output channel;

[0050] The middle input channel performs the following operations sequentially from the input side to the back: first, perform a fusion operation to unify the data stream into 4*4*2, then enter the convolution layer with a convolution kernel of 1*1 and a stride of 1. The non-linear activation layer uses the ReLu function. The data size for the normalization layer is 4 and the number of channels is 2. The mean is normalized to 0.5 and the variance is normalized to 1. The max pooling layer uses 2*2 max pooling processing, and then enters the output channel;

[0051] The output channel includes the following operations sequentially from the input side to the back: first, perform a fusion operation to unify the data stream into 4*4*2, then perform a normalization process in cooperation. The data size for processing is 4 and the number of channels is 2. The mean is normalized to 0.5 and the variance is normalized to 1. Then, immediately perform two convolution operations with a convolution kernel of 2*2 and a stride of 1. After one non-linear activation using the ReLu function, perform two fully connected operations with a kernel size of 1*1. Finally, enter the regression layer 1*1*88 to output a control signal.

[0052] During the training stage of CP-Net, 5000 simulated perfect point spread function (PSF) images are used as the data set, 3000 of which are used as the training set. The wavefront image part in the training set is covered with white squares to simulate signal loss under weak light sources. 1000 are used as the test set and 1000 are used as the validation set. Examples of the simulated perfect PSF images are as Figure 2 shown.

[0053] During the process of simulating the PSF image, various parameters of the adaptive optical system to be corrected are considered, including aperture, wavelength, light intensity, etc. Its calculation formula is executed step by step according to formulas (1), (2), and (3).

[0054] (1)

[0055] Where k is the wave number of the input light source and wl is the wavelength of the input light source.

[0056] (2)

[0057] Wherein, Na represents the numerical aperture, Dia represents the aperture of the adaptive optical system, and f represents the focal length of the optical system.

[0058] The formula for calculating the finally simulated PSF image is as follows:[[]]END]

[0059] (3)

[0060] Wherein, A is the light intensity of the input light source, Besselj( ) represents the Bessel function, and (x, y) are the horizontal and vertical coordinates of each point on the PSF image.

[0061] Subsequently, using the generated perfect PSF image as a benchmark and taking this as the optimization target, the control voltage of the deformable mirror is generated using the stochastic gradient descent method as the reference voltage, which is used as the label for training the CP-Net network to generate the voltage. The formula for generating the control voltage using the stochastic gradient descent method is as follows:[[]]END]

[0062] ;[[]]END]

[0063] Where k is the number of iterations, u is the control voltage of the deformable mirror actuator,[[]]END] is the gain coefficient, which is positive when the evaluation function is optimized towards the minimum value and negative otherwise,[[]]END] is the change amount of the evaluation function,[[]]END] is the applied random perturbation voltage signal and follows the Bernoulli distribution.[[]]END]

[0064] To generate the perfect reference voltage, the evaluation function (Metrics Function) of the stochastic gradient descent method is set as:[[]]END]

[0065]

[0066] Wherein, (x, y) are the horizontal and vertical coordinates of each point on the generated perfect PSF image, (x i , y i ) are the horizontal and vertical coordinates of each point on the actual captured image of the camera in the system. When the value of the evaluation function reaches the order of 0.0001, it can be considered that the PSF of the actual system is already consistent with the perfect PSF. At this time, the obtained control voltage v ir is used as the perfect reference voltage and used as the training label of the CP-Net. The following formula is used as the loss function during the training of the CP-Net:[[]]END]

[0067]

[0068] Wherein, v iIt represents the actual control voltage of the deformable mirror. When the value of the loss function reaches the interval [0.9, 1.1], it is considered that the optimal network training result is achieved.

[0069] After the training is completed, CP-Net can be connected to the control loop of the adaptive optical system to automatically correct the non-common path aberrations in the system. Compared with the existing neural network algorithms, CP-Net only takes more time in training. After arriving at the observation site and starting to run, it can directly eliminate the non-common path aberrations in the system, no longer needs to measure the aberrations or adjust the parameters on-site, and is more suitable for the ground-based large-aperture telescopes that are difficult to adjust.

[0070] The training and application process of the neural network CP-Net is as Figure 3 shown. Using the perfectly simulated PSF images, first calculate the reference voltage control matrix using the stochastic gradient descent algorithm and provide it as the training label data to the CP-Net network proposed by the present invention. After the training is completed, deploy CP-Net in a commercial computer (controller), directly generate the deformable mirror control command according to the real-time image, and then send it to the deformable mirror (DM). The operation control result of the system is as Figure 4 . It can be clearly seen from the figure that after the system runs, it can effectively concentrate the energy and eliminate the influence of the non-common path aberrations in the system on the observation result.

[0071] In summary, in order to ensure that the adaptive optical system is not affected by non-common path aberrations during scientific observations, the present invention proposes a method for correcting non-common path aberrations of an adaptive optical system based on a neural network. Based on the lightweight neural network CP-Net constructed according to the present invention, it automatically corrects the non-common path aberrations in the adaptive optical system before observation, skips the calculation stage of the non-common path aberrations, directly generates the deformable mirror control command, corrects the non-common path aberrations, and automatically makes adjustments during the correction process. It solves the defect that the traditional phase diversity method executes detection and correction separately, greatly improves the degree of intelligence, eliminates the manual parameter adjustment during the correction process and saves the correction time. At the same time, the CP-Net network structure is simple and can run smoothly even on a commercial PC, and can be incorporated into the existing adaptive optical system at low cost.

[0072] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, and improvements made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. Adaptive optical system non-common path aberration correction method based on neural network, characterized in that, It includes the following steps: Step 1: Construct a lightweight neural network CP-Net; the neural network CP-Net includes three input channels, namely upper, middle, and lower channels, and one output channel; where: The upper and lower input channels are the same, and each includes a number of repeated comprehensive layers. Each comprehensive layer sequentially includes a convolutional layer, a non-linear activation layer, a normalization layer, and a max pooling layer from the input side backward; The middle input channel sequentially includes a fusion layer, a convolutional layer, a non-linear activation layer, a normalization layer, and a max pooling layer from the input side backward; The output channel sequentially includes a fusion layer, a normalization layer, two consecutive convolutional layers, a non-linear activation layer, two consecutive fully connected layers, and a regression layer from the input side backward; Step 2: Simulate and generate a perfect point spread function image; Step 3: Based on the perfect point spread function image used as the training set, generate a deformable mirror control voltage matrix as the reference voltage; wherein, the control voltage of the deformable mirror is generated using the stochastic gradient descent method to form the deformable mirror control voltage matrix, which is used as the label for training the CP-Net network to generate voltage; the following evaluation function is used to evaluate the stochastic gradient descent method: ； where, ( x , y ) are the abscissa and ordinate of each point on the generated perfect point spread function image, ( x i , y i ) are the abscissa and ordinate of each point on the actual captured image of the scientific camera in the adaptive optical system; when the value of the evaluation function reaches the preset threshold, the control voltage v ir is the perfect reference voltage, which is used as the training label of the CP-Net network; Step 4: Use the deformable mirror control voltage matrix as the input of the neural network CP-Net for training. When the loss function value is less than the preset minimum value, the output voltage is the perfect reference voltage, and at this time, the training ends; Step 5: Connect the trained neural network CP-Net to the actual control loop, receive the real-time image obtained by the scientific camera, generate the perfect reference voltage, and directly use it to control and change the mirror surface shape of the deformable mirror.

2. The method for correcting non-common path aberrations of an adaptive optical system based on a neural network according to claim 1, wherein In the upper and lower input channels, the convolutional kernel of the convolutional layer is 4*4, the stride is 1, the non-linear activation layer uses the ReLu function, the data size processed by the normalization layer is 8, the number of channels is 2, the mean is normalized to 0.5, the variance is normalized to 1, and the max pooling layer uses 2*2 max pooling processing; In the middle input channel, the fusion layer unifies the data stream into 4*4*2, the convolutional kernel of the convolutional layer is 1*1, the stride is 1, the non-linear activation layer uses the ReLu function, the data size processed by the normalization layer is 4, the number of channels is 2, the mean is normalized to 0.5, the variance is normalized to 1, and the max pooling layer uses 2*2 max pooling processing; In the output channel, the fusion layer unifies the data stream into 4*4*2, the data size processed by the normalization layer is 4, the number of channels is 2, the mean is normalized to 0.5, the variance is normalized to 1, the convolutional kernel of the convolutional layer is 2*2, the stride is 1, the non-linear activation layer uses the ReLu function, the fully connected layer uses 1*1 full connection, and the regression layer outputs a control signal of 1*1*88.

3. The method for correcting the non-common path aberration of the adaptive optical system based on neural network according to claim 1, wherein In step 2, the generation formula of the perfect point spread function image is: ； Among them, A is the light intensity of the input light source, Besselj ( ) represents the Bessel function, ( x , y ) are the horizontal and vertical coordinates of each point on the PSF image, k is the wave number of the input light source, Na represents the numerical aperture.

4. The method for correcting the non-common path aberration of the adaptive optical system based on neural network according to claim 1, characterized in that, In step 3, the perfect point spread function image used as the training set is partially covered to simulate the signal loss under a faint light source.

5. The method for correcting the non-common path aberration of the adaptive optical system based on neural network according to claim 1, characterized in that The formula for generating the control voltage by the stochastic gradient descent method is as follows: ； wherein k is the number of iterations, u is the control voltage of the deformable mirror actuator, is the gain coefficient, which is positive when the evaluation function is optimized towards the minimum value and negative otherwise, is the change in the evaluation function, is the applied random perturbation voltage signal and follows a Bernoulli distribution.

6. The method for correcting the non-common path aberration of the adaptive optical system based on neural network according to claim 1, characterized in that The loss function adopted in step 4 is: ； Among them, v i represents the actual control voltage of the deformable mirror.

7. The method for correcting the non-common path aberration of the adaptive optical system based on neural network according to claim 1, characterized in that Skip the calculation stage of the off-axis aberrations, directly generate the control commands for the deformable mirror, automatically correct the off-axis aberrations in the adaptive optics system, and make automatic adjustments during the correction process.

Images