Wavefront correction method for wavefront-detection-free adaptive optical system based on deep learning

Abstract

The invention discloses a wavefront correction method for a wavefront detection-free adaptive optical system based on deep learning, and the method comprises the steps: S110, constructing a neural network for training a mapping relation from the light intensity of a far-field focal plane to a group of complex conjugate wavefront data corresponding to the far-field focal plane; s120, constructing a data set, taking the far-field focal plane light intensity as the input of a neural network, and taking a pair of complex conjugate wavefront data corresponding to the far-field focal plane light intensity as the standard output of the neural network; s130, the loss of back propagation is processed by using a loss function, and the network learns one of the determined mapping relationships for the relatively small loss of back propagation; and S140, designing a wavefront correction algorithm of the wavefront-detection-free adaptive optical system for the neural network, and realizing closed-loop correction of the wavefront-free adaptive optical system. According to the technical scheme, the wavefront correction speed and precision are greatly improved, and rapid and high-precision wavefront correction of the wavefront-detection-free adaptive optical system is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of wavefront correction for an adaptive optical system without wavefront detection, and in particular to a wavefront correction method for an adaptive optical system without wavefront detection based on deep learning. Background Art

[0002] Adaptive optics technology can compensate for wavefront aberrations and is widely used in the fields of laser transmission, beam purification, and astronomical observation. Traditional adaptive optics systems use wavefront sensors to detect wavefronts and control wavefront correctors to compensate for wavefronts. However, adaptive optics systems based on wavefront sensors are difficult to function in scenarios with strong aberration levels, long horizontal transmission distances, and dim targets, and the system complexity and system cost are high. Wavefront-free adaptive optics technology does not require wavefront sensors for wavefront detection, but directly uses far-field light intensity information for iterative optimization to compensate for wavefront aberrations, making it possible for these scenarios to be applied. Since no wavefront sensors are required, wavefront-free adaptive optics systems have lower system complexity and system cost, and are not affected by non-public optical path aberrations.

[0003] The non-wavefront detection adaptive optical system based on the model-free optimization algorithm takes the performance index of the far-field image as a function of the control parameter, and uses the optimization algorithm to iteratively optimize and improve the performance index. In 1997, MA Vorontsov et al. proposed the stochastic parallel gradient descent algorithm (SPGD), which directly uses far-field information to perform closed-loop correction of aberrations. Subsequent researchers also used other optimization algorithms in non-wavefront detection adaptive optical technology. For example, S. Zommer et al. used the simulated annealing algorithm (SA) in 2006, and Huanhuan Yu et al. used the asymptotic near-end point algorithm (APP) in 2024 to achieve closed-loop correction of the non-wavefront detection adaptive optical system. However, when the aberration is large, the above optimization algorithm is prone to fall into the optimal solution, and the above algorithm requires multiple iterations to achieve aberration compensation, and the real-time performance is poor.

[0004] In recent years, deep learning algorithms have been widely used in the field of adaptive optics due to their advantages such as high real-time performance, high learning ability, and high generalization ability. Some researchers have applied deep learning algorithms to adaptive optics technology without wavefront detection. In 2019, Huimin Ma et al. used AlexNet to estimate the Zernike coefficients based on the focal plane and defocused plane images, which can be used for wavefront correction of adaptive optics systems without wavefront detection without multiple iterations, but the additional defocus camera increases the system complexity and system cost. In 2019, Nishizaki et al. used Xception to reconstruct the wavefront directly from a single light intensity image, but because the far-field focal plane image corresponds to a pair of wavefronts that are complex conjugates of each other with a rotation of 180°, the wavefront reconstruction accuracy of a single focal plane intensity is poor due to this double solution problem.

[0005] Although deep learning-based non-wavefront detection adaptive optics technology has greatly reduced the aberration correction time, when only focal plane information is available, the existing deep learning-based non-wavefront detection adaptive optics technology is difficult to show good correction effects. The existing methods still have certain problems in the closed-loop correction of actual non-wavefront detection adaptive optics systems, which seriously affects the wavefront correction speed and correction accuracy of non-wavefront detection adaptive optics systems. How to achieve fast and high-precision wavefront correction of non-wavefront detection adaptive optics systems using only far-field focal plane images is a problem that still needs to be solved. Summary of the invention

[0006] In view of the shortcomings of the prior art, the present invention provides a wavefront correction method for an adaptive optical system without wavefront detection based on deep learning. By training a neural network, the trained neural network can estimate one set of wavefronts according to the far-field focal plane intensity. Combined with the wavefront correction algorithm for an adaptive optical system without wavefront detection used in a neural network, fast and high-precision wavefront correction of an adaptive optical system without wavefront detection can be achieved.

[0007] The technical solution adopted by the present invention is:

[0008] A wavefront correction method for an adaptive optical system without wavefront detection based on deep learning is implemented by the following steps:

[0009] S110, constructing a neural network to train a mapping relationship from far-field focal plane light intensity to a corresponding set of complex conjugate wavefront data;

[0010] S120, constructing a data set, taking the far-field focal plane light intensity as an input of the neural network, and taking a pair of complex conjugate wavefront data corresponding to the far-field focal plane light intensity as a standard output of the neural network;

[0011] S130, using a loss function to process the back-propagation loss, back-propagating a smaller loss, so that the network learns one of the determined mapping relationships;

[0012] S140, a wavefront correction algorithm for adaptive optical systems without wavefront detection designed for neural networks, realizes closed-loop correction of adaptive optical systems without wavefront detection.

[0013] Beneficial effects:

[0014] The traditional wavefront correction algorithm of the adaptive optical system without wavefront detection realizes the closed-loop correction of the system through multiple iterations. The algorithm is time-consuming and the correction effect is poor under large aberrations. Although the adaptive optical technology without wavefront detection based on deep learning has greatly improved the speed of wavefront correction, since a pair of wavefronts that are rotated 180° in a complex conjugate relationship have the same far-field focal plane intensity, this double solution problem seriously affects the wavefront correction effect. The present invention improves the wavefront reconstruction algorithm based on deep learning, constructs a neural network, and the trained neural network has a high wavefront estimation accuracy. At the same time, a wavefront correction algorithm for the adaptive optical system without wavefront detection is designed for the neural network, which greatly improves the speed and accuracy of wavefront correction and realizes fast and high-precision wavefront correction of the adaptive optical system without wavefront detection.

[0015] By improving the training data set of the neural network and the loss function during the training process, the neural network can accurately estimate one of the corresponding wavefronts based on the far-field focal plane intensity. Combined with the wavefront correction algorithm of the adaptive optical system without wavefront detection for neural networks, there is no additional optical path complexity. Only the far-field focal plane intensity information is needed to achieve high-precision wavefront correction of the adaptive optical system without wavefront detection, and it has high real-time performance, which provides a new method for wavefront correction of the adaptive optical system without wavefront detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 A flow chart of a wavefront correction method for an adaptive optical system without wavefront detection based on deep learning;

[0017] Figure 2 The comparison results of the method of the present invention and the SPGD algorithm on a set of test sets;

[0018] Figure 3 This is the corrected SR comparison result between the method of the present invention and the SPGD algorithm on 1000 test sets. DETAILED DESCRIPTION

[0019] In order to more specifically describe the technical solutions and advantages of the present invention, the present invention will be described in more detail below with reference to the accompanying drawings.

[0020] The present invention is achieved by the following steps:

[0021] S110, constructing a neural network (NN) to train a mapping relationship from far-field focal plane light intensity to a corresponding set of complex conjugate wavefront data;

[0022] S120, constructing a data set, taking the far-field focal plane light intensity as an input of the neural network, and taking a pair of complex conjugate wavefront data corresponding to the far-field focal plane light intensity as a standard output of the neural network;

[0023] S130, using a loss function to process the back-propagation loss, back-propagating a smaller loss, so that the neural network learns one of the determined mapping relationships;

[0024] S140, a wavefront correction algorithm for adaptive optical systems without wavefront detection designed for neural networks, realizes closed-loop correction of adaptive optical systems without wavefront detection.

[0025] The data set mentioned in step S120 uses Zernike coefficients that conform to the Kolmogorov distribution to describe the wavefront data. The process of constructing the data set includes: first randomly generating Zernike coefficients that conform to the Kolmogorov distribution, and inverting the dual-order radial Zernike mode coefficients to obtain their complex conjugate wavefront Zernike coefficients, and simulating and generating a far-field focal plane intensity image according to the Zernike coefficients. A set of data sets includes a far-field focal plane intensity image and a pair of mutually complex conjugate Zernike coefficients. The present invention generates multiple sets of data sets at different aberration levels, and divides the data sets into a training set, a verification set, and a test set, so that the neural network learns the mapping relationship from the far-field focal plane intensity to one of the wavefront data, and uses the test set to verify the accuracy of the neural network.

[0026] A pair of complex conjugate wavefront data has the following relationship ,in, and is a pair of complex conjugate wavefronts, are pupil coordinates.

[0027] In one embodiment, in step S130, the loss function expression is as follows:

[0028] ,

[0029] in, is the Zernike coefficient output by the neural network, and is a pair of mutually complex conjugate Zernike coefficients corresponding to the far-field focal plane intensity, i.e., the standard output. and is the root mean square error between a pair of standard outputs and the neural network output, is the final back-propagation loss of the neural network. is the root mean square error, and its expression is:

[0030] ,

[0031] in, The network output Zernike coefficients of order, The standard output Zernike coefficients of order, is the total order of the Zernike coefficients.

[0032] By optimizing the loss function, the loss corresponding to the complex conjugate wavefront data with smaller loss than the neural network output during training is used as the final back-propagation loss, so that the neural network can learn and update in the direction of the solution, thereby avoiding uncertainty. The neural network eventually converges to a solution in one direction and can accurately estimate one of the corresponding sets of complex conjugate wavefront data based on the far-field focal plane intensity.

[0033] The neural network described above may be a convolutional neural network (CNN), a neural network including an attention mechanism, or other applicable neural networks.

[0034] In step S140, the wavefront correction algorithm of the adaptive optical system without wavefront detection performs wavefront correction on the adaptive optical system according to the output of the neural network: first, an evaluation function of the far-field focal plane intensity is determined, and the evaluation function is the Strehl ratio (SR) or the surrounding energy (EE) or other applicable far-field focal plane intensity image evaluation indicators; at the beginning of each round of iteration, the far-field focal plane intensity evaluation function value is calculated, and the neural network outputs the Zernike coefficient according to the current far-field focal plane intensity. The wavefront correction algorithm of the adaptive optical system without wavefront detection obtains a pair of mutually complex conjugate Zernike coefficients according to the Zernike coefficients output by the neural network, and applies control signals twice according to the pair of Zernike coefficients, and calculates two evaluation function values, and takes the control signal corresponding to the higher evaluation function value as the control signal finally applied in this round of iteration. If the two evaluation function values ​​are both less than the evaluation function value at the beginning of the iteration, the algorithm iteration is stopped to complete the closed-loop correction of the adaptive optical system.

[0035] An example of this correction process is described below. Figure 1The flowchart of the wavefront correction method of the adaptive optical system without wavefront detection based on deep learning is shown in the figure. The process of wavefront correction of the adaptive optical system without wavefront detection includes: firstly determining the evaluation function of the far-field focal plane intensity, which can be the Strehl ratio, the surrounding energy, or other applicable evaluation functions. In the present invention, SR is used as the evaluation function; according to the steps S110-S130, a neural network model is obtained by training, and the initial distorted wavefront is , the wavefront after initial correction , calculate the index based on the far-field focal plane image , using the trained neural network model to reconstruct the wavefront from the far-field focal plane image , forward compensation wavefront , and calculate the far-field index after forward compensation ,according to The complex conjugate wavefront is calculated , compensated complex conjugate wavefront , and calculate the far-field index after complex conjugate compensation .like , , The maximum value is or , then the control signal corresponding to the higher evaluation function value is taken as the control signal applied to the adaptive optical system in this round of iteration, and the compensation wavefront of this round of iteration is obtained. ,make And enter the next iteration, if , , The maximum value is , the algorithm iteration is stopped and the closed-loop correction of the adaptive optical system is completed.

[0036] Figure 2 The comparison results of the algorithm and the SPGD algorithm on a set of test sets. The far-field focal plane intensity SR after neural network correction is 0.9315, which is higher than the 0.8386 obtained by the SPGD algorithm, and the wavefront residual after neural network correction also has a smaller peak value (PV) and root mean square (RMS). The neural network algorithm only needs 3 rounds of iterations to converge. The single output of the neural network is only about 1 ms after TensorRT acceleration, while the SPGD algorithm converges after about 2000 rounds. The neural network algorithm also has a great advantage in the speed of wavefront correction. Figure 3The following is the comparison of the corrected SR between the algorithm and the SPGD algorithm on 1000 test sets. It can be seen that the average SR of the algorithm on the 1000 test sets is greater than 0.8, and it shows good generalization on data with different aberration levels. The corrected far-field focal plane intensity has a higher SR, which is better than the traditional SPGD algorithm.

[0037] The above description is only a specific implementation of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be understood and thought of by anyone familiar with the technology within the technical scope disclosed by the present invention should be covered within the protection scope of the present invention.

Claims

1. A wavefront correction method for an adaptive optical system without wavefront detection based on deep learning, characterized in that: This is accomplished by following these steps: S110, constructing a neural network to train a mapping relationship from far-field focal plane light intensity to a corresponding set of complex conjugate wavefront data; S120, constructing a data set, taking the far-field focal plane light intensity as an input of the neural network, and taking a pair of complex conjugate wavefront data corresponding to the far-field focal plane light intensity as a standard output of the neural network; S130, using a loss function to process the back-propagation loss, back-propagating a smaller loss, so that the network learns one of the determined mapping relationships; S140, a wavefront correction algorithm for adaptive optical systems without wavefront detection designed for neural networks, realizes closed-loop correction of adaptive optical systems without wavefront detection.

2. The wavefront correction method for an adaptive optical system without wavefront detection based on deep learning according to claim 1, characterized in that: The data set includes multiple data sets, one of which includes a far-field focal plane intensity image and a pair of wavefront data. The pair of wavefront data has the following relationship: , and is a pair of complex conjugate wavefronts, is the pupil coordinate, and the wavefront data is composed of Zernike coefficients that conform to the Kolmogorov distribution; the data set is divided into training set, validation set and test set. The training set and validation set are used by the neural network to learn the mapping relationship from the far-field focal plane intensity to one of the wavefronts, and the test set is used to verify the accuracy of the neural network.

3. The wavefront correction method for an adaptive optical system without wavefront detection based on deep learning according to claim 1, characterized in that: The loss function expression is as follows: ， in, is the Zernike coefficient output by the neural network, and is a pair of mutually complex conjugate Zernike coefficients corresponding to the far-field focal plane intensity, i.e., the standard output. and is the root mean square error between a pair of standard outputs and the neural network output, is the final back-propagation loss of the network, is the root mean square error, and its expression is: ， in, The network output Zernike coefficients of order, The standard output Order Zernike, is the total order of the Zernike coefficients.

4. The wavefront correction method for an adaptive optical system without wavefront detection based on deep learning according to claim 1, characterized in that: The neural network is a convolutional neural network.

5. The wavefront correction method for an adaptive optical system without wavefront detection based on deep learning according to claim 1, characterized in that: In step S140, the wavefront correction algorithm of the adaptive optical system without wavefront detection performs wavefront correction on the adaptive optical system according to the output of the neural network: first, the evaluation function of the far-field focal plane intensity is determined, and the evaluation function is the Strehl ratio or the surrounding energy; at the beginning of each round of iteration, the far-field focal plane intensity evaluation function value is calculated, and the neural network outputs the Zernike coefficient according to the current far-field focal plane intensity. The wavefront correction algorithm of the adaptive optical system without wavefront detection obtains a pair of complex conjugate Zernike coefficients according to the Zernike coefficients output by the neural network, and applies control signals twice according to the pair of Zernike coefficients, and calculates two evaluation function values, and takes the control signal corresponding to the higher evaluation function value as the control signal finally applied in this round of iteration. If the two evaluation function values ​​are both less than the evaluation function value at the beginning of the iteration, the algorithm iteration is stopped to complete the closed-loop correction of the adaptive optical system.

Images