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

Abstract

The application discloses a wavefront correction method for a wavefront detection-free adaptive optical system based on deep learning, comprising the following steps: S110, constructing a neural network to train a mapping relationship from far-field focal plane light intensity to corresponding complex conjugate wavefront data; S120, constructing a data set, taking the far-field focal plane light intensity as the input of the 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, using a loss function to process the loss of back propagation, and making the network learn a certain mapping relationship by taking a smaller loss; and S140, designing a wavefront correction algorithm for the neural network of the wavefront detection-free adaptive optical system to realize closed-loop correction of the wavefront detection-free adaptive optical system. According to the technical scheme, the wavefront correction speed and precision are greatly improved, and fast and high-precision wavefront correction of the wavefront detection-free adaptive optical system is realized.

Description

Technical Field

[0001] This invention relates to the field of wavefront correction technology for wavefront-free adaptive optics systems, and more particularly to a wavefront correction method for wavefront-free adaptive optics systems based on deep learning. Background Technology

[0002] Adaptive optics technology can compensate for wavefront aberrations and has wide applications in laser transmission, beam cleanup, and astronomical observation. Traditional adaptive optics systems use wavefront sensors to detect the wavefront and control wavefront correctors to compensate for it. However, wavefront sensor-based adaptive optics systems are difficult to operate effectively in scenarios with strong aberrations, long horizontal transmission distances, or faint targets, and also have high system complexity and cost. Wavefront-detection-free adaptive optics technology eliminates the need for wavefront sensors, directly using far-field light intensity information for iterative optimization to compensate for wavefront aberrations, making applications possible in these scenarios. Because it does not require wavefront sensors, wavefront-detection-free adaptive optics systems have lower system complexity and cost, and are unaffected by non-common optical path aberrations.

[0003] Model-free optimization algorithms for wavefront-detection-free adaptive optics systems use the performance metrics of far-field images as a function of control parameters, iteratively optimizing these metrics to improve performance. In 1997, MA Vorontsov et al. proposed the Stochastic Parallel Gradient Descent (SPGD) algorithm, directly utilizing far-field information for closed-loop aberration correction. Subsequent researchers have applied other optimization algorithms to wavefront-detection-free adaptive optics. For example, in 2006, S. Zommer et al. used Simulated Annealing (SA), and in 2024, Huanhuan Yu et al. used the Asymptotic Near Endpoint (APP) algorithm to achieve closed-loop correction for wavefront-detection-free adaptive optics systems. However, when aberrations are large, these optimization algorithms are prone to getting trapped in the optimal solution, and they require multiple iterations to achieve aberration compensation, resulting in poor real-time performance.

[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 capacity, and high generalization ability. Some researchers have applied deep learning algorithms to wavefront-free adaptive optics technology. In 2019, Huimin Ma et al. used AlexNet to estimate Zernike coefficients from focal plane and defocus plane images, which could be used for wavefront correction in wavefront-free adaptive optics systems without multiple iterations. However, the additional defocus camera increased the system complexity and cost. In 2019, Nishizaki et al. used Xception to directly reconstruct the wavefront from a single intensity image. However, because the far-field focal plane image corresponds to a pair of wavefronts that are complex conjugates rotated 180°, the wavefront reconstruction accuracy of a single focal plane intensity image is poor due to this double-solution problem.

[0005] While deep learning-based wavefront-detectionless adaptive optics (WDI) techniques significantly reduce aberration correction time, they struggle to achieve satisfactory correction results when only focal plane information is available. Existing methods still present challenges in closed-loop correction within practical WDI systems, severely impacting wavefront correction speed and accuracy. Achieving fast and high-precision wavefront correction for WDI systems using only far-field focal plane images remains a critical challenge. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a wavefront correction method for wavefront-free adaptive optics systems based on deep learning. By training a neural network, the trained neural network can estimate a set of wavefronts based on the light intensity at the far-field focal plane. Combined with a wavefront correction algorithm for wavefront-free adaptive optics systems used in the neural network, fast and high-precision wavefront correction of wavefront-free adaptive optics systems can be achieved.

[0007] The technical solution adopted in this invention is:

[0008] A wavefront correction method for a wavefront-free adaptive optics system based on deep learning is implemented through the following steps:

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

[0010] S120, construct a dataset, use the far-field focal plane light intensity as the input of the neural network, and use a pair of complex conjugate wavefront data corresponding to the far-field focal plane light intensity as the standard output of the neural network.

[0011] S130 uses a loss function to process the backpropagation loss. The smaller loss during backpropagation allows the network to learn a specific mapping relationship.

[0012] S140 is designed as a wavefront correction algorithm for wavefront-free adaptive optics systems for neural networks, enabling closed-loop correction of wavefront-free adaptive optics systems.

[0013] Beneficial effects:

[0014] Traditional wavefront correction algorithms for wavefront-free adaptive optics systems achieve closed-loop correction through multiple iterations, resulting in time-consuming algorithms and poor correction performance under large aberrations. While deep learning-based wavefront-free adaptive optics significantly improves wavefront correction speed, the double-solution problem—where a pair of wavefronts rotated 180° and conjugate to each other have the same far-field focal plane intensity—severely affects the wavefront correction effect. This invention improves the deep learning-based wavefront reconstruction algorithm by constructing a neural network. The trained neural network exhibits high wavefront estimation accuracy. Simultaneously, a wavefront correction algorithm for wavefront-free adaptive optics systems is designed using this neural network, significantly improving both the speed and accuracy of wavefront correction, achieving fast and high-precision wavefront correction for wavefront-free adaptive optics systems.

[0015] By improving the training dataset and loss function of the neural network during training, the neural network can accurately estimate one of the corresponding wavefronts based on the light intensity of the far-field focal plane. Combined with the wavefront correction algorithm of the wavefront-free adaptive optics system used for the neural network, without adding extra optical path complexity, it can achieve high-precision wavefront correction of the wavefront-free adaptive optics system with only far-field focal plane light intensity information, and has high real-time performance. This provides a new method for wavefront correction of the wavefront-free adaptive optics system. Attached Figure Description

[0016] Figure 1 This is a flowchart of a wavefront correction method for a wavefront-free adaptive optics system based on deep learning.

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

[0018] Figure 3 The results show the corrected SR comparison between the method of this invention and the SPGD algorithm on 1000 test sets. Detailed Implementation

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

[0020] This invention is achieved through the following steps:

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

[0022] S120, construct a dataset, use the far-field focal plane light intensity as the input of the neural network, and use a pair of complex conjugate wavefront data corresponding to the far-field focal plane light intensity as the standard output of the neural network.

[0023] S130 uses a loss function to process the backpropagation loss. The smaller loss during backpropagation allows the neural network to learn a specific mapping relationship.

[0024] S140 is designed as a wavefront correction algorithm for wavefront-free adaptive optics systems for neural networks, enabling closed-loop correction of wavefront-free adaptive optics systems.

[0025] The dataset mentioned in step S120 uses Zernike coefficients conforming to the Kolmogorov distribution to describe wavefront data. The process of constructing the dataset includes: firstly, randomly generating Zernike coefficients conforming to the Kolmogorov distribution, and then inverting the coefficients of the even-order radial Zernike mode to obtain their complex conjugate wavefront Zernike coefficients. Based on the Zernike coefficients, a far-field focal plane intensity image is simulated and generated. Each dataset contains one far-field focal plane intensity image and a pair of complex conjugate Zernike coefficients. This invention generates multiple datasets at different aberration levels and divides the datasets into training, validation, and test sets, enabling the neural network to learn the mapping relationship from far-field focal plane intensity to one of the wavefront data points, and using 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 It is a pair of complex conjugate wavefronts. These are pupil plane coordinates.

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

[0028] ,

[0029] in, The Zernike coefficients are the output of the neural network. and These are a pair of complex conjugate Zernike coefficients corresponding to the light intensity at the far-field focal plane, i.e., the standard output. and It is the root mean square error of a pair of standard outputs and neural network outputs. It is the final backpropagation loss of the neural network. The root mean square error is expressed as follows:

[0030] ,

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

[0032] By optimizing the loss function, the loss corresponding to the complex conjugate wavefront data with a small loss in the neural network output during training is used as the loss for the final backpropagation. This allows the neural network to learn and update in the direction of this solution, thereby avoiding uncertainty. The neural network eventually converges to the solution in one of the directions, and can accurately estimate one set of complex conjugate wavefront data based on the light intensity at the far-field focal plane.

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

[0034] In step S140, the wavefront correction algorithm for the wavefront-free adaptive optics system performs wavefront correction on the adaptive optics system based on the output of the neural network: First, the evaluation function of the far-field focal plane intensity is determined. The evaluation function is either the Strelby ratio (SR), the circumferential energy (EE), or other applicable far-field focal plane intensity image evaluation index. At the beginning of each iteration, the far-field focal plane intensity evaluation function value is calculated. The neural network outputs Zernike coefficients based on the current far-field focal plane intensity. The wavefront correction algorithm for the wavefront-free adaptive optics system obtains a pair of complex conjugate Zernike coefficients based on the Zernike coefficients output by the neural network. Based on this pair of Zernike coefficients, two control signals are applied respectively, and two evaluation function values ​​are calculated. The control signal corresponding to the higher evaluation function value is taken as the final control signal applied in this iteration. If both evaluation function values ​​are less than the evaluation function value at the beginning of the iteration, the algorithm iteration is stopped, and the closed-loop correction of the adaptive optics system is completed.

[0035] The following is an example describing the correction process. Figure 1This is a flowchart of a wavefront correction method for a wavefront-detection-free adaptive optics system based on deep learning. The wavefront correction process of the wavefront-detection-free adaptive optics system wavefront correction algorithm includes: first, determining the evaluation function for the far-field focal plane light intensity; the evaluation function can be the Strel ratio, the circumferential energy, or other applicable evaluation functions; in this invention, SR is used as the evaluation function; and then training a neural network model according to steps S110-S130, with the initial distortion wavefront being... The wavefront after initial correction Calculate the index based on the far-field focal plane image. The wavefront is reconstructed from the far-field focal plane image using a trained neural network model. Positive compensation wavefront And calculate the far-field index after positive compensation. ,according to The complex conjugate wavefront was calculated. Compensating 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 final control signal applied to the adaptive optics system in this iteration, thus obtaining the compensation wavefront of this iteration. ,make And proceed to the next iteration, if , , The maximum value is If the algorithm stops iterating, the closed-loop correction of the adaptive optics system is completed.

[0036] Figure 2 The results show a comparison between this algorithm and the SPGD algorithm on a test set. The far-field focal plane light intensity SR after neural network correction is 0.9315, which is higher than the 0.8386 obtained by the SPGD algorithm. Furthermore, the wavefront residual after neural network correction also has a smaller peak value (PV) and root mean square value (RMS). The neural network algorithm converges after only 3 iterations, and the single output of the neural network, after TensorRT acceleration, is only about 1 ms, while the SPGD algorithm converges after about 2000 iterations. The neural network algorithm also has a significant advantage in wavefront correction speed. Figure 3The results show a comparison of the corrected SR values ​​of this algorithm and the SPGD algorithm on 1000 test sets. It can be seen that the mean SR of this algorithm on the 1000 test sets is greater than 0.8, demonstrating good generalization ability across data at different aberration levels. The corrected far-field focal plane light intensity has a higher SR, which is superior to the traditional SPGD algorithm.

[0037] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any transformations or substitutions that can be understood or conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A wavefront correction method for a wavefront-free adaptive optics system based on deep learning, characterized in that, This can be achieved through the following steps: S110, construct a neural network to train the mapping relationship from the light intensity of the far-field focal plane to its corresponding set of complex conjugate wavefront data; S120, construct a dataset, using the far-field focal plane light intensity as the input to the neural network, and using a pair of complex conjugate wavefront data corresponding to the far-field focal plane light intensity as the standard output of the neural network; the dataset contains multiple datasets, one of which contains a far-field focal plane light intensity image and a pair of wavefront data, which have the following relationship. , and It is a pair of complex conjugate wavefronts. These are pupil plane coordinates. The wavefront data consists of Zernike coefficients that conform to the Kolmogorov distribution. The dataset is divided into a training set, a validation set, and a test set. The training set and validation set are used for the neural network to learn the mapping relationship from the light intensity of the far-field focal plane to one of the wavefronts. The test set is used to verify the accuracy of the neural network. S130, the loss function is used to process the backpropagation loss. The smaller loss during backpropagation allows the network to learn a specific mapping relationship. The expression for the loss function is as follows: ， in, The Zernike coefficients are the output of the neural network. and These are a pair of complex conjugate Zernike coefficients corresponding to the light intensity at the far-field focal plane, i.e., the standard output. and It is the root mean square error of a pair of standard outputs and neural network outputs. It is the final backpropagation loss of the network. The root mean square error is expressed as follows: ， in, The first output of the network Zernike coefficients of order, The first standard output Zernike, The total order of the Zernike coefficients; S140 designs a wavefront correction algorithm for a wavefront-free adaptive optics system using a neural network, achieving closed-loop correction of the wavefront-free adaptive optics system. The wavefront correction algorithm performs wavefront correction on the adaptive optics system based on the output of the neural network: First, it determines the evaluation function of the far-field focal plane intensity, which can be either the Strell ratio or the circumferential energy. At the beginning of each iteration, the evaluation function value of the far-field focal plane intensity is calculated. The neural network outputs Zernike coefficients based on the current far-field focal plane intensity. The wavefront correction algorithm obtains a pair of complex conjugate Zernike coefficients based on the Zernike coefficients output by the neural network. Two control signals are applied based on this pair of Zernike coefficients, and two evaluation function values ​​are calculated. The control signal corresponding to the higher evaluation function value is taken as the final control signal applied in this iteration. If both evaluation function values ​​are less than the evaluation function value at the beginning of the iteration, the algorithm iteration stops, completing the closed-loop correction of the adaptive optics system.

2. The wavefront correction method for a wavefront-free adaptive optics system based on deep learning according to claim 1, characterized in that: The neural network is a convolutional neural network.

Images