CN118409427A - Sophia-based random parallel gradient descent self-adaptive optical wavefront correction method, system and electronic equipment - Google Patents

Abstract

The invention discloses a self-adaptive optical wavefront correction method, a system and electronic equipment based on the random parallel gradient descent of Sophia, wherein the method calculates the signal performance index variation quantity generated by an image sensor according to the random disturbance generated in the iteration of a voltage control signal of a wavefront corrector; the SPGD algorithm is improved based on Sopiha optimizers, and gradients are calculated by using the variation of the performance indexes and random disturbance; constructing a first-order momentum and a second-order momentum in the improved SPGD algorithm; and adding weight parameters and gradient shearing into an algorithm for constructing first-order momentum and second-order momentum, and updating a voltage control signal of the wavefront corrector in a mode of combining the first-order momentum and the second-order momentum to realize gain coefficient self-adaption. The scheme of the invention can improve the system convergence accuracy, accelerate the system convergence speed and reduce the probability of sinking into local extremum.

Description

Stochastic parallel gradient descent adaptive optical wavefront correction method, system and electronic equipment based on Sophia

Technical Field

The present invention relates to the field of adaptive optics, and in particular to a random parallel gradient descent adaptive optics wavefront correction method, system and electronic equipment based on Sophia.

Background technique

Adaptive optics (AO) is an optical system technology that aims to correct aberrations caused by non-uniform materials or defects in optical components. Adaptive optics systems usually include wavefront sensors, wavefront correctors, and wavefront controllers. They analyze the data from the wavefront sensor in real time and use feedback mechanisms to control the wavefront corrector to deform and compensate for the distorted wavefront.

The existence of wavefront sensors in traditional adaptive optical systems makes adaptive optical systems expensive and complex in structure. For fields that require miniaturized structures, traditional adaptive optical systems can no longer meet application requirements. Wavefront-sensor-less adaptive optical systems (Wavefront-sensor-less AO) are an adaptive optical technology that does not rely on wavefront sensors. It uses adaptive algorithms or model methods to achieve wavefront correction without measuring wavefront distortion. At present, this technology is widely used in wavefront correction in various application scenarios, such as optical microscopy, free-space optical communications, space detection and measurement, astronomical target observation, retinal imaging, etc.

Whether the adaptive optical system without wavefront sensing can achieve the expected correction effect in practical applications depends on the system control algorithm, and the control algorithm involves the iterative solution of control parameters. The system first initializes the wavefront corrector of the adaptive optical system and other parameters required by the correction algorithm. The image sensor is used to measure the far-field spot corresponding to the wavefront distortion, and the system performance indicators, such as the square sum of the far-field spot intensity distribution, Strehl ratio and average radius, and ring energy, are calculated based on the far-field spot information. The control algorithm calculates the control signal required by the wavefront corrector according to the change of the performance index value, and applies it to the wavefront corrector to correct the wavefront distortion; iterate the above process until the predetermined correction criterion or convergence standard is met. In the iterative process, different control methods can be used, such as genetic algorithm (GA), simulated annealing algorithm (SA), particle swarm optimization algorithm (PSO) and stochastic parallel gradient descent algorithm (SPGD), etc., which can effectively adjust the control parameters to minimize the wavefront residual.

The wavefront sensor-free adaptive optical system based on the stochastic parallel gradient descent algorithm SPGD is a wavefront correction technology that combines stochastic gradient descent (SGD) and parallel computing. Its main idea is to use parallel computing to accelerate the gradient calculation process, and use stochastic gradient descent to update the control parameters. Since SPGD adopts the idea of stochastic gradient descent, the gradient updated at each node is based on randomly selected samples, which may cause fluctuations or oscillations in the algorithm during the iteration process, making the convergence speed unstable; the conventional SPGD algorithm uses a fixed gain coefficient, and has poor performance in correcting wavefront distortion in strong turbulence, is prone to falling into local extremes, and converges slowly; SPGD is not suitable for the optimization problem of strong turbulence phase distortion, and its correction ability is not as good as other optimization algorithms, and its application scenarios are limited.

Summary of the invention

The purpose of the present invention is to address the above problems and propose a random parallel gradient descent adaptive optical wavefront correction method, system and electronic equipment based on Sophia, which adopts an adaptive gain adjustment strategy, has better correction capability for wavefront distortion of turbulence of different intensities, reduces the probability of falling into local extreme values, improves the convergence speed, and enhances the generalization ability of the system.

To achieve the above object, the technical solution provided by the present invention is:

A random parallel gradient descent adaptive optical wavefront correction method based on Sophia, comprising the following steps:

Initialize the parameters, and calculate the change in the signal performance index corresponding to the image sensor according to the random perturbation of the voltage control signal of the wavefront corrector in the iteration;

The gradient is calculated from the change of signal performance index and random disturbance, and the Sopiha optimizer is used to improve the SPGD algorithm according to the gradient to perform wavefront correction;

The first-order momentum and the second-order momentum are constructed in the improved SPGD algorithm. The weight parameter is added to the SPGD algorithm with the first-order momentum and the second-order momentum constructed, and the voltage control signal of the wavefront corrector is updated by the combination of the first-order momentum and the second-order momentum to achieve the adaptive gain coefficient.

The initial value of the SPGD algorithm is set, and the change in the signal performance index corresponding to the image sensor is calculated according to the random disturbance of the voltage control signal of the wavefront corrector in the iteration, specifically:

Set the initial value of the SPGD algorithm, assume that the voltage control signal of the wavefront corrector is u(k) = {u ₁ ,u ₂ ,···,u _N }, where N represents the number of driver units of the wavefront corrector. At the kth iteration, a random disturbance {Δu} is generated, and the corresponding disturbances Δu and -Δu in two directions are applied to the wavefront corrector respectively; the controller collects the signal of the image sensor and calculates the performance indicators J(u+Δu) and J(u-Δu), and obtains the signal performance indicator change ΔJ(k):

ΔJ(k)=J(u+Δu)-J(u-Δu).

The calculation of the gradient based on the signal performance index change and the random disturbance specifically includes:

The first-order gradient g(k) is calculated using the product of the signal performance index change ΔJ(k) and the random disturbance Δu:

g(k)＝ΔJ(k)·sign(Δu);

Wherein, sign() is a sign function, whose function is to take the positive or negative sign of a number; the sign of ΔJ(k)·sign(Δu) is determined by the optimization direction of the signal performance index. If the optimization direction of the signal performance index is forward, it is positive, otherwise it is negative.

The construction of first-order momentum and second-order momentum in the improved SPGD algorithm is specifically as follows:

The calculation formula for first-order momentum is:

m(k+1)=β ₁ ·m(k)+(1-β ₁ )·g(k);

Where β ₁ is the set hyperparameter, m(k) represents the first-order momentum of the kth iteration, and m(k+1) represents the first-order momentum of the k+1th iteration;

Calculate the second-order gradient using the formula:

In the formula, represents the second-order gradient of k iterations;

Construct the second-order momentum, and the calculation formula is:

Where β ₂ is the set hyperparameter, h(k) represents the second-order momentum of the kth iteration, and h(k+1) represents the second-order momentum of the k+1th iteration.

The Sopiha optimizer updates the voltage parameters of the wavefront corrector according to the combined optimization of the first-order gradient and the second-order gradient, finds the optimal solution, and completes the wavefront correction.

The weight parameter is added to the SPGD algorithm that constructs the first-order momentum and the second-order momentum, and the voltage control signal of the wavefront corrector is updated by combining the first-order momentum and the second-order momentum to achieve adaptive gain coefficient, specifically:

Add the weight parameter λ to achieve adaptive gain coefficient and improve the convergence speed. At this time, the voltage control signal formula of the wavefront corrector is updated as follows:

u(k)=u(k)+λ·α(k)·u(k);

In the formula, λ is the set parameter, α(k) is the adaptive learning rate, and the reference expression is α(k) = 0.05/(1+0.001*k);

The voltage control signal of the wavefront corrector is updated by combining the first-order momentum and the second-order momentum to achieve adaptive gain coefficient and improve convergence speed. The updated voltage control signal of the wavefront corrector is obtained as follows:

u(k+1)＝u(k)+α(k)·clip(m/max(γ·h(k),ε),ρ(k))

Wherein, u(k+1) represents the voltage signal of the k+1th iteration; ρ(k) is the adaptive clipping range parameter, the reference expression is 0.5*(1+cos((π*k)/T)), T is the total number of iterations; clip(m/max(γ·h(k),ε),ρ(k)) is the clipping function, which is used to limit m/max(γ·h(k),ε) within the range of -ρ(k)～ρ(k); the parameter ε is 10 ^-8 ;

The updated voltage control signal of the wavefront corrector is applied to the wavefront corrector through the amplifier to complete the current iteration; the above correction process is repeated until the end condition is met; the end condition is a certain number of iterations or the performance index reaches a fixed value.

The present invention also protects a system used in the random parallel gradient descent adaptive optical wavefront correction method based on Sophia, comprising:

The gradient descent module is used to initialize parameters and calculate the change in the signal performance index corresponding to the image sensor according to the random perturbation of the voltage control signal of the wavefront corrector in the iteration; the gradient is calculated from the change in the signal performance index and the random perturbation, and the Sopiha optimizer is used to improve the SPGD algorithm according to the gradient to perform wavefront correction;

The gain coefficient adaptive module is used to construct the first-order momentum and the second-order momentum in the improved SPGD algorithm; a weight parameter is added to the SPGD algorithm that constructs the first-order momentum and the second-order momentum, and the voltage control signal of the wavefront corrector is updated by combining the first-order momentum and the second-order momentum to achieve gain coefficient adaptation.

The present invention also protects an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the random parallel gradient descent adaptive optical wavefront correction method based on Sophia as described above is implemented.

The beneficial effects of the present invention are:

The present invention integrates the deep learning Sophia optimizer into the conventional SPGD algorithm, approximates the product of the micro-change amount and the disturbance amount of the system performance index as the gradient, calculates the first-order gradient and the second-order gradient, and uses the first-order momentum, second-order momentum and adaptive parameters of the gradient to control the direction and step size of the gradient descent, realizes the adaptation of the gain coefficient, improves the convergence speed of the algorithm, and reduces the probability of falling into the local extreme value; adopts a single loop structure and vector parallel operation to reduce the calculation amount and running time of the controller, greatly reduces the number of iterations, that is, the number of far-field camera sampling, and further improves the system correction speed.

In aberration-free target imaging, the light beam will form an Airy disk at the imaging position, reaching the diffraction limit, while in an aberrated optical system, the far-field spot is diffuse; the conventional SPGD algorithm uses a fixed gain coefficient for random gradient descent. When the wavefront aberration changes randomly, it is easy to fall into a local extreme value, the convergence speed is slow, and the convergence is not stable enough. The solution of the present invention solves the above problems of conventional SPGD, and has the advantages of adaptive learning rate characteristics and parallel control of SPGD. The micro-change amount of the system performance index is approximated as the gradient, the first-order mean and second-order mean of the gradient are calculated, and the weight decay and adaptive shear limit update mechanism are introduced to control the direction and step size of the gradient descent in combination, so as to realize the adaptive adjustment of the gain coefficient, improve the system convergence accuracy, accelerate the system convergence speed, and reduce the probability of falling into a local extreme value.

The Sopiha optimizer is a second-order optimizer. In order to improve and optimize the effect of gradient descent, the present invention uses the Sopiha optimizer to improve the SPGD algorithm, so as to improve the adaptive learning rate adjustment and implement a more accurate update strategy for the wavefront corrector voltage signal, thereby improving the convergence speed and convergence accuracy of the SPGD algorithm and achieving a more efficient wavefront correction effect.

The algorithm of the present invention can further accelerate the convergence speed, reduce the calculation amount of the control parameters of the wavefront corrector and the sampling times of the image sensor information, and improve the real-time performance and effectiveness of the adaptive optical system without wavefront sensing.

Specifically, compared with the prior art, the solution of the present invention has the following advantages:

(1) Compared with traditional adaptive optics systems, the wavefront correction system based on Sophia's SPGD optimization algorithm does not require a wavefront sensor;

(2) Compared with the wavefront correction system based on the conventional SPGD optimization algorithm, the gradient descent strategy is improved, and the mechanisms of gain coefficient adaptation, weight attenuation and adaptive shear limit update are realized, which makes the system converge faster, reduces the probability of falling into local extreme values, and can perform wavefront correction more efficiently.

(3) The wavefront correction system is based on Sophia's SPGD optimization algorithm, and has strong adaptability to wavefront distortion under different turbulence intensities and good correction effect on strong turbulence.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a system structure diagram of the Sophia-based random parallel gradient descent adaptive optical wavefront correction method of the present invention.

FIG. 2 is an algorithm flow chart of the Sophia-based random parallel gradient descent adaptive optical wavefront correction method of the present invention.

FIG3 is an example of wavefront distortion to be corrected in the present invention, wherein (a) is the far-field spot diagram before correction, (b) is the far-field spot diagram after correction, and (c) is a comparison of the convergence of the control algorithm.

Detailed ways

The above contents of the present invention are further described in detail below in the form of embodiments, but this should not be understood as the scope of the above subject matter of the present invention being limited to the following embodiments, and all technologies realized based on the above contents of the present invention belong to the scope of the present invention.

The present invention provides a random parallel gradient descent adaptive optical wavefront correction method based on Sophia, comprising the following steps:

Step S1: Initialize parameters, and calculate the change in the signal performance index corresponding to the image sensor according to the random disturbance of the voltage control signal of the wavefront corrector in the iteration;

Step S2: Calculate the gradient from the signal performance index change and the random disturbance, and use the Sopiha optimizer to improve the SPGD algorithm according to the gradient to perform wavefront correction;

Step S3: construct the first-order momentum and the second-order momentum in the improved SPGD algorithm; add weight parameters to the SPGD algorithm that constructs the first-order momentum and the second-order momentum, and update the voltage control signal of the wavefront corrector by combining the first-order momentum and the second-order momentum to achieve gain coefficient adaptation.

The present invention also provides a system adopted by the Sophia-based random parallel gradient descent adaptive optical wavefront correction method, comprising:

The gradient descent module is used to initialize parameters and calculate the change in the signal performance index corresponding to the image sensor according to the random perturbation of the voltage control signal of the wavefront corrector in the iteration; the gradient is calculated from the change in the signal performance index and the random perturbation, and the Sopiha optimizer is used to improve the SPGD algorithm according to the gradient to perform wavefront correction;

The gain coefficient adaptive module is used to construct the first-order momentum and the second-order momentum in the improved SPGD algorithm; a weight parameter is added to the SPGD algorithm that constructs the first-order momentum and the second-order momentum, and the voltage control signal of the wavefront corrector is updated by combining the first-order momentum and the second-order momentum to achieve gain coefficient adaptation.

The present invention also provides an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the random parallel gradient descent adaptive optical wavefront correction method based on Sophia as described above is implemented.

The present invention is further described in detail below in conjunction with specific embodiments:

As shown in FIG1 , a wavefront sensor-free adaptive wavefront correction system based on Sophia's SPGD optimization algorithm is mainly composed of a wavefront corrector, a lens, an image sensor and a controller. In the specific implementation scheme, a deformable mirror is used as an example of the wavefront corrector, a CCD is used as an example of the image sensor, the mean radius (MR) of the far-field light is used as an example of the system performance indicator, and the Strehl ratio (SR) is used as an example of the system performance observation indicator. The above examples are only specific embodiments of the present invention, and the present invention is not limited to the above embodiments.

The wavefront distortion is reflected by the deformable mirror and enters the CCD camera to obtain the far-field light spot information; the controller reads the far-field light spot image and calculates the performance index MR based on the far-field light spot information.

Sophia’s SPGD optimization algorithm is used to update the voltage control signal and then drive the deformable mirror to correct the wavefront distortion. The algorithm flow is shown in Figure 2.

The specific implementation steps are as follows:

Step 1: Set the initial value of the control algorithm; assume that the current voltage of the wavefront corrector is u(k) = {u ₁ ,u ₂ ,···,u _N }, where N represents the number of driver units of the wavefront corrector. At the kth iteration, a small random voltage disturbance {Δu} is generated, and the corresponding disturbances Δu and -Δu in two directions are applied to the deformable mirror respectively. The controller collects the CCD signal and calculates the performance indicators J(u+Δu) and J(u-Δu), and obtains the signal performance indicator change ΔJ(k):

ΔJ(k)＝J(u+Δu)-J(u-Δu);

Wherein, k represents the number of iterations, that is, the voltage control signal of the wavefront corrector is updated for the kth time, and the voltage control signal of the wavefront corrector is updated iteratively. k is usually sequentially increasing, representing the order of iterations, and each iteration k+1 is based on the result of the previous iteration k;

Step 2: Calculate the first-order gradient g(k) by multiplying the change in the signal performance index ΔJ(k) and the random disturbance Δu;

g(k)＝ΔJ(k)·sign(Δu);

Where sign() is a sign function, which is used to take the positive or negative sign of a number; the sign and performance of ΔJ(k)·sign(Δu) are determined by the optimization direction of the performance index. If the performance index moves in the direction of maximum, it is positive, otherwise it is negative;

Step 3: Improve the conventional SPGD algorithm based on the Sopiha optimizer and add the first-order momentum. The calculation formula of the first-order momentum is:

m(k+1)=β ₁ ·m(k)+(1-β ₁ )·g(k);

Where β ₁ is a set hyperparameter with a reference value of 0.9; m(k) represents the first-order momentum m of the kth iteration, and m(k+1) represents the first-order momentum of the k+1th iteration;

Step 4: Calculate the second-order gradient The approximate calculation formula is

The Sopiha optimizer updates the voltage parameters of the wavefront corrector according to the combined optimization of the first-order gradient and the second-order gradient, so that it can find the optimal solution faster and more accurately and complete the wavefront correction.

Step 5: Add a hyperparameter β ₂ to construct the second-order momentum h, calculated as

Where β ₂ is a hyperparameter with a reference value of 0.99; h(k) represents the second-order momentum of the kth iteration, and h(k+1) represents the second-order momentum of the k+1th iteration;

Among them, β ₁ and β ₂ are hyperparameters, which are the decay rates of two moving averages respectively; the reference values of β 1 and β ₂ are mainly obtained through experimental tests, and the values with the best effects are selected as reference values; when starting to measure, the recommended default values can be used, for example, β ₁ = 0.9, β ₂ = 0.999.

Step 6: Add weight parameter λ to achieve adaptive gain coefficient and improve convergence speed. At this time, the control voltage signal update formula is:

u(k)＝u(k)+λ·α(k)·u(k)

Where λ is a hyperparameter, the reference value is 0.01, α(k) is the adaptive learning rate, and the reference expression is α(k) = 0.05/(1+0.001*k);

λ is mainly used to increase the regularization effect. The weight decay coefficient is usually set in the range of 10-4 to 10-2. The specific value of λ should be adjusted based on the experimental performance.

The expression of α(k) is obtained according to the inverse time decay strategy of learning rate decay in the gradient descent algorithm:

The values of α ₀ and β are measured experimentally;

Step 7: Further update the control signal by combining the first-order momentum and the second-order momentum to achieve adaptive gain coefficients, improve convergence speed, and avoid falling into local extreme solutions. The final control voltage signal update formula is:

u(k+1)＝u(k)+α(k)·clip(m/max(γ·h(k),ε),ρ(k))

Wherein, u(k+1) represents the voltage signal of the k+1th iteration; ρ(k) is the adaptive clipping range parameter, and its reference expression is 0.5*(1+cos((π*k)/T)), where T is the total number of iterations; clip(m/max(γ·h(k),ε),ρ(k)) is the clipping function, which is used to limit m/max(γ·h(k),ε) to the range of -ρ(k)～ρ(k); the parameter ε is 10 ^-8 to avoid the denominator being zero during the algorithm iteration process; the expression of ρ(k) is obtained according to the cosine attenuation strategy of the learning rate attenuation in the gradient descent algorithm:

Among them, the value of α ₀ is measured by experiment;

Step 8: The updated voltage control signal is applied to the wavefront calibrator through the amplifier to complete the current iteration, and the correction process from the first step to the seventh step is repeated until the end condition is met. The end condition can be a certain number of iterations or the performance index reaches a fixed value.

The above description is only a preferred embodiment of the present invention and does not limit the present invention in any form. Any technician familiar with the profession, without departing from the scope of the technical solution of the present invention, according to the technical essence of the present invention, any simple modification, equivalent replacement and improvement made to the above embodiment still falls within the protection scope of the technical solution of the present invention.

Claims (8)

1. A random parallel gradient descent adaptive optical wavefront correction method based on Sophia, characterized in that it includes the following steps: Initialize the parameters, and calculate the change in the signal performance index corresponding to the image sensor according to the random perturbation of the voltage control signal of the wavefront corrector in the iteration; The gradient is calculated from the change of signal performance index and random disturbance, and the Sopiha optimizer is used to improve the SPGD algorithm according to the gradient to perform wavefront correction; The first-order momentum and the second-order momentum are constructed in the improved SPGD algorithm. The weight parameters and gradient clipping are added to the SPGD algorithm with the first-order momentum and the second-order momentum constructed. The voltage control signal of the wavefront corrector is updated by the combination of the first-order momentum and the second-order momentum to achieve gain coefficient adaptation. 2. The Sophia-based random parallel gradient descent adaptive optical wavefront correction method according to claim 1 is characterized in that: the initial value of the SPGD algorithm is set, and the change in the signal performance index corresponding to the image sensor is calculated according to the random perturbation of the voltage control signal of the wavefront corrector in the iteration, specifically: Set the initial value of the SPGD algorithm, assume that the voltage control signal of the wavefront corrector is u(k) = {u ₁ ,u ₂ ,…,u _N }, where N represents the number of driver units of the wavefront corrector. At the kth iteration, a random disturbance {Δu} is generated, and the corresponding disturbances Δu and -Δu in two directions are applied to the wavefront corrector respectively; the controller collects the signal of the image sensor and calculates the performance indicators J(u+Δu) and J(u-Δu), and obtains the signal performance indicator change ΔJ(k): ΔJ(k)=J(u+Δu)-J(u-Δu). 3. The random parallel gradient descent adaptive optical wavefront correction method based on Sophia according to claim 2 is characterized in that: the gradient is calculated based on the signal performance index change and the random disturbance, specifically including: The first-order gradient g(k) is calculated using the product of the signal performance index change ΔJ(k) and the random disturbance Δu: g(k)＝ΔJ(k)·sign(Δu); Wherein, sign() is a sign function, which is used to obtain the positive or negative sign of a number; the sign of ΔJ(k)·sign(Δu) is determined by the optimization direction of the signal performance index. 4. The Sophia-based random parallel gradient descent adaptive optical wavefront correction method according to claim 3 is characterized in that: the first-order momentum and the second-order momentum are constructed in the improved SPGD algorithm, specifically: The calculation formula for first-order momentum is: m(k+1)=β ₁ ·m(k)+(1-β ₁ )·g(k); Where β ₁ is the set hyperparameter, m(k) represents the first-order momentum of the kth iteration, and m(k+1) represents the first-order momentum of the k+1th iteration; Calculate the second-order gradient using the formula: In the formula, represents the second-order gradient of k iterations; Construct the second-order momentum, and the calculation formula is: Where β ₂ is the set hyperparameter, h(k) represents the second-order momentum of the kth iteration, and h(k+1) represents the second-order momentum of the k+1th iteration. 5. The Sophia-based random parallel gradient descent adaptive optical wavefront correction method according to claim 4 is characterized in that: the weight parameter and gradient clipping are added to the SPGD algorithm that constructs the first-order momentum and the second-order momentum, and the voltage control signal of the wavefront corrector is updated by combining the first-order momentum and the second-order momentum to achieve gain coefficient adaptation, specifically: Add the weight parameter λ to achieve adaptive gain coefficient and improve convergence speed. At this time, the voltage control signal formula of the wavefront corrector is updated as follows: u(k)=u(k)+λ·α(k)·u(k); Where α(k) is the adaptive learning rate; The voltage control signal of the wavefront corrector is updated by combining the first-order momentum and the second-order momentum to achieve adaptive gain coefficient and improve convergence speed. The updated voltage control signal of the wavefront corrector is obtained as follows: u(k+1)＝u(k)+α(k)·clip(m/max(γ·h(k),ε),ρ(k)) Where u(k+1) represents the voltage signal of the k+1th iteration; ρ(k) is the adaptive clipping range parameter, T is the total number of iterations; clip(m/max(γ·h(k),ε),ρ(k)) is the clipping function, which is used to limit m/max(γ·h(k),ε) to the range of -ρ(k) to ρ(k); ε is a parameter to avoid the denominator being zero in the iteration. 6. According to the Sophia-based random parallel gradient descent adaptive optical wavefront correction method according to claim 1, it is characterized in that: the updated voltage control signal of the wavefront corrector is applied to the wavefront corrector through the amplifier to complete the current iteration; the above correction process is repeated until the end condition is met; the end condition is a certain number of iterations or the performance index reaches a fixed value. 7. The system used in the Sophia-based random parallel gradient descent adaptive optical wavefront correction method according to any one of claims 1 to 6, characterized in that it comprises: The gradient descent module is used to initialize parameters and calculate the change in the signal performance index corresponding to the image sensor according to the random perturbation of the voltage control signal of the wavefront corrector in the iteration; the gradient is calculated from the change in the signal performance index and the random perturbation, and the Sopiha optimizer is used to improve the SPGD algorithm according to the gradient to perform wavefront correction; The gain coefficient adaptive module is used to construct the first-order momentum and the second-order momentum in the improved SPGD algorithm; a weight parameter is added to the SPGD algorithm that constructs the first-order momentum and the second-order momentum, and the voltage control signal of the wavefront corrector is updated by combining the first-order momentum and the second-order momentum to achieve gain coefficient adaptation. 8. An electronic device, characterized in that it comprises: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the Sophia-based stochastic parallel gradient descent adaptive optical wavefront correction method is implemented as described in any one of claims 1 to 6.