Wavefront-free detection adaptive optics method suitable for variable imaging targets

Abstract

The application discloses a wavefront-free detection adaptive optical method suitable for variable imaging targets, and comprises the following steps: loading a phase screen by using a first spatial light modulator; collecting in-focus and out-of-focus images of collimated laser after the phase screen and then after an out-of-focus grating; pre-processing the in-focus and out-of-focus images collected by the camera; extracting normalized fine features and structural focusing features from the pre-processed in-focus and out-of-focus images; completing data set making of the variable imaging target; making a training set, a verification set and a test set; building an efficient deep neural network, with the input being feature images of the normalized fine features and feature images of the structural focusing features; based on the output of the deep neural network, restoring a near-field distortion wavefront, solving a conjugate phase corresponding to the near-field distortion wavefront, and loading the conjugate phase on a second spatial light modulator to realize phase correction of the distortion wavefront. According to the technical scheme, the method can be suitable for imaging scenes with random changes of targets.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of wavefront correction of extended targets and deep learning, and in particular to a wavefront detection adaptive optical method suitable for variable imaging targets. BACKGROUND

[0002] Image-based wavefront sensing is a technique for measuring wavefront errors, which uses far-field spot quality as an objective function and employs continuous iterative optimization to retrieve the wavefront phase. This method eliminates the need for specialized wavefront sensors, has the advantages of simple system, high optical efficiency, consistent control performance and image quality evaluation, and is currently widely used in research fields such as inertial confinement fusion, microscope imaging, human eye imaging, optical tracking and free-space laser communication. In recent years, deep learning has developed rapidly, and due to its inherent advantage of not requiring iteration or optimization, it has attracted the attention of scholars.

[0003] Researchers have shown that the most advanced convolutional neural network can not only estimate the wavefront represented by Zernike coefficients from point source intensity images, but also from intensity images of specified extended objects. In recent years, some scholars have used the MNIST (Modified National Institute of Standards and Technology, initiated by the United States National Institute of Standards and Technology) handwritten digit dataset as an imaging target, and then used a neural network to achieve wavefront restoration of the handwritten digit dataset. Some scholars have also used neural networks to achieve wavefront reconstruction of satellite images. However, the phase retrieval method based on a single image may have multiple solutions, and it is easy to stagnate during the iteration process, which is caused by the fact that the actual near-field complex amplitude f(x,y) and its rotated 180-degree conjugate complex amplitude f'(x,y) have the same light intensity distribution in the far field. Therefore, the algorithm is easy to converge to one of the local minimum pseudo-solution or two global minimum ambiguous solutions. In order to overcome the multiple solution problem in phase retrieval technology, Gonsalves and Chidlaw first proposed the basic idea of phase diversity (PD) method in 1979, which achieves the uniqueness of the solution of the phase retrieval method by adding a pair of collected information to increase the constraint condition. In addition to the multiple solution and stagnation phenomenon, most of the current methods are tailored for specific types of extended objects or a single extended object, which will lead to the fact that such methods cannot achieve high-precision wavefront reconstruction in different extended object scenarios. Such deep learning-based extended target wavefront restoration methods have limitations in generalization, especially for imaging targets that change at any time. SUMMARY

[0004] The technical problems solved by the present application are: in order to solve the problems of weak generalization of the existing wavefront recovery deep learning network for extended targets, and possible data set error caused by the traditional phase difference method, the present application proposes a wavefront detection adaptive optical system suitable for variable imaging targets, which realizes high-precision and efficient data acquisition, and realizes the wavefront recovery and correction of variable targets.

[0005] The technical scheme adopted by the present application to solve the above problems is: a wavefront detection adaptive optical method suitable for variable imaging targets, comprising:

[0006] Step 1: randomly generate 4-35 order Zernike coefficients on the computer side using a simulation program, make a predetermined number of phase screens, the surface topography of each phase screen conforms to the two-dimensional distribution of the Kolmogorov power spectrum, and use a first spatial light modulator to load the predetermined number of phase screens in turn to simulate the dynamic phase distortion of atmospheric turbulence of light, the first spatial light modulator is a phase type spatial light modulator;

[0007] Step 2: after the first spatial light modulator receives collimated laser with a center wavelength of 520nm, the collimated laser is phase modulated, the in-focus and out-of-focus images of the collimated laser after the phase screen of the first spatial light modulator and the defocus grating are collected by the camera, and the corresponding near-field wavefront data is recorded in sequence, the in-focus and out-of-focus images include in-focus images, positive out-of-focus images and negative out-of-focus images;

[0008] Step 3: preprocessing is performed on the in-focus and out-of-focus images collected by the camera, which includes: first, the slider matching method is used to realize registration of the positive and negative out-of-focus images, and then N2D-GAN is used to denoise the positive and negative out-of-focus images;

[0009] Step 4: the in-focus and out-of-focus images after preprocessing are subjected to normalization fine feature and structure focusing feature extraction to form feature images of normalized fine features and structure focusing features, and the feature images and corresponding wavefront aberration labels are used as samples, which are used to make a training set in step 6;

[0010] Step 5: use a projector to make random size and type of extended targets, and execute steps 3-4 for variable imaging targets to complete the data set making of variable imaging targets;

[0011] Step 6: Select the first predetermined number of samples of the point target obtained in steps 2-4 as a training set; randomly select the second predetermined number of samples in the data set of the variable target made by the projector in step 5 as a validation set, and the remaining third predetermined number of samples as a test set, wherein the training set is used for the deep neural network to train and learn the mapping relationship between the feature image and the corresponding near-field wave surface, and the validation set and the test set are used to measure the accuracy and generalization of the method;

[0012] Step 7: Configure the deep learning environment and build an efficient deep neural network, the input of which is the normalized fine feature feature image and the structural focus feature feature image, and the output is the 4-35 order Zernike coefficient. Compare the Zernike coefficient predicted by the deep neural network with the Zernike coefficient corresponding to the real distortion wavefront label, calculate the loss value using the L2 loss function to promote the parameter update of the deep neural network;

[0013] Step 8: Based on the output of the deep neural network, the near-field distortion wavefront is restored, and the conjugate phase corresponding to the obtained near-field distortion wavefront is solved, and the conjugate phase is loaded onto the second spatial light modulator for correction, so that the phase correction of the distortion wavefront is realized.

[0014] The beneficial effects of the present application compared with the prior art are:

[0015] The present application provides a detailed design scheme for the key component of the defocus grating, which simplifies the optical system and avoids the problems existing in the traditional phase difference acquisition method.

[0016] The present application proposes normalized fine features (NFF) and structural focus features (SFF) which are irrelevant to the imaging target but accurately correspond to the wavefront aberration. These two features provide a more accurate and robust representation of the wavefront aberration.

[0017] The present application deeply studies the image registration algorithm and the noise model, and uses the slider matching method and N2D-GAN to preprocess the collected far-field images to minimize the interference with the subsequent feature extraction method.

[0018] The present application uses a lightweight ADENet to realize efficient and high-precision mapping between features and wavefronts. Through testing, the single wavefront reconstruction inference time is less than 3ms. BRIEF DESCRIPTION OF DRAWINGS

[0019] Figure 1 A schematic process diagram of a wavefront detection adaptive optical method suitable for variable imaging targets is proposed for the present application;

[0020] Figure 2 A flowchart of the slider registration algorithm used in the embodiment of the present application;

[0021] Figure 3 The schematic diagram of the denoising neural network N2D-GAN architecture used for the embodiments of the present application;

[0022] Figure 4 The normalized fine feature example diagram, wherein (a) is the sharpness feature, (b) is the power feature, (c) is the feature obtained after multiplication operation of the sharpness and power features, (d) is the feature when the square root times is 1, (e) is the feature when the square root times is 4, and (f) is the feature when the square root times is 8;

[0023] Figure 5 The comparison of the normalized fine feature and the structure focusing feature extraction results under the noise condition;

[0024] Figure 6 The two kinds of feature extraction results after replacing different targets under the same atmospheric turbulence condition;

[0025] Figure 7 The schematic diagram of a wavefront-free detection adaptive optical system suitable for a variable imaging target proposed by the present application;

[0026] Figure 8 The wavefront correction example diagram of the method proposed by the present application, wherein (a) is the far-field image before correction, and (b) is the far-field image after correction. DETAILED DESCRIPTION

[0027] In order to make the purpose, technical scheme and advantages of the present application more clear and explicit, the present application is further described in detail below in combination with specific embodiments and with reference to the drawings.

[0028] Figure 7 The schematic diagram of a wavefront-free detection adaptive optical system suitable for a variable imaging target proposed by the present application. The adaptive optical system performs the adaptive optical method according to the embodiments of the present application. As shown in Figure 7 , the hardware facilities required for the method are as follows: one imaging camera, two spatial light modulators, one 520nm laser, one projector, one 50:50 beam splitter, one adjustable diaphragm, one 520nm filter, one lens with a focal length of 200mm, the center wavelength of the incident light is set to , the focal length , the entrance pupil radius , the defocus amount , the defocus grating diffraction transverse distance , the defocus grating component is designed and made according to the above parameters, and the adaptive optical system based on the phase difference method is built.

[0029] In the adaptive optical system based on the phase difference method, after the light beam enters the imaging system, it can pass through the diaphragm, the beam splitter, and then the defocus grating for light splitting. After the defocus grating, an imaging lens can be arranged in sequence, and the in-focus and out-of-focus images can be displayed on a camera at the same time, thereby avoiding the problems of multiple camera simultaneous frequency triggering and collection delay that may exist in the traditional phase difference collection method.

[0030] It is necessary to collect the far-field images on the in-focus and out-of-focus planes by using the defocus grating. According to the definition, the defocus amount refers to the defocus capability of the grating, which is equal to the additional optical path difference introduced by the ±1st order diffracted wavefront of the pupil edge, is the focal length of the off-axis Fresnel zone plate, is the radius of the defocus grating, is the off-axis displacement of the defocus grating pupil center relative to the center of the Fresnel zone plate, is the central wavelength of the incident light, is the grating period of the pupil center.

[0031] By using the central wavelength of the incident light, the focal length of the lens, the pupil radius , the defocus amount , and the lateral distance , the design parameters of the defocus grating can be intuitively and comprehensively established, and the focal length and the off-axis displacement of the defocus grating can be calculated in sequence. In addition, it is necessary to determine the phase step, step width and step height of the defocus grating. In an embodiment, considering the complexity of manufacturing, the step is set to two equal parts, and in order to achieve equal and as high as possible efficiency of the ±1st order diffraction efficiency, a binary phase type defocus grating is adopted, and the phase step height is set to 0.96π.

[0032] Figure 1 A schematic process diagram of a wavefront-free detection adaptive optical method suitable for variable imaging targets is provided. As shown in Figure 1 , the wavefront-free detection adaptive optical method suitable for variable imaging targets comprises the following steps:

[0033] Step 1: Randomly generate 4-35 order Zernike coefficients on a computer using a simulation program, make a predetermined number of phase screens, the surface topography of each phase screen conforms to the two-dimensional distribution of the Kolmogorov power spectrum, and use the first spatial light modulator to load the predetermined number of phase screens in sequence to simulate the dynamic phase distortion of atmospheric turbulence. The first spatial light modulator is a phase-type spatial light modulator. As described above, the hardware facilities for performing the method include two spatial light modulators, the first spatial light modulator and the second spatial light modulator are both phase-type spatial light modulators. The predetermined number can be 20000.

[0034] Step 2: After the first spatial light modulator receives collimated laser with a center wavelength of 520 nm, it performs phase modulation on the collimated laser. The in-focus and out-of-focus images of the collimated laser after passing through the phase screen of the first spatial light modulator and then passing through the defocus grating are collected by the camera, and the corresponding near-field wavefront data is recorded in sequence. The collimated laser is a point target. The in-focus and out-of-focus images are the images of the laser emitted by the projector after phase modulation by the first spatial light modulator, impinging on the phase screen and then imaging by the defocus grating. The in-focus and out-of-focus images are also called far-field pictures, which include in-focus images, positive out-of-focus images and negative out-of-focus images.

[0035] The reason for fixing the imaging target as a point target for network training is that the point target can be more easily controlled and generated compared to an extended target. Therefore, in order to reduce the complexity and time cost of data set creation, a point target is used to make a data set for training in the network training stage.

[0036] Step 3: Preprocess the in-focus and out-of-focus images collected by the camera, which includes first performing slide block matching to register the positive and negative out-of-focus images, and then using N2D-GAN to denoise the positive and negative out-of-focus images. This preprocessing can reduce the impact on subsequent feature extraction.

[0037] The reason for performing image registration and denoising on the in-focus and out-of-focus images is that: (1) After obtaining the in-focus and out-of-focus images generated by the defocus grating, it is crucial to accurately segment the point images of the positive and negative defocus images. Once there is some deviation in positioning, it is inevitable to introduce some tilt errors in the data set, leading to differences between the feature images and the labels, thereby affecting the wavefront reconstruction. (2) In practical applications, feature extraction is obtained by operating on real images, and noise has a great influence on feature images. In order to establish an accurate and stable mapping relationship between the feature images and the wavefront, accurate simulation and effective noise removal are the key to image preprocessing.

[0038] Step 4: Normalized fine feature (NFF) and structure focus feature (SFF) are extracted from the pre-processed in-focus and out-of-focus images to form the feature images of NFF and SFF, and the feature images and corresponding wavefront aberration labels are used as samples to make the training set in step 6. The size of the two feature images can be 340 340. The corresponding wavefront aberration label is the 4-35 order Zernike coefficient.

[0039] In this step, the normalized fine feature (NFF) which fuses the sharpness and power information at the same time, and the structure focus feature (SFF) which enhances the anti-noise performance are introduced. The combination of the two features can provide more accurate and robust representation of wavefront aberration, so that target-independent wavefront detection can be realized.

[0040] Step 5: A variable imaging target of random size and type is made by a projector, and steps 3-4 are performed on the variable imaging target to complete the data set making of the variable imaging target. The variable imaging target is randomly generated by the projector and can be any picture, and is variable, not fixed. In one example, the extended target can be 3000 groups. The means of using the projector to make the extended target of random size and random type can be: connecting the projector to the computer, and then playing the selected picture to realize the output of the variable imaging target. In order to keep the center wavelength consistent with the point target, a filter is placed in the optical path to filter out light in the remaining wavelength range.

[0041] Step 6: The first predetermined number of samples of the point target obtained in steps 2-4 are selected as the training set; the second predetermined number of samples in the data set of the variable target made by the projector in step 5 are randomly selected as the validation set, and the remaining third predetermined number of samples are selected as the test set. The training set is used for network training and learning the mapping relationship between the feature images and the corresponding near-field distortion wavefront, and the validation set and the test set are used to measure the accuracy and generalization of the method. The first predetermined number can be 20000, the second predetermined number can be 2000, and the third predetermined number can be 1000.

[0042] Step 7: Configure the deep learning environment and build an efficient deep neural network. The input of the deep neural network is the feature images of the normalized fine feature and the structure focus feature, and the output is the 4-35 order Zernike coefficient. The predicted Zernike coefficient of the deep neural network is compared with the Zernike coefficient label corresponding to the real distortion wavefront, and the L2 loss function is used to calculate the loss value to promote the parameter update of the deep neural network. The deep neural network uses Adam optimizer, and the initial learning rate is set to 0.0001.

[0043] The input of the network is to merge two feature images in the channel, and then send the merged feature image into the network for training. The loss value is calculated by comparing the Zernike coefficients output by the neural network with the Zernike coefficient labels corresponding to the real distortion wavefront. The loss function adopts L2 loss function. The means of obtaining the Zernike coefficient labels corresponding to the real distortion wavefront belongs to the technical content familiar to those skilled in the art, which will not be described herein. The parameters updated by the deep neural network include the weight coefficients corresponding to the nodes of the neural network.

[0044] Step 8: Based on the output of the deep neural network, the near-field distortion wavefront is restored, and the conjugate phase corresponding to the obtained near-field distortion wavefront is solved, and the conjugate phase is loaded on the second spatial light modulator for correction, so as to realize the phase correction of the distortion wavefront. The way of restoring the near-field distortion wavefront through the output of the deep neural network is well known to those skilled in the art, and will not be described herein.

[0045] During the training process, the loss value is continuously reduced, and when the loss is almost unchanged, it represents that the network training is mature. Only a set of fine features and structure focusing feature images are input into the network, and the network can output the near-field wavefront information corresponding to the sample, and then the wavefront correction is realized. The training duration is about 2 hours, and the inference time is about 2.9 ms. Tests on various imaging targets (symbols, numbers and letters) have achieved high accuracy.

[0046] A specific implementation of step 3 will be described in detail below. Step 3: The in-focus and out-of-focus images collected by the camera are preprocessed, which includes: first, the slide block matching method is used to realize the registration of the positive and negative out-of-focus images, and then the N2D-GAN is used to denoise the positive and negative out-of-focus images.

[0047] The specific implementation of step 3 will be described in detail below. Step 3: The in-focus and out-of-focus images collected by the camera are preprocessed, which includes: first, the slide block matching method is used to realize the registration of the positive and negative out-of-focus images, and then the N2D-GAN is used to denoise the positive and negative out-of-focus images. Figure 2 The flowchart of the slide block traversal method mentioned in step 3 is described below. The core of this algorithm is to use the positive and negative out-of-focus spot images without aberration to realize the positioning of the aberration spot image, and only one positioning is needed to realize the registration of all images. First, about step one, without introducing additional aberration, define an N A sliding window of size N (N = 340 in an example) that matches the pixel size and resolution of the actual CCD camera exactly. Then a circular mask matrix is defined, where the area inside the circle is set to 1 and outside to 0. The radius of the circular mask can be determined based on the number of pixels in the simulation or actual imaging. Then the window is traversed over the left and right half of the in-focus and out-of-focus images respectively. Step two, the pixel values within the window are summed and the results are recorded and compared in turn. The maximum light intensity sum and its corresponding sliding window coordinates are recorded each time the window is moved. After traversing the images, the positions of the maximum sums in the left and right half of the in-focus and out-of-focus images correspond to the positions of the positive and negative out-of-focus plane far field spots to be located respectively, thus completing the registration of the positive and negative out-of-focus images. Step three, recording the coordinates of the two sliding windows can achieve image registration in other imaging targets and different wavefront aberration scenarios.

[0048] The following describes an embodiment of step 3. Figure 3 The denoising process mentioned in step 3 is described. Figure 3 The denoising neural network N2D-GAN (N2D-Generative Adversarial Network) architecture used in the embodiment of the present application is shown in the figure. N2D-GAN includes a noise estimation subnetwork and a non-blind denoising subnetwork, which can achieve high-precision blind denoising of real images. In order to improve the robustness and practicality of the deep denoising model, a new denoising method for real images is proposed. As shown in the figure, the main idea of the network is to input the noisy image (x) into the generator to generate a denoised image (x'), and then use the denoised image and the clean, noise-free images in the training set to train the discriminator. This process continues until the discriminator can hardly distinguish between real and fake images, indicating that the training is mature. Figure 3

[0049] An embodiment of step 4 is described below. Step 4: Extract the normalized fine feature and structure focusing feature from the pre-processed in-focus and out-of-focus images to form a feature image of the normalized fine feature and a feature image of the structure focusing feature, and use the feature images and the corresponding wavefront aberration labels as samples to make a training set in step 7, where the size of the two feature images is 340 340, and the corresponding wavefront aberration label is the 4-35 order Zernike coefficient.

[0050] In step 4, the normalized fine feature combines the power feature and the sharpness feature, where the power feature matrix and the sharpness feature matrix can be calculated according to the frequency spectrum of the positive and negative out-of-focus images:

[0051] ​​​ ,

[0052] ,

[0053] ,

[0054] ,

[0055] wherein, and are the spectrum of the positive and negative defocus images, respectively, and are the complex conjugate functions of the spectrum of the positive and negative defocus images, is the point spread function of the adaptive optics system, is the point spread function corresponding to the negative defocus image, is the point spread function corresponding to the positive defocus image, is the Fourier transform operation, is the conjugate operation. Figure 4 are the normalized fine features, wherein (a) is the sharpness feature, (b) is the power feature, (c) is the feature obtained by multiplying the sharpness and power features, (d) is the feature with the square root number of 1, (e) is the feature with the square root number of 4, and (f) is the feature with the square root number of 8. As shown in FIG. 4, the power feature and the sharpness feature can be captured simultaneously by the multiplication operation. Therefore, the multiplication operation is used first to enhance the effective information contained in the features. Figure 4

[0056] In addition, the square root operation can further enhance the details of the blur, as shown in FIG. 5. As can be seen from the box regions in (d) and (e) in FIG. 5, the feature image contains more details after the fourth root operation, and the impact of local blur can be reduced to some extent. However, using an insufficient or excessive number of square root operations can lead to uneven distribution of feature values, as shown in (f) in FIG. 5. In this case, a large part of the data can be concentrated in a small range, which can cause the neural network to ignore some key features, resulting in information loss. In order to further select a suitable square root number, we statistically analyzed the average variance values of 1000 different square root numbers. The results show that the variance of the features obtained after the fourth root operation is the highest, indicating that the feature distribution is the most uniform at this time. Therefore, the normalized fine features are defined as: Figure 4 Figure 4 Figure 4

[0057] ,

[0058] ​​​​However, in the presence of signal and noise, their frequency domain representation usually has some overlap. Within these overlapping frequency bands, noise will mask or affect the signal, making it more difficult to accurately distinguish or identify the signal part in these frequency bands. However, in the time domain, the signal tends to be concentrated in the main structure of the image, while the noise is more dispersed. When the frequency domain features are inverse transformed to the time domain, the image structure and the main part of the signal are usually more concentrated, while the noise is relatively dispersed, so the effect of noise on the time domain features will be relatively small. The above-mentioned frequency domain normalized fine features are converted to time domain features:

[0059] ,

[0060] wherein denotes the inverse Fourier transform. Figure 5 The results of the proposed frequency domain features and the corresponding time domain features in the presence of noise are shown, wherein SSIM (Structural Similarity) structural similarity is an index for measuring the similarity of two images, and the higher the index, the more similar the two images. We can see that they are more sensitive to noise in the frequency domain, and relatively less sensitive to noise in the time domain. Figure 5 The frequency domain and time domain feature results of two groups of different wavefront aberrations under noiseless and noisy conditions are shown.

[0061] Therefore, the time domain features obtained by introducing the inverse Fourier transform based on the frequency domain features are considered, and the two features are combined to enhance the robustness of the method to noise. Since the proposed frequency domain features combine power features and sharpness features, they enrich the feature details, and the data information is distributed between 0 and 1, so the frequency domain features are defined as normalized fine features (NFF). The time domain features focus on the overall structure and main features of the image, and the effect of noise is weakened, so they are defined as structure-focused features (SFF). Both of these features can eliminate the imaging target in the calculation process. As shown in Figure 6 the feature extraction results before and after changing the imaging target are completely consistent under the same atmospheric turbulence.

[0062] In step 7, the deep neural network used is: Attention-Driven Efficiency Network (ADENet), Figure 1The architecture of the network is shown, where BatchNormalization (BN) represents the batch normalization operation, and leakyrelu represents the activation function. ADENet is an advanced neural network architecture based on the CBAM (Convolutional Block Attention Module) attention mechanism. The CBAM attention mechanism endows ADENet with excellent adaptability, which can intelligently perceive and focus on the features in the image, making the network pay more attention to the feature regions that are effective for the task. At the same time, ADENet adopts three efficient modules, which are carefully designed to maintain high efficiency while ensuring the depth and complexity of the network, making it better to realize the nonlinear fitting of features and near-field wavefronts.

[0063] After the correction of the imaging target, the wavefront correction effect is evaluated using NIQE (Natural Image Quality Evaluator). NIQE is an index for evaluating image quality, which is a no-reference (or blind) evaluation index, that is, it does not need to refer to the original high-quality image when evaluating image quality. Figure 8 Part of the correction results of the test set are shown, and the lower NIQE value means better image quality and more natural images, where (a) is the far-field image without wavefront correction, and (b) is the far-field image after wavefront correction. From Figure 8 It can be seen from the images before and after correction in the

[0064] In the specification provided herein, a large number of specific details are described. However, it can be understood that the embodiments of the present application can be practiced without these specific details. In some examples, well-known methods, structures and techniques are not shown in detail in order not to obscure the understanding of the present specification.

[0065] Although the present application is described in terms of a limited number of embodiments, those skilled in the art, with the benefit of the above description, will appreciate that other embodiments can be conceived within the scope of the application described herein. In addition, it should be noted that the language used in the present specification is mainly selected for readability and instructional purposes, and is not selected to explain or limit the subject matter of the present application.

Claims

1. A wavefront-free probing adaptive optics method suitable for variable imaging targets, characterized in that, The method comprises the following steps: Step 1: randomly generate 4-35 order Zernike coefficients on a computer using a simulation program, make a predetermined number of phase screens, the surface topography of each phase screen conforms to the two-dimensional distribution of the Kolmogorov power spectrum, and use a first spatial light modulator to sequentially load the predetermined number of phase screens to simulate the dynamic phase distortion of atmospheric turbulence, the first spatial light modulator is a phase-type spatial light modulator; Step 2: after the first spatial light modulator receives collimated laser with a center wavelength of 520nm, the collimated laser is phase modulated, the in-focus and out-of-focus images of the collimated laser after passing through the phase screen of the first spatial light modulator and then passing through the defocus grating are collected by a camera, and the corresponding near-field wavefront data is recorded, the in-focus and out-of-focus images include in-focus images, positive out-of-focus images and negative out-of-focus images; Step 3: preprocessing is performed on the in-focus and out-of-focus images collected by the camera, which includes: first, the slider matching method is used to realize registration of the positive and negative out-of-focus images, and then the N2D-GAN is used to denoise the positive and negative out-of-focus images; Step 4: after the preprocessed in-focus and out-of-focus images, the normalized fine features and structural focusing features are extracted to form feature images of the normalized fine features and feature images of the structural focusing features, and the feature images and corresponding wavefront aberration labels are used as samples, which are used to make a training set in step 6; Step 5: a projector is used to make variable imaging targets with random size and type, and steps 3-4 are performed on the variable imaging targets to complete the data set making of the variable imaging targets; Step 6: the first predetermined number of samples of the point targets obtained in steps 2-4 are selected as a training set; a random second predetermined number of samples in the data set of the variable imaging targets made by the projector in step 5 are selected as a validation set, and the remaining third predetermined number of samples are selected as a test set, wherein the training set is used for deep neural network training and learning the mapping relationship between the feature images and the corresponding near-field wavefronts, and the validation set and the test set are used to measure the accuracy and generalization of the method; Step 7: configure a deep learning environment and build an efficient deep neural network, the input of the deep neural network is the feature images of the normalized fine features and the feature images of the structural focusing features, and the output is the 4-35 order Zernike coefficient, compare the Zernike coefficient predicted by the deep neural network with the Zernike coefficient label corresponding to the real distortion wavefront, calculate the loss value using the L2 loss function to promote the parameter update of the deep neural network; Step 8: based on the output of the deep neural network, the near-field distortion wavefront is restored, and the conjugate phase corresponding to the near-field distortion wavefront is solved, and the conjugate phase is loaded into the second spatial light modulator for correction, so that the phase correction of the distortion wavefront is realized.

2. The wavefront-free detection adaptive optics method suitable for variable imaging targets according to claim 1, characterized in that, The slider matching method of step 3 for realizing registration of the positive and negative out-of-focus images comprises: First, without introducing additional aberrations, define a sliding window of size N*N, N is any natural number, the window is completely matched with the actual camera pixel size and resolution; Then a circular mask matrix is defined, in which the area inside the circle is set to 1 and the area outside the circle is set to 0, and the radius of the circular mask is determined based on the number of pixels in the simulation or actual imaging; Then the window is traversed on the left half and the right half of the in-focus and out-of-focus images respectively, the pixel values in the window are summed, and the results are recorded and compared in turn, and the current maximum light intensity sum and the coordinates of the sliding window are recorded once the window moves once, after traversing the image, the positions of the maximum sums of the left half and the right half of the in-focus and out-of-focus images correspond to the positions of the far-field light spots of the positive and negative defocus planes to be positioned respectively, so that the registration of the positive and negative defocus images is completed.

3. The wavefront probe free adaptive optics method suitable for variable imaging targets according to claim 1, characterized in that, The denoising of the positive and negative defocus images by the N2D-GAN in step 3 comprises: The N2D-GAN comprises a noise estimation subnetwork and a non-blind denoising subnetwork, the noise image is input into the generator to generate a denoised image, then the denoised image and the clean and noise-free image in the training set are used to train the discriminator, and the training process is performed until the discriminator cannot distinguish between real and fake images.

4. The wavefront probe free adaptive optics method suitable for variable imaging targets according to claim 1, characterized in that, In step 4, the normalized fine features combine both the power features and the sharpness features, where the power feature matrix M P and the sharpness feature matrix M S are computed from the spectral images of the positive and negative defocus images, respectively. where I d+ (u,v) and I d- (u,v) are the spectra of the positive and negative defocus images, respectively, and are the complex conjugate functions of the spectra of the positive and negative defocus images, psf is the point spread function of the adaptive optical system, psf d- is the point spread function corresponding to the negative defocus image, psf d+ is the point spread function corresponding to the positive defocus image, is the Fourier transform operation, and * is the conjugate operation. The normalized fine feature is defined as: The frequency domain normalized fine feature is converted into a time domain feature to obtain a structure focusing feature. wherein denotes the inverse Fourier transform.

5. The wavefront probe free adaptive optics method suitable for variable imaging targets according to claim 1, characterized in that, In step 7, the deep neural network used is an efficient network ADENet based on an attention mechanism.

Images