Wavefront distortion correction method based on residual attention network

Abstract

The present invention relates to the technical field of communication systems, and in particular to a wavefront distortion correction method based on a residual attention network. The present invention adopts a residual network as a backbone network, constructs a multi-scale residual hybrid attention network structure, uses a convolution operation to convert a light intensity image into a feature map for backward propagation; extracts features in a distributed manner through convolution kernels of different scales, and uses an attention mechanism to improve the network's recognition rate of damaged light spot features; finally, a network loss function is designed to obtain a zernike coefficient that conforms to the actual wavefront aberration. The present invention can quickly and accurately reconstruct the turbulent phase through a residual attention network, thereby effectively correcting the distortion; through the setting of a hybrid attention network, the accuracy can be greatly improved with less increase in time complexity. The method of the present invention has the characteristics of high precision, good real-time performance and strong flexibility, which provides a guarantee for the practical application of deep learning in AO systems.

Description

Technical Field

[0001] The present invention relates to the technical field of communication systems, and in particular to a wavefront distortion correction method based on a residual attention network. Background Art

[0002] Light beams carrying orbital angular momentum (OAM) have become the research object of many scholars due to their unique optical properties. At present, Tbit / s-level data transmission has been achieved by multiplexing OAM light beams under laboratory conditions. However, in free space optical communications (FSO) systems, due to the change in atmospheric refractive index caused by atmospheric turbulence, the laser beam transmitted in the atmosphere has random amplitude fluctuations and phase distortion, which leads to the performance degradation of wireless optical communication systems. Adaptive optics (AO) is a technology that can improve the performance of communication systems by correcting the dynamic wavefront distortion of light beams. AO has become a research hotspot in wireless optical communication systems.

[0003] AO systems generally include wavefront sensors, wavefront controllers, and wavefront correctors. Depending on whether the AO system uses a wavefront sensor, it can be divided into conventional AO and AO without wavefront detection. Conventional AO systems have been widely used in astronomical observations and biomedical fields. However, in wireless optical communications with strong turbulence and in long-distance atmospheric laser transmission applications, the application of conventional adaptive optical control technology based on wavefront sensors is limited. AO systems without wavefront sensors can directly use the detected light intensity information to design control algorithms and generate the control signals required by the wavefront corrector, that is, to reconstruct the wavefront based on the distorted light intensity image.

[0004] Classic wavefront reconstruction methods without wavefront detection include Gerchberg-Saxton phase recovery algorithm, random parallel gradient descent algorithm, simulated annealing algorithm and genetic algorithm, etc. Most of these algorithms need to be solved by iterative calculation, and it is difficult to achieve real-time wavefront reconstruction. With the rapid development of deep learning (DL), wavefront detection without wavefront detection based on deep learning uses the light intensity image captured by the CCD camera as the input of the neural network, and the wavefront aberration or zernike coefficient as the output, and then converts the output into a control signal, and finally controls the deformable mirror to achieve wavefront correction. Paine et al. (Paine SW, Fienup JR. Machine learning for improved image-based wavefront sensing. Opt Lett. 2018 Mar 15; 43 (6): 1235-1238.) applied the deep learning method of point spread function to achieve wavefront reconstruction, and obtained a lower error than random reconstruction, but the real-time performance for larger images was poor. Tian et al. (Tian Q, Lu C, Liu B, et al. DNN-based aberration correction in a wavefront sensorless adaptive optics system. Opt Express. 2019Apr 15; 27(8): 10765-10776) proposed a model based on deep neural network to solve the problem of too many iterations of existing search algorithms, so as to obtain the Zernike coefficient for wavefront correction and effectively reduce the time cost. These studies have shown the effectiveness of deep neural network models. In order to reduce the computational complexity, Nishizaki et al. (Yohei Nishizaki, Matias Valdivia, Ryoichi Horisaki, et al. Deep learning wavefront sensing [J], Opt. Express, 2019, 27, 240-251) used a convolutional neural network model to estimate the wavefront Zernike coefficient directly from a single light intensity image, and experimentally verified the pre-processing methods such as overexposure, defocus and scattering, thereby improving the network estimation performance.Zhai et al. (Yanwang Zhai, Shiyao Fu, Jianqiang Zhang, et al., Turbulence aberration correction for vector vortex beams using deep neural networks on experimental data[J], Opt. Express 2020, 28, 7515-7527) also used a convolutional neural network model to learn the mapping relationship between the distorted vortex light intensity distribution and the first 20 Zernike coefficients, and shortened the calculation time. Inspired by the phase difference method, Ma et al. (Ma Huimin, Lin Haiqiu, Qiao Yan, et al. Numerical study of adaptive optics compensation based on convolutional neural networks[J]. Optics Communications, 2019, 433: 283-289) proposed to use the intensity maps of the focal plane and the defocused plane as the neural network input, and use their corresponding Zernike coefficients as labels to output the wavefront Zernike coefficients. Ma et al. (Ma Huimin, Jiao Jun, Qiao Yan, et al. A wavefront restoration method based on deep learning of light intensity images [J]. Progress in Laser and Optoelectronics, 2020, 57 (08): 255-264) also explored the impact of the consistency of training data on the wavefront restoration performance. The simulation results show that the model using a single training data set is not robust. Wu et al. (Wu Y, Guo Y, Bao H, et al. Sub-Millisecond Phase Retrieval for Phase-Diversity Wavefront Sensor. Sensors (Basel). 2020 Aug 28; 20 (17): 4877) also collected the real phase generated by the spatial light modulator based on the idea of ​​phase difference, and used a convolutional neural network to establish a mapping between the real image and its zernike coefficient. In addition, due to the small number of model parameters, the real-time performance of the model was significantly improved.Zhang et al. (ZhangYang, He Yulong, Ning Yu, et al. Deep learning method of far-field spot inversion wavefront phase[J]. Infrared and Laser Engineering, 2021, 50(08): 278-287) verified the feasibility of applying deep learning methods to reconstruct wavefronts, using a residual network to take the far-field intensity map as input and the wavefront zernike coefficient as output, and verified the strong robustness of the deep residual network.

[0005] The above literature shows that compared with traditional wavefront detection algorithms, deep convolutional neural networks have shown superiority in non-wavefront detection AO systems, but there are still problems with insufficient perception and expression of light spot features. As the network continues to deepen, gradient disappearance, explosion, and even overfitting are prone to occur, making it difficult to complete real-time and accurate reconstruction tasks. Excessive pooling operations will also lead to the loss of small target information, resulting in excessive reconstruction errors. Summary of the invention

[0006] The purpose of the present invention is to provide a wavefront distortion correction method based on a residual attention network, which is used to solve the problems of insufficient perception and expression of light spot features in the existing AO system technology. As the network continues to deepen, gradient disappearance, explosion and even overfitting are prone to occur, making it difficult to complete real-time and accurate reconstruction tasks. Excessive pooling operations will also lead to the loss of small target information, resulting in excessive reconstruction errors.

[0007] In order to achieve the above object, the present invention adopts the following technical solutions:

[0008] The wavefront distortion correction method based on residual attention network includes the following steps:

[0009] The residual network is used as the backbone network, and a hybrid attention structure is constructed based on the residual network. The convolution operation is used to convert the light intensity image into a feature map for backward propagation.

[0010] Distributed feature extraction is performed through convolution kernels of different scales, and the attention mechanism is used to improve the network's recognition rate of damaged light spot features and enhance the network's ability to express light intensity image features;

[0011] Design the network loss function to obtain the Zernike coefficient that conforms to the actual wavefront aberration.

[0012] Furthermore, the residual network is a ResNet50 model, which first converts the input light intensity map into a feature image through a 7×7 downsampling convolution operation, and then uses a maximum pooling operation with a filter size of 3×3 to reduce the parameters of the feature map and reduce its dimension to 1 / 4 of the size of the initial image.

[0013] Furthermore, the hybrid attention structure includes a spatial attention mechanism and a dynamic selection mechanism network, uses a convolution operation to extract feature mapping information, and adds a residual skip connection structure to allocate computing resources to the model according to the importance of the input features.

[0014] Furthermore, the dynamic selection mechanism network includes three parts: decomposition, fusion and selection.

[0015] Further, in the decomposition part:

[0016] A two-stream network convolution structure of 3×3 and 5×5 is selected, expressed as:

[0017]

[0018]

[0019] where X∈R h×w×C is the input feature map, f represents the convolution operation; the 5×5 convolution uses the hole convolution operation;

[0020] In the fusion section:

[0021] The feature map is added by the summation operation, and then the information carried by each channel of the convolution layer is statistically analyzed by the global average pooling to obtain s. Finally, the feature dimension reduction is performed to obtain z, which is expressed as:

[0022]

[0023]

[0024] z=F fc (s) = δ(B(W s ))

[0025] where F gp is the global pooling operation, F fc is a fully connected operation, δ is a linear rectification function, B is a BatchNormalization layer, and W s ∈R d×C , d = max(C / r,L) represents the feature dimension after full connection, r is the compression factor, and L represents the minimum value of d;

[0026] In the Selection section:

[0027] The softmax function is used to calculate the weight distribution of each channel, expressed as:

[0028]

[0029]

[0030] Where a and b are the soft attention weight matrices of the two channels respectively, and V is the output matrix of SKNet.

[0031] Furthermore, the spatial attention mechanism is used to weight the features of the target area, find the connection between the features in the network, and make the feature extraction network selectively focus on the target area containing important information; it is expressed as:

[0032]

[0033]

[0034] Where c represents the concatenation operation, σ is the sigmoid activation function, and F avg With F max Represent the average pooling operation and the maximum pooling operation performed along the channel axis respectively.

[0035] Furthermore, the network loss function is designed, mainly including:

[0036] The expressions of wavefront peak-to-valley value and wavefront root mean square value in the AO system evaluation index are:

[0037] PV=max(Δφ(ρ,θ))-min(Δφ(ρ,θ))

[0038]

[0039] where Δφ is the wavefront aberration, is its mean, ρ and θ are the polar coordinates of the pupil plane;

[0040] Combining the two, the MSE loss function is generated, which is expressed as:

[0041]

[0042] L=λ MSE L MSE +λ PV L PV +λ RMS L RMS

[0043] Where z act and z preare the actual and estimated zernike coefficients respectively, with a length of N, λ being the weight of each loss, and L being the weighted sum of each loss function. The reduction of the loss function represents the improvement of the reconstruction accuracy.

[0044] The present invention has at least the following beneficial effects:

[0045] The present invention adopts a residual network as the backbone network, and on this basis constructs a multi-scale residual hybrid attention network structure, uses a convolution operation to convert the light intensity image into a feature map for backward propagation; secondly, the features are distributedly extracted through convolution kernels of different scales, and the attention mechanism is used to improve the network's recognition rate of the damaged light spot features, so as to enhance the network's ability to express the light intensity image features; finally, a network loss function combined with realistic evaluation indicators is designed to obtain the zernike coefficient that conforms to the actual wavefront aberration. The present invention can quickly and accurately reconstruct the turbulent phase through the residual attention network, thereby effectively correcting the distortion; through the setting of the hybrid attention network, the accuracy can be greatly improved with less increase in time complexity. The method of the present invention has the characteristics of high precision, good real-time performance and strong flexibility, which provides a guarantee for the practical application of deep learning in AO systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying any creative work.

[0047] Figure 1 This is the schematic diagram of the AO system without wavefront detection;

[0048] Figure 2 This is the overall structure diagram of the network;

[0049] Figure 3 This is a diagram of the mixed attention structure. DETAILED DESCRIPTION

[0050] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0051] 1. Vortex beam atmospheric transmission model

[0052] See also Figure 1 , Figure 1 This is the schematic diagram of the AO system without wavefront detection.

[0053] In deep learning-based AO systems without wavefront detection, neural network models are usually used instead of traditional correction algorithms to invert wavefront aberrations or their zernike coefficients from distorted light intensity images. Figure 1 In the process, the spatial light modulator (SpatialLight Modulator) adds a spiral phase to the Gaussian beam to turn it into a vortex beam, which is distorted when passing through atmospheric turbulence. At this time, the distorted wavefront is divided into two parts by the polarization beam splitter (Polarization Beam Splitter). One part enters the wavefront correction, and after the CCD camera captures the light intensity information, it is sent to the PC end equipped with the neural network model (Neural Network). The wavefront coefficient is reconstructed according to the wavefront distortion information, thereby generating an output drive voltage to control the deformable mirror (Deformable Mirror) to compensate the wavefront, and the other part is used for information transmission.

[0054] In free-space optical communications, vortex beams have continuous spiral wavefronts compared to ordinary beams, and the phase in the direction of beam propagation is uncertain, which is the phase singularity. Among vortex beams, the Laguerre-Gaussian (LG) beam is a representative beam, and its optical field expression in free space is:

[0055]

[0056] Where p represents the radial index of the LG beam; r is the radial distance from the spatial point to the transmission axis; is the azimuth; w 0 is the beam waist radius; is the Laguerre polynomial, l is the topological charge of the vortex beam, which can be any integer; k = 2π / λ is the wave number; z 0 =kw 0 / λ is the Rayleigh distance.

[0057] The present invention assumes that the light beam emission point is located at z = 0, and for the convenience of research without affecting the results, p = 0, then the expression of the emitted light beam is:

[0058]

[0059] According to the above formula, the transmission process of LG beam in atmospheric turbulence can be simulated. When the light wave passes through multiple phase screens, only the phase changes while the amplitude remains unchanged. Based on the Fresnel formula, the light field expression when the transmission distance is Δz can be obtained as:

[0060]

[0061] where F -1 and F represent inverse Fourier transform and Fourier transform respectively; φ(r,zi ) represents the atmospheric turbulence phase screen function; k x and k y denote the spatial wavenumbers in the x and y directions, respectively, and

[0062] Therefore, it is only necessary to simulate the phase screen model of atmospheric turbulence to calculate the size of the light field at the receiving end. There are two commonly used atmospheric turbulence phase screen simulations. The first is the power spectrum inversion method, which uses the power spectrum density function to filter the complex Gaussian random matrix, and then obtains the distortion phase of atmospheric turbulence through inverse Fourier transform; the second is the Zernike polynomial method, which uses orthogonal Zernike polynomial expansion as the basis function to simulate the wavefront, and each item has a clear physical meaning. By analyzing the size of the coefficient, the size of various aberration components can be known. The present invention uses the Zernike polynomial defined by Noll to simulate the turbulence phase screen φ(r), which can be expressed by the following formula:

[0063]

[0064]

[0065]

[0066] Where r and θ are polar coordinates, n and m are radial and angular orders respectively, which are always integers and satisfy the following conditions: m≤n, n-|m|=even. The first term of the Zernike polynomial represents translation, which has no effect on the image quality. A higher-order polynomial can represent higher frequency components. Therefore, the present invention does not consider the first term, and selects a 36-order polynomial to generate an atmospheric turbulence phase screen that conforms to the actual situation.

[0067] According to the Noll matrix, we can deduce that any a in the Zernike polynomial i and a j The covariance matrix of is as follows:

[0068]

[0069] in D is the aperture of the optical system, r 0 is the atmospheric coherence length, δ mm' is the Kronecker function, and Γ is the gamma function. After calculating the covariance matrix C, we can perform singular value decomposition on it to obtain:

[0070] C=UVU T (8)

[0071] Where V is a diagonal matrix and U is a unitary matrix. TTo calculate the vector b, as shown in the following formula, the covariance of b is a diagonal matrix, and the vector b conforms to the Gaussian random distribution:

[0072] b=U T ·a (9)

[0073] Taking it as the coefficient of the expansion of the Zernike polynomial, the turbulent phase can be expressed as follows:

[0074]

[0075] According to the above formula, the coefficients before the Zernike polynomial can be determined, and the atmospheric turbulence phase screen that conforms to the Kolmogorov turbulence theory can be simulated, and the influence of atmospheric turbulence on the light field transmission can be studied. The spiral spectrum is usually used to evaluate the dispersion of the vortex beam. The complex amplitude of the vortex beam is in the harmonic function. Zhang Cheng's linear space is expanded as follows:

[0076]

[0077] where a l (r,z) is the amplitude of the lth order harmonic. l By integrating, we can get the lth order harmonic energy C l ,Right now

[0078]

[0079] The total energy U of the vortex beam can be expressed as:

[0080]

[0081] where ε 0 is the vacuum dielectric constant. The spiral spectrum is:

[0082]

[0083] The spiral spectrum, also called relative power or l-th-order mode purity, can clearly indicate the energy contribution of each OAM mode of the vortex beam.

[0084] 2. Residual Attention Network Model:

[0085] 2.1 Residual Network Model

[0086] Since the captured intensity map contains wavefront aberration information, the model based on deep convolutional neural network can extract the wavefront zernike coefficient from the distorted intensity map, and even directly extract the atmospheric turbulence phase. Its weight sharing feature can achieve real-time wavefront reconstruction while reducing the amount of learning parameters. However, blindly increasing the network depth is not only prone to overfitting, but also very easy to fall into local optimum. In addition, the convolution operation will cause differences in the features extracted between the related pixels due to the different sizes of the receptive fields. These differences will cause large reconstructed aberrations. In response to the above problems, the present invention uses the idea of ​​residual network to construct an identity mapping of the feature map, so that the subsequent layers can learn the residual of the network, reduce the risk of overfitting, and combine the global context information, add the attention mechanism, establish connections between features, and increase the wavefront reconstruction ability of the network. Wavefront reconstruction system based on residual network such as Figure 2 As shown:

[0087] The backbone network is the ResNet50 model. First, the input light intensity map is converted into a feature image through a 7×7 downsampling convolution operation. The large-scale convolution kernel reduces the amount of calculation while retaining the feature information as much as possible. Then, a maximum pooling operation with a filter size of 3×3 is used to reduce the parameters of the feature map and reduce its dimension to 1 / 4 of the size of the initial image. Then the hybrid attention module is embedded in the ResNet model to fully capture the feature information. Each hybrid attention module contains two convolution operations and one hybrid attention calculation operation. The number of channels for each convolution operation is shown in the figure. Finally, the features are mapped from the 3D space to the 1D space through the fully connected layer, and Dropout is added to prevent overfitting. The final output is the zernike coefficient corresponding to the wavefront aberration in the input light intensity map.

[0088] 2.2 Hybrid Attention Structure

[0089] In order to make up for the limited receptive field of deep ResNet, lack of cross-channel interaction, and easy loss of small target information, the present invention adds a hybrid attention structure that combines the spatial attention (SA) mechanism with the dynamic selection mechanism network (Selective Kernel Neural Networks, SKNet) in the network. It mainly uses convolution operations to extract feature mapping information, and adds a residual skip connection structure to reasonably allocate computing resources for the model according to the importance of the input features. The hybrid attention structure proposed in the present invention includes SKNet with a channel attention mechanism (CA) combined with a spatial attention mechanism, and its structure is as follows Figure 3 shown.

[0090] For light spots of different sizes in the light intensity image, the scale of their features also changes. If the convolution kernel of the same size is used, the ability to express the feature information of small light spots gradually weakens, resulting in an increase in the reconstruction aberration error. To address this, the present invention uses a multi-scale feature extraction network SKNet to enable the model to enhance the expression of shallow features, which is beneficial to the regression of smaller targets in the network and allows the model to focus more on the target itself.

[0091] SKNet can be divided into three parts: decomposition, fusion, and selection. In the decomposition part, convolution kernels of different sizes are selected for feature mapping operations. Although the multi-branch structure can mine richer semantic information and enhance the network's expressiveness, it also increases the difficulty of network training. Therefore, the present invention uses a dual-stream network structure of 3×3 and 5×5 sizes, which is expressed by formula 10:

[0092]

[0093]

[0094] where X∈R h×w×C is the input feature map, and f represents the convolution operation. This not only allows the multi-branch structure to obtain the feature representation of the high and low frequency parts of the intensity image, but also avoids the difficulty of training convergence caused by too many network branches. The 5×5 convolution uses a dilated convolution operation. Dilated convolution is also a type of convolution operation. The biggest difference from ordinary convolution is the difference in the receptive field of the two. Its advantage is that it does not reduce the resolution of the image while obtaining a relatively large receptive field of view, thereby reducing the computational complexity of the network. Different receptive fields can obtain multi-scale features of the intensity map, which helps to improve the accuracy of predicting the zernike coefficient.

[0095] In the fusion part, the feature map is added by the summation operation, and then the information carried by each channel of the convolution layer is statistically analyzed by the global average pooling to obtain s, and finally the feature dimension reduction is performed to obtain z. The overall process is shown in the following formula:

[0096]

[0097]

[0098] z=F fc (s) = δ(B(W s )) (19)

[0099] where F gp is the global pooling operation, F fc is a fully connected operation, δ is a linear rectified function (Rectified linear unit, relu), B is a Batch Normalization layer, Ws ∈R d×C , d = max(C / r, L) represents the feature dimension after full connection, r is the compression factor, and L represents the minimum value of d. The fusion operation aggregates the information from these two paths to obtain a global comprehensive representation of the selective weights, that is, the part that calculates the weight of each convolution kernel. At the same time, the global attention module is used to integrate deep and shallow features to alleviate the problem of small spot information loss caused by the low resolution of deep features.

[0100] Select the part and use the softmax function to calculate the weight distribution of each channel. The calculation method is as follows:

[0101]

[0102]

[0103] Where a and b are the soft attention weight matrices of the two channels respectively, and V is the output matrix of SKNet. The selection operation is based on the selection weight aggregation, that is, the new feature map obtained after the convolution kernels with different weights are calculated, which enables the model to adaptively adjust the size of the neuron receptive field.

[0104] Due to the existence of the pooling layer, the amount of high-frequency information will be further reduced, resulting in the weaker expression ability of deep features for small targets. The present invention introduces a spatial attention mechanism to weight the features of the target area, find the connection between features in the network, and enable the feature extraction network to selectively focus on the target area containing important information. The specific process is shown in the following formula:

[0105]

[0106]

[0107] Where c represents the concatenation operation, σ is the sigmoid activation function, and F avg With F max They represent the average pooling operation and the maximum pooling operation performed along the channel axis respectively. Unlike channel attention, spatial attention can distinguish high-frequency and low-frequency information, which is complementary to channel attention. The feature map after optimizing spatial information can effectively express the feature similarity between pixels, and spatial attention can obtain contextual semantic information well.

[0108] 2.3 Loss Function

[0109] In order to make the reconstruction of the Zernike coefficient of the wavefront aberration more reasonable, the present invention combines the commonly used wavefront evaluation indicators to improve the loss function of the network. Commonly used AO system evaluation indicators include wavefront peak to valley (PV) and wavefront root mean square (RMS). PV is used to represent the difference between the highest and lowest points of the wavefront, and RMS represents the degree of deviation of the measured beam wavefront from the ideal wavefront. The formula is as follows.

[0110] PV=max(Δφ(ρ,θ))-min(Δφ(ρ,θ)) (24)

[0111]

[0112] where Δφ is the wavefront aberration, is its mean, ρ and θ are the polar coordinates of the pupil plane. It can be seen from the formula that the wavefront information reflected by the PV value is not comprehensive enough, while the RMS value pays more attention to the overall information of the wavefront. Combining the two can more accurately represent the wavefront situation. Therefore, the loss function combined with MSE is shown below.

[0113]

[0114] L=λ MSE L MSE +λ PV L PV +λ RMS L RMS (27)

[0115] Where z act and z pre are the actual and estimated zernike coefficients respectively, with a length of N, λ being the weight of each loss, and L being the weighted sum of each loss function. The reduction of the loss function represents the improvement of the reconstruction accuracy.

[0116] Therefore, combined with the above, we can know that:

[0117] Aiming at the problem that atmospheric turbulence causes vortex beam distortion, which leads to reduced FSO communication quality, the present invention proposes an AO system combined with a residual attention network to invert the wavefront phase, thereby achieving effective wavefront correction. The trained model establishes an accurate mapping relationship between the first 36-order Zernike coefficients and the distorted light intensity distribution. Experimental simulations were carried out under different turbulence conditions, and the obtained wavefront residual PV values ​​and RMS values ​​were both small, indicating that the network has strong robustness. The phase reconstructed by the predicted Zernike coefficients has a high similarity to the actual phase, and the corrected light intensity distribution has been significantly improved. Then, the effectiveness of the hybrid attention network in the task of reconstructing the wavefront phase was verified, and the accuracy was the highest with less increase in time complexity. The high accuracy, good real-time performance and strong flexibility of the residual attention network provide guarantees for the practical application of deep learning in the AO system.

[0118] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions only describe the principles of the present invention. The present invention may be subject to various changes and improvements without departing from the spirit and scope of the present invention. These changes and improvements fall within the scope of the present invention. The scope of protection claimed by the present invention is defined by the attached claims and their equivalents.

Claims

1. A wavefront distortion correction method based on residual attention network, characterized in that: The following steps are involved: The residual network is used as the backbone network, and a hybrid attention structure is constructed based on the residual network. The convolution operation is used to convert the light intensity image into a feature map for backward propagation. Distributed feature extraction is performed through convolution kernels of different scales, and the attention mechanism is used to improve the network's recognition rate of damaged light spot features and enhance the network's ability to express light intensity image features; Design the network loss function to obtain the Zernike coefficient that meets the actual wavefront aberration; The hybrid attention structure includes a spatial attention mechanism and a dynamic selection mechanism network, uses convolution operations to extract feature mapping information, and adds a residual skip connection structure to allocate computing resources to the model according to the importance of the input features; The dynamic selection mechanism network includes three parts: decomposition, fusion and selection; In the decomposition section: A two-stream network convolution structure of 3×3 and 5×5 is selected, expressed as: where X∈R h×w×C is the input feature map, f represents the convolution operation; the 5×5 convolution uses the hole convolution operation; In the fusion section: The feature map is added by the summation operation, and then the information carried by each channel of the convolution layer is statistically analyzed by the global average pooling to obtain s. Finally, the feature dimension reduction is performed to obtain z, which is expressed as: z=F fc (s)=δ(B(W s )) where F gp is the global pooling operation, F fc is a fully connected operation, δ is a linear rectification function, B is a Batch Normalization layer, and W s ∈R d×C , d = max(C / r,L) represents the feature dimension after full connection, r is the compression factor, and L represents the minimum value of d; In the Selection section: The softmax function is used to calculate the weight distribution of each channel, expressed as: Where a and b are the soft attention weight matrices of the two channels respectively, and V is the output matrix of SKNet.

2. The wavefront distortion correction method based on residual attention network according to claim 1, characterized in that: The residual network is a ResNet50 model, which first converts the input light intensity map into a feature image through a 7×7 downsampling convolution operation, and then uses a maximum pooling operation with a filter size of 3×3 to reduce the parameters of the feature map and reduce its dimension to 1 / 4 of the size of the initial image.

3. The wavefront distortion correction method based on residual attention network according to claim 1, characterized in that: The spatial attention mechanism is used to weight the features of the target area, find the connection between the features in the network, and make the feature extraction network selectively focus on the target area containing important information; it is expressed as: Where c represents the concatenation operation, σ is the sigmoid activation function, and F avg With F max Represent the average pooling operation and the maximum pooling operation performed along the channel axis respectively.

4. The wavefront distortion correction method based on residual attention network according to claim 1, characterized in that: in, Designing the network loss function mainly includes: The expressions of wavefront peak-to-valley value and wavefront root mean square value in the AO system evaluation index are: PV=max(Δφ(ρ,θ))-min(Δφ(ρ,θ)) where Δφ is the wavefront aberration, is its mean, ρ and θ are the polar coordinates of the pupil plane; Combining the two, the MSE loss function is generated, which is expressed as: L=λ MSE L MSE +λ PV L PV +λ RMS L RMS Where z act and z pre are the actual and estimated zernike coefficients respectively, with a length of N, λ being the weight of each loss, and L being the weighted sum of each loss function. The reduction of the loss function represents the improvement of the reconstruction accuracy.

Images