Wavefront correction method for AO system based on shufflenet network

Abstract

The present application relates to a kind of AO system wavefront correction methods based on ShuffleNet network, comprising the following steps: data set is generated by Matlab platform simulation, the data set includes several corresponding Zernike coefficients as label photo;ShuffleNet network is built by python platform, and the data set is as the input of network, and ShuffleNet network is trained, and wavefront correction model is obtained after training;The picture to be wavefront corrected is input into the wavefront correction model, and the corrected picture is obtained.The present application applies ShuffleNet network to the wavefront correction in adaptive optical system, so that the algorithm structure of the wavefront sensing based on ShuffleNet is simple, time delay is short, simple and easy to realize, and does not need multiple iterations, real-time is higher, and the requirement of equipment is low, cost is low, and engineering realization is easy.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of adaptive optical systems, and particularly relates to an AO system wavefront correction method based on a ShuffleNet network. BACKGROUND

[0002] With the development of communication devices, free space optical communication (FSOC) has the advantages of high transmission rate, strong anti-interference ability, wide frequency band and large capacity, and gradually enters the public's field of vision. Since free space optical communication is a communication mode taking light as a carrier and taking atmosphere as a medium, its performance is affected by atmospheric turbulence and is greatly affected by weather conditions. In windy, snowy weather or extreme sand weather conditions, the aerosol in the atmosphere seriously reduces the laser power. In clear weather conditions, the turbulence effect of the atmosphere also causes the light intensity fluctuation and wavefront distortion of the optical signal. The irregular fluctuation of air density in the atmosphere layer is called atmospheric turbulence. The influence of turbulence on beam transmission is called turbulence effect. The formation of atmospheric turbulence fading is due to the uneven air density under the irradiation of sunlight, which causes the refractive index of the medium in the light beam propagation path to be different, and the light signal produces scattering. Atmospheric turbulence causes the wavefront distortion and random fluctuation of the light beam in the FSOC system, mainly manifested as light intensity flicker, speckle, beam drift, expansion, angle of arrival fluctuation, which leads to the increase of the signal-to-noise ratio and the bit error rate of the FOSC system, seriously affects the mixing efficiency and coupling efficiency of the system, and makes the performance of the FSOC system decline. In summary, atmospheric turbulence seriously restricts the development of FSOC technology. Common techniques for suppressing atmospheric turbulence and compensating performance include adaptive optics (AO) technology, which is one of the recognized, effective and practical ways to suppress the influence of atmospheric turbulence on the FSOC system. At present, the AO system has been widely used in astronomical observation, wireless laser communication, laser quality correction, human eye retina imaging, biomedical imaging and other fields, and can achieve good correction of wavefront aberration.

[0003] Conventional adaptive optics system is composed of wavefront sensor, wavefront controller and wavefront corrector. The wavefront sensor is mainly used for measuring wavefront aberration, which is an important part of AO system. The wavefront controller can obtain the control quantity according to the wavefront aberration obtained by the wavefront sensor. The wavefront corrector corrects the wavefront aberration. The sensor is the most important part of the AO system, and the most widely used is the Shack-Hartmann wavefront sensor (SH-WFS). The S-H wavefront sensor is composed of a two-dimensional microlens array, a matching lens and a CCD. When the light is incident on the S-H wavefront sensor, the two-dimensional microlens array divides the incident light into several sub-wavefront arrays, and each part of the light wave converges on the sub-aperture focal point to form a sub-aperture spot array image. When the incident light is ideal plane light, the sub-wavefront spot imaging will focus on the center position of the sub-aperture. When the incident light is affected by atmospheric turbulence, the position information can be obtained, and the phase information of the incident light can be restored by calculation. Generally speaking, the more the number of sub-apertures, the more the generated sub-aperture spot array, and the better the compensation effect of the light beam. However, with the increase of the number of sub-apertures, the light intensity entering each lens is reduced, and even it may be lower than the light detection threshold of the lens array, resulting in a decrease in the utilization rate of the light beam. The S-H wavefront sensor is more suitable for scenes with high total power. Due to the evolution of atmospheric turbulence with a frequency of 1 kHz, the hardware processing speed of the AO system is extremely high. Due to the existence of the wavefront sensor and the wavefront controller, the traditional AO is not a real-time system, and there are problems such as long time delay, high equipment cost and large size. When the above problems occur, it will seriously affect the reliability of the FSOC system.

[0004] Sensor Less Adaptive Optics (SLAO) is a technique based on image-based methods to retrieve phase information from intensity measurements. The SLAO system is composed of a tilt correction subsystem and a high-order aberration correction subsystem. After the laser carrier signal passes through the tilt correction subsystem, the large tilt and jitter will be corrected, and the laser beam will be transmitted approximately along the optical axis. The high-order aberration correction subsystem includes a deformable mirror, an algorithm operator and a high-speed camera. The algorithm operator has a great influence on the performance of the SLAO system.

[0005] According to whether the optimization method is based on a point spread function, WFS-less AO techniques are divided into two categories: model-based and model-free. Model-based algorithms mainly include the Jacopo Antonello (JA) algorithm based on a second-order polynomial model, the Martin J.Booth algorithm based on a spherical model, and a convolutional neural network algorithm. Model-free algorithms include the SA algorithm, the MSFL algorithm, and the like.

[0006] The model-based method is a point spread function-based optimization algorithm, and can be specially applied to a WFS-less AO system. The main principle is to calculate the Zernike mode coefficient of an introduced aberration through a received end focal plane light intensity distribution.

[0007] The SA algorithm is a global search algorithm, which can accept a solution with a performance index worse than a current solution according to a certain probability, so as to avoid a local optimal solution. After a sufficient number of iteration searches, a global optimal solution is finally obtained. In the search process, the performance index may gradually decrease, but as long as a sufficient search time is passed, the global optimal solution can be converged. The MSFL algorithm greatly improves the search speed of the algorithm by using the mechanism of parallel search of multiple individuals, and further improves the real-time performance of the system on the basis of maintaining the global optimization characteristics.

[0008] The SA algorithm solves the local optimization problem of the stochastic parallel gradient descent algorithm (SPGD), but the SA algorithm and the MSFL algorithm are generally based on a random optimization search mechanism, and do not have a more explicit search purpose except for a cost function (objective function). The real-time performance is low, and multiple iterations are required, so the search speed has always been a technical bottleneck of this type of method. In addition, the model-based method needs strong technical equipment support, the equipment is complex, the engineering cost is high, and it is difficult to apply to engineering practice in a short time, and engineering implementation is difficult. SUMMARY

[0009] To solve the problems of the traditional adaptive optical system, such as multiple light splitting, being suitable for a scene with large total light power, complex structure, long time delay, high equipment cost, large volume, and the like, the present application provides an AO system wavefront correction method based on a ShuffleNet network. The method has low requirements for technical equipment and low cost, and can obtain the required result at one time through a pre-trained model, and can well solve the problems of the SA algorithm and the MSFL algorithm.

[0010] The present application adopts the following technical solutions:

[0011] An AO system wavefront correction method based on a ShuffleNet network, comprising the following steps:

[0012] Step 1: a data set is generated by simulation through a Matlab platform, wherein the data set comprises a plurality of photos, and each photo takes a corresponding Zernike coefficient as a label;

[0013] Step 2: a ShuffleNet network is built through a python platform, the data set is taken as an input of the ShuffleNet network, the ShuffleNet network is trained, and a wavefront correction model is obtained after training;

[0014] Step 3: a picture to be wavefront corrected is input into the wavefront correction model, and a corrected picture is obtained.

[0015] The wavefront correction method based on the ShuffleNet network has the advantages that the ShuffleNet network is applied to wavefront correction of an adaptive optical system, so that an algorithm structure of the wavefront sensing-free AO system based on the ShuffleNet is simple, time delay is short, and implementation is simple and easy; the wavefront correction model is obtained after the ShuffleNet network is trained by using the data set, the picture to be corrected can be corrected once through the wavefront correction model, and real-time performance is higher; and the wavefront correction method based on the ShuffleNet network has low requirements on equipment, low cost and easy engineering implementation. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 FIG. 1 is a structural schematic diagram of an FSOC system of an adaptive optical system including a wavefront sensor;

[0017] Figure 2 FIG. 2 is a structural schematic diagram of an FSOC system of an adaptive optical system including no wavefront sensor;

[0018] Figure 3 FIG. 3 is a flowchart of the AO system wavefront correction method based on the ShuffleNet network;

[0019] Figure 4 FIG. 4 is a learning rate curve diagram;

[0020] Figure 5 FIG. 5 is an accuracy rate curve diagram of a training set;

[0021] Figure 6 FIG. 6 is a loss value curve diagram of the training set;

[0022] Figure 7 FIG. 7 is an accuracy rate curve diagram of a test set;

[0023] Figure 8 loss value curve of the test set;

[0024] Figure 9 contrast diagram of a picture to be wavefront corrected before and after correction. DETAILED DESCRIPTION

[0025] As Figure 1 FSOC system block diagram is shown, first laser is modulated by modulator, then is emitted into the atmosphere by optical amplifier, and is affected by atmospheric turbulence in the atmosphere. Due to a large number of gas molecules and aerosols in the atmospheric turbulence, a certain influence will be generated on the laser signal, mainly manifested as that the phase and amplitude of the laser will change, resulting in phase tilt and high-order aberration, then the correction of the AO system is carried out, and finally into the receiving end, in which the local oscillator is used to change the frequency of the mixer in the receiver, and then the optical energy is amplified through a light amplifier, and then corresponding demodulation is carried out, and the transmitted information is obtained after digital signal processing. The AO is mainly used to compensate for the influence of atmospheric turbulence, and the traditional AO system includes a wavefront sensor, a wavefront corrector and a wavefront controller. The wavefront sensor commonly used is a Shack-Hartmann wavefront sensor, hereinafter referred to as an S-H wavefront sensor. After the light signal is subjected to beam shrinking, beam expanding and a beamsplitter, part of the light beam is coupled into a single-mode optical fiber, and the other part of the light beam enters the S-H wavefront sensor. The S-H wavefront sensor gives the average phase tilt of each sub-aperture. However, the light intensity decrease caused by the sub-aperture beamsplitting of the S-H wavefront sensor in the traditional AO system will lead to the decrease of the system performance, and even the light intensity is lower than the threshold value, resulting in the failure of the system. Therefore, the adaptive optical system without a wavefront sensor appears.

[0026] FSOC system including an adaptive optical system without a wavefront sensor is shown in FIG. 2, which is different from the FSOC system shown in Figure 1 The difference between the adaptive optical system without a wavefront sensor and the traditional adaptive optical system with a wavefront sensor is that the high-speed CCD camera (charge-coupled device camera) replaces the traditional wavefront sensor, and the wavefront does not need to be reconstructed, which can replace the traditional AO system. The light intensity distribution captured by the CCD camera is fed back to the adaptive optical system without a wavefront sensor algorithm control module, the control module transmits the control signal to the wavefront corrector, and the wavefront corrector adjusts the angle position of the deformable mirror to compensate for the wavefront distortion.

[0027] FSOC system including an adaptive optical system without a wavefront sensor is shown in FIG. 2, which is different from the FSOC system shown in Figure 3 The flow chart of the AO system wavefront correction method based on the ShuffleNet network provided by the embodiment of the application is shown in FIG. 2, and the method specifically includes the following steps:

[0028] Step 1: Generate a dataset including several photos by Matlab platform, each photo is labeled with its corresponding Zernike coefficients.

[0029] In FSOC, due to the existence of atmospheric turbulence and the non-idealism of communication system, the light beam is disturbed in the process of propagation, causing various aberrations. The present application uses Zernike function to describe the aberration. Zernike coefficients represent the wavefront aberration as a linear combination of a set of orthogonal modes, each corresponding to an aberration mode. Considering that Zernike polynomials are two-dimensional functions on a circular domain, their low-order forms are consistent with Seidel aberrations in optics. Within the unit circle, Zernike polynomials can be expressed in polar form, defined as radial and angular parts, which can be expressed as:

[0030]

[0031] Where m and n represent the spatial frequency variation of Zernike mode, is a real-valued radial polynomial, ρ is the length of the vector from the origin to the point (X, Y), and θ is the angle between the vector ρ and the X-axis counterclockwise, is the expression of real-valued radial polynomial when m = 0. Formula (1) can be expressed as:

[0032]

[0033] Here, the expression of the first 15 Zernike modes is given, as shown in Table 1.

[0034] As can be seen from Table 1, Zernike polynomials have orthogonal characteristics within the unit circle, which can be expressed as:

[0035]

[0036] Where Z i (ρ,θ) is the i-th order Zernike polynomial, Z j (ρ,θ) is the j-th order Zernike polynomial, and δ ij is the Kronecker symbol.

[0037] As can be seen from formula (3), the integral of the product of Zernike polynomials in the same mode within the unit circle is zero.

[0038] Table 1: Expression of the first 15 Zernike modes

[0039]

[0040]

[0041] The wavefront aberration caused by the laser passing through the atmospheric turbulence is described by the Zernike coefficients. It is known that a0 in the Zernike coefficient is a constant term, representing a piston term coefficient, a1 and a2 are tilt coefficients, and a3 is a defocus coefficient. The tilt correction system is not considered in the present application, that is, the coefficients a0-a3 are not considered. According to the theory of Noll, the low-order modes of the wavefront aberration caused by atmospheric turbulence account for the majority, so the Zernike coefficients a4-a8 will be studied, and the higher-order Zernike coefficients are not considered.

[0042] Further, the process of constructing the data set comprises the following steps:

[0043] Step 1.1: The a4-a8 in the Zernike coefficient is divided into 486 categories according to the size range of the coefficient.

[0044] In this step, the a4-a8 in the Zernike coefficient is classified according to the following ranges, respectively, wherein a4 is divided into 6 categories, and a5-a8 is divided into 3 categories: a4∈[-0.9,-0.6)∪[-0.6,-0.3)∪[-0.3,0)∪[0,0.3)∪[0.3,0.6)∪[0.6,0.9], a5∈[-0.5,-0.2)∪[-0.2,0.2)∪[0.2,0.5), a6∈[-0.5,-0.2)∪[-0.2,0.2)∪[0.2,0.5), a7∈[-0.3,-0.1)∪[-0.1,0.1)∪[0.1,0.3), a8∈[-0.3,0.1)∪[-0.1,0.1)∪[0.1,0.3). A total of 6*3*3*3*3=486 categories are divided.

[0045] Step 1.2: Two photos with the same absolute value of defocus are regarded as the same photo, the number of categories is halved, and 243 categories are obtained.

[0046] Since the point spread function based on defocus has similar characteristics in both positive and negative cases and has the same absolute value, two photos with the same absolute value of defocus can be regarded as the same photo under the condition that other terms Zernike coefficients are the same. That is, the number of categories is halved, and there are 243 categories.

[0047] Step 1.3: For the 243 categories obtained, use the Matlab platform to simulate N photos with a resolution of 64*64 under each category, and each photo is labeled with the range of a4-a8 in its Zernike coefficient.

[0048] In this step, N pictures with resolution of 64*64 are simulated under each category using Matlab platform, for example, N=2000, in which 1800 pictures in each category are used as training set and 200 pictures are used as test set, a total of 486000 pictures. Each picture is labeled with the range of its Zerinike coefficients a4-a8.

[0049] Step 1.4: The pictures and their labels under all categories together constitute a data set. The construction of the data set is completed at this time.

[0050] Step 2: Build ShuffleNet network through python platform, and use the data set as the input of ShuffleNet network to train ShuffleNet network, and obtain a trained network, i.e. wavefront correction model, which is the above wavefront sensor algorithm control module.

[0051] ShuffleNet network is a CNN network. CNN network is considered to be a good picture processing tool, which can reduce the dimension of a large amount of data in a picture, and is a kind of deep learning. There are several basic structures in CNN, which are convolution layer (Convolutions), subsampling, full connection layer and activation function. The purpose of convolution operation is to extract different features of input, and some convolution layers may only extract some low-level features such as edges, lines and angles. More layers of network can extract more complex features from low-level features. Pooling layer samples the features extracted by convolution layer. Pooling layer is generally divided into two kinds: maximum pooling layer and average pooling layer. Pooling layer reduces the complexity of subsequent network, improves the calculation precision and prevents overfitting. Full connection layer flattens the learned features and finally applies to model task.

[0052] ShuffleNet is a lightweight network, which provides efficient basic architecture for devices with limited computing power. It reduces the computational complexity of the network through group convolution, but at the same time brings the problem of information not flowing within each group, which reduces the feature extraction ability of the network. Therefore, the concept of channel shuffle is introduced. Channel shuffle reorganizes the information within each group, and the information can flow between different groups. In this way, on the basis of lower computational complexity and simpler network structure, the real-time performance and accuracy are improved.

[0053] The ShuffleNet structure is specifically: first, a 3*3 convolution and a maxpool layer are used, then three stages of shuffle units are used, for each stage, the first basic unit adopts Stride=2, so that the feature map width and height are each reduced by half, and the channel number is doubled. The following basic units are all Stride=1, and the feature map and the channel number remain unchanged. For the basic unit, the bottleneck layer is that the channel number of the 3x3 convolution layer is 1 / 4 of the output channel number, which is the same as the design concept of the residual unit. The ShuffleNet network structure used in the embodiment is shown in Table 2.

[0054] Table 2 ShuffleNet network structure

[0055]

[0056] Step 3: input the picture to be wavefront corrected into the trained network, that is, the wavefront correction model, to obtain the corrected picture, and complete the wavefront correction.

[0057] The present application also provides experimental results of wavefront correction experiments using the wavefront correction method of the present application. The laser wavelength used in this experiment is 523mm, and the focal length is 30cm. The experimental results are shown in Figures 4-9 , wherein Figure 4 is a learning rate curve diagram; Figure 5 is an accuracy rate curve diagram of the training set, and from the diagram, it can be seen that the accuracy rate of the training set is more than 97%; Figure 6 is a loss value curve diagram of the training set, and from the diagram, it can be seen that the loss value of the training set converges to 0.06; Figure 7 is an accuracy rate curve diagram of the test set, and from the diagram, it can be seen that the accuracy rate of the test set is more than 91.1%; Figure 8 is a loss value curve diagram of the test set, and from the diagram, it can be seen that the loss value of the test set converges to about 0.2. The experimental results show that the wavefront correction model in the embodiment has superior performance. Figure 9 is a comparison diagram of the picture to be wavefront corrected before correction Figure 9 (a) and Figure 9 (c), and after correction Figure 9 (b) and Figure 9 (d) using the wavefront correction method provided in the embodiment, and from the diagram, it can be seen that the present application can accurately correct the picture to be wavefront corrected, and the correction effect is excellent.

[0058] The application applies the ShuffleNet network to wavefront correction of an adaptive optical system, so that the wavefront sensing-free algorithm based on the ShuffleNet has simple structure, short time delay and simple realization, the wavefront correction method of the application trains the pre-built ShuffleNet network by using the constructed data set, and the wavefront correction model can be obtained after the training, the picture to be corrected can be corrected once by using the wavefront correction model, and the real-time performance is higher; meanwhile, the AO system wavefront correction method based on the ShuffleNet network has low requirement for equipment, low cost and easy engineering realization.

[0059] The technical features of the above-described embodiments can be combined in any manner, and to make the description concise, all possible combinations of the technical features in the above-described embodiments are not described, but as long as the combinations of the technical features do not exist, they should be considered as the scope of the description.

[0060] The above-described embodiments only express several implementation manners of the application, and the description is relatively specific and detailed, but it should not be understood as a limitation on the scope of the patent right of the application. It should be pointed out that, for ordinary skilled persons in the art, some modifications and improvements can be made without departing from the concept of the application, and these all belong to the protection scope of the application. Therefore, the protection scope of the patent right of the application should be subject to the appended claims.

Claims

1. A method for wavefront correction of an AO system based on a ShuffleNet network, characterized in that, The method comprises the following steps: Step 1: generating a data set through a Matlab platform, the data set comprising a plurality of photos, each photo being labeled with corresponding Zernike coefficients; step 1 comprising the following steps: Step 1.1: dividing a4-a8 in the Zernike coefficients into 486 categories according to the size range of the coefficients; In step 1.1, a4-a8 in the Zernike coefficients are classified according to the following ranges respectively: a4∈[-0.9,-0.6)∪[-0.6,-0.3)∪[-0.3,0)∪[0,0.3)∪[0.3,0.6)∪[0.6,0.9], a5∈[-0.5,-0.2)∪[-0.2,0.2)∪[0.2,0.5), a6∈[-0.5,-0.2)∪[-0.2,0.2)∪[0.2,0.5), a7∈[-0.3,-0.1)∪[-0.1,0.1)∪[0.1,0.3), a8∈[-0.3,-0.1)∪[-0.1,0.1)∪[0.1,0.3), a total of 486 categories; Step 1.2: regarding two photos with the same absolute value of defocus as the same photo, halving the number of categories, and obtaining 243 categories; Step 1.3: for the obtained 243 categories, using the Matlab platform to simulate N photos with a resolution of 64*64 under each category, each photo being labeled with the range to which a4-a8 in the Zernike coefficients belong; Step 1.4: the photos and labels under all categories jointly forming a data set; Step 2: building a ShuffleNet network through a python platform, and training the ShuffleNet network by taking the data set as the input of the ShuffleNet network, and obtaining a wavefront correction model after training; Step 3: inputting a picture to be wavefront-corrected into the wavefront correction model, the wavefront correction model transmitting a control signal to a wavefront corrector, the wavefront corrector adjusting the angle position of a deformable mirror to compensate for wavefront distortion, and obtaining a corrected picture.

2. The AO system wavefront correction method based on ShuffleNet network according to claim 1, wherein, N=2000, wherein 1800 photos in each category are taken as a training set, and 200 photos are taken as a test set.

3. The AO system wavefront correction method based on ShuffleNet network according to claim 1, wherein, The ShuffleNet network comprises 3*3 convolution and maxpool layers, three stages of shuffle units, for each stage, the first basic unit adopting a stride value of 2, and the other basic units adopting a stride value of 1.

Images