A vortex light wavefront distortion correction method based on a deep multi-branch compensation network

Abstract

The present application relates to a kind of vortex light wavefront distortion correction method based on deep multi-branch compensation network.Vortex light is a kind of special light field with helical wave front, and a variety of mode single state vortex light self-interference can generate superposition state vortex light.The method designs a kind of deep multi-branch compensation network, selects the first ten Zernike polynomials to numerically simulate atmospheric turbulence phase, according to the principle of different Zernike polynomial representing different aberration characteristics, branch learns the mapping relationship between vortex light distortion intensity map and atmospheric turbulence phase screen.In conceptual experiment, it can be realized under a variety of turbulence intensity, once the trained network is input with a distorted vortex light intensity distribution, only a very short time network can be more accurately predicted corresponding atmospheric turbulence compensation screen.The method can be used to correct the distorted vortex light interfered by phase, and the correction effect is good.The present application has good flexibility, rapidity and robustness, and is simple to operate, and has strong generalization ability.

Description

Technical Field

[0001] The present invention relates to a method for correcting vortex light wavefront distortion based on a deep multi-branch compensation network. By using a branched training network to accurately extract different types of aberration features from an input distortion intensity distribution map, the multi-branch compensation network can more accurately learn the mapping relationship between the distortion intensity map and the atmospheric turbulence compensation screen. Once a distorted vortex light intensity distribution map is fed into the trained network, the network can quickly predict the turbulence compensation screen's effect on the vortex light. Technical Background

[0002] The vortex phenomenon in the light field was first discovered by Boivin, Dow and Wolf in 1967 near the focal plane of a lens group. In 1973, Bryngdahl first explored the experimental method of preparing vortex light. In 1979, Vaughan and Willets successfully prepared vortex light using a continuous laser. In 1990, Yu and Bazgenov V first completed the preparation of vortex light using the grating method. In 1992, L. Allen discovered the phase factor under paraxial conditions. The vortex beam has an orbital angular momentum, where l is the topological charge of the vortex light orbital angular momentum, is the azimuth angle; each photon carries The orbital angular momentum of To reduce the Planck constant, the angular phase factor indicates that during the propagation of vortex light, if the light beam propagates for one cycle, the wavefront rotates exactly once around the optical axis, and the phase also changes by 2πl accordingly.

[0003] Vortex light, a novel structured beam with a spiral wavefront, has important applications in optical communications, particle micromanipulation, motion detection, and optical micromeasurement. Laguerre-Gaussian light is a typical example of vortex light. The photons in the beam possess both spin angular momentum (SAM) and orbital angular momentum (OAM), with the topological charge determining the magnitude of the OAM. A complete single-state Laguerre-Gaussian beam has a circular intensity distribution and a hollow dark core. The region of zero intensity at the center of the beam is defined as a phase singularity. Vortex beams can be divided into two categories based on the type of phase singularity: one in which the deflection direction of the light field is uniform but the phase of the singularity is uncertain, known as phase vortex light; the other in which the polarization direction of the singularity is uncertain, known as vector vortex light. Laguerre-Gaussian light is a type of phase vortex light. The superposition of multiple single-mode vortex light produces a superposition state vortex light with a different intensity and phase distribution from single-state vortex light.

[0004] The preparation of vortex light is the basis for conducting vortex light research. Common preparation methods include mode conversion method, computer-generated holography method, spatial light modulator method, Q-plate method and matrix spiral phase plate method. Under laboratory conditions, the spatial light modulator method is a commonly used preparation method. The spatial light modulator controls the electric field to cause changes in the spatial phase or amplitude image of the liquid crystal display, thereby writing certain information into the light wave and modulating the light wave. A holographic pattern of superposition state vortex light is prepared by complex amplitude control technology and loaded into the spatial light modulator. The spatial light modulator is irradiated with a beam of linearly polarized Gaussian light, and the output light is a superposition state vortex beam.

[0005] In long-distance transmission of superposition-state vortex light, the atmospheric temperature and pressure cause random fluctuations in the atmospheric refractive index, leading to wavefront distortion and amplitude fluctuations. This in turn causes vortex light beam jitter, intensity flickering, and crosstalk between adjacent modes, severely impacting beam quality. This can also severely affect the accuracy of vortex light target parameter measurements. Therefore, using specific wavefront correction techniques to compensate for this, maintaining good beam quality even during long-distance vortex light transmission, is crucial for applications in detection and communications.

[0006] Currently, adaptive optics is a common method for correcting vortex light wavefront distortion. This technique uses components such as wavefront sensors, controllers, and wavefront correctors to compensate for vortex light distortion caused by atmospheric turbulence. While adaptive optics can effectively restore the transmitted vortex light wavefront, adaptive optics systems often have highly complex structures. Correcting vortex light distortion in the presence of atmospheric turbulence generally employs traditional adaptive optics techniques without wavefront sensors. One approach involves using a stochastic-parallel-gradient-descent (SPGD) optimization algorithm to control a wavefront corrector to correct vortex light, forming an iterative feedback loop. Another approach involves a turbulence pre-compensation scheme for vortex light based on the Gerchberg-Saxton (GS) algorithm. While both algorithms can correct vortex light distortion caused by atmospheric turbulence, they often require lengthy iterations and are prone to falling into local optimal solutions. These algorithms also suffer from drawbacks such as long processing times and difficulty converging. As the number of vortex light modes and turbulence intensity increase, the difficulty of implementing these algorithms while maintaining a certain compensation effect increases. The deep multi-branch compensation network proposed in this patent is a vortex light wavefront compensation solution with short time consumption and low complexity.

[0007] The correction of vortex light wavefront distortion is of great significance for expanding applications. In the actual application of vortex light in the field of long-distance target detection, the temperature and pressure of the atmosphere cause random fluctuations in the atmospheric refractive index, resulting in wavefront distortion and amplitude fluctuations of the light wave, causing vortex light beam jitter, light intensity flickering, and crosstalk between adjacent order modes, which seriously affects the beam quality. The accuracy of vortex light target parameter measurement will also be seriously affected. Therefore, a certain wavefront correction technology is used to compensate and correct the wavefront distortion, so that the vortex light can maintain good beam quality even in long-distance transmission. This is very important for the application of vortex light in detection, communication and other aspects. The deep multi-branch compensation network proposed in this patent performs multi-branch training on the input distorted vortex light intensity map, accurately extracts features by type, deeply learns the mapping relationship between the distortion intensity map and the turbulence compensation screen, and more accurately predicts the turbulence compensation screen to correct the distorted vortex light.

[0008] Correcting vortex light wavefront distortion is a prerequisite for its practical application. This patent proposes a novel vortex light compensation scheme: a deep multi-branch compensation network is designed to accurately learn the features of vortex light distortion intensity maps. The trained network can predict the corresponding turbulence compensation screen, and the purity of the distorted vortex light pattern after compensation can reach over 80% under various turbulence intensities. In a proof-of-concept experiment, a dataset was collected to train the deep multi-branch network, and the trained network was used to correct the distorted vortex light, achieving significant compensation results. Summary of the Invention

[0009] The technical problem solved by the present invention is that the current development of vortex light detection technology is strongly affected by atmospheric turbulence, and the vortex light is severely distorted, resulting in crosstalk between adjacent vortex light modes and poor beam quality. Therefore, in order to promote the practical application of vortex light detection technology, vortex light wavefront distortion correction technology is particularly necessary. To solve the problems of long time consumption and high system complexity of vortex light correction technology, this patent proposes a vortex light wavefront distortion correction method based on a deep multi-branch compensation network. This method has good flexibility, rapidity and robustness, and is suitable for various occasions and harsh conditions. It proposes to extract the type of features of the distorted vortex light intensity map, and more accurately learn the mapping relationship between the distorted vortex light intensity map and the turbulence phase screen, so as to more accurately predict the atmospheric turbulence compensation screen to correct the vortex light distortion. In the proof-of-concept experiment, it can be achieved that the network can accurately predict the corresponding atmospheric turbulence compensation screen in a very short time. This method can be used to correct distorted vortex light affected by phase interference, and the correction effect is good. This method has good flexibility, rapidity and robustness, is easy to operate and has strong generalization ability.

[0010] The technical solution of the present invention is:

[0011] The present invention relates to a method for correcting vortex light wavefront distortion based on a deep multi-branch compensation network. The method can predict the turbulence compensation screen by simply inputting a distortion intensity map into the network. The method mainly includes the following steps:

[0012] (1) The first ten Zernike coefficients are used to numerically simulate atmospheric turbulence, encode the vortex light hologram disturbed by atmospheric turbulence and load it into a spatial light modulator. Linearly polarized Gaussian light is used to illuminate the pure phase spatial light modulator to prepare the distorted vortex light intensity map.

[0013] (2) After the distorted vortex light passes through the beam collimation and filtering system, a CCD camera is used to detect and collect the beam intensity distribution along the beam propagation path.

[0014] (3) Use the data set collected from the experiment and make it into a suitable size for input into the network.

[0015] (4) Based on the deep residual network, the distortion vortex light features are preliminarily extracted, and then the three distortion features are divided into three branches for training. The branches are as follows: the first branch is the coefficients of the 1st, 2nd, and 3rd order Zernike polynomials that characterize the translation and tilt phenomena; the second branch is the coefficients of the 4th, 5th, and 6th order Zernike polynomials that characterize the defocus and astigmatism phenomena; the third branch is the coefficients of the 7th, 8th, 9th, and 10th order Zernike polynomials that characterize the coma and high-order coma phenomena;

[0016] (5) After hundreds of iterations, the trained network learns the mapping relationship between the input distorted vortex light intensity map and the atmospheric turbulence screen, thereby predicting the ten Zernike polynomial coefficients in step (1) that characterize the atmospheric turbulence compensation screen and compensating the distorted vortex light.

[0017] The principle of the present invention is:

[0018] Laguerre-Gaussian light is a typical vortex light and is a set of solutions to the paraxial wave equation in the cylindrical coordinate system. When the propagation distance z = 0, its complex amplitude can be expressed in the cylindrical coordinate system as:

[0019]

[0020] Where U is the wave vector of Laguerre-Gaussian light, is the cylindrical coordinate, r is the polar diameter, is the polar angle, m is the topological charge, p is the radial node number, ω0 is the beam waist radius of the fundamental mode Gaussian light, is the Laguerre polynomial, i is the imaginary unit, and π is the circumference of the circumference of the circle. The mode parameters of the vortex light include the topological charge and the number of radial nodes.

[0021] In order to express it concisely and retain the characteristics of vortex light, equation (1) can be shortened to:

[0022] E1=Aexp(imφ) (2)

[0023] E2=Aexp(-imφ) (3)

[0024] E=E1+E2 (4)

[0025] Where E1 and E2 represent the intensity vectors of the two single-state vortex light respectively, E represents the intensity vector of the superposition state vortex light, A represents the intensity distribution, i is the imaginary unit, m is the topological charge, and φ is the phase.

[0026] In the context of using the rotational Doppler effect of superposition state vortex light with large topological charge number to detect long-range targets, ±20-order superposition state vortex light is selected as the research object (i.e., m=20). The present invention designs a superposition state vortex light wavefront distortion correction method based on a deep multi-branch compensation network. The method uses the first ten Zernike polynomials to perform numerical simulation on the phase disturbance of atmospheric turbulence. The first ten Zernike coefficients are used as labels for the network supervised learning of the distorted superposition state vortex light intensity map. The intensity map and the corresponding labels constitute a data set input into the network for iterative training. In order to allow the network to focus more on the characteristics of different wavefront distortions and achieve the effect of more accurate feature extraction and improved network prediction ability, the present invention proposes a deep multi-branch compensation network model, such as Figure 2 The model consists of a deep feature extraction module and a multi-branch compensation screen parameter prediction module. The deep feature extraction module contains five residual modules, while the multi-branch compensation screen parameter prediction module consists of three branches after the residual module. These modules undergo convolutional layers, average pooling layers, and fully connected layers to produce three one-dimensional vectors: 1*3, 1*3, and 1*4. These vectors correspond to the predicted values ​​of the 1st, 2nd, and 3rd order Zernike coefficients, the 4th, 5th, and 6th order Zernike coefficients, and the 7th, 8th, 9th, and 10th order Zernike coefficients, respectively. During network training, the input is the intensity distribution of the distorted vortex light captured by a CCD camera, and the output is a 1*10 vector of Zernike polynomial coefficients representing atmospheric turbulence. The convolution kernel extracts features, allowing the network to continuously learn the mapping between the distorted vortex light intensity distribution and the atmospheric turbulence screen. The parameters of each layer are continuously adjusted to minimize the loss function. Once fully trained, the network can predict the atmospheric turbulence compensation screen and perform distortion correction on the vortex light. The network selects the mean square error (MSE) between the predicted value and the actual value of the first ten Zernike coefficients, weights them, and then adds them together as the loss function Loss, as shown in the following formula:

[0027]

[0028] Loss=αLoss1+βLoss2+(1-α-β)Loss3 (6)

[0029] where a i represents the true value of the i-th order Zernike coefficient, represents the predicted value of the i-th order Zernike coefficient, α and β are the weight coefficients of the loss function for the three coefficient groups, Loss1 represents the mean square error of the 1st, 2nd, and 3rd coefficients, Loss2 represents the mean square error of the 4th, 5th, and 6th coefficients, and Loss3 represents the mean square error of the 7th, 8th, 9th, and 10th coefficients. The network optimizes the parameters of each layer using Loss as the objective function until the optimal value is achieved for each layer when the Loss is minimized. Comparing the light field pattern purity before and after correction shows a significant improvement in pattern purity, demonstrating a significant compensation effect.

[0030] The mode purity of vortex light is defined as the mode fraction of any light field Ψ, which can be calculated as:

[0031]

[0032] in,

[0033]

[0034]

[0035] in, represents the Laguerre-Gaussian mode. In this paper, the ratio of the desired vortex light mode to the total mode is defined as mode purity. Equation (7) is the orbital angular momentum spectrum expanded in angular space. In this paper, the topological charge number l is between -30 and 30 when measuring mode purity.

[0036] The main advantages of the solution of the present invention are:

[0037] (1) Highly efficient and simple, easy to operate, it only takes a very short time to compensate for the distorted vortex light, with a simple structure and short time consumption.

[0038] (2) It has a wide range of applications, strong flexibility and robustness, and the multi-branch network structure can accurately learn the mapping relationship between vortex light of various distorted forms and turbulent phase screens, improve prediction accuracy, and adapt to different turbulence intensities. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 This is a flow chart of a method for correcting vortex light wavefront distortion based on a deep multi-branch compensation network;

[0040] Figure 2 This is the structure diagram of the deep multi-branch compensation network;

[0041] Figure 3 This is the schematic diagram of the experimental platform;

[0042] Figure 4Comparison of vortex light mode purity under different turbulence intensities before and after correction Specific implementation plan

[0043] The present invention uses symmetrical superposition state vortex light as the measurement object and the implementation object is a spatial light modulator. The specific implementation steps are as follows:

[0044] First, a hologram of symmetrical superposition state vortex light is encoded and loaded onto a pure phase spatial light modulator (SLM). A helium-neon laser (NEWPORT N-LHP-151) emits a collimated Gaussian beam with a wavelength of 632.8 nm after being collimated using a linear polarizer (LP), a half-wave plate (HWP) and a telescope composed of two lenses (L1, L2). The combination of LP and HWP is used to rotate the laser polarization state along the long display axis of the spatial light modulator (SLM) and adjust the power of the incident light on the spatial light modulator (SLM). The SLM (UPOLABS HDSLM80R) precisely modulates the incident light by loading the above-mentioned hologram. The aperture (AP) is then used to select the first order diffraction of the beam to avoid other stray light. The CCD camera (NEWPORT LBP2) records the intensity distribution after L4, as shown Figure 3 shown.

[0045] For example, a distorted symmetric superposition state Laguerre-Gaussian beam hologram with a topological charge of ±20 is first obtained through multi-parameter joint control technology. In the process of making the hologram, the blazed grating prevents part of the unmodulated light from mixing into the required distorted superposition state vortex light, so that the required beam is diffracted to the first order and the unmodulated light is kept at the zero order. The unmodulated light is caused by the gap in the SLM liquid crystal arrangement. Although the use of SLM for complex amplitude modulation sacrifices the phase depth, it allows us to radially modulate the incident light field to generate the distorted superposition state vortex light eigenmode instead of the hypergeometric mode. The encoded hologram is loaded onto the SLM to obtain the distorted superposition state vortex light intensity distribution map, which is combined with the ten Zernike coefficients that characterize atmospheric turbulence as labels to form a data set. The present invention has collected a total of three data sets under turbulence intensities, namely D / r0 = 2.67, 5, and 7.27, where D is the receiving aperture diameter, r0 is the atmospheric coherence length, and D / r0 characterizes the atmospheric turbulence intensity.

[0046] As shown in formula (6), the [αβ1-α-β] coefficient is a weight combination of multiple branches in the network model. In the method proposed in this paper, it directly affects which type of feature extraction and learning the network focuses on, and thus affects the network's compensation effect. It is a key parameter for the experimental results. Therefore, this paper roughly selects several groups of [αβ1-α-β] coefficient combinations based on the degree of attention the network pays to each type of feature for comparative experiments. This paper extracts 6,000 small samples from 18,000 sets of data sets collected in the laboratory, and randomly selects 2,000 sets for each of the three turbulence intensities. The training set and test set are divided into 9:1. Different weight combinations are selected for comparative experiments. The experimental results show that the network prediction error is minimized when the combination of α = 0.3 and β = 0.5, thereby determining the network target loss function.

[0047] The collected dataset is fed into the network for training. The deep multi-branch compensation network consists of a deep feature extraction module and a multi-branch compensation screen parameter prediction module. The deep feature extraction module contains five residual modules, and the multi-branch compensation screen parameter prediction module includes three branches after the residual module. These are processed through convolutional layers, average pooling layers, and fully connected layers to produce three one-dimensional vectors of 1*3, 1*3, and 1*4, corresponding to the predicted values ​​of the 1st, 2nd, and 3rd order Zernike coefficients, the 4th, 5th, and 6th order Zernike coefficients, and the 7th, 8th, 9th, and 10th order Zernike coefficients, respectively. During network training, the input of the network model is the distorted vortex light intensity distribution captured by a CCD camera, and the output is a 1*10 vector of Zernike polynomial coefficients representing atmospheric turbulence. The convolution kernel extracts features, allowing the network to continuously learn the mapping between the distorted vortex light intensity distribution and the atmospheric turbulence screen, and continuously adjusts the parameters of each layer to reduce the loss function. After mature training, the network can predict the atmospheric turbulence compensation screen and perform distortion correction on vortex light. Figure 4 As shown in the figure, the purity of the vortex light pattern under different turbulence intensities after correction can reach more than 80%. The intensity map of the vortex light after correction is more uniform, the correction effect is significant, efficient and easy to operate.

[0048] In addition, the spatial light modulator has certain limitations on the incident angle and power of the light beam, so the specific optical path design must be carried out according to the actual situation of the laboratory.

[0049] The contents not described in detail in this specification belong to the prior art known to those skilled in the art.

Claims

1. A method for correcting vortex light wavefront distortion based on a deep multi-branch compensation network, characterized by: The network does not require repeated iterations. Instead, it only needs to input a distorted vortex light intensity map into the network trained with the sample data set. The network can then predict the phase compensation screen corresponding to the distorted vortex light and compensate for the distorted vortex light. The process includes the following steps: (1) The first ten Zernike coefficients were used to numerically simulate atmospheric turbulence, and the vortex light hologram disturbed by atmospheric turbulence was encoded and loaded into a spatial light modulator. The distorted vortex light intensity map was prepared by irradiating the pure phase spatial light modulator with linearly polarized Gaussian light. (2) After the distorted vortex light passes through the beam collimation and filtering system, a CCD camera is used to detect and collect the beam intensity distribution along the beam propagation path; (3) Use the data set collected from the experiment and make it into a suitable size to input into the network; (4) Based on the deep residual network, the distortion vortex light features are preliminarily extracted, and then the three distortion features are divided into three branches for training. The branches are as follows: the first branch is the 1st, 2nd, and 3rd order Zernike polynomial coefficients that characterize the translation and tilt phenomena; the second branch is the 4th, 5th, and 6th order Zernike polynomial coefficients that characterize the defocus and astigmatism phenomena; the third branch is the 7th, 8th, 9th, and 10th order Zernike polynomial coefficients that characterize the coma and high-order coma phenomena; (5) After hundreds of iterations, the trained network learns the mapping relationship between the input distorted vortex light intensity map and the atmospheric turbulence compensation screen, thereby predicting the ten Zernike coefficients in step (1) that characterize the atmospheric turbulence compensation screen and compensating for the distorted vortex light.

2. The method for correcting wavefront distortion of vortex light based on a deep multi-branch compensation network according to claim 1, characterized in that: The multi-branch compensation network has strong generalization ability and can be used to correct other types of distorted light fields by changing the sample data set. The input distorted image features can be effectively extracted by adjusting the network branch structure and branch parameters according to the aberration characteristics.

Images