A direct correction method for distorted vortex beams based on image translation algorithm

Abstract

The application discloses a distortion vortex beam direct correction method based on an image translation algorithm. The method directly corrects the distorted vortex beam without generating a compensation phase screen to compensate the distorted light beam, eliminates the pair matching requirement limitation of manufacturing a training data set, reduces the number requirement of the training data set, simplifies the preparation work before model training, and has excellent correction performance. The method does not depend on phase information to correct the distorted light intensity, corrects the distorted light intensity by using a machine learning algorithm between the distorted light intensity diagram and the lossless light intensity diagram at a receiving end, does not require any optical phase compensation element, saves the cost of optical hardware, and is convenient to integrate.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of free space optical communication and artificial intelligence, and particularly relates to a vortex beam direct correction method using an image translation algorithm. BACKGROUND

[0002] The vortex beam is also called an orbital angular momentum (OAM) beam, and the wave front thereof is spiral-shaped. Because there is a phase singularity in the center of the beam, the central light intensity of the vortex beam is zero, and the beam distribution is annular like a doughnut. In 1992, Allen et al. of Leiden University in the Netherlands studied the characteristics of Laguerre-Gaussian beams and proved that each photon of the Laguerre-Gaussian beam carries orbital angular momentum. In recent years, the application of vortex beams in the field of optical communication has become one of the research hotspots at home and abroad. In the application of optical communication, the orbital angular momentum of the vortex beam provides new dimensional resources for the spatial domain of light waves. Because the OAM vortex beams with different orbital angular momentum mode values are orthogonal to each other, the application of vortex beams to the field of optical communication can greatly improve the channel capacity and spectral efficiency of the optical communication system. Considering that the transmission medium is the atmosphere, the phase and light intensity distribution of the beam will be distorted after transmission through the atmospheric turbulence, and therefore the existence of atmospheric turbulence will deteriorate the performance of the optical communication system using vortex beams. The existing distortion compensation methods for vortex beams transmitted through atmospheric turbulence, whether traditional adaptive optics methods or machine learning combined with adaptive optics methods, need to use optical phase compensation elements at the receiving end to compensate for the distorted phase. The optical elements are high in cost, large in size and not easy to integrate. The adaptive optics method based on machine learning uses a supervised learning algorithm model, and the training data volume of such a model is large and additional data labels need to be made. The present application aims to solve the problems in the prior art and provide a distortion vortex beam direct correction method based on an image translation algorithm. SUMMARY

[0003] The present application aims to solve the problems in the prior art and provide a distortion vortex beam direct correction method based on an image translation algorithm. The method does not need to generate a compensation phase screen first and then compensate for the distorted beam, but directly corrects the distorted vortex beam, eliminates the requirement of paired matching for making a training data set, reduces the requirement for the number of training data sets, simplifies the preparation work before model training, and has excellent correction performance.

[0004] The technical solution adopted by the present application to solve the technical problems is as follows:

[0005] A distortion vortex beam direct correction method based on an image translation algorithm, and the steps thereof are as follows:

[0006] Step (1), modeling the transmission of vortex beams in atmospheric turbulence medium, using the method of light wave transmission numerical simulation, respectively obtaining the simulated light intensity distribution diagram of vortex beams before transmission (target domain) and the distorted light intensity distribution diagram after transmission (source domain).

[0007] The simulated light intensity distribution diagram of vortex beams is made into a data set, including a training set containing training model parameters and a test set testing model performance. When making the data set, it is not necessary to pair each distorted light intensity distribution diagram with its corresponding light intensity distribution diagram before transmission.

[0008] Step (2), establishing an image translation algorithm model, taking the light intensity distribution diagram of vortex beams before transmission (target domain) and the distorted light intensity distribution diagram after transmission (source domain) in the training set as the input of the image translation algorithm model for training, so that the network learns the mapping relationship from the source domain to the target domain.

[0009] Step (3), inputting the test data set into the trained image translation algorithm model, the algorithm model corrects the distorted light intensity distribution diagram of vortex beams according to the learned mapping relationship from the source domain to the target domain, directly generating the undistorted light intensity distribution diagram of vortex beams, which does not rely on the phase information, so that the receiving end does not need optical phase compensation elements, with low cost and easy integration.

[0010] Further, the data set is obtained by the following way: establishing a multi-turbulence phase screen simulation of atmospheric turbulence transmission environment, generating data by light wave transmission numerical simulation calculation of diffraction intensity; respectively controlling the turbulence intensity and transmission distance to generate the distorted light intensity distribution diagram of vortex beams after transmission under different conditions. When making the data set, it is not necessary to pair each distorted light intensity distribution diagram with its corresponding light intensity distribution diagram before transmission, thereby reducing the complexity of making the data set and obtaining the training set and the test set.

[0011] Further, the image data input into the image translation algorithm model needs to be pre-processed into 256x256, 3-channel RGB images.

[0012] The image translation algorithm model is a variant based on the generative adversarial network model, called CUT network, which is composed of a 15-layer network generator and a 5-layer network discriminator.

[0013] The generator of the CUT network is composed of 15 layers of network, the first 3 layers are convolutional layers, the middle 9 layers are residual blocks, followed by two deconvolutional layers, and the last layer is a convolutional layer. The first and last convolutional layers use a convolution kernel size of 7x7, and their step size is 1; the second and third convolutional layers use a convolution kernel size of 3x3, and their step size is 2; the nine residual blocks in the middle use a convolution kernel size of 3x3, and their step size is 1; the two deconvolutional layers use a convolution kernel size of 3x3, and their step size is 2.

[0014] The discriminator of the CUT network is composed of 5 layers of networks, the first 4 layers are all convolutional layer combinations of the previous layer instance standardization layer, the convolution kernel size used is 4*4, the stride is 2, and the ReLU function is used as the activation function; the last layer is a convolutional layer, the convolution kernel size used is 4*4, the stride is 1, and the Sigmoid function is used as the activation function.

[0015] The data set includes data of different turbulence intensities and different transmission distances. The three cases represent strong turbulence, medium-intensity turbulence and weak turbulence respectively, and 1000 distorted light intensity maps and 1000 undamaged light intensity maps are collected under each turbulence intensity. When the turbulence intensity is

[0016] The light intensity distribution map is a matrix in the image translation algorithm model, and the purpose of training the image translation algorithm model is to learn the matrix mapping transformation relationship between the source domain and the target domain.

[0017] The beneficial effects of the present application are as follows:

[0018] The present application simplifies the process of making data sets, and does not need to make corresponding labels for each distorted light intensity map, but only collects distorted light intensity maps and undamaged light intensity maps, which can be classified and stored.

[0019] The method of the present application does not depend on phase information to correct the distorted light intensity, and through the relationship between the distorted light intensity map and the undamaged light intensity map, the distorted light intensity is corrected at the receiving end by using a machine learning algorithm, without any optical phase compensation element, saving the cost of optical hardware and facilitating integration. BRIEF DESCRIPTION OF DRAWINGS

[0020] In order to further explain the present application, the image data used in the examples of the present application will be briefly introduced below, and it is proposed that the following legends are only some embodiments of the present application.

[0021] Figure 1 is the conceptual flowchart of the scheme proposed by the present application;

[0022] Figure 2 is the image translation algorithm flowchart used by the present application;

[0023] Fig. 3(a) and Fig. 3(b) are network model structure diagrams used in the scheme proposed by the present application, including a generator and a discriminator;

[0024] Figure 4 is the effect diagram and performance index PSNR of the present application example under different turbulence intensity conditions when the transmission distance is 1 km;

[0025] Figure 5 is the effect diagram and performance index PSNR of the present application example under different turbulence intensity conditions when the transmission distance is 1 km;

[0026] English interpretation of the figure: Original Vortex Beam: vortex beam before transmission; Distorted Vortex Beam: distorted vortex beam; Corrected Vortex Beam: corrected vortex beam; Atmospheric Turbulence Propagation: atmospheric turbulence transmission; Convolutional Layer: convolutional layer; Deconvolutional Layer: deconvolutional layer; Residual Block: residual block; Encoder: encoder; Decoder: decoder; Generator: generator; Discriminator: discriminator; Filter: filter (convolution kernel); Stride: stride. DETAILED DESCRIPTION

[0027] In order to make the purpose, technical scheme and advantages of the present application clearer, the present application will be further described below in combination with the drawings and examples.

[0028] ​The present inventors understand that, when they come into the field of free space optical communication transmission using vortex beams, the atmospheric turbulence cannot be ignored as a major factor degrading the performance of this communication system. Existing solutions mainly focus on adaptive optics, one of which is a traditional solution using optical components such as a Shack-Hartmann wavefront sensor; the other is a technical solution based on deep learning computer vision technology without complex optical equipment. The traditional solution is divided into a probe beam scheme and a non-probe beam scheme according to whether the probe beam is used to indirectly obtain the distortion phase information. The probe beam scheme transmits coaxial beams of a fundamental mode Gaussian beam and a vortex beam at the sending end, and indirectly obtains the interference of the vortex beam caused by the turbulence according to the distortion of the fundamental mode Gaussian beam. After the CCD camera captures the light intensity distribution map of the fundamental mode Gaussian beam, the terminal uses, for example, the GS algorithm to obtain the compensation phase screen for distortion correction. The non-probe beam scheme needs to send the vortex beam before transmission and the distorted vortex beam into the GS algorithm iteration, and then obtain the compensation phase screen by subtracting the phase distribution of the undistorted vortex beam from the phase distribution output by the algorithm.

[0029] The above-mentioned schemes are all indirect correction processes. In actual scenarios, after obtaining the distorted light intensity distribution map at the receiving end, the compensation phase screen needs to be obtained first, and then the compensated light intensity distribution is obtained. The improvements are as follows:

[0030] (1) The traditional adaptive optics technology, whether using a fundamental mode Gaussian beam as a probe beam, needs to use numerous complex optical components, and the cost of these optical hardware is high.

[0031] (2) The method of obtaining a phase compensation screen for correction, whether based on traditional adaptive optics or machine learning, is an indirect method, which increases the cost in terms of computation and time cost of compensation correction.

[0032] (3) The existing method of using machine learning for distortion compensation, whether using image recognition principles or image translation principles, uses models belonging to the supervised learning category. When preparing the data set for training, the paired principle of supervised learning needs to be followed, which increases the difficulty of preparing the training data set. Moreover, the performance of these models also has a higher requirement for the data volume of the data set. In summary, the cost of training the model is high.

[0033] In view of the problems of existing compensation for the distortion of vortex beams after transmission through atmospheric turbulence based on traditional adaptive optical technology or machine learning technology based on supervised learning, the present application proposes a method for directly correcting the distorted vortex beams to generate repaired vortex beams, adopts a generative network model based on unsupervised learning to directly generate the corrected vortex beam intensity distribution map, solves the problems of complex system and high cost of using optical components, solves the problem of complex training preparation work of using machine learning technology based on supervised learning, and most importantly, solves the problem of using the indirect compensation method of compensation phase screen, simplifies the steps, reduces the calculation amount and time consumption.

[0034] The concept of the present application is:

[0035] (1) The problem of compensating for the distortion of vortex beams after transmission through atmospheric turbulence is essentially transformed into the problem of obtaining the mapping relationship of two matrices, so that direct compensation can be performed without indirect compensation through a compensation phase screen.

[0036] (2) The condition restriction of the training set pairing in the machine learning technology using supervised learning is removed, the data amount requirement of the model used is greatly reduced, and the performance is ensured.

[0037] The specific process of actual implementation is as follows: first, determine the atmospheric turbulence power spectrum model used and generate a random phase screen representing atmospheric turbulence by using the power spectrum inversion method to simulate atmospheric turbulence; then, calculate the intensity distribution map of the vortex beam after transmission through atmospheric turbulence by using the method of light wave transmission numerical simulation; under the condition of respectively controlling the turbulence intensity and the transmission distance, a certain amount of data samples are generated. The CUT network model for correcting the distorted light intensity is trained to learn the mapping relationship between the distorted light intensity and the light intensity before transmission, and can directly generate the corrected vortex beam intensity distribution map. Both the complex system and the high cost of using optical components are solved, and the problem of complex training preparation work of using machine learning technology based on supervised learning is solved, and most importantly, the problem of using the indirect compensation method of compensation phase screen is solved, the process is simplified, the calculation efficiency is improved, and the calculation time is saved.

[0038] The network model used in the present application is named CUT, which is a variant of the generative adversarial network model belonging to the category of unsupervised learning, as shown in Figure 1 The specific details of the vortex beam distortion correction scheme based on the network in the implementation are as follows:

[0039] Step (1) models the transmission of vortex beams in atmospheric turbulence medium, and respectively obtains the simulated vortex beam intensity distribution map before transmission (target domain) and the distorted light intensity distribution map after transmission (source domain) by using the method of light wave transmission numerical simulation.

[0040] The fixed transmission distance is 1km, and 800 distorted light intensity maps and 800 undistorted light intensity maps are collected respectively with turbulence intensity of The fixed turbulence intensity is 0.2, and 800 distorted light intensity maps and 800 undistorted light intensity maps are collected respectively with transmission distance of 1km, 2km and 4km. After removing the duplicates, a total of 8000 are collected to form the training set data.

[0041] Step (2) constructs the CUT network model as shown in FIG. 3(a) and FIG. 3(b), which includes a generator and a discriminator:

[0042] 2a) The generator of the CUT network is composed of a 15-layer network, the first 3 layers are convolutional layers, the middle 9 layers are residual blocks, followed by two deconvolutional layers, and the last layer is a convolutional layer. The first and last convolutional layers use a convolution kernel size of 7x7, and their stride is 1; the second and third convolutional layers use a convolution kernel size of 3x3, and their stride is 2; the nine residual blocks in the middle use a convolution kernel size of 3x3, and their stride is 1; the two deconvolutional layers use a convolution kernel size of 3x3, and their stride is 2.

[0043] 2b) The discriminator of the CUT network is composed of a 5-layer network, the first 4 layers are all convolutional layers combined with an instance normalization layer of the previous layer, using a convolution kernel size of 4x4 and a stride of 2, using a ReLU function as the activation function; the last layer is a convolutional layer, using a convolution kernel size of 4x4 and a stride of 1, using a Sigmoid function as the activation function.

[0044] Step (3) preprocesses the data set to adjust it to 256x256, 3 channels (RGB image). Input the CUT network model for training, the generator will generate the corrected results according to the distorted vortex beam intensity distribution, and the discriminator will distinguish it from the vortex beam intensity distribution diagram not disturbed by turbulence transmission. The discriminator should distinguish the real vortex beam intensity distribution diagram as true, and distinguish the light intensity distribution diagram generated by the generator as false; the purpose of the generator is to make the generated vortex beam intensity distribution diagram not disturbed by turbulence transmission be judged as true by the discriminator. Through the continuous confrontation between the generator and the discriminator, the corrected results generated by the generator are more and more close to the vortex beam intensity distribution not disturbed by turbulence transmission.

[0045] Figure 2 The image translation algorithm flowchart is shown in FIG. 1. After the training starts, the distorted light intensity map E o and the source field light intensity map E i are input, the corresponding feature information is automatically extracted by the CUT network, and the reconstructed light intensity map U o ​The loss function is calculated by using the source field intensity map and the reconstructed intensity map. If the loss function does not converge, the input intensity map is repeated, the loss function is calculated, and the training process is ended until convergence. After the loss function converges, the weight parameters in the model are fixed, the weight parameters are output and saved, and the trained model is obtained. The distorted light intensity after atmospheric turbulence transmission is input into the trained model, and the trained image translation algorithm model can give the generated corrected and compensated vortex beam intensity distribution.

[0046] Step (4) verifies the correction performance of the CUT network using the data of the test set, and the measurement standard is PSNR.

[0047] The standard is as follows:

[0048] Greater than 40 dB Very good 30 dB - 40 dB Good 20 dB - 30 dB Fair, barely acceptable Less than 20 dB Very poor, unacceptable

[0049] Effect reference Figure 4 、 Figure 5 .

[0050] The vortex light is transmitted through different intensity atmospheric turbulence media for 1km, and the correction result of the present scheme is as shown in Figure 4 From top to bottom, the atmospheric refractive index structure constant is respectively represent strong atmospheric turbulence, medium intensity turbulence, and weak turbulence; the received distorted light intensity is the first column, the second column is the corrected light intensity, and the third column is the lossless light intensity transmitted. When the atmospheric turbulence is strong turbulence, the PSNR of the corrected light intensity map and the lossless light intensity map is 44.6dB; under the condition of medium intensity turbulence, the PSNR of the corrected light intensity map and the lossless light intensity map is 43.6dB; under the condition of weak turbulence, the PSNR of the corrected light intensity map and the lossless light intensity map is 23.8dB, which is caused by overfitting of the model.

[0051] The vortex light is transmitted through different intensity atmospheric turbulence media for 1km, 2km, and 4km, and the correction result of the present scheme is as shown in Figure 5 From top to bottom, the transmission distance is 1km, 2km, and 4km; from left to right, the received distorted light intensity, the corrected light intensity, and the lossless light intensity at the sending end are shown. When the transmission distance is 1km, the PSNR of the corrected light intensity map and the lossless light intensity map is 43.6dB; when the transmission distance increases to 2km, the PSNR of the corrected light intensity map and the lossless light intensity map is 38.4dB; when the transmission distance is 4km, the PSNR of the corrected light intensity map and the lossless light intensity map is 40.4dB.

[0052] The above-described embodiments are only the preferred ones of the present application, and are not intended to limit the present application. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the present application. Therefore, any technical solutions obtained by equivalent replacement or equivalent transformation shall fall within the protection scope of the present application.

Claims

1. A method for direct correction of distorted vortex beams based on an image translation algorithm, characterized in that: Here are the steps: Step (1) modeling the transmission of the vortex beam in the atmospheric turbulent medium, using a light wave transmission numerical simulation method to obtain a simulated vortex beam intensity distribution map before transmission and a distorted intensity distribution map after transmission; The simulated vortex beam intensity distribution map is compiled into a data set, including a training set for training model parameters and a test set for testing model performance. When compiling the data set, there is no need to pair each distorted intensity distribution map with its corresponding intensity distribution map before transmission. Step (2), establishing an image translation algorithm model, using the vortex beam intensity distribution map before transmission and the distorted intensity distribution map after transmission in the training set as inputs to the image translation algorithm model for training, so that the network learns the mapping relationship from the source domain to the target domain; Step (3), inputting the test data set into the trained image translation algorithm model, the algorithm model corrects the distorted vortex beam intensity distribution map according to the mapping relationship between the source domain and the target domain, and directly generates an undistorted vortex beam intensity distribution map; The data set is obtained by establishing an atmospheric turbulence transmission environment simulated by multiple turbulence phase screens, calculating the diffracted light intensity through numerical simulation of light wave transmission to generate data; controlling the turbulence intensity and transmission distance respectively to generate intensity distribution maps of the vortex light beam distorted after transmission under different conditions; when preparing the data set, there is no need to match each distorted intensity distribution map with its corresponding intensity distribution map before transmission, thereby reducing the complexity of data set preparation and obtaining training and test sets; The image translation algorithm model is a variant of the generative adversarial network model, called the CUT network, which consists of a 15-layer generator and a 5-layer discriminator. The generator of the CUT network consists of a 15-layer network. The first three layers are convolutional layers, the middle nine layers are residual blocks, followed by two deconvolutional layers, and the last layer is a convolutional layer. The convolution kernel size used in the first and last convolutional layers is 7×7, and their stride is 1; the convolution kernel size used in the second and third convolutional layers is 3×3, and their stride is 2; the convolution kernel size of the nine residual blocks in the middle is 3×3, and their stride is 1; the convolution kernel size of the two deconvolutional layers is 3×3, and their stride is 2; The discriminator of the CUT network consists of a 5-layer network. The first 4 layers are convolutional layers combined with the previous instance normalization layer. The convolution kernel size used is 4×4, the stride is 2, and the ReLU function is used as the activation function. The last layer is a convolutional layer. The convolution kernel size used is 4×4, the stride is 1, and the Sigmoid function is used as the activation function.

2. The method for directly correcting distorted vortex beams based on an image translation algorithm according to claim 1, characterized in that: For the image data input into the image translation algorithm model for training, it needs to be pre-processed into 256×256, 3-channel RGB images.

3. The method for directly correcting distorted vortex beams based on an image translation algorithm according to claim 1, characterized in that: The data set includes data of different turbulence intensities and different transmission distances. When the transmission distance is fixed at 1 km, the atmospheric refractive index structure constants are Three cases, representing strong turbulence, medium turbulence and weak turbulence, respectively, collect 1000 distorted light intensity images and 1000 lossless light intensity images under each turbulence intensity; when the turbulence intensity is The transmission distances are 1km, 2km, and 4km respectively, and 1000 distorted light intensity images and 1000 lossless light intensity images are collected under each transmission distance; the dataset contains a total of 10,000 image data, of which 8,000 are training set data and 2,000 are test set data.

4. The method for directly correcting distorted vortex beams based on an image translation algorithm according to claim 3, characterized in that: The light intensity distribution map is a matrix in the image translation algorithm model. The purpose of training the image translation algorithm model is to learn and obtain the matrix mapping transformation relationship between the source domain and the target domain.

Images