Vortex light transmission wavefront correction method based on deep learning, medium and equipment
By directly training vortex optical distortion compensation through generative adversarial networks, the problems of high cost of optical hardware and training difficulties in existing technologies are solved, and high-speed communication without wavefront reconstruction is achieved.
Patent Information
- Application Number
- CN202210882648.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-07-26
- Publication Date
- 2026-02-13
- Estimated Expiration
- 2042-07-26
AI Technical Summary
Existing deep learning-based vortex wavefront correction techniques require probe beams and numerous optical hardware resources, and require wavefront reconstruction before correction, resulting in high costs and training difficulties, and failing to meet the needs of high-speed communication.
Generative adversarial networks are used for end-to-end turbulence distortion compensation. The mapping relationship between the distorted OAM intensity map and the target OAM intensity map is directly trained, skipping wavefront reconstruction, and solving the turbulence distortion compensation problem through image style transfer.
It saves optical hardware resources, improves compensation rate, meets high-speed transmission requirements, and simplifies the training process without requiring wavefront reconstruction.
Smart Images
Figure CN115239552B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of optical communication, and more specifically, to a deep learning-based method, medium, and device for wavefront correction of vortex optical transmission. Background Technology
[0002] In free-space vortex optical communication systems, the orbital angular momentum (OAM) vortex intensity map captured at the receiver is closely related to the spiral wavefront distribution. Free-space atmospheric turbulence introduces a perturbation phase into the vortex beam during transmission, leading to wavefront distortion and intermodal crosstalk over a certain distance. This results in dispersion of the spiral spectrum distribution at the receiver, severely impacting optical communication performance. To improve the robustness of the vortex optical communication system to atmospheric turbulence distortion and ensure high-precision OAM mode identification, dynamic compensation of the distorted beam is necessary. Summary of the Invention
[0003] The purpose of this invention is to overcome the shortcomings of the prior art and provide a deep learning-based vortex optical transmission wavefront correction method, medium and device that can still ensure excellent distortion compensation performance without reconstructing the wavefront and saving optical hardware.
[0004] The objective of this invention is achieved through the following solution:
[0005] A deep learning-based wavefront correction method for vortex optical transmission includes the following steps:
[0006] S1, obtain orbital angular momentum intensity maps before and after turbulence distortion to obtain training samples;
[0007] S2, Construct a deep learning model, input training samples into the deep learning model for training; the deep learning model includes a generative adversarial network, based on the end-to-end distortion compensation of the generative adversarial network, skips wavefront reconstruction, directly trains the mapping relationship between the distorted OAM intensity map and the target OAM intensity map, treats the two images as two image styles, transforms turbulence distortion compensation into an image style transfer problem, and obtains the mapping relationship between the distorted OAM intensity map and the corresponding target OAM intensity map according to the trained deep learning model;
[0008] S3, after training, the deep learning model directly outputs the compensated orbital angular momentum intensity map, which can ensure the wavefront correction distortion compensation performance of vortex light transmission without wavefront reconstruction and by saving optical hardware.
[0009] Further, in step S1, the training samples are obtained in the following way: First, a vortex optical communication system is constructed, and experimental data is generated through physical formula simulation; then, an OAM state set with a unified image style is selected, and the distorted OAM light intensity map captured by the receiver and its corresponding undistorted target OAM light intensity map are combined into an image pair to obtain the training samples for deep learning; the unified image style includes a petal-shaped distribution image style.
[0010] Furthermore, in step S2, the generative adversarial network includes a generator and a discriminator.
[0011] Furthermore, in step S2, the input of the deep learning model is the distorted OAM light intensity map captured by the receiver, and the output is the corresponding compensated OAM light intensity map.
[0012] Further, step S2 includes the following sub-steps: uniformly adjusting the size of the OAM light intensity map to 256*256*3, horizontally flipping the input image with a set probability, and uniformly distributing and initializing the network weight coefficients using the Xavier method.
[0013] Further, in step S2, when the generative adversarial network is a Pix2pix network, the following sub-steps are included:
[0014] The Pix2pix network consists of a generator with 16 layers and a discriminator with 5 layers.
[0015] A set of OAM states with a uniform image style is selected. The set of OAM states consists of 10 superimposed OAM states: {1, -1}, {2, -2}, {3, -3}, {4, -4}, {5, -5}, {6, -6}, {1, -2}, {2, -3}, {3, -4}, and {4, -5}. The distorted OAM intensity map captured by the receiver is quantized to form a matrix, which is used as the input of the Pix2pix network.
[0016] The corresponding compensated OAM intensity map is quantized to form another matrix, which is used as the output of the Pix2pix network.
[0017] The Pix2pix network is trained based on the input and output matrices to obtain the mapping relationship between the distorted OAM intensity map and the corresponding target OAM intensity map.
[0018] Further, in step S2, when the generative adversarial network is a CycleGAN network, the following sub-steps are included:
[0019] The CycleGAN network consists of a generator with 16 layers and a discriminator with 5 layers.
[0020] A set of OAM states with a unified image style is selected. The set of OAM states consists of 10 superimposed OAM states: {1, -1}, {2, -2}, {3, -3}, {4, -4}, {5, -5}, {6, -6}, {1, -2}, {2, -3}, {3, -4}, and {4, -5}. The distorted OAM intensity map captured by the receiver is quantized to form a matrix, which is used as the input of the CycleGAN network.
[0021] The corresponding compensated OAM intensity map is quantized to form another matrix, which is used as the output of the CycleGAN network.
[0022] The CycleGAN network is trained based on the input matrix and the output matrix to obtain the mapping relationship between the distorted OAM intensity map and the corresponding target OAM intensity map.
[0023] Furthermore, the generation of experimental data through physical formula simulation includes using random phase screen simulation to simulate atmospheric turbulence and generating experimental data through phase screen theoretical simulation.
[0024] A readable storage medium includes a memory on which a program is stored, which, when loaded by a processor, executes the deep learning-based vortex optical transmission wavefront correction method as described above.
[0025] A computer device includes a memory and a processor, wherein a program is stored in the memory, and when the program is loaded by the processor, it executes the deep learning-based vortex optical transmission wavefront correction method as described above.
[0026] The beneficial effects of this invention include:
[0027] This invention solves the technical problems of existing deep learning-based vortex wavefront correction techniques, which require probe beams and numerous optical hardware resources, as well as the need to reconstruct the wavefront before correction. The proposed end-to-end turbulence distortion compensation method based on generative adversarial networks directly outputs the compensated OAM intensity map, solving the problems of high cost of optical systems and training difficulties caused by large differences between the intensity distribution and the wavefront phase image, while also meeting the requirements of high-speed transmission.
[0028] The embodiments of the present invention not only save optical hardware resources, but also eliminate the need to reconstruct the wavefront before wavefront correction, directly outputting the compensated OAM intensity map, further improving the compensation rate and meeting the requirements of high-speed transmission.
[0029] Since the distorted OAM intensity map is similar to the target OAM intensity map, the training in this embodiment is simpler and more efficient than the mapping relationship between the training intensity distribution and the wavefront phase.
[0030] Methods that reconstruct wavefronts based on Zernike coefficients calculated using convolutional neural networks may overlook much high-frequency information and cannot be applied to non-circular aperture reconstruction, thus limiting wavefront reconstruction capabilities. The U-Net architecture used in Pix2pix and CycleGAN embodiments of this invention is more suitable for phase retrieval tasks. Attached Figure Description
[0031] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0032] Figure 1 This is a flowchart illustrating the steps of the vortex optical distortion compensation method based on generative adversarial networks in an embodiment of the present invention.
[0033] Figure 2 This is a schematic diagram of the vortex optical distortion compensation technology based on Pix2pix in an embodiment of the present invention;
[0034] Figure 3 This is a schematic diagram of the vortex optical distortion compensation technology based on CycleGAN in an embodiment of the present invention;
[0035] Figure 4 This is a graph showing the variation of MSSIM during the training process under different turbulence intensities according to an embodiment of the present invention.
[0036] The English definitions in the image are as follows: Skip connection; Generator; Discriminator; Ground truth; Fake / Real; Loss. Detailed Implementation
[0037] All features disclosed in all embodiments of this specification, or steps in all methods or processes implied in the disclosure, may be combined and / or extended or replaced in any way, except for mutually exclusive features and / or steps.
[0038] The inventors of this invention went through the following process in researching the problem and searching for corresponding solutions:
[0039] First, the inventors of this invention recognized that traditional vortex wavefront correction techniques are mainly divided into two categories. The first category is a vortex wavefront correction system with a probe beam and a wavefront sensor. At the transmitting end, a fundamental mode Gaussian beam is added as a probe beam and coaxially transmitted with the vortex beam. At the receiving end, the probe beam is separated by a polarization beam splitter (PBS) and input to a wavefront detector. A charge-coupled device (CCD), such as a CCD camera, captures the bright spot distribution formed by the convergence of the Gaussian beam through a microlens array. The wavefront distribution of the Gaussian beam can be inferred from the spot distribution. Since the separated Gaussian beam and the vortex beam pass through the same atmospheric turbulence channel and have the same wavefront distortion, the wavefront corrector can be controlled to compensate for the distortion of the vortex beam based on the wavefront distribution of the probe beam. The second type is a wavefront correction system without probe beams and wavefront sensors. After being affected by atmospheric turbulence distortion, the wavefront corrector is controlled by memoryless iterative algorithms such as simulated annealing, stochastic parallel gradient descent, and genetic algorithms to correct the OAM light intensity distribution map captured by the CCD camera. The entire working process forms a closed loop.
[0040] Existing deep learning-based vortex wavefront correction techniques are mostly improvements on the first type of vortex wavefront correction system mentioned above. They do not use a wavefront detector to reconstruct the probe beam wavefront distribution, but directly train the mapping relationship between the Gaussian intensity map captured by the CCD camera and the turbulence (Zernike turbulence intensity coefficient or turbulence phase map). At the receiving end, the fundamental mode Gaussian intensity map captured by the CCD camera is input into the algorithm model, which outputs the wavefront phase distribution, and uses this to control the wavefront corrector to compensate for distortion in the vortex beam.
[0041] In 2019, Tian Qinghua's team at Beijing University of Posts and Telecommunications used a convolutional neural network to map distributed features learned from the OAM light intensity distribution to a sample label space, outputting the first 2-400 Zernike coefficients and converting them into control signals for spatial light modulators for correction. Also in 2019, Liu's team proposed a CNN-based scheme for predicting turbulent phase screens, training the mapping relationship between the intensity distribution of vortex beams and the turbulent phase. This scheme demonstrated strong generalization ability and could predict turbulent phase screens not present in the training set. In 2020, Gao Chunqing's team at Beijing Institute of Technology proposed a CNN-based wavefront reconstruction technique, training the mapping relationship between the vortex light intensity map and the first 20 Zernike coefficients. The core of these deep learning-based wavefront correction techniques is the wavefront reconstruction algorithm, which reconstructs the wavefront based on a probe beam by training the mapping relationship between the probe beam intensity map and the turbulent Zernike coefficients.
[0042] The core of the aforementioned deep learning-based wavefront correction scheme lies in the indirect, non-iterative wavefront reconstruction algorithm, which only achieves end-to-end wavefront reconstruction but does not achieve end-to-end distortion compensation.
[0043] Its shortcomings are as follows:
[0044] (1) Whether it is the coaxial transmission of vortex beam and probe beam at the transmitter, the separation of probe beam by the beam splitter at the receiver, or the control of the spatial light modulator to add reverse wavefront phase, a lot of optical hardware is used, which requires a complex optical system and is expensive.
[0045] (2) It requires multi-step calculations, including wavefront reconstruction and wavefront correction, which hinders further improvement of communication speed.
[0046] (3) Due to the large difference between the light intensity distribution and the wavefront phase image, it is not easy to ensure the correspondence between the light intensity distribution and the wavefront phase when training the mapping relationship between the light intensity distribution and the wavefront phase, making training difficult.
[0047] To address the challenges of existing deep learning-based vortex wavefront correction techniques, which require probe beams, extensive optical hardware resources, and wavefront reconstruction before correction, this invention proposes a generative adversarial network-based end-to-end turbulence distortion compensation solution. This solution directly outputs the compensated OAM intensity map, aiming to solve the problems of high optical system costs and training difficulties caused by significant differences between the intensity distribution and the wavefront phase image, while also meeting the requirements for high-speed communication transmission.
[0048] The inventive concept of this invention includes:
[0049] 1) By eliminating wavefront reconstruction, the mapping relationship between the distorted OAM intensity map and the target OAM intensity map can be directly trained, which simplifies the process and better meets the high-speed transmission requirements of communication.
[0050] 2) Treat the two images of vortex light before and after turbulent distortion as two different styles, and transform the turbulent distortion compensation problem into a style conversion problem.
[0051] In the specific implementation process, the steps are as follows: Atmospheric turbulence can be simulated using random phase screen simulation to obtain orbital angular momentum intensity maps (OAM) before and after turbulence distortion, thus obtaining training samples; based on end-to-end distortion compensation technology using generative adversarial networks, wavefront reconstruction is skipped, treating the two images as two different image styles, transforming turbulence distortion compensation into an image style transfer problem; the mapping relationship between the distorted OAM intensity map and the corresponding target OAM intensity map is obtained based on the trained deep learning model; the trained model can directly output the compensated orbital angular momentum intensity map. This approach ensures wavefront correction distortion compensation performance for vortex light transmission without requiring wavefront reconstruction and saving optical hardware.
[0052] In an embodiment, such as Figure 1 As shown, it may include the following steps:
[0053] Step 1: Construct training samples.
[0054] Step 2: Build a deep learning model.
[0055] Step 3: Input the training set into the deep learning model for training.
[0056] Step 4: Obtain the mapping relationship between the distorted OAM intensity map and the corresponding target OAM intensity map based on the trained deep learning model.
[0057] Step 1 includes the following sub-steps: The training set samples are obtained as follows. First, a vortex optical communication system is constructed, and experimental data is generated through physical formula simulation. Then, a set of OAM states with a uniform image style (petal-shaped distribution) is selected, and the distorted OAM light intensity map captured by the receiver and its corresponding undistorted target OAM light intensity map are combined to form image pairs, thus obtaining the training samples for deep learning.
[0058] Step 2 includes the sub-step: the deep learning model is a generative adversarial network consisting of a generator and a discriminator.
[0059] Step 3 includes the following sub-steps: The input to the deep learning model is the distorted OAM light intensity map captured by the receiver, and the output is the corresponding compensated OAM light intensity map. The size of the OAM light intensity map is uniformly adjusted to 256*256*3. The deep learning model is trained based on this input.
[0060] Step 4 includes the sub-step of obtaining the mapping relationship between the distorted OAM intensity map and the corresponding target OAM intensity map based on the trained deep learning model.
[0061] In one embodiment of the present invention, a vortex optical distortion compensation technology based on Pix2pix is provided. Figure 2 The diagram below illustrates a Pix2pix-based vortex optical distortion compensation technique, including the process of constructing a Pix2pix network to implement the method of this invention as follows:
[0062] Step 1, construct training samples, which includes the following sub-steps:
[0063] Step 1a): Construct a vortex optical communication system and generate experimental data through physical formula simulation;
[0064] Step 1b): Select a set of OAM states with a uniform image style (petal-shaped distribution). In this embodiment, 10 OAM superposition states were selected: {1, -1}, {2, -2}, {3, -3}, {4, -4}, {5, -5}, {6, -6}, {1, -2}, {2, -3}, {3, -4}, and {4, -5}.
[0065] Step 1c): Combine the distorted OAM light intensity map captured by the receiver with its corresponding undistorted target OAM light intensity map to form an image pair, thus obtaining training samples for deep learning.
[0066] Step 2, constructing a deep learning model, specifically includes the following sub-steps:
[0067] Step 2a): The deep learning model is a generative adversarial network Pix2pix, consisting of a generator and a discriminator;
[0068] Step 2b): Each sub-layer of the generator and discriminator of Pix2pix consists of three parts: a convolutional layer, a batch normalization layer, and an activation function layer. The kernel size of the convolutional layer is 4*4.
[0069] Step 2c): The generator contains a total of 16 layers and adopts a structure similar to U-net. Among them, layers E1 to E8 perform downsampling operations on the input OAM light intensity distribution features through convolution, and the inter-layer activation function is LeakyReLU.
[0070] Step 2d): Layers D8 to D1 perform upsampling on the input OAM light intensity distribution features through deconvolution, and ReLU is selected as the inter-layer activation function;
[0071] Step 2e): The corresponding upsampling E layer and downsampling D layer (e.g., E1 and D1 layers) adopt a skip connection to pass the high-frequency information learned in the encoder layer to the corresponding decoder layer, so as to prevent the loss of important high-frequency image information during the downsampling process.
[0072] Step 2f): The OAM intensity map output by the generator is normalized and then sent to the discriminator to determine its authenticity in conjunction with the target OAM intensity map.
[0073] Step 2g): The discriminator consists of 5 layers, and the LeakyReLU function is selected as the inter-layer activation function.
[0074] Step 3, input the training set into the deep learning model for training, which includes the following sub-steps:
[0075] Step 3a): The input to the deep learning model is the distorted OAM light intensity map captured by the receiver, and the output is the corresponding compensated OAM light intensity map.
[0076] Step 3b): The size of the OAM light intensity map is uniformly adjusted to 256*256*3, and the input image is horizontally flipped with a certain probability. The network weight coefficients are initialized uniformly using the Xavier method.
[0077] Step 3c): The input to the E1 layer is 256*256*3. After the "convolution-batch normalization-function activation" operation, the size of the output feature map is 128*128*3.
[0078] Step 3d): From layer E2 to layer E8, the feature image size is halved with each layer. The final output feature map size of layer E8 is 1*1*256, which is then input into layer D8. From layer D8 to layer D1, the feature image size is halved with each layer. The final output size of layer D8 is 256*256*3, restoring the original image size.
[0079] Step 3e): The OAM light intensity map generated by the generator is input into the discriminator. The first four layers of the discriminator are used to extract features. Then, a convolutional layer with a kernel size of 4*4 is used to perform downsampling to reduce the resolution of the feature map to 1*1. The final output vector has a size of 1*1*2, which represents the discrimination result.
[0080] Step 4: Obtain the mapping relationship between the distorted OAM intensity map and the corresponding target OAM intensity map based on the trained deep learning model.
[0081] In another embodiment of the present invention, a vortex optical distortion compensation technique based on CycleGAN (Generative Recurrent Adversarial Network) is provided. Figure 3 A schematic diagram of vortex optical distortion compensation technology based on CycleGAN is given. For details not disclosed in this embodiment, please refer to the relevant steps in other embodiments. The difference is that for vortex optical turbulence distortion compensation, the image x in the distortion domain X is the distorted OAM light intensity map captured by the receiver, and the image y in the target domain Y is the undistorted target OAM light intensity map.
[0082] During training, the distorted OAM intensity map x is first passed through the generator G in the X-domain to Y-domain direction, and then through the generator F in the Y-domain to X-domain to obtain the reconstructed OAM intensity map F(G(x)) in the distortion domain, making it as close as possible to the original image x.
[0083] The same process applies to the migration from the target domain Y to the distortion domain X, making the reconstructed target OAM intensity map as close as possible to the original image y, with the reconstruction loss based on L1 loss.
[0084] Two unidirectional GAN discriminators distinguish between the reconstructed image and the original image within their respective domains.
[0085] The technical effects of the embodiments of the present invention can be further illustrated by the following simulations: Figure 4 As shown, this example's mixed dataset contains 10,000 OAM light intensity images, divided into training and test sets in an 8:2 ratio. The training set contains 8,000 images, and the test set contains 2,000 images. The image size is 256*256*3.
[0086] The mean structural similarity (MSSIM) was used to evaluate the compensation effects of the Pix2pix-based and CycleGAN-based vortex optical distortion compensation techniques. Under both turbulent environments, the MSSIM of the compensated OAM intensity map gradually increased during training, approaching 1 after training, indicating that the compensated OAM intensity map gradually approached the target OAM intensity map. The comparison shows that the compensated OAM intensity map is very close to the target OAM intensity map, indicating a good distortion compensation effect.
[0087] Example 1
[0088] A deep learning-based wavefront correction method for vortex optical transmission includes the following steps:
[0089] S1, obtain orbital angular momentum intensity maps before and after turbulence distortion to obtain training samples;
[0090] S2, Construct a deep learning model, input training samples into the deep learning model for training; the deep learning model includes a generative adversarial network, based on the end-to-end distortion compensation of the generative adversarial network, skips wavefront reconstruction, directly trains the mapping relationship between the distorted OAM intensity map and the target OAM intensity map, treats the two images as two image styles, transforms turbulence distortion compensation into an image style transfer problem, and obtains the mapping relationship between the distorted OAM intensity map and the corresponding target OAM intensity map according to the trained deep learning model;
[0091] S3, after training, the deep learning model directly outputs the compensated orbital angular momentum intensity map, which can ensure the wavefront correction distortion compensation performance of vortex light transmission without wavefront reconstruction and by saving optical hardware.
[0092] Example 2
[0093] Based on Example 1, in step S1, the training samples are obtained in the following way: First, a vortex optical communication system is constructed, and experimental data is generated through physical formula simulation; then, a set of OAM states with a unified image style is selected, and the distorted OAM light intensity map captured by the receiver and its corresponding undistorted target OAM light intensity map are combined into image pairs to obtain training samples for deep learning; the unified image style includes a petal-shaped distribution image style.
[0094] Example 3
[0095] Based on Example 1, in step S2, the generative adversarial network includes a generator and a discriminator.
[0096] Example 4
[0097] Based on Example 1, in step S2, the input of the deep learning model is the distorted OAM light intensity map captured by the receiver, and the output is the corresponding compensated OAM light intensity map.
[0098] Example 5
[0099] Based on Example 1, step S2 includes the following sub-steps: uniformly adjusting the size of the OAM light intensity map to 256*256*3, horizontally flipping the input image with a set probability, and uniformly distributing and initializing the network weight coefficients using the Xavier method.
[0100] Example 6
[0101] Based on Example 1, in step S2, when the generative adversarial network is a Pix2pix network, the following sub-steps are included:
[0102] The Pix2pix network consists of a generator with 16 layers and a discriminator with 5 layers.
[0103] A set of OAM states with a uniform image style is selected. The set of OAM states consists of 10 superimposed OAM states: {1, -1}, {2, -2}, {3, -3}, {4, -4}, {5, -5}, {6, -6}, {1, -2}, {2, -3}, {3, -4}, and {4, -5}. The distorted OAM intensity map captured by the receiver is quantized to form a matrix, which is used as the input of the Pix2pix network.
[0104] The corresponding compensated OAM intensity map is quantized to form another matrix, which is used as the output of the Pix2pix network.
[0105] The Pix2pix network is trained based on the input and output matrices to obtain the mapping relationship between the distorted OAM intensity map and the corresponding target OAM intensity map.
[0106] Example 7
[0107] Based on Example 1, when the generative adversarial network is a CycleGAN network in step S2, the following sub-steps are included:
[0108] The CycleGAN network consists of a generator with 16 layers and a discriminator with 5 layers.
[0109] A set of OAM states with a unified image style is selected. The set of OAM states consists of 10 superimposed OAM states: {1, -1}, {2, -2}, {3, -3}, {4, -4}, {5, -5}, {6, -6}, {1, -2}, {2, -3}, {3, -4}, and {4, -5}. The distorted OAM intensity map captured by the receiver is quantized to form a matrix, which is used as the input of the CycleGAN network.
[0110] The corresponding compensated OAM intensity map is quantized to form another matrix, which is used as the output of the CycleGAN network.
[0111] The CycleGAN network is trained based on the input matrix and the output matrix to obtain the mapping relationship between the distorted OAM intensity map and the corresponding target OAM intensity map.
[0112] Example 8
[0113] Based on Example 2, the generation of experimental data through physical formula simulation includes using random phase screen simulation to simulate atmospheric turbulence and generating experimental data through phase screen theoretical simulation.
[0114] Example 9
[0115] A readable storage medium includes a memory on which a program is stored, which, when loaded by a processor, executes the deep learning-based vortex optical transmission wavefront correction method as described in any one of Examples 1 to 8.
[0116] Example 10
[0117] A computer device includes a memory and a processor, wherein a program is stored in the memory, and when the program is loaded by the processor, it executes the deep learning-based vortex optical transmission wavefront correction method as described in any one of Examples 1 to 8.
[0118] The units described in the embodiments of the present invention can be implemented in software or hardware, and the described units can also be located in a processor. The names of these units do not necessarily limit the specific unit itself.
[0119] According to one aspect of this application, a computer program product or computer program is provided, comprising computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the methods provided in the various alternative implementations described above.
[0120] In another aspect, this application also provides a computer-readable medium, which may be included in the electronic device described in the above embodiments; or it may exist independently and not assembled into the electronic device. The computer-readable medium carries one or more programs, which, when executed by the electronic device, cause the electronic device to perform the methods described in the above embodiments.
[0121] All parts not covered in this invention are the same as or can be implemented using existing technologies.
[0122] The above technical solution is only one embodiment of the present invention. For those skilled in the art, based on the application methods and principles disclosed in the present invention, it is easy to make various types of improvements or modifications, and not limited to the methods described in the above specific embodiments of the present invention. Therefore, the methods described above are only preferred and are not restrictive.
[0123] In addition to the examples above, other embodiments may be obtained by those skilled in the art based on the above disclosure or by making modifications using knowledge or technology in related fields. The features of each embodiment may be interchanged or replaced. Modifications and changes made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.
Claims
1. A deep learning based vortex light transmission wavefront correction method, characterized in that, The method comprises the following steps: S1, obtaining an undistorted and distorted orbital angular momentum (OAM) light intensity map to obtain a training sample; S2, constructing a deep learning model, inputting the training sample into the deep learning model for training; the deep learning model comprises a generative adversarial network (GAN), and the GAN is used for end-to-end distortion compensation, skipping wavefront reconstruction, directly training a mapping relationship between a distorted OAM light intensity map and a target OAM light intensity map, regarding the two kinds of images as two image styles, converting the distortion compensation into an image style conversion problem, and obtaining the mapping relationship between the distorted OAM light intensity map and the corresponding target OAM light intensity map according to the trained deep learning model; S3, directly outputting a compensated OAM light intensity map by the trained deep learning model, so that the performance of vortex light transmission wavefront correction and distortion compensation can be ensured without wavefront reconstruction and optical hardware. In step S2, when the GAN is a Pix2pix network, the following sub-steps are included: The Pix2pix network comprises a generator comprising a 16-layer network and a discriminator comprising a 5-layer network; An OAM state set with a unified image style is selected, the OAM state set is {1, -1}, {2, -2}, {3, -3}, {4, -4}, {5, -5}, {6, -6}, {1, -2}, {2, -3}, {3, -4}, {4, -5}, and a distorted OAM light intensity map captured at a receiving end is quantized to form a matrix as an input of the Pix2pix network; A corresponding compensated OAM light intensity map is quantized to form another matrix as an output of the Pix2pix network; The Pix2pix network is trained according to the input matrix and the output matrix, so that a mapping relationship between the distorted OAM light intensity map and the corresponding target OAM light intensity map is obtained. When the GAN is a CycleGAN network, the following sub-steps are included: The CycleGAN network comprises a generator comprising a 16-layer network and a discriminator comprising a 5-layer network; An OAM state set with a unified image style is selected, the OAM state set is {1, -1}, {2, -2}, {3, -3}, {4, -4}, {5, -5}, {6, -6}, {1, -2}, {2, -3}, {3, -4}, {4, -5}, and a distorted OAM light intensity map captured at a receiving end is quantized to form a matrix as an input of the CycleGAN network; A corresponding compensated OAM light intensity map is quantized to form another matrix as an output of the CycleGAN network; The CycleGAN network is trained according to the input matrix and the output matrix, so that a mapping relationship between the distorted OAM light intensity map and the corresponding target OAM light intensity map is obtained.
2. The deep learning based vortex optical transmission wavefront correction method according to claim 1, wherein, In step S1, the samples of the training samples are obtained by: firstly, constructing a vortex optical communication system, generating experimental data through physical formula simulation; then, selecting an OAM mode set with a unified image style, and forming an image pair by combining the distorted OAM light intensity map captured by the receiving end and the corresponding undistorted target OAM light intensity map, to obtain the training samples for deep learning; the unified image style includes a petal-shaped distribution image style.
3. The deep learning based vortex optical transmission wavefront correction method according to claim 1, wherein, In step S2, the generative adversarial network includes a generator and a discriminator.
4. The deep learning based vortex optical transmission wavefront correction method according to claim 1, wherein, In step S2, the input of the deep learning model is the distorted OAM light intensity map captured by the receiving end, and the output is the corresponding compensated OAM light intensity map.
5. The deep learning based vortex optical transmission wavefront correction method according to claim 1, wherein, In step S2, it includes a sub-step of: uniformly adjusting the size of the OAM light intensity map to 256*256*3, horizontally flipping the input image at a set probability, and initializing the network weight coefficients uniformly using the Xavier method.
6. The deep learning based vortex optical transmission wavefront correction method according to claim 2, wherein, The experimental data is generated by physical formula simulation, which includes simulating atmospheric turbulence by random phase screen simulation, and generating experimental data by phase screen theory simulation.
7. A readable storage medium, characterized by, The computer device comprises a memory, and a program is stored on the memory, and when the program is loaded by the processor, the program executes the vortex optical transmission wavefront correction method based on deep learning as claimed in any one of claims 1-6.
8. A computer device, comprising: The computer device comprises a memory and a processor, and a program is stored on the memory, and when the program is loaded by the processor, the program executes the vortex optical transmission wavefront correction method based on deep learning as claimed in any one of claims 1-6.