Orbital angular momentum modal recognition method and system based on transfer learning
Patent Information
- Application Number
- CN202410490917.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-04-23
- Publication Date
- 2026-10-30
- Estimated Expiration
- 2044-04-23
AI Technical Summary
然而，在自由空间信道传输时，轨道角动量模态会受到大气湍流的影响，并且大气湍流环境是在多变、不确定的，接收端解码时导致在未知大气湍流强度下不能有效识别轨道角动量模态信息
[0028]本发明将CBAM注意力机制模块引入到ResNet50网络模型中，结合迁移学习的方式对轨道角动量模态进行识别，这种方法能够有效识别在未知大气湍流强度中传输后的涡旋光束轨道角动量模态，并且训练时间短，识别准确率高，具有较高的泛化能力，提高了自由空间光通信性能。
Smart Images
Figure CN118365845B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of free-space optical communication technology and relates to a method and system for orbital angular momentum mode recognition based on transfer learning. Background Technology
[0002] With the development of science and technology, people have fully entered the information age. In this information world, information transmission systems are a crucial component. As the amount of information transmitted continues to increase, larger-capacity communication systems are needed.
[0003] Orbital angular momentum (OAM) is an inherent property of vortex beams with helical wavefronts, and information is transmitted by establishing a mapping relationship between OAM modes and digital signals. However, during transmission in free-space channels, OAM modes are affected by atmospheric turbulence, and the atmospheric turbulence environment is variable and uncertain. This leads to the inability to effectively identify OAM mode information at the receiver when the intensity of atmospheric turbulence is unknown. Therefore, it is of great significance to study a method that can effectively identify OAM modes after transmission under unknown atmospheric turbulence intensity. Summary of the Invention
[0004] The purpose of this invention is to provide a method and system for identifying orbital angular momentum modes based on transfer learning, which addresses the generalization capability of vortex beam identification. By combining transfer learning with an improved ResNet50 network model, an optimal model is trained. The optimal model is then used to identify the orbital angular momentum modes of vortex beams transmitted under unknown atmospheric turbulence intensities, effectively improving the identification accuracy of orbital angular momentum modes of vortex beams transmitted under unknown atmospheric turbulence intensities.
[0005] The first aspect of this invention provides a method for orbital angular momentum mode recognition based on transfer learning, the method specifically comprising:
[0006] Obtain the Laguerre-Gaussian beam dataset, which contains orbital angular momentum mode information;
[0007] Perform image preprocessing on the dataset;
[0008] Construct a transfer learning model, which is an improved ResNet50 network model;
[0009] The preprocessed training set is input into the transfer learning model for training. The performance of the trained model is verified by the validation set, and the optimal model is saved.
[0010] Obtain a test set, test the optimal model using the test set, and identify the orbital angular momentum mode of the vortex beam.
[0011] The acquisition of the Laguerre-Gaussian beam dataset includes:
[0012] The intensity distribution map of a Laguerre-Gaussian beam with different orbital angular momentum modes was generated by numerical simulation. The Laguerre-Gaussian beam with different orbital angular momentum modes is a Laguerre-Gaussian beam composed of 16 different orbital angular momentum modes with a radial exponent of 0 and a topological charge of {-4, -1, +2, +5}.
[0013] Several random phase screens are generated based on the power spectrum inversion method. Different atmospheric turbulence intensities are represented by changing the atmospheric refractive index structure constant. Random phase screens are inserted into the propagation path of the Laguerre-Gaussian beam to simulate the atmospheric turbulence channel. The transmitted Laguerre-Gaussian beam dataset is obtained through the atmospheric turbulence channel, and the dataset is divided into training set and validation set in a 7:3 ratio.
[0014] The power spectrum inversion method includes:
[0015] Generate a complex Gaussian random matrix C N×N In the frequency domain, the mean is 0 and the variance is 1; according to the modified von Karman spectral model φ n (κ) yields the phase perturbation σ(κ); using the phase perturbation σ(κ) to adjust C N×N Filtering is performed to obtain the matrix. right Performing an inverse Fourier transform and then taking the real part yields the atmospheric turbulence phase screen.
[0016] The image preprocessing of the dataset includes:
[0017] The images in the training set are randomly cropped to 224×224 pixels and horizontally rotated to increase the diversity of the training data and improve the robustness and generalization ability of the model. The images in the validation set are cropped at the center of the image to a size of 224×224. The images in the training set and validation set are converted into tensor data types. The images in the training set and validation set are then normalized.
[0018] The improved ResNet50 network model is based on the ResNet50 network model trained on the ImageNet dataset, which introduces the CBAM attention mechanism module. The CBAM attention mechanism module is embedded between the fourth block and the average pooling layer in the ResNet50 network model. The CBAM attention mechanism module can improve the model's ability to perceive important features in the input data, thereby improving the model's recognition accuracy.
[0019] The step of inputting the preprocessed training set into the transfer learning model for training includes:
[0020] The improved ResNet50 network model is loaded, and the fully connected layers in the network model are modified so that the output of the fully connected layers corresponds to the number of Laguerre-Gaussian beam categories. Combined with transfer learning, the parameters of all layers above the CBAM attention mechanism module are frozen so that only the parameters in the CBAM attention mechanism module and the fully connected layers are updated during training.
[0021] The process of testing the optimal model with a test set involves inputting the test set into the optimal model for prediction, and using a confusion matrix to visualize the prediction results.
[0022] A second aspect of the present invention provides an orbital angular momentum mode recognition system based on transfer learning, the system specifically comprising:
[0023] The acquisition module is used to acquire the intensity distribution map of the Laguerre-Gaussian beam after transmission in the atmospheric turbulence channel;
[0024] The preprocessing module is used to perform preprocessing operations on the acquired Laguerre-Gaussian beam intensity distribution map;
[0025] The training module is used to input the preprocessed Laguerre-Gaussian beam intensity distribution map into the transfer learning model for training to obtain the optimal model.
[0026] The identification module is used to input the real-time acquired Laguerre-Gaussian beam intensity distribution map after transmission through the atmospheric turbulence channel into the optimal model for orbital angular momentum mode identification.
[0027] The beneficial effects of this invention are as follows:
[0028] This invention introduces the CBAM attention mechanism module into the ResNet50 network model and combines it with transfer learning to identify orbital angular momentum modes. This method can effectively identify the orbital angular momentum modes of vortex beams transmitted in unknown atmospheric turbulence intensities. It also has short training time, high recognition accuracy, and high generalization ability, thus improving the performance of free-space optical communication. Attached Figure Description
[0029] Figure 1 This is a flowchart illustrating a method for identifying orbital angular momentum modes based on transfer learning according to the present invention.
[0030] Figure 2 The intensity distribution of a Laguerre-Gaussian beam with 16 superimposed orbital angular momentum modes is shown.
[0031] Figure 3 shows the curves of the recognition accuracy of CNN1 and CNN3, and CNN2 and CNN3 on the validation set as a function of training rounds.
[0032] Figure 4 shows the confusion matrices of CNN1, CNN2, and CNN3, respectively.
[0033] Figure 5 This is a schematic diagram of the structure of an orbital angular momentum mode recognition system based on transfer learning according to the present invention. Detailed Implementation
[0034] To better illustrate the present invention, a detailed description is now provided in conjunction with specific embodiments and accompanying drawings. For ease of description, the original ResNet50 network model is referred to as CNN1, the ResNet50 network model based on transfer learning is referred to as CNN2, and the improved ResNet50 network model based on transfer learning is referred to as CNN3. The training method of CNN2 is to update the parameters of the fully connected layers in the ResNet50 network model only during training.
[0035] This invention discloses a method and system for identifying orbital angular momentum modes based on transfer learning. The method includes the following steps: acquiring a Laguerre-Gaussian beam dataset, wherein the vortex beam contains orbital angular momentum mode information; preprocessing the dataset; constructing a transfer learning model, wherein the transfer learning model is an improved ResNet50 network model; inputting the preprocessed training set into the transfer learning model for training, verifying the performance of the trained model using a validation set, and saving the optimal model; acquiring a test set, testing the optimal model using the test set, and identifying the orbital angular momentum modes of the vortex beam. This invention can effectively identify the orbital angular momentum modes of vortex beams transmitted in unknown atmospheric turbulence intensities, and has a short training time, high recognition accuracy, and high generalization ability, thus improving the performance of free-space optical communication.
[0036] Example 1:
[0037] like Figure 1 The diagram shows a flowchart illustrating a transfer learning-based method for identifying orbital angular momentum modes. This method effectively identifies orbital angular momentum modes transmitted under unknown atmospheric turbulence intensity, has a short training time, high accuracy, and high generalization ability. The specific implementation steps include:
[0038] Step 1: Obtain the Laguerre-Gaussian beam dataset, which contains orbital angular momentum mode information;
[0039] Step 2: Perform image preprocessing on the dataset;
[0040] Step 3: Construct a transfer learning model, which is an improved ResNet50 network model;
[0041] Step 4: Input the preprocessed training set into the transfer learning model for training, verify the performance of the trained model through the validation set, and save the optimal model.
[0042] Step 5: Obtain the test set, test the optimal model using the test set, and identify the orbital angular momentum mode of the vortex beam.
[0043] Step 1 specifically includes:
[0044] The complex amplitude expression for a Laguerre-Gaussian beam propagating along the z-axis is:
[0045]
[0046] Where l is the topological charge number and φ is the azimuth angle. w0 is the waist radius. Let λ be the Rayleigh length, λ be the wavelength, and k = 2π / λ be the wavenumber. When any number of Laguerre-Gaussian beams are superimposed, the resulting optical field intensity can be expressed as:
[0047] I = |u LG1 (r,φ,z)+u LG2 (r,φ,z)+…u LGn (r,φ,z)| 2 (2)
[0048] Combining equations (1) and (2), numerical simulations were used to generate intensity distribution maps of Laguerre-Gaussian beams with superimposed orbital angular momentum modes. These Laguerre-Gaussian beams with superimposed orbital angular momentum modes consist of 16 different orbital angular momentum modes superimposed, each with a radial exponent of 0 and topological charge numbers of {-4, -1, +2, +5}. The intensity distribution maps of these 16 Laguerre-Gaussian beams are shown below. Figure 2 As shown;
[0049] Several random phase screens are generated based on the power spectrum inversion method. Different atmospheric turbulence intensities are represented by changing the atmospheric refractive index structure constant. Random phase screens are inserted into the propagation path of the Laguerre-Gaussian beam to simulate the atmospheric turbulence channel. The transmitted Laguerre-Gaussian beam dataset is obtained through the atmospheric turbulence channel, and the dataset is divided into training set and validation set in a 7:3 ratio.
[0050] The power spectrum inversion method includes:
[0051] Generate a complex Gaussian random matrix C N×N In the frequency domain, the mean is 0 and the variance is 1; according to the modified von Karman spectral model φ n (κ) yields the phase perturbation σ(κ); using the phase perturbation σ(κ) to adjust C N×N Filtering is performed to obtain the matrix. right By performing an inverse Fourier transform and then taking the real part, the atmospheric turbulence phase screen can be obtained.
[0052] in, κ is the atmospheric refractive index structure constant, describing the intensity of atmospheric turbulence, κ is the space wavenumber, κ0 = 2π / l0, κ m = 5.92 / l0, where l0 is the internal scale and L0 is the external scale. N is the size of the phase screen, Δx is the grid spacing, and Δz is the distance between the phase screens.
[0053] In this invention, when simulating Laguerre-Gaussian beam transmission, λ = 1550 nm, w0 = 0.06 m, l0 = 0.0003 m, L0 = 50 m, N = 500, Δz = 100 m, and the collected dataset is from transmission over 1 km in an atmospheric turbulence channel, with atmospheric turbulence intensities of [missing information]. and The dataset consists of a mixed dataset with 200 images for each dataset, for a total of 3200 images. The training set contains 2240 images, and the validation set contains 960 images.
[0054] Step 2 specifically includes:
[0055] The images in the training set are randomly cropped to 224×224 pixels and horizontally rotated to increase the diversity of the training data and improve the robustness and generalization ability of the model. The images in the validation set are cropped at the center of the image to a size of 224×224. The images in the training set and validation set are converted into tensor data types. The images in the training set and validation set are then standardized.
[0056] Step 3 specifically includes:
[0057] Based on the ResNet50 network model trained on the ImageNet dataset, a CBAM attention mechanism module is introduced. The CBAM attention mechanism module is embedded between the fourth block and the average pooling layer in the ResNet50 network model. The CBAM attention mechanism module can improve the model's ability to perceive important features in the input data, thereby improving the model's recognition accuracy.
[0058] Step 4 specifically includes the following:
[0059] The improved ResNet50 network model is loaded, and the fully connected layers in the network model are modified so that the output of the fully connected layers corresponds to the number of Laguerre-Gaussian beam categories. Combined with transfer learning, the parameters of all layers above the CBAM attention mechanism module are frozen, so that only the parameters in the CBAM attention mechanism module and the fully connected layers are updated during training. The model performance after each training round is verified through the validation set, and the optimal model is saved after all training rounds are completed.
[0060] The hardware configuration for this invention experiment was an Intel Core i5-9300H CPU with a clock speed of 2.40GHz, an NVIDIA GeForce GTX 1650 GPU with 4.00GB of video memory, a Windows 10 x64 operating system, a torch = 1.7.1 and a python = 3.6.13 runtime environment, 50 training epochs, a training batch size of 26, an Adam optimizer, an initial learning rate of 0.001, which decreased by 50% every 5 training epochs thereafter, and a cross-entropy loss function.
[0061] After training, the highest accuracy rates of CNN1, CNN2, and CNN3 on the validation set were 100%, 97.1%, and 97.8%, respectively. Figures 3(a) and 3(b) show the curves of the recognition accuracy of CNN1, CNN2, and CNN3 on the validation set as a function of training epochs. In Figure 3(a), the solid line represents CNN3 and the dashed line represents CNN1; in Figure 3(b), the solid line represents CNN3 and the dashed line represents CNN2. Figure 3(a) shows that although CNN1 has a higher accuracy than CNN3, CNN3 converges 20 epochs faster than CNN1. Figure 3(b) shows that CNN2 and CNN3 have roughly the same convergence speed, but CNN3's overall accuracy is higher than CNN2's.
[0062] This invention selected three evaluation metrics—maximum accuracy, number of network training parameters, and training time—to compare CNN1 and CNN3, as shown in Table 1.
[0063] Table 1
[0064]
[0065]
[0066] As shown in Table 1, although the highest verification accuracy of CNN1 is 2.2% higher than that of CNN3, the number of network training parameters of CNN1 is 42 times more than that of CNN3, and the training speed of CNN3 is more than twice as fast as that of CNN1. Considering all factors, the performance of CNN3 of this invention is the best.
[0067] Step 5 specifically includes the following:
[0068] Following the steps in step 1, 1520 new images were collected as the test set. To demonstrate the model's generalization ability, the test set includes images containing atmospheric turbulence intensity. exist The mixed dataset of intervals was used to input the test set into the optimal model saved after training for prediction. The prediction results were visualized using a confusion matrix. The prediction results of the confusion matrix are shown in Figures 4(a), 4(b), and 4(c), where Figures 4(a), 4(b), and 4(c) are the prediction results of CNN1, CNN2, and CNN3, respectively. The horizontal axis represents the predicted class, and the vertical axis represents the true class. From the confusion matrix, it can be seen that the test set accuracy of CNN1 and CNN2 is 99.9% and 96.7%, respectively, and the test set accuracy of CNN3 is 97.8%, which shows high generalization ability.
[0069] Example 2:
[0070] This invention provides an orbital angular momentum mode recognition system based on transfer learning, specifically including:
[0071] The acquisition module is used to acquire the intensity distribution map of the Laguerre-Gaussian beam after transmission in the atmospheric turbulence channel;
[0072] The preprocessing module is used to perform preprocessing operations on the acquired Laguerre-Gaussian beam intensity distribution map;
[0073] The training module is used to input the preprocessed Laguerre-Gaussian beam intensity distribution map into the transfer learning model for training to obtain the optimal model.
[0074] The identification module is used to input the real-time acquired Laguerre-Gaussian beam intensity distribution map after transmission through the atmospheric turbulence channel into the optimal model for orbital angular momentum mode identification.
[0075] like Figure 5 As shown, in the acquisition module, the Gaussian beam generated by the laser is passed through a spatial light modulator (SLM) loaded with a phase hologram to generate a Laguerre-Gaussian beam with multiple superimposed orbital angular momentum modes. After passing through a simulated atmospheric turbulence channel, the intensity distribution map of the transmitted Laguerre-Gaussian beam is obtained by a Charge Coupled Device (CCD).
[0076] In the preprocessing module, the obtained Laguerre-Gaussian beam intensity distribution map is randomly cropped to a size of 224×224 and horizontally rotated, then converted to tensor data type and normalized.
[0077] In the training module, the preprocessed Laguerre-Gaussian beam intensity distribution map is input into the improved ResNet50 network model for training to obtain the optimal model.
[0078] In the identification module, the Laguerre-Gaussian beam intensity distribution map, which is obtained in real time by CCD and transmitted through the atmospheric turbulence channel, is preprocessed and then input into the optimal model for orbital angular momentum mode identification.
[0079] Note that the above description is merely a preferred embodiment of the present invention and the technical principles employed. Those skilled in the art will understand that the present invention is not limited to the specific embodiments described herein, and various obvious changes, readjustments, and substitutions can be made without departing from the scope of protection of the present invention. Therefore, although the present invention has been described in detail through the above embodiments, the present invention is not limited to the above embodiments, and may include many other equivalent embodiments without departing from the concept of the present invention, the scope of which is determined by the scope of the appended claims.
Claims
1. A method for orbital angular momentum mode recognition based on transfer learning, characterized in that, The orbital angular momentum mode recognition method based on transfer learning includes: Obtain the Laguerre-Gaussian beam dataset, which contains orbital angular momentum mode information; Perform image preprocessing on the dataset; Construct a transfer learning model, which is an improved ResNet50 network model; The preprocessed training set is input into the transfer learning model for training. The performance of the trained model is verified by the validation set, and the optimal model is saved. Obtain a test set, test the optimal model using the test set, and identify the orbital angular momentum mode of the vortex beam; The improved ResNet50 network model is based on the ResNet50 network model trained on the ImageNet dataset, which introduces the CBAM attention mechanism module. The CBAM attention mechanism module is embedded between the fourth block and the average pooling layer in the ResNet50 network model. The CBAM attention mechanism module improves the model's ability to perceive important features in the input data, thereby improving the model's recognition accuracy. The step of inputting the preprocessed training set into the transfer learning model for training includes: The improved ResNet50 network model is loaded, and the fully connected layers in the network model are modified so that the output of the fully connected layers corresponds to the number of Laguerre-Gaussian beam categories. Combined with transfer learning, the parameters of all layers above the CBAM attention mechanism module are frozen, so that only the parameters in the CBAM attention mechanism module and the fully connected layers are updated during training.
2. The orbital angular momentum mode recognition method based on transfer learning according to claim 1, characterized in that, The acquisition of the Laguerre-Gaussian beam dataset includes: The intensity distribution map of a Laguerre-Gaussian beam with different orbital angular momentum modes was generated by numerical simulation. The Laguerre-Gaussian beam with different orbital angular momentum modes is a Laguerre-Gaussian beam composed of 16 different orbital angular momentum modes with a radial exponent of 0 and a topological charge of {-4, -1, +2, +5}. Several random phase screens are generated based on the power spectrum inversion method. Different atmospheric turbulence intensities are represented by changing the atmospheric refractive index structure constant. Random phase screens are inserted into the propagation path of the Laguerre-Gaussian beam to simulate the atmospheric turbulence channel. The transmitted Laguerre-Gaussian beam dataset is obtained through the atmospheric turbulence channel, and the dataset is divided into training set and validation set in a 7:3 ratio.
3. The orbital angular momentum mode recognition method based on transfer learning according to claim 2, characterized in that, The power spectrum inversion method includes: Generate a complex Gaussian random matrix In the frequency domain, the mean is 0 and the variance is 1; according to the modified von Karman spectral model Phase perturbation ; Utilizing phase perturbation right Filtering is performed to obtain the matrix. ;right Performing an inverse Fourier transform and then taking the real part yields the atmospheric turbulence phase screen.
4. The orbital angular momentum mode recognition method based on transfer learning according to claim 1 or 3, characterized in that, The image preprocessing of the dataset includes: The images in the training set are randomly cropped to 224×224 pixels and horizontally rotated to increase the diversity of the training data and improve the robustness and generalization ability of the model. The images in the validation set are cropped at the center of the image to a size of 224×224. The images in the training set and validation set are converted into tensor data types. The images in the training set and validation set are then standardized.
5. The orbital angular momentum mode recognition method based on transfer learning according to claim 4, characterized in that, The process of testing the optimal model with a test set involves inputting the test set into the optimal model for prediction, and using a confusion matrix to visualize the prediction results.
6. A system for identifying orbital angular momentum modes based on transfer learning, characterized in that, The system is used to implement the method as described in claim 1, comprising: The acquisition module is used to acquire the intensity distribution map of the Laguerre-Gaussian beam after transmission in the atmospheric turbulence channel; The preprocessing module is used to perform preprocessing operations on the acquired Laguerre-Gaussian beam intensity distribution map; The training module is used to input the preprocessed Laguerre-Gaussian beam intensity distribution map into the transfer learning model for training to obtain the optimal model; The identification module is used to input the real-time acquired Laguerre-Gaussian beam intensity distribution map after transmission through the atmospheric turbulence channel into the optimal model for orbital angular momentum mode identification.