A LG beam recognition method based on physical feature information fusion neural network

Abstract

The present invention discloses a method for identifying LG beams based on a physical feature information fusion neural network. The method comprises: 1) modulating and expanding an incident light beam through a polarization filter, then irradiating it onto a spatial light modulator to generate a Laguerre-Gaussian LG beam; 2) focusing the generated LG beam using a lens to produce a Fourier transform, and collecting a Fraunhofer diffraction image of the LG beam using a CCD camera; and 3) inputting the collected Fraunhofer diffraction image of the LG beam into an improved neural network structure for processing, thereby accurately identifying multiple physical feature information within the LG beam. Based on the unique distinguishability of fractional-order LG beams in diffraction images, the method generates three outputs for the identification of three physical feature parameters, which are output at different stages of the neural network based on the difficulty of displaying the spot features. Furthermore, the method combines deep and shallow layer fusion techniques within the neural network structure to achieve fusion of physical information features, resulting in more stable and accurate LG beam identification results.

Description

Technical Field

[0001] The present invention relates to the technical field of cross-application of artificial intelligence and physical models, and in particular to a LG light beam recognition method based on physical feature information fusion neural network. Background Art

[0002] Traditional techniques for extracting LG beam feature information suffer from high human error and low efficiency. The accuracy of identification relies heavily on the distinct intensity characteristics of the diffraction spot, a significant drawback for both experimental and instrumental measurements. The rise of deep neural networks offers hope for overcoming these shortcomings. Deep neural networks can not only accurately identify integer vortex beams, but also fractional vortex beams with very small intervals. However, while deep neural networks demonstrate unique advantages in vortex beam pattern recognition, the resolution limitations of optical communication systems mean that the intervals between adjacent orbital angular momentum patterns cannot be infinitesimally small. Therefore, accurately identifying the information carried by light at the receiving end has become a key technology for increasing optical communication capacity. Summary of the Invention

[0003] The purpose of the present invention is to address the shortcomings of the existing technology and provide an LG beam identification method based on a physical feature information fusion neural network. Unlike the traditional one-to-one identification forward training and identification of subtle topological charge changes, the present invention observes that fractional-order LG beams have special distinguishable characteristics in the diffraction image, thereby abandoning the previous reliance on topological charge identification and instead identifying the size of the light waist. By utilizing the physical fusion deep neural network optimization design, a robust neural network architecture with low data training and a certain degree of interpretability is designed, thereby achieving high-precision and high-speed intelligent identification of vortex beams.

[0004] In order to achieve the above-mentioned purpose, the present invention adopts the following technical solutions.

[0005] A LG beam recognition method based on physical feature information fusion neural network includes the following steps:

[0006] Step S1, modulating and expanding an incident light beam through a polarization filter and then irradiating the spatial light modulator to generate a Laguerre-Gaussian (LG) beam;

[0007] Step S2: using a lens to focus the LG beam generated in step S1 to generate a Fourier transform, and collecting a Fraunhofer diffraction image of the LG beam using a CCD camera;

[0008] Step S3: Input the collected LG beam Fraunhofer diffraction image into the improved neural network structure for processing, and accurately identify the multi-physical feature information in the LG beam through the optimized loss function.

[0009] Specifically, the polarization filter in step S1 includes a rotatable linear polarizer and a 1 / 2 wave plate combination, which is used to control the intensity and polarization direction of the LG light beam; the spatial light modulator has a built-in composite modulation function of a computer-generated hologram and a spiral phase plate, which is used to synchronously control the wavefront phase and intensity distribution of the LG light beam.

[0010] Specifically, the lens in step S2 includes a biconvex lens pair with a focal length ratio of 1:1 and another biconvex lens. The biconvex lens pair forms a 4f system, which retains the initial beam radius of the LG beam. , and then the first-order diffraction spot is filtered out at the focal plane of the first biconvex lens to extract a pure LG beam. The subsequent two biconvex lenses focus at the focal plane to produce a Fourier transform to form a Fraunhofer diffraction image containing topological charge information; the pixel size of the CCD camera is less than 1 / 5 of the diffraction ring spacing of the Fraunhofer diffraction image.

[0011] Specifically, in step S2, the LG beam generated in step S1 is focused by a lens to generate a Fourier transform. The equation of the LG beam at Z=0 is expressed as:

[0012] ;

[0013] In the above formula, represents the LG beam at Z=0, is the polar coordinate on the incident plane, z is the propagation distance, r is the radial coordinate in polar coordinates; is the generalized Laguerre polynomial, is the radial index, is the topological charge of the LG beam; is the LG beam radius; is the radial intensity distribution of the LG beam; is an imaginary unit; is the azimuth in polar coordinates;

[0014] The LG beam passes through the lens and forms a Fraunhofer diffraction spot at the focal plane. The expression of the Fraunhofer diffraction spot can be simplified as the Fourier transform of the incident light field:

[0015] ;

[0016] In the above formula, is the Fourier transform of the incident light field; is the Fourier transform representation; Corresponding to The frequency domain coordinates of .

[0017] Specifically, the improved neural network structure in step S3 includes a convolutional layer, a pooling layer, a feature information fusion layer and a fully connected layer;

[0018] The convolution layer is used to extract local features through the convolution kernel, expand the receptive field by inserting holes in the convolution kernel, and enhance the ability to extract global and local information of the image;

[0019] The pooling layer includes a maximum pooling layer and an adaptive pooling layer. The maximum pooling layer is used to reduce the spatial size of the feature while retaining important feature information. The adaptive pooling layer is used to uniformly pool inputs of any size into a fixed size to adapt to different input sizes.

[0020] The physical feature information fusion layer is used to fuse the input matrix information. The input feature matrix is ​​divided into three branches. Each branch is subjected to nonlinear transformation by matrix dot multiplication and Sigmoid function. Then, the features of each branch are fused at different nodes. The features of multiple branches are merged by matrix addition operation. The fused features are then subjected to ReLU activation function to output a more tightly fused and more stable feature matrix, thereby enhancing the model's ability to recognize and express input information.

[0021] The fully connected layer is used to map the extracted features to the output space to obtain the final prediction result.

[0022] Furthermore, the convolution layer includes a dilated convolution layer and a standard convolution layer. The dilated convolution layer expands the receptive field of the convolution kernel through a preset dilation rate, so that the dilated convolution layer can capture periodic features in the input data that span a larger physical scale while avoiding the introduction of additional parameters; the standard convolution layer uses a dense convolution kernel receptive field to finely extract detailed features such as the phase gradient between adjacent pixels and the singularity of the vortex center; the dilated convolution layer and the standard convolution layer independently process the input through a parallel branch structure, and finally fuse the two feature maps through channel splicing, thereby enhancing the ability to extract global and local information of the image.

[0023] Specifically, the improved neural network structure in step S3 improves the robustness of the neural network in complex multi-task scenarios by optimizing the loss function. The design process of the loss function is as follows:

[0024] First calculate the k Cross entropy loss in each task scenario :

[0025] ;

[0026] In the above formula, C is the number of categories; Represents a sample x The true label ofn The probability of each category; For samples x In the n The predicted probability of each category;

[0027] Then introduce the variance parameter Dynamically adjust the loss weight of each task, dividing the cross entropy loss in each task scenario by the corresponding variance parameter , and then introduce the logarithmic term of the variance parameter Get the final loss function :

[0028] ;

[0029] In the above formula, For the k Cross entropy loss in each task scenario; For the k Learnable uncertainty parameters for each task;

[0030] The logarithmic term Used to prevent gradient explosion, if the variance parameter Too big, It will be very small. Introducing the logarithmic term of the variance parameter can avoid this gradient explosion. Specifically:

[0031] Usually the program is initially set =0, that is =1, if Too large (difficult task) → gradient is negative → increase → Exponential increase → reduce the weight of the task ;like Too small (easy task) → gradient is positive → decrease → Exponential decrease → Increase the weight of the task .

[0032] Compared with the prior art, the present invention has the following beneficial effects:

[0033] 1. The method of the present invention is based on the special distinguishable characteristics of fractional-order LG beams in diffraction images. It will identify three physical characteristic parameters and make three outputs. According to the difficulty of displaying the spot characteristics, the output is output at different stages of the neural network. In addition, it combines the deep and shallow layer fusion technology of the neural network structure to achieve the fusion of physical information features, making the prediction results of the neural network more stable and accurate.

[0034] 2. The method of the present invention uses innovative binary phase encoding technology to efficiently generate LG beams with specific orbital angular momentum using only two phase states, 0 and π. Compared with traditional multi-order phase modulation schemes, this method significantly reduces the performance requirements of the spatial light modulator (SLM), enabling ordinary SLM devices to achieve high-quality LG beam output.

[0035] 3. The calculation process of the method of the present invention uses analytical expressions to directly generate holograms, avoiding complex iterative optimization algorithms and increasing the calculation speed by several times. It is particularly suitable for experimental scenarios that require real-time dynamic control. At the same time, the binary phase structure has strong anti-interference ability and higher tolerance to calibration errors of the spatial light modulator (SLM).

[0036] 4. The method of the present invention is different from the traditional method of identifying subtle topological charge changes. By simultaneously identifying three parameters, it not only reduces the difficulty of task identification, but also expands the information capacity exponentially. Compared with the traditional single topological charge identification method, it reduces the amount of neural network training. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 This is a flow chart of a LG beam recognition method based on physical feature information fusion neural network of the present invention;

[0038] Figure 2 is a schematic diagram of the principle of collecting the Fraunhofer diffraction image of the LG beam according to the present invention;

[0039] Figure 3 Schematic diagram of the neural network structure in an embodiment of the present invention

[0040] Figure 4 is a schematic structural diagram of the physical feature information fusion layer in an embodiment of the present invention;

[0041] Figure 5 In the embodiment of the present invention Schematic diagram of the confusion matrix of the test set;

[0042] Figure 6 In the embodiment of the present invention Schematic diagram of the confusion matrix of the test set;

[0043] Figure 7 In the embodiment of the present invention Schematic diagram of the confusion matrix of the test set;

[0044] Figure 8 3 is a spatial scatter diagram of the recognition accuracy of the LG beam Fraunhofer diffraction image in an embodiment of the present invention. DETAILED DESCRIPTION

[0045] In order to facilitate those skilled in the art to understand and implement the present invention, each step of the method proposed in the present invention is described in detail below. It should be understood that these embodiments are only used to illustrate the present invention and are not intended to limit the scope of the present invention. In addition, it should be understood that after reading the content taught by the present invention, those skilled in the art can make various changes or modifications to the present invention, and these equivalent forms also fall within the scope defined by the claims appended hereto.

[0046] Example

[0047] like Figure 1 As shown, the present invention discloses a LG beam recognition method based on physical feature information fusion neural network, comprising the following steps:

[0048] Step S1, modulating and expanding an incident light beam through a polarization filter and then irradiating the spatial light modulator to generate a Laguerre-Gaussian (LG) beam;

[0049] Step S2: using a lens to focus the LG beam generated in step S1 to generate a Fourier transform, and collecting a Fraunhofer diffraction image of the LG beam using a CCD camera;

[0050] Step S3: Input the collected LG beam Fraunhofer diffraction image into the improved neural network structure for processing, and accurately identify the multi-physical feature information in the LG beam through the optimized loss function.

[0051] Specifically, such as Figure 2 As shown, the polarization filter in step S1 includes a rotatable linear polarizer and a 1 / 2 wave plate combination, which is used to control the intensity and polarization direction of the LG light beam; the spatial light modulator has a built-in composite modulation function of a computer-generated hologram and a spiral phase plate, which is used to synchronously control the wavefront phase and intensity distribution of the LG light beam.

[0052] Furthermore, the lens in step S2 includes a biconvex lens pair with a focal length ratio of 1:1 and another biconvex lens, and the biconvex lens pair forms a 4f system, that is, the initial beam radius of the LG beam is retained. , and then the first-order diffraction spot is filtered out at the focal plane of the first biconvex lens to extract a pure LG beam. The subsequent two biconvex lenses focus at the focal plane to produce a Fourier transform to form a Fraunhofer diffraction image containing topological charge information; the pixel size of the CCD camera is less than 1 / 5 of the diffraction ring spacing of the Fraunhofer diffraction image.

[0053] Specifically, in step S2, the LG beam generated in step S1 is focused by a lens to generate a Fourier transform. The equation of the LG beam at Z=0 is expressed as:

[0054] ;

[0055] In the above formula, represents the LG beam at Z=0, is the polar coordinate on the incident plane, z is the propagation distance, r is the radial coordinate in polar coordinates; is the generalized Laguerre polynomial, is the radial index, is the topological charge of the LG beam; is the LG beam radius; is the radial intensity distribution of the LG beam; is an imaginary unit; is the azimuth in polar coordinates;

[0056] The LG beam passes through the lens and forms a Fraunhofer diffraction spot at the focal plane. The expression of the Fraunhofer diffraction spot can be simplified as the Fourier transform of the incident light field:

[0057] ;

[0058] In the above formula, is the Fourier transform of the incident light field; is the Fourier transform representation; Corresponding to The frequency domain coordinates of .

[0059] Specifically, the improved neural network structure described in step S3 is as follows: Figure 3 As shown in Figure 1, the improved neural network structure includes convolutional layer, pooling layer, physical feature information fusion layer and fully connected layer;

[0060] The convolution layer is used to extract local features through the convolution kernel, expand the receptive field by inserting holes in the convolution kernel, and enhance the ability to extract global and local information of the image;

[0061] The pooling layer includes a maximum pooling layer and an adaptive pooling layer. The maximum pooling layer is used to reduce the spatial size of the feature while retaining important feature information. The adaptive pooling layer is used to uniformly pool inputs of any size into a fixed size to adapt to different input sizes.

[0062] like Figure 4As shown in the figure, the physical feature information fusion layer is used to fuse the input matrix information. The input feature matrix is ​​divided into three branches. Each branch is subjected to nonlinear transformation by matrix dot multiplication and Sigmoid function respectively. Then, the features of each branch are fused at different nodes. The features of multiple branches are merged by matrix addition operation. The fused features are then subjected to ReLU activation function to output a feature matrix that is more tightly fused and has stronger stability, thereby enhancing the model's ability to recognize and express input information.

[0063] The fully connected layer is used to map the extracted features to the output space to obtain the final prediction result.

[0064] Furthermore, the convolution layer includes a dilated convolution layer and a standard convolution layer. The dilated convolution layer expands the receptive field of the convolution kernel through a preset dilation rate, so that the dilated convolution layer can capture periodic features in the input data that span a larger physical scale while avoiding the introduction of additional parameters; the standard convolution layer uses a dense convolution kernel receptive field to finely extract detailed features such as the phase gradient between adjacent pixels and the singularity of the vortex center; the dilated convolution layer and the standard convolution layer independently process the input through a parallel branch structure, and finally fuse the two feature maps through channel splicing, thereby enhancing the ability to extract global and local information of the image.

[0065] Specifically, the improved neural network structure in step S3 improves the robustness of the neural network in complex multi-task scenarios by optimizing the loss function. The design process of the loss function is as follows:

[0066] First calculate the k Cross entropy loss in each task scenario :

[0067] ;

[0068] In the above formula, C is the number of categories; Representation sample x The true label of n The probability of each category; For samples x In the n The predicted probability of each category;

[0069] Then introduce the variance parameter Dynamically adjust the loss weight of each task, dividing the cross entropy loss in each task scenario by the corresponding variance parameter , and then introduce the logarithmic term of the variance parameter Get the final loss function :

[0070] ;

[0071] In the above formula, For the k Cross entropy loss in each task scenario; For the k Learnable uncertainty parameters for each task.

[0072] The technical effect of the method of the present invention is further illustrated below through a specific example.

[0073] The dataset used in this example is a Matlab simulation dataset; the deep neural network training environment is a CPU 12th Gen Intel(R) Core(TM) i7-12700KF 3.60 GHz, 32GB RAM, and a GPU NVIDIA GeForce RTX3080. The program is implemented using Python 3.9 and PyTorch in the Windows 11 22000.1696 operating system.

[0074] See for example Figure 3 、 Figure 4 The neural network structure and physical feature information fusion layer structure shown in the figure input the Fraunhofer diffraction image of the LG beam into the neural network for training. The neural network is divided into deep and shallow layers. The shallow layer output channel has obvious feature information. and Parameters, deep input is shallow and The output of the channel is combined with the output of the original network, and the three-channel information is finally identified by the fully connected layer for three parameters.

[0075] like Figure 5 、 Figure 6 、 Figure 7 The three parameters are shown respectively P、 、L Schematic diagram of the confusion matrix of the test set. From the analysis of the confusion matrix, it can be concluded that the neural network model proposed in this invention is P Label, Tags and L The high values ​​on the diagonal of the confusion matrix indicate that the model can accurately identify most samples, while the low values ​​on the non-diagonal line indicate that there are fewer misclassifications, indicating that the method of the present invention has high accuracy and excellent recognition ability in image classification tasks.

[0076] like Figure 8 The figure shows the spatial scatter plot of the LG beam Fraunhofer diffraction image recognition accuracy. Figure 8The red points represent the data that were misclassified, and the green points represent the data that were correctly identified by all three parameters. The spatial scatter plot further verifies that the predicted results and actual values ​​between the labels are highly consistent, the number of misclassified points is very small, and the Figure 8 It can be clearly seen that the uniformity of the test set selected by the method of the present invention is good, indicating that the method of the present invention has good stability and accuracy in identifying the physical feature information in the LG beam.

[0077] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other manner. Any person skilled in the art may utilize the above-disclosed technical content to modify or modify the present invention into equivalent embodiments. However, any simple modifications, equivalent variations, and modifications to the above embodiments that do not depart from the technical content of the present invention and are based on the technical essence of the present invention remain within the scope of protection of the present invention.

Claims

1. A LG beam recognition method based on physical feature information fusion neural network, characterized in that: The following steps are involved: Step S1, modulating and expanding an incident light beam through a polarization filter and then irradiating the spatial light modulator to generate a Laguerre-Gaussian (LG) beam; Step S2: using a lens to focus the LG beam generated in step S1 to generate a Fourier transform, and collecting a Fraunhofer diffraction image of the LG beam using a CCD camera; The lens is used to focus the LG beam generated in step S1 to generate a Fourier transform. The equation of the LG beam at Z=0 is expressed as: ； In the above formula, represents the LG beam at Z=0, is the polar coordinate on the incident plane, z is the propagation distance, r is the radial coordinate in polar coordinates; is the generalized Laguerre polynomial, is the radial index, is the topological charge of the LG beam; is the LG beam radius; is the radial intensity distribution of the LG beam; is an imaginary unit; is the azimuth in polar coordinates; The LG beam passes through the lens and forms a Fraunhofer diffraction spot at the focal plane. The expression of the Fraunhofer diffraction spot can be simplified as the Fourier transform of the incident light field: ； In the above formula, is the Fourier transform of the incident light field; is the Fourier transform representation; Corresponding to Frequency domain coordinates of Step S3: inputting the collected LG beam Fraunhofer diffraction image into the improved neural network structure for processing, and realizing accurate recognition of multi-physical feature information in the LG beam through the optimized loss function; The improved neural network structure includes a convolutional layer, a pooling layer, a physical feature information fusion layer and a fully connected layer; The convolution layer is used to extract local features through the convolution kernel, expand the receptive field by inserting holes in the convolution kernel, and enhance the ability to extract global and local information of the image; The pooling layer includes a maximum pooling layer and an adaptive pooling layer. The maximum pooling layer is used to reduce the spatial size of the feature while retaining important feature information. The adaptive pooling layer is used to uniformly pool inputs of any size into a fixed size to adapt to different input sizes. The physical feature information fusion layer is used to fuse the input matrix information. The input feature matrix is ​​divided into three branches. Each branch is subjected to nonlinear transformation by matrix dot multiplication and Sigmoid function. Then, the features of each branch are fused at different nodes. The features of multiple branches are merged by matrix addition operation. The fused features are then subjected to ReLU activation function to output a more tightly fused and more stable feature matrix, thereby enhancing the model's ability to recognize and express input information. The fully connected layer is used to map the extracted features to the output space to obtain the final prediction result; The improved neural network structure improves the robustness of the neural network in complex multi-task scenarios by optimizing the loss function. The design process of the loss function is as follows: First calculate the k Cross entropy loss in each task scenario : ； In the above formula, C is the number of categories; Represents a sample x The true label of n The probability of each category; For samples x In the n The predicted probability of each category; Then introduce the variance parameter Dynamically adjust the loss weight of each task, dividing the cross entropy loss in each task scenario by the corresponding variance parameter , and then introduce the logarithmic term of the variance parameter Get the final loss function : ； In the above formula, For the k Cross entropy loss in each task scenario; For the k Learnable uncertainty parameters for each task.

2. The LG beam recognition method based on physical feature information fusion neural network according to claim 1 is characterized in that: The polarization filter in step S1 includes a rotatable linear polarizer and a 1 / 2 wave plate combination, which is used to control the intensity and polarization direction of the LG beam; the spatial light modulator has a built-in composite modulation function of a computer-generated hologram and a spiral phase plate, which is used to synchronously control the wavefront phase and intensity distribution of the LG beam.

3. The LG beam recognition method based on physical feature information fusion neural network according to claim 1 is characterized in that: The lens in step S2 includes a biconvex lens pair with a focal length ratio of 1:1 and another biconvex lens. The biconvex lens pair forms a 4f system, which preserves the initial beam radius of the LG beam. , and then the first-order diffraction spot is filtered out at the focal plane of the first biconvex lens to extract a pure LG beam. The subsequent two biconvex lenses focus at the focal plane to produce a Fourier transform to form a Fraunhofer diffraction image containing topological charge information; the pixel size of the CCD camera is less than 1 / 5 of the diffraction ring spacing of the Fraunhofer diffraction image.

4. The LG beam recognition method based on physical feature information fusion neural network according to claim 1 is characterized in that: The convolution layer includes a dilated convolution layer and a standard convolution layer. The dilated convolution layer expands the receptive field of the convolution kernel through a preset dilation rate, so that the dilated convolution layer can capture periodic features across larger physical scales in the input data while avoiding the introduction of additional parameters; the standard convolution layer uses the dense receptive field of the convolution kernel to finely extract the detailed features of the phase gradient between adjacent pixels and the singularity of the vortex center; the dilated convolution layer and the standard convolution layer independently process the input through a parallel branch structure, and finally fuse the two feature maps through channel splicing, thereby enhancing the ability to extract global and local information of the image.

Images