Vector light field polarization state transient solution method based on convolutional neural network

Abstract

The application discloses a kind of vector light field polarization state transient solution methods based on convolutional neural network, specific steps are as follows: construct the CNN for vector light field polarization state transient solution;Rotating wave plate experimental system is built, and the data set required for CNN training is obtained using rotating wave plate experimental system and CNN is trained;After improving rotating wave plate experimental system, vector light field polarization state transient solution system is obtained, using the system to obtain the light intensity diagram of transient vector light field, the light intensity diagram of modulated transient vector light field, and the polarization direction and ellipticity of the corresponding position of the output obtained from the light intensity diagram of transient vector light field are input into the CNN of training completion, to reconstruct vector light field vector light field polarization state.The application can high-precision instantaneous solve vector light field polarization state from vector light field light intensity diagram and modulated vector light field light intensity diagram, without rotating wave plate multiple times, with higher measurement efficiency and measurement accuracy, and has wide application prospect in the field of large bandwidth laser communication.

Description

Technical Field

[0001] This invention belongs to the field of laser communication, specifically relating to a transient solution method for the polarization state of a vector optical field based on a convolutional neural network. Background Technology

[0002] In today's increasingly competitive modernization of military equipment, lasers, with their advantages of high security, high transmission speed, and high resolution, are replacing radio as the information carrier for next-generation communication and radar systems. However, ordinary lasers suffer from small bandwidth, low information transmission efficiency, and are easily interfered with by environmental factors such as clouds, fog, and turbulence during atmospheric transmission, hindering their application in military fields such as naval formation communication, airborne ground-to-ground radar, and satellite-to-ground communication. Therefore, the demand for high-bandwidth, high-fidelity, high-performance laser communication and lidar systems in the military field is growing rapidly. Achieving high-bandwidth, high-fidelity information transmission is crucial to the performance of such systems. Vector light, with its non-uniformly distributed polarization state (meaning it can have a different polarization state at every spatial location), theoretically allows for an infinitely large bandwidth for communication systems, and is less affected by clouds, fog, and turbulence during transmission. Therefore, vector light, as an ideal information carrier, is increasingly being applied to high-performance laser communication, lidar, and other optoelectronic systems.

[0003] However, the application of vector light in military optoelectronic systems is currently constrained, hindering large-scale deployment. The main reason is the lack of high-precision, high-efficiency detection methods in vector light field polarization measurement, making it impossible to quickly and accurately characterize the polarization state of the vector light field. This delays the development and performance improvement of military optoelectronic systems such as naval formation laser communication systems, airborne ground-to-ground lidar, and satellite-to-ground laser communication systems. Therefore, it is essential to focus on researching high-precision, high-efficiency measurement technologies for vector light fields to ensure the research and production progress and quality of high-performance laser communication, lidar, and other military optoelectronic systems in my country. Summary of the Invention

[0004] The purpose of this invention is to provide a transient solution method for the polarization state of a vector optical field based on a convolutional neural network, which solves the problem that the polarization state of the receiver is difficult to solve quickly and accurately in the process of vector laser communication.

[0005] The technical solution to achieve the objective of this invention is: a transient solution method for the polarization state of a vector light field based on a convolutional neural network, comprising the following steps:

[0006] Step 1: Construct a convolutional neural network (CNN) for transient solution of vector light field polarization state.

[0007] Step 2: Build a rotating waveplate experimental system and use the rotating waveplate experimental system to obtain the dataset required for CNN training. The dataset includes a vector light field intensity map, a modulated vector light field intensity map, and the polarization direction and ellipticity of the corresponding position of the vector light field.

[0008] Step 3: Use the vector light field intensity map and the modulated vector light field intensity map in the dataset as input to the CNN, and the polarization direction and ellipticity of the corresponding position of the vector light field in the dataset as output to train the CNN and obtain the trained CNN.

[0009] Step 4: After improving the rotating waveplate experimental system, a transient solution system for the polarization state of the vector light field is obtained. The transient vector light field intensity map and the modulated transient vector light field intensity map are obtained using the transient vector light field intensity map and the modulated transient vector light field intensity map. The transient vector light field intensity map and the modulated transient vector light field intensity map are input into the trained CNN and the polarization direction and ellipticity of the corresponding position of the transient vector light field are output.

[0010] Step 5: Reconstruct the polarization state of the vector light field using the polarization direction and ellipticity at the corresponding position of the transient vector light field.

[0011] Compared with the prior art, the significant advantages of this invention are:

[0012] (1) This invention is based on CNN for polarization state calculation. After training the CNN with the training set, it can directly convert the input light intensity map into the corresponding polarization direction and ellipticity without rotating the waveplate multiple times, thus realizing the transient detection of the polarization state of the vector light field.

[0013] (2) This invention uses CNN to calculate the polarization state, uses the method of rotating waveplate multiple times to create a dataset, and uses the dataset to train the CNN network to achieve high-precision detection of the polarization state of the vector light field, with a polarization direction accuracy of ±0.5° and an ellipticity accuracy of ±0.5°. Attached Figure Description

[0014] Figure 1 This is a flowchart of the transient solution method for vector light field polarization state based on convolutional neural networks according to the present invention.

[0015] Figure 2 This is the network structure diagram of the present invention.

[0016] Figure 3 This is a structural diagram of the attention module of the present invention.

[0017] Figure 4 This invention is a transient solution system for the polarization state of a vector light field.

[0018] Figure 5 This invention relates to a rotating waveplate experimental system. Detailed Implementation

[0019] The present invention will now be described in detail with reference to the accompanying drawings.

[0020] Combination Figure 1 The transient solution method for the polarization state of a vector light field based on a convolutional neural network described in this invention can directly calculate the polarization state of the vector light field from the light intensity diagram without requiring multiple waveplate rotations. The method includes the following steps:

[0021] Step 1: Construct a CNN for transient solution of vector light field polarization state;

[0022] Step 1-1 Figure 2 The network structure described in step one is shown. The specific structure of the CNN is as follows: The CNN includes an upper channel, a lower channel, 6 convolutional blocks, 3 attention modules, 1 vectorization module, and 1 fully connected module. The vector light field intensity map and the modulated vector light field intensity map are input into the CNN network and then superimposed. The superimposed images are sent to two channels respectively. In the upper channel, the images are output after passing through 3 convolutional blocks in sequence, where each convolutional block includes a convolutional layer, a batch normalization layer, and an activation layer arranged sequentially. In the lower channel, the images are output after passing through 3 attention modules and 1 convolutional block in sequence. The output images of the two channels are superimposed, passed through two convolutional blocks in sequence, and then the output images are vectorized and output through a fully connected module to obtain the polarization direction and ellipticity of the corresponding position of the vector light field.

[0023] Steps 1-2 Figure 3 The attention module of this invention is illustrated. The attention module includes a grid layer, a max pooling layer, an average pooling layer, a convolutional layer, and an activation layer. After input to the image attention module, the image first passes through the grid layer, then through max pooling and average pooling respectively. The output images from the two pooling operations are superimposed and then passed through one convolutional layer and one activation layer to obtain an attention image. The attention image is then multiplied by the input image and output. The grid layer is used to manually intervene in the attention map; by superimposing a pre-defined vector light polarization grid onto the input image, the accuracy of the output can be improved.

[0024] Step 2: Build a rotating waveplate experimental system and use it to obtain the dataset required for CNN training. This dataset includes a vector light field intensity map, a modulated vector light field intensity map, and the polarization direction and ellipticity at the corresponding positions of the vector light field, as detailed below:

[0025] Step 2-1 Figure 5 The diagram shows a rotating waveplate experimental system, consisting from left to right of a rotatable quarter-wave plate, a linear polarizer along the y-axis (fast axis), and a CCD camera. This system can obtain polarization state information of a vector light field. Below, we will use the Jones matrix to explain the principle of obtaining polarization state information using the rotating waveplate method.

[0026] The Jones matrices of elliptically polarized light with polarization direction θ and ellipticity ψ, a quarter-wave plate with a fast axis at an angle α to the x-axis, and a linear polarizer with a fast axis parallel to the y-axis are as follows:

[0027]

[0028] Where i represents an imaginary number.

[0029] Therefore, the final emitted electric field is:

[0030]

[0031] Where E(α) is the magnitude of the electric field received by the CCD camera.

[0032] Therefore, the light intensity value detected by the CCD camera is I(α)=|E(α)| 2 After simplification, it can be written as:

[0033]

[0034] If we use the discrete Fourier series to represent I(α) If expressed in the form of , then there are 4 non-zero coefficients, which are:

[0035]

[0036]

[0037]

[0038]

[0039] These coefficients can be represented by light intensity values:

[0040]

[0041] Where the Fourier series n = 2 or 4, N is the total number of rotations of the quarter-wave plate, and I(α) k Let θ be the light intensity value captured by the CCD camera during the k-th rotation of the quarter-wave plate. It is quite clear that the values ​​of θ and ψ can be calculated using these four coefficients:

[0042]

[0043] In the formula, the adjustment coefficient m is a freely selectable integer. Its function is to limit θ and ψ to the corresponding ranges of 0°≤θ≤180° and -45°≤ψ≤45°. The obtained θ and ψ are the polarization direction and ellipticity of the corresponding position of the vector light field.

[0044] Step 3: Use the vector light field intensity map and the modulated vector light field intensity map in the dataset as input to the CNN, and the polarization direction and ellipticity of the corresponding position of the vector light field in the dataset as output to train the CNN and obtain the trained CNN.

[0045] In this invention, mean squared error (MSE) is used as the loss function Loss, defined as follows:

[0046]

[0047] Where C is the number of samples, and the sample numbers j = 1, 2, ..., C, x j For output data, y j This is true data.

[0048] Step 4: After improving the rotating waveplate experimental system, a transient solution system for the polarization state of the vector light field is obtained (e.g., Figure 4 The transient vector light field intensity map and the modulated transient vector light field intensity map are obtained using a vector light field polarization state transient solution system. These are then input into a trained CNN to output the polarization direction and ellipticity at the corresponding positions of the transient vector light field, as detailed below:

[0049] Step 4-1: Set a beam splitter with a splitting ratio of 1:1 in front of the quarter-wave plate of the rotating wave plate experimental system, adjust the quarter-wave plate so that the fast axis is along the x-axis and fix it, and adjust the linear polarizer so that the transmission axis is along the y-axis to obtain the transient solution system of vector light field polarization state.

[0050] Step 4-2: After the vector light enters the system, it is first split into two beams by a beam splitter with a splitting ratio of 1:1. One beam is directly received by the CCD camera to obtain the transient vector light field intensity map; the other beam passes through a quarter-wave plate along the fast axis along the x-axis and a linear polarizer along the transmission axis along the y-axis before being received by the CCD camera to obtain the modulated transient vector light field intensity map.

[0051] Step 4-3: Input the transient vector light field intensity map and the modulated transient vector light field intensity map into the trained CNN, and output the polarization direction and ellipticity of the corresponding position of the transient vector light field.

[0052] Step 5: Reconstruct the polarization state of the vector light field using the polarization direction and ellipticity at the corresponding position of the transient vector light field.

[0053] In summary, this invention provides a high-precision solution for the polarization state information of a vector light field using transient vector light field intensity maps and modulated transient vector light field intensity maps, without requiring multiple waveplate rotations. It boasts high measurement efficiency and accuracy, and has broad application prospects in fields such as wide-bandwidth laser communication.

Claims

1. A transient solution method for the polarization state of a vector light field based on a convolutional neural network, characterized in that, Includes the following steps: Step 1: Construct a CNN for transient solution of vector light field polarization state; Step 2: Build a rotating waveplate experimental system and use the rotating waveplate experimental system to obtain the dataset required for CNN training. The dataset includes a vector light field intensity map, a modulated vector light field intensity map, and the polarization direction and ellipticity of the corresponding position of the vector light field. Step 3: Use the vector light field intensity map and the modulated vector light field intensity map in the dataset as input to the CNN, and the polarization direction and ellipticity of the corresponding position of the vector light field in the dataset as output to train the CNN and obtain the trained CNN. Step 4: After improving the rotating waveplate experimental system, a transient solution system for the polarization state of the vector light field is obtained. This system is used to acquire the transient vector light field intensity map and the modulated transient vector light field intensity map. These maps are then input into the trained CNN, which outputs the polarization direction and ellipticity at the corresponding positions of the transient vector light field. The details are as follows: Step 4-1: Set a beam splitter with a splitting ratio of 1:1 in front of the quarter-wave plate of the rotating wave plate experimental system, adjust the quarter-wave plate so that the fast axis is along the x-axis and fix it, and adjust the linear polarizer so that the transmission axis is along the y-axis to obtain the transient solution system of vector light field polarization state. Step 4-2: After the vector light to be measured enters the transient solution system for the polarization state of the vector light field, it is first split into two beams by a beam splitter with a splitting ratio of 1:

1. One beam is directly received by the CCD camera to obtain the transient vector light field intensity map; the other beam passes through a quarter-wave plate along the fast axis along the x-axis and a linear polarizer along the transmission axis along the y-axis before being received by the CCD camera to obtain the modulated transient vector light field intensity map. Step 4-3: Input the transient vector light field intensity map and the modulated transient vector light field intensity map into the trained CNN, and output the polarization direction and ellipticity of the corresponding position of the transient vector light field; Step 5: Reconstruct the polarization state of the vector light field using the polarization direction and ellipticity at the corresponding position of the transient vector light field.

2. The transient solution method for vector light field polarization state based on convolutional neural network according to claim 1, characterized in that, In step 1, the CNN includes an upper channel, a lower channel, 6 convolutional blocks, 3 attention modules, 1 vectorization module, and 1 fully connected module, with the following specific structure: The vector light field intensity map and the modulated vector light field intensity map are input into the CNN and then superimposed. The superimposed image is then fed into two channels. The upper channel passes through three convolutional blocks in sequence before being output. The lower channel passes through three attention modules and one convolutional block in sequence before being output. The output images of the two channels are superimposed and then passed through the remaining two convolutional blocks in sequence. The output image is then vectorized in the vectorization module and output through the fully connected module to obtain the polarization direction and ellipticity at the corresponding position of the vector light field. The attention module includes a grid layer, a max pooling layer, an average pooling layer, a convolutional layer, and an activation layer. The input image first passes through the grid layer, then through the max pooling layer and the average pooling layer respectively. The output images of the two pooling operations are superimposed and then passed through one convolutional layer and one activation layer to obtain the attention image. The attention image is multiplied by the input image and then output. The grid layer is used to manually intervene in the attention map by superimposing a pre-set vector light polarization grid with the input image to improve the accuracy of the output.

3. The transient solution method for vector light field polarization state based on convolutional neural network according to claim 1, characterized in that, In step 2, a rotating waveplate experimental system is built, and the dataset required for CNN training is obtained using this system. This dataset includes a vector light field intensity map, a modulated vector light field intensity map, and the polarization direction and ellipticity at the corresponding positions of the vector light field. The specific steps are as follows: Step 2-1: Arrange a rotatable quarter-wave plate, a linear polarizer along the y-axis in the fast axis direction, and a CCD camera in sequence along the optical path to form a rotating wave plate experimental system; Step 2-2, polarization direction is Ellipticity is After the elliptically polarized light passes through the rotating waveplate experimental system, the CCD camera receives the modulated vector light field intensity map, the intensity of which is: ， In the formula, Polarization direction For ellipticity, Let be the angle between the fast axis of the waveplate and the x-axis. For the present The light intensity value at the specified value; Will Using Discrete Fourier Series If expressed in the form of , then there are 4 non-zero coefficients, which are respectively , , , : ， These coefficients are represented by light intensity values: ， Where the Fourier order n = 2 or 4, and N is the total number of rotations of the quarter-wave plate. The light intensity value collected by the CCD camera when the quarter-wave plate rotates for the kth time; Therefore, it can be concluded that and The relationship between the value and these four coefficients: ， In the formula, the adjustment coefficient m is a freely chosen integer, and its function is to adjust the... and Limited to the corresponding range , obtained and That is, the polarization direction and ellipticity of the vector light field at the corresponding position.

Images