Optical orbital angular momentum recognition method and system based on binary neural network model

Abstract

The application provides an optical orbital angular momentum recognition method and system based on a binary neural network model, and aims to solve the technical problem that the OAM recognition method based on a spatial light modulator deep learning method has high calculation complexity, large memory occupation, and is difficult to meet the application requirements of lightweight and real-time of space laser communication. The application provides an optical orbital angular momentum recognition method based on a binary neural network model, which realizes the recognition of different topological charge OAM by constructing a binary neural network model. Compared with the traditional OAM recognition method based on the spatial light modulator deep learning method, the application adopts binary weights and binary activation functions, and the parameter storage requirement is significantly reduced. At the same time, the binary neural network model adopted by the application adopts PopCount processing instead of convolution operation, reduces the calculation amount, enhances the real-time performance, and can meet the requirement of fast recognition in the application scene of inter-satellite communication.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of space laser communication, and particularly relates to an optical orbital angular momentum identification method and system based on a binary neural network model. BACKGROUND

[0002] With the deepening of space exploration activities and the growing demand for global information transmission, space laser communication technology has attracted widespread attention due to its high speed, large capacity, low delay and other advantages. As an important physical quantity of light waves, optical orbital angular momentum (OAM) has unique application value in the field of optical communication. By utilizing the diversity of OAM, multiplexing can be achieved, thereby greatly improving the capacity and efficiency of optical communication. Precise identification of OAM is one of the key technologies for its application, therefore, it is of great significance to study new efficient OAM identification methods.

[0003] Existing OAM identification methods mainly include interference method, coordinate transformation method and deep learning method based on spatial light modulator. The interference method utilizes the number of spiral fringes or misaligned fringes generated by the interference of OAM beams with plane waves or spherical waves to realize OAM identification; the coordinate transformation method is a method based on mathematical transformation, which realizes OAM identification by mapping the complex amplitude field of the light beam to a transformed space; the deep learning method based on spatial light modulator utilizes deep neural networks to extract features and classify OAM images, thereby realizing OAM identification. The deep learning method based on spatial light modulator has the advantages of wide recognition range and high recognition accuracy, but its computational complexity is high, and the memory occupation is large, which is difficult to meet the application requirements of lightweight and real-time of space laser communication. SUMMARY

[0004] The purpose of the present application is to solve the technical problem that the OAM identification method based on the deep learning method of spatial light modulator has high computational complexity, large memory occupation, and is difficult to meet the application requirements of lightweight and real-time of space laser communication, and to provide an optical orbital angular momentum identification method and system based on a binary neural network model.

[0005] In order to achieve the above purpose, the technical solution provided by the present application is as follows:

[0006] An optical orbital angular momentum identification method based on a binary neural network model, characterized in that it comprises the following steps:

[0007] S1, a binary neural network model is built;

[0008] The binary neural network model comprises an input layer, at least one processing block and an output layer connected in turn according to input and output; the input layer is used for receiving the interference image data of the OAM and performing binary processing on the interference image data of the OAM; the processing block comprises a binary convolution layer, a normalization layer and a binary activation layer arranged in turn; the binary convolution layer is used for performing convolution operation on the binary-processed interference image data through XOR and PopCount processing to extract feature information; the normalization layer is used for performing normalization processing on the data output by the binary convolution layer; the binary activation layer is used for performing nonlinear mapping on the data output by the normalization layer; and the output layer is used for outputting the final recognition result of the OAM.

[0009] S2, interference image data of different OAMs is collected and preprocessed;

[0010] S3, a training data set and a test data set are established, and the preprocessed interference image data of the OAMs is input into the training data set and the test data set in proportion;

[0011] S4, training of the binary neural network model;

[0012] The training parameters of the binary neural network model are set, the OAM interference image data in the training data set is input into the binary neural network model to train the binary neural network model, the weights of the binary neural network model are iteratively updated, and a trained binary neural network model is obtained;

[0013] S5, OAM recognition;

[0014] The OAM interference image data in the test data set is input into the trained binary neural network model in step S4 for classification and recognition to obtain an OAM recognition result.

[0015] Further, in step S2, the interference image data of the OAM is obtained by the following method:

[0016] The wavefront phase of the signal light and the local oscillator light is modulated respectively to obtain modulated signal light and local oscillator light, the modulated signal light and local oscillator light are combined, and the interference image data of the OAM is obtained by detection.

[0017] Further, in step S4, the OAM interference image data in the training data set is input into the binary neural network model to train the binary neural network model, specifically:

[0018] After rotating, stretching and twisting the training data set, the OAM interference image data in the training data set is input into the binary neural network model in random order, and the binary neural network model is trained by using the back propagation algorithm and the Adam adaptive learning rate optimization method.

[0019] Further, in step S1, the binary neural network model adopts a binary function as an activation function; and the number of the processing blocks is four.

[0020] Further, in step S3, the interference image data of the preprocessed OAM is respectively input into the training data set and the test data set in a ratio of 7:3.

[0021] Further, in step S2, interference image data of different OAMs is collected, and filtering and down-sampling processing is sequentially performed on the interference image data.

[0022] In addition, the application further provides an optical orbital angular momentum recognition system based on a binary neural network model, which is used to implement the above-mentioned optical orbital angular momentum recognition method based on a binary neural network model, and comprises a spatial light modulation unit, a spatial light interference unit and an OAM recognition unit; the spatial light modulation unit is used to perform wavefront phase modulation on generated signal light and emit the signal light; the spatial light interference unit comprises a receiving optical antenna, a local laser, a second collimating mirror, a second polarizer, a beam splitter and a CCD detection module; the receiving optical antenna is used to receive and converge the signal light emitted by the spatial light modulation unit; the second collimating mirror and the second polarizer are sequentially arranged on a laser emission light path of the local laser, and are used to sequentially perform collimation and polarization state filtering on the emitted local light;

[0023] The special feature is that the spatial light interference unit further comprises a second control module and a second wavefront modulation module; the second control module is used to generate a wavefront phase distribution to be modulated onto the local light; an output end of the second control module is connected to an input end of the second wavefront modulation module, and the second wavefront modulation module is located on a light path of the local light filtered by the second polarizer, and is used to load the generated wavefront phase distribution to be modulated onto the local light onto the second wavefront modulation module, so as to realize wavefront phase modulation on the local light; the beam splitter is simultaneously located on the light paths of the modulated local light and the signal light transmitted by the receiving optical antenna, and is used to converge the modulated local light and the received signal light; the CCD detection module (16) is located on an emission light path of the beam splitter, and is used to detect OAM interference images of the signal light and the local light;

[0024] The OAM recognition unit comprises an image acquisition module, an image preprocessing module and a classification recognition module.

[0025] The input end of the image acquisition module is connected with the output end of the CCD detection module, and is used for acquiring OAM interference image data; the input end of the image preprocessing module is connected with the output end of the image acquisition module, and is used for preprocessing the acquired OAM interference image data; the input end of the classification and identification module is connected with the output end of the image preprocessing module, and a binary neural network model is arranged in the classification and identification module, which is used for extracting feature information of the OAM interference image data through the binary neural network model and integrating and classifying the feature information to obtain an OAM identification result.

[0026] The binary neural network model comprises an input layer, at least one processing block and an output layer connected in sequence according to input and output; the input layer is used for receiving the interference image data of the OAM and performing binary processing on the interference image data of the OAM; the processing block comprises a binary convolution layer, a normalization layer and a binary activation layer arranged in sequence; the binary convolution layer is used for performing convolution operation on the binary-processed interference image data through XOR and PopCount processing to extract feature information; the normalization layer is used for performing normalization processing on the data output by the binary convolution layer; the binary activation layer is used for performing nonlinear mapping on the data output by the normalization layer; and the output layer is used for outputting the final identification result of the OAM.

[0027] Further, the spatial light modulation unit comprises a signal laser, an optical power amplifier, a first collimating mirror, a first polarizer, a first control module, a first wavefront modulation module, a reflecting mirror and a transmitting optical antenna.

[0028] The signal laser is used for generating signal light based on continuous laser; the optical power amplifier, the first collimating mirror and the first polarizer are arranged in sequence on the exit light path of the signal light, and are used for sequentially performing power amplification, collimation and polarization state filtering on the signal light; the first control module is used for generating a wavefront phase distribution to be modulated onto the signal light; the output end of the first control module is connected with the input end of the first wavefront modulation module, and the first wavefront modulation module is located on the light path of the signal light filtered by the first polarizer, and is used for loading the generated wavefront phase distribution to be modulated onto the signal light onto the first wavefront modulation module to realize wavefront phase modulation on the signal light; the reflecting mirror is located on the light path of the signal light modulated by the first wavefront modulation module, and is used for changing the direction of the light beam; and the transmitting optical antenna is located on the reflection light path of the reflecting mirror, and is used for transmitting the modulated signal light after beam processing.

[0029] The first wavefront modulation module is a reflective spatial light modulator (SLM) or a transmissive liquid crystal phased array.

[0030] The second wavefront modulation module is a reflective spatial light modulator (SLM) or a transmissive liquid crystal phased array.

[0031] Furthermore, the image preprocessing module includes a filtering module and a downsampling module;

[0032] The input end of the filtering module is connected to the output end of the image acquisition module, and is used to filter the input OAM interferometric image data;

[0033] The input of the downsampling module is connected to the output of the filtering module, and is used to extract and interpolate the filtered OAM interferometric image data.

[0034] The input of the classification and recognition module is connected to the output of the downsampling module.

[0035] Furthermore, the filtering module employs a Gaussian filter, a mean filter, or a median filter to perform filtering; the downsampling module employs nearest neighbor interpolation or bilinear interpolation to perform interpolation processing of the OAM interferometric image data.

[0036] The binary neural network model uses a binarization function as the activation function; the number of processing blocks is four.

[0037] The advantages of this invention compared to the prior art are as follows:

[0038] 1. This invention provides an optical orbital angular momentum (OAM) identification method based on a binary neural network model. By constructing a binary neural network model, it achieves the identification of OAMs with different topological charge numbers. Compared with the traditional OAM identification method based on deep learning of spatial light modulators, this invention adopts binary weights and binary activation functions, which significantly reduces the parameter storage requirements and is suitable for resource-constrained space laser communication systems. At the same time, the binary neural network model uses PopCount processing instead of convolution operations, which reduces the amount of computation and enhances real-time performance, meeting the needs of rapid identification in application scenarios such as inter-satellite communication.

[0039] 2. The optical orbital angular momentum identification method based on a binary neural network model provided by the present invention first performs binarization processing on the interferometric image data of OAM input to the binary neural network model, which has high recognition accuracy and good robustness for noise in the interferometric image data of OAM.

[0040] 3. The optical orbital angular momentum identification system based on a binary neural network model provided by this invention achieves wavefront phase modulation of the local oscillator light through a second wavefront modulation module and a second control module, thereby realizing the removal or reduction of the topological charge number of the local oscillator light OAM. Only a small topological charge number needs to be identified to obtain the topological charge number result of the signal light received at the receiving end, reducing the complexity of identifying large topological charge numbers of OAM and providing better flexibility. At the same time, the binary neural network model set in the OAM identification unit greatly reduces the amount of computation and enhances real-time performance. BRIEF DESCRIPTION OF DRAWINGS

[0041] Figure 1 is a structural schematic diagram of an embodiment of an optical orbital angular momentum recognition system based on a binary neural network model of the application;

[0042] Figure 2 is a structural schematic diagram of an OAM recognition unit in an embodiment of the application;

[0043] Figure 3 is an interference image of signal light and local oscillator light that has not undergone wavefront phase modulation when the OAM topological charge number of the signal light is 1, 2, and 3, respectively, in an embodiment of the application, wherein (a) is an interference image of signal light and local oscillator light that has not undergone wavefront phase modulation when the OAM topological charge number of the signal light is 1, (b) is an interference image of signal light and local oscillator light that has not undergone wavefront phase modulation when the OAM topological charge number of the signal light is 2, and (c) is an interference image of signal light and local oscillator light that has not undergone wavefront phase modulation when the OAM topological charge number of the signal light is 3;

[0044] Figure 4 is an intensity distribution diagram and a phase distribution diagram of signal light that has undergone wavefront phase modulation in an embodiment of the application, wherein (a) is an intensity distribution diagram of signal light that has undergone wavefront phase modulation, and (b) is a phase distribution diagram of signal light that has undergone wavefront phase modulation;

[0045] Figure 5 is an intensity distribution diagram and a phase distribution diagram of local oscillator light that has undergone wavefront phase modulation in an embodiment of the application, wherein (a) is an intensity distribution diagram of local oscillator light that has undergone wavefront phase modulation, and (b) is a phase distribution diagram of local oscillator light that has undergone wavefront phase modulation;

[0046] Figure 6 is an interference image of signal light that has undergone wavefront phase modulation and local oscillator light that has undergone wavefront phase modulation and a phase distribution diagram thereof in an embodiment of the application, wherein (a) is an interference image of signal light that has undergone wavefront phase modulation and local oscillator light that has undergone wavefront phase modulation, and (b) is a phase distribution diagram of the interference image;

[0047] Figure 7 is a classification result schematic diagram of a confusion matrix obtained by performing a classification experiment on OAM interference image data of a test data set using a binary neural network model in an embodiment of the application.

[0048] Specific reference signs are as follows:

[0049] 1 - signal laser; 2 - optical power amplifier; 3 - first collimating mirror; 4 - first polarizer; 5 - first control module; 6 - first wavefront modulation module; 7 - reflecting mirror; 8 - transmitting optical antenna; 9 - receiving optical antenna; 10 - local oscillator laser; 11 - second collimating mirror; 12 - second polarizer; 13 - second control module; 14 - second wavefront modulation module; 15 - beam splitter; 16 - CCD detection module. DETAILED DESCRIPTION

[0050] In order to make the advantages and characteristics of the present application clearer, the present application will be further described in detail below in combination with the drawings and specific embodiments.

[0051] As shown in Figure 1 , an optical orbital angular momentum recognition system based on a binary neural network model includes a spatial light modulation unit, a spatial light interference unit, and an OAM recognition unit. The spatial light modulation unit serves as a transmitting end and includes a signal laser 1, an optical power amplifier 2, a first collimating mirror 3, a first polarizer 4, a first control module 5, a first wavefront modulation module 6, a reflecting mirror 7, and a transmitting optical antenna 8. The spatial light interference unit serves as a receiving end and includes a receiving optical antenna 9, a local oscillator laser 10, a second collimating mirror 11, a second polarizer 12, a second control module 13, a second wavefront modulation module 14, a beam splitter 15, and a CCD detection module 16.

[0052] The signal laser 1 is used to generate stable continuous laser, and the generated continuous laser is used as signal light. The optical power amplifier 2, the first collimating mirror 3, and the first polarizer 4 are sequentially arranged on the outgoing light path of the signal light. The optical power amplifier 2 is used to amplify the power of the signal light to meet the spatial transmission requirements of a certain distance; the first collimating mirror 3 is used to collimate the signal light after power amplification and radiate the collimated signal light from the optical fiber to the space; and the first polarizer 4 is used to filter the polarization state of the collimated signal light.

[0053] The first control module 5 is used to generate a wavefront phase distribution to be modulated onto the signal light. The output end of the first control module 5 is connected to the input end of the first wavefront modulation module 6, and the first wavefront modulation module 6 is located on the light path of the signal light after polarization state filtering by the first polarizer 4. The first wavefront modulation module 6 is used to load the generated wavefront phase distribution to be modulated onto the signal light onto the first wavefront modulation module 6, so as to realize wavefront phase modulation of the signal light. The reflecting mirror 7 is located on the light path of the signal light after modulation by the first wavefront modulation module 6, and is used to change the direction of the light beam. The transmitting optical antenna 8 is located on the reflection light path of the reflecting mirror, and is used to transmit the modulated signal light after beam processing.

[0054] The receiving optical antenna 9 receives and converges the signal light emitted by the transmitting optical antenna 8 and transmitted through the transmission channel. The local laser 10 is used to generate local light, and the second collimating mirror 11 and the second polarizer 12 are sequentially arranged on the outgoing light path of the local light. The second collimating mirror 11 is used to collimate the local light and couple it from the optical fiber into space; and the second polarizer 12 is used to filter the polarization state of the collimated local light.

[0055] The second control module 13 is arranged in a computer and is used to generate a wavefront phase distribution to be modulated onto the local light. The output end of the second control module 13 is connected to the input end of the second wavefront modulation module 14, and the second wavefront modulation module 14 is located on the light path of the local light filtered by the second polarizer 12. The second wavefront modulation module 14 is used to load the generated wavefront phase distribution to be modulated onto the local light onto the second wavefront modulation module 14, so as to realize wavefront phase modulation of the local light. The beam splitter 15 is located on the light paths of the modulated local light and the signal light emitted by the receiving optical antenna 9 at the same time, and is used to converge the modulated local light and the received signal light. At this time, the signal light and the local light will produce coaxial interference. The CCD detection module 16 is located on the outgoing light path of the beam splitter 15 and is used to detect the interference image of the signal light and the local light.

[0056] The first wavefront modulation module 6 and the second wavefront modulation module 14 can adopt, but are not limited to, reflective spatial light modulators (SLMs), transmissive liquid crystal phased arrays, etc. In the embodiment, the first wavefront modulation module 6 and the second wavefront modulation module 14 both adopt reflective spatial light modulators, which are denoted as SLM1 and SLM2 respectively.

[0057] The OAM recognition unit is arranged in a computer and includes an image acquisition module, an image preprocessing module, and a classification recognition module. The input end of the image acquisition module is connected to the output end of the CCD detection module 16, and is used to acquire OAM interference image data. The image preprocessing module includes a filtering module and a downsampling module. The input end of the filtering module is connected to the output end of the image acquisition module, and is used to filter the input OAM interference image data. Specifically, a Gaussian filter or a mean filter or a median filter can be used to realize filtering. The input end of the downsampling module is connected to the output end of the filtering module, and is used to extract and interpolate the filtered OAM interference image data. Specifically, a nearest neighbor interpolation method or a bilinear interpolation method can be used to realize interpolation processing of the OAM interference image data. The input end of the classification recognition module is connected to the output end of the image preprocessing module, and a binary neural network model is arranged inside the classification recognition module. The binary neural network model is used to extract feature information of the OAM interference image data and integrate and classify the feature information, so as to obtain an OAM recognition result. In the embodiment, the second control module 13 and the OAM recognition unit use the same computer for operation.

[0058] AsFigure 2 As shown, the binary neural network model includes an input layer, four processing blocks and an output layer connected in turn according to input and output. The input layer is used to receive the interference image data of the OAM, and performs binary processing on the interference image data of the OAM, and converts it into image data with only two gray scales of black and white, so as to facilitate subsequent processing. The application can realize binary processing by using a global threshold method or a local threshold method or an adaptive threshold method. The processing block includes a binary convolution layer, a normalization layer and a binary activation layer arranged in turn. The binary convolution layer is used to perform XOR operation on the interference image data after binary processing, and count the number of 1 in the XOR result through PopCount processing, so as to realize the function of convolution operation and extract feature information from the interference image data of the OAM; the normalization layer is used to perform normalization processing on the data output by the binary convolution layer; the binary activation layer is used to perform nonlinear mapping on the data output by the normalization layer. The output layer is used to output the final recognition result of the OAM.

[0059] The signal light generated in the embodiment is a vortex plane wave, and the local light is a spherical wave. When the vortex plane wave of the signal light and the spherical wave of the local light perform coaxial interference, the OAM topological charge number of the signal light is l1, the OAM topological charge number of the local light is l2, and the topological charge number of the OAM interference image generated after the interference of the signal light and the local light is l3. Due to the spiral phase structure of the vortex beam, the number of spiral interference fringes in the interference image is l3, and the orbital angular momentum mode can be effectively identified through the number and rotation direction of the spiral fringes in the interference image.

[0060] Specifically, in the embodiment, the electric field E sig of the signal light vortex plane wave is:

[0061]

[0062] The electric field E loc of the local light spherical wave in the embodiment is:

[0063]

[0064] Wherein, A sig represents the amplitude of the signal light vortex plane wave, A loc represents the amplitude of the local light spherical wave, k is the wave number, ω0 is the beam waist radius, r 2 =(x-x0) 2 +(y-y0) 2 , x0 and y0 are the horizontal coordinate and vertical coordinate of the center of the local light spherical wave respectively, and z is the axial coordinate during coaxial interference. i is the imaginary unit, is the phase of the signal light.

[0065] When no wavefront phase modulation is applied to the local oscillator, the OAM topological charge l2 of the local oscillator is 0, and the electric field of the local oscillator degenerates as follows:

[0066]

[0067] The intensity distribution I of the interference pattern on the z=0 plane is:

[0068]

[0069] As shown in the above equation, the interference fringes in the interference image on the plane where z≠0 are spiral-shaped. The number of interference fringes l3 in the figure is equal to the OAM topological charge l1 of the signal light minus the OAM topological charge l2 of the local oscillator light, i.e., l3 = l1 - l2. Since l2 = 0 in this embodiment, the rotation direction of the interference fringes is related to the sign of the OAM topological charge l1 of the vortex plane wave of the signal light. When l1 is negative, the interference fringes rotate clockwise; when l1 is positive, the interference fringes rotate counterclockwise.

[0070] like Figure 3 The figures show interference images of the signal light and the local oscillator light when the OAM topological charge l1 of the signal light is 1, 2, and 3, respectively. (a) shows the interference image of the signal light and the local oscillator light without wavefront phase modulation when the OAM topological charge l1 is 1; (b) shows the interference image of the signal light and the local oscillator light without wavefront phase modulation when the OAM topological charge l1 is 2; and (c) shows the interference image of the signal light and the local oscillator light without wavefront phase modulation when the OAM topological charge l1 is 3. In this embodiment, the wavelength λ of both the signal light and the local oscillator light is set to 1550 nm, the horizontal and vertical coordinates of the spatial range of the observed interference image are set to [-100λ, +100λ], and the phase center distance between the received signal light and SLM2 is 500λ.

[0071] Figure 4 The images show the intensity and phase distribution of the signal light after wavefront phase modulation. Figure 5 The images show the intensity and phase distribution of the local oscillator light after wavefront phase modulation. Figure 6 The image shows the interference pattern and phase distribution of the signal light and the local oscillator light after wavefront phase modulation. It can be seen that the OAM topological charge number l1 = 3 after wavefront phase modulation of the signal light, and l2 = 1 after wavefront phase modulation of the local oscillator light. The number of interference fringes in the resulting interference pattern is l3 = l1 - l2 = 3 - 1 = 2, which reduces the 3rd-order OAM topological charge number to a 2nd-order OAM topological charge number, effectively reducing the complexity of identifying large OAM topological charges.

[0072] On the basis of the optical orbital angular momentum recognition system based on the binary neural network model, the application provides an optical orbital angular momentum recognition method based on a binary neural network model, which specifically comprises the following steps:

[0073] S1. Building a binary neural network model.

[0074] S1.1, building the binary neural network model, wherein the number of processing blocks can be set according to experience, and then adjusted according to the output result of the binary neural network model.

[0075] S1.2, determining the number of layers, dimension, convolution kernel size and other hyperparameters of each layer of the binary neural network model: in this embodiment, the number of processing blocks is four, a binary convolution layer uses a 9x9 convolution kernel for convolution operation, and a normalization layer is used to normalize the data output by the binary convolution layer. In this embodiment, the full connection layer is realized in the form of the convolution layer in the fourth processing block.

[0076] S1.3, selecting an activation function: in this embodiment, a binary function is used as the activation function to complete the nonlinear mapping, that is, f(x) = sign(x); wherein x is an input value, f(x) is an output value of the activation function, and the binary function limits the output value to between 0 and 1.

[0077] The number of processing blocks and the size of the convolution kernel in the binary neural network structure of the application can be determined as needed. The embodiments of the application can also use different training data sets, such as OAM with different topological charges, OAM interference image data under different noise conditions, etc.

[0078] S2, collecting interference image data of different OAM and pre-processing the same.

[0079] The interference image data of different OAM is collected by the CCD detection module 16, and the collected OAM interference image data is filtered and down-sampled by the image preprocessing module to improve the robustness and generalization ability of the algorithm, specifically including:

[0080] a) filtering: filtering the input OAM interference image data to remove noise interference in the image. Wherein the filter can be selected from but not limited to a Gaussian filter, a mean filter, and a median filter.

[0081] b) down-sampling: extracting and interpolating the filtered OAM interference image data to reduce the amount of image data and improve the calculation efficiency. Wherein the interpolation method includes but is not limited to nearest neighbor interpolation and bilinear interpolation.

[0082] S3, establishing a training data set and a test data set, and inputting the interference image data after the preprocessing in step S2 into the training data set and the test data set respectively in proportion, and marking the OAM of the interference image data in the training data set.

[0083] In the embodiment, the signal light generates a vortex plane wave, the local light generates a spherical wave, the interference image data of the interference between the vortex plane wave with the topological charge number of 1-8 and the spherical wave is used as the collected data of the sample data set, the topological charge number corresponding to the light intensity distribution diagram is used as the label, and the data in the sample data set is divided into the training data set and the test data set in the proportion of 7:3.

[0084] S4, training of the binary neural network model;

[0085] S4.1, setting training parameters: setting training optimization strategy, initial learning rate, iteration number and other hyperparameters.

[0086] The training optimization strategy adopts the Adam adaptive learning rate optimization method, and all the learnable parameters in the binary neural network model can be updated through the setting of the training optimization strategy. The initial learning rate is set to 0.01, which is the step length of adjusting the parameters in the training process of the binary neural network model; the iteration number is set to 100 times. In addition, a learning rate scheduler is set, which reduces the learning rate when the loss of the training set of the binary neural network model does not improve in two cycles, so that the new learning rate is half of the current learning rate. The above parameter setting is beneficial to the overfitting of the binary neural network model in the training process and accelerates the convergence of the model.

[0087] S4.2, model training.

[0088] After rotating, stretching and twisting the training data set, the interference image data of the OAM in the training data set is input into the binary neural network model in a random order, the binary neural network model is trained by using the back propagation algorithm and the Adam adaptive learning rate optimization method, the weights of the binary neural network model are iteratively updated, and the trained binary neural network model is obtained.

[0089] S5, OAM identification;

[0090] The OAM interference image data in the test data set is input into the trained binary neural network model, the OAM interference image data in the test data set is classified and identified, and the OAM identification result is obtained, that is, the estimated value of the topological charge number l3 of the OAM interference image in the test data set is obtained, and in the embodiment, the value of l3 is between 1 and 8.

[0091] Figure 7A classification result diagram of a confusion matrix obtained by using the binary neural network model of the embodiment of the present application to classify the OAM interference images of the test data set. The experimental results show that the optical orbital angular momentum recognition method and system based on the binary neural network model can effectively recognize OAMs with different topological charges and has a high recognition accuracy.

[0092] Meanwhile, the present application does not need complex calculations in convolution and nonlinear processing, has high calculation efficiency, and the weights and activation functions of the binary neural network model are binary, so the memory occupation is also significantly reduced compared with the traditional deep learning method.

[0093] The above is only used to illustrate the technical solutions of the present application, not to limit them. For ordinary skilled persons in the art, the specific technical solutions described in the above embodiments can be modified, or some technical features can be replaced equivalently, without changing the essence of the corresponding technical solutions out of the scope of the technical solutions protected by the present application.

Claims

1. A method for optical orbital angular momentum recognition based on a binary neural network model, characterized in that, The method comprises the following steps: S1, building a binary neural network model; The binary neural network model comprises an input layer, at least one processing block and an output layer connected in sequence according to input and output; the input layer is used for receiving interference image data of OAM and performing binary processing on the interference image data of OAM; the processing block comprises a binary convolution layer, a normalization layer and a binary activation layer arranged in sequence; the binary convolution layer is used for performing convolution operation on the binary processed interference image data through XOR and PopCount processing to extract feature information; The normalization layer is used for performing normalization processing on the data output by the binary convolution layer; and the binary activation layer is used for performing nonlinear mapping on the data output by the normalization layer; The output layer is used for outputting the final recognition result of OAM; S2, collecting interference image data of different OAM and pre-processing the same; S3, establishing a training data set and a test data set, and inputting the pre-processed interference image data of OAM into the training data set and the test data set in proportion; S4, training of the binary neural network model; The training parameters of the binary neural network model are set, the OAM interference image data in the training data set is input into the binary neural network model to train the binary neural network model, the weights of the binary neural network model are iteratively updated, and a trained binary neural network model is obtained; S5, OAM recognition; The OAM interference image data in the test data set is input into the trained binary neural network model in step S4 for classification and recognition to obtain an OAM recognition result.

2. The optical orbital angular momentum recognition method based on the binary neural network model according to claim 1, wherein in step S2, the interference image data of OAM is obtained by the following method: The wavefront phase of the signal light and the local oscillator light is modulated respectively to obtain modulated signal light and local oscillator light, the modulated signal light and local oscillator light are combined, and the interference image data of OAM is obtained by detection.

3. The optical orbital angular momentum recognition method based on the binary neural network model according to claim 1 or 2, wherein in step S4, the OAM interference image data in the training data set is input into the binary neural network model to train the binary neural network model, specifically as follows: After the training data set is rotated, stretched and twisted, the OAM interference image data in the training data set is input into the binary neural network model in random order, and the binary neural network model is trained by using a back propagation algorithm and an Adam adaptive learning rate optimization method.

4. The optical orbital angular momentum recognition method based on the binary neural network model according to claim 3, wherein in step S1, the binary neural network model uses a binary function as an activation function; The number of processing blocks is four.

5. The optical orbital angular momentum recognition method based on the binary neural network model according to claim 4, wherein in step S3, the pre-processed interference image data of OAM is input into the training data set and the test data set in proportion of 7:

3. Step S2 is specifically as follows: ​ ​ ​ 6. The method of claim 5, wherein the method is based on a binary neural network model. ​ Collect interference image data of different OAMs, and sequentially perform filtering and down-sampling processing on the interference image data.

7. An optical orbital angular momentum recognition system based on a binary neural network model, configured to implement the method of recognizing optical orbital angular momentum based on a binary neural network model according to any one of claims 1-6, comprising a spatial light modulation unit, a spatial light interference unit, and an OAM recognition unit; the spatial light modulation unit is configured to modulate the wavefront phase of the generated signal light and emit the signal light; the spatial light interference unit comprises a receiving optical antenna (9), a local laser (10), a second collimating mirror (11), a second polarizer (12), a beam splitter (15), and a CCD detection module (16); the receiving optical antenna (9) is configured to receive and converge the signal light emitted by the spatial light modulation unit; the second collimating mirror (11) and the second polarizer (12) are sequentially arranged on the laser emission path of the local laser (10) and are configured to sequentially collimate and polarize the emitted local light; characterized in that the spatial light interference unit further comprises a second control module (13) and a second wavefront modulation module (14); the second control module (13) is configured to generate a wavefront phase distribution to be modulated onto the local light; the output end of the second control module (13) is connected to the input end of the second wavefront modulation module (14), and the second wavefront modulation module (14) is located on the light path of the local light filtered by the second polarizer (12) and is configured to load the generated wavefront phase distribution to be modulated onto the local light onto the second wavefront modulation module (14) to modulate the wavefront phase of the local light; the beam splitter (15) is located on the light paths of the modulated local light and the received signal light transmitted by the receiving optical antenna (9) and is configured to combine the modulated local light and the received signal light; the CCD detection module (16) is located on the emission path of the beam splitter (15) and is configured to detect the OAM interference image of the signal light and the local light; the OAM recognition unit comprises an image acquisition module, an image preprocessing module, and a classification and recognition module; the input end of the image acquisition module is connected to the output end of the CCD detection module (16) and is configured to acquire OAM interference image data; the input end of the image preprocessing module is connected to the output end of the image acquisition module and is configured to preprocess the acquired OAM interference image data; the input end of the classification and recognition module is connected to the output end of the image preprocessing module, and the classification and recognition module is internally provided with a binary neural network model and is configured to extract feature information of the OAM interference image data through the binary neural network model, integrate and classify the feature information, and obtain an OAM recognition result. The binary neural network model comprises an input layer, at least one processing block and an output layer connected in sequence according to input and output; the input layer is used for receiving the interference image data of the OAM and performing binary processing on the interference image data of the OAM; the processing block comprises a binary convolution layer, a normalization layer and a binary activation layer arranged in sequence; the binary convolution layer is used for performing convolution operation on the binary-processed interference image data through XOR and PopCount processing to extract feature information; the normalization layer is used for performing normalization processing on the data output by the binary convolution layer; and the binary activation layer is used for performing nonlinear mapping on the data output by the normalization layer; the output layer is used for outputting the final recognition result of the OAM.

8. The optical orbital angular momentum recognition system based on the binary neural network model according to claim 7, wherein: the spatial light modulation unit comprises a signal laser (1), an optical power amplifier (2), a first collimating mirror (3), a first polarizer (4), a first control module (5), a first wavefront modulation module (6), a reflecting mirror (7) and a transmitting optical antenna (8); the signal laser (1) is used for generating signal light based on continuous laser; the optical power amplifier (2), the first collimating mirror (3) and the first polarizer (4) are arranged in sequence on the exit light path of the signal light, and are used for sequentially performing power amplification, collimation and polarization state filtering on the signal light; the first control module (5) is used for generating a wavefront phase distribution to be modulated onto the signal light; the output end of the first control module (5) is connected to the input end of the first wavefront modulation module (6), and the first wavefront modulation module (6) is located on the light path of the signal light filtered by the first polarizer (4), and is used for loading the generated wavefront phase distribution to be modulated onto the signal light onto the first wavefront modulation module (6) to realize wavefront phase modulation on the signal light; the reflecting mirror (7) is located on the light path of the signal light modulated by the first wavefront modulation module (6), and is used for changing the direction of the light beam, and the transmitting optical antenna (8) is located on the reflection light path of the reflecting mirror (7), and is used for transmitting the modulated signal light after beam processing; the first wavefront modulation module (6) is a reflective spatial light modulator or a transmissive liquid crystal phased array; the second wavefront modulation module (14) is a reflective spatial light modulator or a transmissive liquid crystal phased array.

9. The optical orbital angular momentum recognition system based on the binary neural network model according to claim 8, wherein: the image preprocessing module comprises a filtering module and a downsampling module; the input end of the filtering module is connected to the output end of the image acquisition module, and is used for filtering the input OAM interference image data; the input end of the downsampling module is connected to the output end of the filtering module, and is used for extracting and interpolating the filtered OAM interference image data; the input end of the classification and recognition module is connected to the output end of the downsampling module.

10. The optical orbital angular momentum recognition system based on the binary neural network model according to claim 9, wherein: The filter module adopts a Gaussian filter or a mean filter or a median filter to realize filtering; The downsampling module adopts a nearest neighbor interpolation method or a bilinear interpolation method to realize interpolation processing of the OAM interference image data; The binary neural network model adopts a binary function as an activation function; The number of the processing blocks is four.

Images