All-optical vortex beam orbital angular momentum mode identification method and system

Abstract

The invention discloses an all-optical vortex beam orbital angular momentum mode identification method and system, and relates to the field of optical information processing and optical computing, and the method comprises the steps: S1, constructing a multi-layer architecture all-optical diffraction network; s2, dividing the output detection surface after the diffraction network into a plurality of preset space regions; s3, the vortex light beams carrying orbital angular momentum OAM information sequentially penetrate through the all-optical diffraction network, spatial evolution of a light field in the propagation process is completed, and vortex light classified according to intensity distribution is output; s4, mapping the topological charges of the vortex light to different sub-regions of the detection surface in an intensity distribution form; and S5, performing intensity acquisition on each area on the detection surface through a detector, and determining an OAM mode type of the input light field based on the mode number corresponding to the area with the maximum light intensity. The method is used for performing high-robustness identification on the orbital angular momentum vortex light mode of the vortex light in the presence of a random phase scattering medium.

Description

Technical Field

[0001] This invention relates to the fields of optical information processing and optical computing. More specifically, this invention relates to a method and system for all-optical vortex beam orbital angular momentum pattern recognition. Background Technology

[0002] With the rapid development of technologies such as optical communication, quantum information processing, particle manipulation, and high-resolution imaging, the orbital angular momentum (OAM) carried by photons has become one of the important degrees of freedom for optical field multiplexing and transmission and high-dimensional information encoding. OAM vortex beams, due to their unique helical phase front, possess infinite-dimensional orthogonal mode characteristics and can be used to construct high-capacity, low-interference parallel channels, showing great potential in applications such as free-space communication, image recognition, and optical encryption. However, in practical applications, OAM mode recognition faces many challenges, especially given the inevitable phase perturbations and scattering media (such as atmospheric turbulence, biological tissue, and complex structures) in the optical field propagation path. These random phase perturbations can severely damage the wavefront structure of the OAM beam, leading to a decrease in mode recognition rate or even failure. Therefore, achieving highly robust and accurate OAM mode recognition, especially rapid recognition in unknown or uncontrollable scattering environments, has become a crucial problem that urgently needs to be solved in the field of optical communication and information processing.

[0003] Currently, most mainstream OAM (Optical Mode Imaging) recognition schemes rely on digital image processing and electronic neural networks, such as convolutional neural networks (CNNs) and fully connected neural networks (FCNs). These schemes use scattered intensity images as input and then perform subsequent image reconstruction and classification using a computer terminal. However, these electronic computing architectures suffer from two main bottlenecks. First, the computing speed is limited by the von Neumann architecture, leading to latency issues when handling high-dimensional data processing and real-time recognition tasks. Second, these systems are typically energy-intensive and complex, making it difficult to achieve low power consumption, miniaturization, and light-speed processing. To improve the real-time performance and energy efficiency of OAM modality recognition, researchers have recently proposed optical neural networks (ONNs) or diffractive deep neural networks (D2NNs). These networks use light fields as information carriers, employing multiple passive diffraction layers to modulate the incident light wavefront, enabling adaptive encoding and decoding of target information in space. This approach offers advantages such as light-speed computing, zero-power inference, and high parallelism, and has been widely used in tasks such as image recognition, edge detection, and target classification, demonstrating processing capabilities unmatched by traditional electronic neural networks. Despite this, most existing D2NN models are designed for scenarios with clear structures and stable optical paths, making them difficult to apply directly to real-world environments with strong scattering or unknown disturbances. On one hand, after passing through a random phase diffuser, the original spiral wavefront structure of the vortex light mode is disrupted, and traditional optical transformations such as Fourier transforms and interferometry cannot accurately restore its original modal information. On the other hand, since D2NNs are passive modulation systems, their recognition performance is highly susceptible to mechanical assembly errors and manufacturing defects, such as translational and axial shifts between diffraction layers, and local optical phase defects, which can lead to recognition failure in severe cases. Furthermore, there is currently a lack of a unified network training strategy to resist the combined disturbances of scattering, interference, and assembly errors. In the presence of a scattering medium, diffraction layer parameters trained using only a single OAM sample often lack generalization ability for unseen scattering conditions. In experimental assembly, the ideal diffraction layer position inevitably deviates from the actual system, placing higher demands on the model's robustness. Therefore, constructing an all-optical OAM recognition system with anti-scattering, anti-error, and high robustness is one of the current technical bottlenecks in all-optical network research and OAM pattern recognition.

[0004] Based on the above challenges, there is an urgent need to propose an all-optical diffraction network scheme for fast and efficient identification of OAM vortex optical modes. This scheme should fully integrate the optimization capabilities of deep learning with the parallel transmission characteristics of optical systems, be independent of electronic computing resources, achieve nonlinear mapping from phase to intensity through physical layer modulation, and effectively resist the effects of manufacturing errors and assembly misalignments, providing a theoretical and engineering foundation for the practical applications of future optical communication, optical computing, and intelligent sensing systems. Summary of the Invention

[0005] One object of the present invention is to solve at least the above-mentioned problems and / or defects, and to provide at least the advantages described below.

[0006] To achieve these objectives and other advantages of the present invention, a fully optical method for recognizing the orbital angular momentum pattern of a vortex beam is provided, comprising: S1. Construct a multi-layered all-optical diffraction network, which includes multiple diffraction modulation layers. Each diffraction modulation layer is trained based on the principle of optical diffraction neural networks. Each diffraction modulation layer includes several optical neurons, and the size of each neuron is designed to be 8 micrometers. The phase value of each neuron is constrained to between 0 and 2 using the Softmax function. ; S2. Divide the output detection surface after the diffraction network into several preset spatial regions. The number of each spatial region corresponds to one of the OAM modes trained and recognized by the all-optical diffraction network. S3. The vortex beam carrying orbital angular momentum (OAM) information passes through the all-optical diffraction network in sequence. By continuously controlling the wavefront through the fixed phase modulation mode in each diffraction modulation layer, the spatial evolution of the light field during propagation is completed, and the output is vortex light classified by intensity distribution. S4. The vortex light output through the all-optical diffraction network has its topological charge number mapped onto different sub-regions of the detector surface in the form of intensity distribution. S5. The intensity of each region on the detection surface is collected by the detector, and the OAM mode type of the input light field is determined based on the mode number corresponding to the region with the maximum light intensity.

[0007] Preferably, in S1, each diffraction modulation layer is trained using a deep learning framework through an optical diffraction neural network, and the input of the deep learning framework is set to a vortex light wave carrying different orbital angular momentum (OAM) modes, and the target output of the deep learning framework is to maximize the energy of a specific region corresponding to each OAM mode on the detection plane. During the training process using a deep learning framework, a loss function is used to minimize the error between the network's predicted output and the target distribution.

[0008] Preferably, in S1, random errors are introduced during each iteration of optimization of each diffraction modulation layer, so that the phase distribution after training is superimposed with error surface patterns in a uniform distribution form, so that when each diffraction modulation layer converges the signal of the input target to the corresponding sub-region after the iteration optimization is completed, the mechanical alignment error between the diffraction modulation layers and the manufacturing error of the diffraction modulation layer are overcome. The random errors include: translational deviations of each modulation layer in space and phase shift errors of neurons.

[0009] Preferably, the random error is obtained by modeling the random phase scattering medium using the following formula: in, D ( x , y The random height of the medium in the optical path is represented by Δn, where Δn represents the emissivity of the medium, λ represents the wavelength of the coherent light source, and t represents the wavelength of the light source. D [x,y] represents the transmission coefficient of the random scattering medium.

[0010] In S2, the spatial regions are configured as 12, and the OAM modes corresponding to each number are as follows: arrive ,in, The topological charge number represents the orbital angular momentum of a vortex light.

[0011] An optical system employing an all-optical vortex beam orbital angular momentum pattern recognition method includes: A continuous-wave laser that provides a stable monochromatic coherent light source; A polarizer is placed at the exit of the continuous laser to construct the main optical path; Set up a beam splitter in the main optical path; A vortex phase loader that works in conjunction with one of the sub-optical paths of the beam splitter; An all-optical diffraction network that works in conjunction with another sub-optical path of the beam splitter; The detector is located on the output plane behind the all-optical diffraction network.

[0012] The present invention has at least the following beneficial effects: Firstly, this invention constructs an all-optical diffraction network structure that does not require electronic computing. By propagating and interfering with the light field between multiple passive modulators, it achieves direct identification of orbital angular momentum (OAM) vortex light modes, significantly reducing energy consumption and system latency, and possessing natural advantages such as light-speed processing and zero-power inference.

[0013] Secondly, this invention introduces a random phase scatterer into the optical path for training. This scattering medium simulates the complex environment in the optical path, enhancing the network's adaptability and generalization ability to complex perturbation environments. It can maintain high recognition accuracy under unknown scattering conditions, solving the problem of strong environmental dependence of traditional methods.

[0014] Third, this invention employs a robustness enhancement strategy, simulating manufacturing and assembly errors during the training phase, enabling the network to exhibit excellent stability and fault tolerance in actual physical construction. Furthermore, the network can achieve a nonlinear transformation from phase information to intensity spatial distribution, resulting in output intensity with clear spatial separability, facilitating rapid readout by various detectors and significantly simplifying the recognition process.

[0015] In summary, this invention combines high efficiency, reliability, and practicality in OAM identification tasks, and has broad application prospects and promotional value.

[0016] Other advantages, objectives and features of the present invention will become apparent in part from the following description, and in part from those skilled in the art through study and practice of the invention. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the optical system of the present invention; Among them, continuous laser-1, polarizer-2, beam splitter-3, vortex phase loader-4, diffraction modulation layer I-5, diffraction modulation layer II-6, diffraction modulation layer III-7, diffraction modulation layer IV-8, and CCD-9. Detailed Implementation

[0018] The present invention will now be described in further detail with reference to the accompanying drawings, so that those skilled in the art can implement it based on the description.

[0019] This invention proposes an all-optical diffraction network structure and its training method that do not require electronic computation, enabling an OAM pattern recognition system and method under random phase perturbation conditions, comprising the following steps: Step 1: This step first designs multiple diffraction modulation layers based on the principle of optical diffraction neural networks, constructing a multi-layered all-optical network system. Each modulation surface consists of several optical neurons, which act on the incident light field in the form of phase modulation and have a trainable phase response function. During the training phase, using a deep learning framework (such as the error backpropagation algorithm), the input is set as a vortex light wave carrying different orbital angular momentum (OAM) modes, and the target output is to maximize the energy of a specific region corresponding to each OAM mode on the detection plane. The error between the network's predicted output and the target distribution is minimized through a loss function (such as cross-entropy loss). To improve the significance of the light intensity distribution within the detection region, physical constraints or structural priors are introduced; here, a joint loss function is used. L total ,and L total The representation method is as follows: in, α Represents the weighting coefficient, and α The value ranges from 0 to 1; the log cross-entropy loss function L CE To ensure recognition accuracy, and L CE The representation method is as follows: In the above formula, C represents the number of categories. t i When the true category is the i-th category, t i =1, other categories t j =0, z i This represents the logit value output by the model.

[0020] Second item L percep is the perceptual loss function used to improve the saliency of the light intensity distribution on the probe surface. Its specific expression is: Where N represents the number of sample pixels, and α is the scaling factor. X Indicates the prediction result. Y Indicates a label.

[0021] After training, the phase distribution parameters of each modulation layer are fixed, thus forming a passive all-optical diffraction network that can directly perform OAM recognition without electronic computation. To design the physical parameters of the diffraction network, each neuron is designed to be 8 micrometers in size and function as an independent modulated unit. This unit emits optical signals to the next diffraction modulation layer while simultaneously receiving optical signals emitted by all neurons in the previous layer. Considering that phase modulation alone has better modulation capability and higher diffraction efficiency than amplitude modulation, the Softmax function is used during optimization to constrain the phase value of each neuron to between 0 and 2. .

[0022] Since this scheme employs multiple cascaded diffraction modulation layers, mechanical alignment errors between layers become a challenge affecting the final inference performance, especially in the visible light band where such errors are highly sensitive. To avoid deviations between the actual deployment of the optical diffraction modulation layers and the theoretical design, which could lead to a decrease in overall optical inference performance, this step introduces a "robustness enhancement mechanism" during the training phase to effectively improve the network's tolerance to mechanical assembly and manufacturing errors. Specifically, in each iteration of training, the translational deviations of each modulation layer in space (such as lateral δx, longitudinal δy, and axial δz) are randomly simulated, where: Here, U represents a random uniform distribution used to control manufacturing errors, ∆x represents random lateral displacement, ∆y represents longitudinal displacement in the y-direction, and ∆z represents axial displacement between modulation layers. ∆x, ∆y, and ∆z are all used to control errors during mechanical assembly. These errors are superimposed on the ideal position and ideal phase surface in a uniform distribution, simulating the effects of inaccurate modulator positioning or phase surface manufacturing errors in real systems. In the training of the diffraction network after introducing random errors, the optimization objective is not only to correctly converge the input target signal to the corresponding sub-region but also to overcome mechanical alignment errors between diffraction modulation layers and manufacturing errors of the diffraction modulation layers. By incorporating these perturbations into the training, the network can learn the corresponding optimal response under biased conditions, thereby forming an optical mapping capability that is robust to structural and process errors, significantly improving the stability and recognition reliability of the system in subsequent actual physical construction. The purpose of training optimization is to concentrate the energy of vortex light carrying orbital angular momentum into the corresponding detection area as much as possible. To achieve this, the maximum value of the optical intensity signal in 12 sub-regions is selected, and the category of the input target is determined by the position corresponding to the maximum value. After 50 training iterations, the phase modulation parameters of the diffraction network are fixed, and a physical network is formed. The modulation layer parameters can be physically manufactured / loaded using a liquid crystal spatial light modulator or metasurface structure, thereby achieving all-optical target category determination. Step 2: In order to achieve all-optical recognition of vortex light patterns, the topological charge of the vortex light is mapped onto the detection surface in the form of intensity distribution. Specifically, the detection surface is divided into 12 detection sub-regions, each of which represents the corresponding vortex light pattern.

[0023] Step 3: To achieve the adaptive capability of the diffraction network to the random scattering medium in the optical path, a random phase scatterer is deliberately introduced into the optical path for training. That is, during the training process, a random scattering medium is used to simulate the complex situation in the optical path. Specifically, the random phase scattering medium is modeled as follows: in, D ( x , y The random height of the medium in the optical path is represented by Δn, and the randomness follows a uniform distribution. Δn represents the refractive index of the medium, λ represents the wavelength of the coherent light source, and t represents the wavelength of the light source. D [x,y] represents the transmission coefficient of the random scattering medium.

[0024] In each training round, n different phase scattering media are randomly generated, so that the diffraction network must not only learn the correct mapping of the input target to the detector surface, but also overcome the negative impact of the scattering media in the optical path.

[0025] Step 4: After training, the all-optical diffraction network can be directly used in the inference stage without further computation. The system sequentially passes the distorted beam carrying OAM information through multiple trained diffraction modulation layers. Each layer continuously modulates the wavefront according to its fixed phase modulation mode, realizing the spatial evolution of the light field during propagation. Although each layer's diffraction is linear optical propagation, the entire network structure achieves a nonlinear mapping of the input phase information through deep design, transforming the originally indistinguishable phase structure into differences in spatial intensity distribution. This "phase-intensity" nonlinear transformation is the core of the recognition mechanism of this invention, enabling different OAM modes to correspond to different position regions on the output plane, thereby achieving spatially distinguishable light intensity output and providing a basis for subsequent recognition. Step 5: In this step, the output detector surface after the diffraction network is divided into several preset spatial regions. Each sub-region corresponds to a specific OAM mode number identified by the network during training (e.g., arrive After the input vortex light is modulated by the network, its output intensity distribution on the detection plane exhibits a clear spatial concentration, meaning the energy corresponding to the OAM mode is mainly concentrated in the region corresponding to that mode. By setting up a photosensitive detector array or using a camera to collect intensity in different zones, it is possible to determine in real time which region has the highest light intensity, thereby inferring the OAM mode type of the input light field. This discrimination process does not involve electronic computation and can be achieved solely through light intensity distribution measurement, thus possessing the advantages of low latency, high parallelism, and all-optical recognition.

[0026] An optical system, mainly comprising: (1) A continuous laser 1 is used to provide a stable monochromatic coherent light source with a wavelength of 532 nm or other selectable green light band. The beam generated by the laser propagates along the main optical path and is the incident source of the entire diffraction recognition optical system. This light source ensures high spatial coherence and stability, which is the basis for subsequent spatial modulation and interference.

[0027] (2) Polarizer 2 is placed at the exit of the continuous laser to convert the laser output from the continuous laser into linearly polarized light to meet the polarization direction requirements of the all-optical diffraction network. The polarizer can also effectively filter out non-linearly polarized components, improving modulation efficiency and system stability.

[0028] (3) The beam splitter 3 is disposed in the main optical path to split the incident laser beam into two paths. One path continues to be transmitted in a straight line to the subsequent system, while the other path is reflected onto the vortex phase loading path, thereby realizing the requirements of beam path control, reflection, and beam combining. In some embodiments, the beam splitter can be a 45° inclined beam splitter prism or a cubic beam splitter, which can be designed according to different wavelengths to achieve efficient beam splitting.

[0029] (4) The vortex phase loader 4 is used to load orbital angular momentum (OAM) information onto the laser beam. This loader can take the form of a spiral phase plate, a spatial light modulator, or a coded DOE, etc., to convert the input light field into a vortex beam with a specific OAM mode. The loaded mode is... arrive Integer values ​​within the range correspond to different spiral phase structures.

[0030] (5) Diffraction modulation layers I5, II6, III7, and IV8 constitute a four-layer all-optical diffraction network, and each diffraction modulation layer uses a spatial light modulator to perform layer-by-layer phase modulation on the distorted vortex light wave. Each layer is a passive phase surface containing multiple programmable neuron structures, which have been trained through deep learning to obtain the optimal phase distribution. The light field propagates between each layer in free space (Fresnel approximation), gradually completing the nonlinear mapping from phase to intensity, forming a discriminative light intensity output mode.

[0031] The interlayer spacing between the aforementioned diffraction modulation layers I5, II6, III7, and IV8 can be selected from 5cm to 10cm, specifically determined through simulation optimization to ensure sufficient diffraction propagation and adequate neuronal interconnectivity. Furthermore, because phase perturbation and interlayer offset are introduced during the training process of each diffraction modulation layer, the network robustness is improved, ensuring that recognition performance is maintained even under manufacturing errors in actual systems.

[0032] (6) The CCD detector 9 is located on the output plane behind the last diffraction modulation layer and is used to receive the intensity distribution of the modulated output light. This plane is divided into several spatial regions according to the recognition task, and each sub-region corresponds to an OAM mode tag. When the system is running, the OAM mode carried by the input light field can be determined by judging which region on the CCD has the greatest light intensity.

[0033] In actual operation, the system first emits linearly polarized light from a laser, which is selectively guided by beam splitter 3 and then loaded with the spiral phase of the target OAM mode by a phase loader. After random perturbation, the light enters a four-layer diffraction optical network, where the light field diffracts and propagates layer by layer and is affected by the phase of the modulator, ultimately forming a spatial intensity pattern with classification characteristics on the output surface. The detector samples the intensity of this pattern and outputs the corresponding OAM mode recognition result through region matching.

[0034] The above solution is merely an illustration of a preferred example and is not limited thereto. When implementing this invention, appropriate substitutions and / or modifications can be made according to the user's needs.

[0035] Although embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. It can be applied to various fields suitable for the present invention. Other modifications can be readily made by those skilled in the art. Therefore, without departing from the general concept defined by the claims and their equivalents, the present invention is not limited to the specific details and examples shown and described herein.

Claims

1. A fully optical method for recognizing the orbital angular momentum pattern of a vortex beam, characterized in that, include: S1. Construct a multi-layered all-optical diffraction network, which includes multiple diffraction modulation layers. Each diffraction modulation layer is trained based on the principle of optical diffraction neural networks. Each diffraction modulation layer includes several optical neurons, and the size of each neuron is designed to be 8 micrometers. The phase value of each neuron is constrained to between 0 and 2 using the Softmax function. ; S2. Divide the output detection surface after the diffraction network into several preset spatial regions. The number of each spatial region corresponds to one of the OAM modes trained and recognized by the all-optical diffraction network. S3. The vortex beam carrying orbital angular momentum (OAM) information passes through the all-optical diffraction network in sequence. By continuously controlling the wavefront through the fixed phase modulation mode in each diffraction modulation layer, the spatial evolution of the light field during propagation is completed, and the output is vortex light classified by intensity distribution. S4. The vortex light output through the all-optical diffraction network has its topological charge number mapped onto different sub-regions of the detector surface in the form of intensity distribution. S5. The intensity of each region on the detection surface is collected by the detector, and the OAM mode type of the input light field is determined based on the mode number corresponding to the region with the maximum light intensity.

2. The all-optical vortex beam orbital angular momentum pattern recognition method as described in claim 1, characterized in that, In S1, each diffraction modulation layer is trained using a deep learning framework through an optical diffraction neural network. The input of the deep learning framework is set to vortex light waves carrying different orbital angular momentum (OAM) modes. The target output of the deep learning framework is to maximize the energy of a specific region corresponding to each OAM mode on the detection plane. During the training process using a deep learning framework, a loss function is used to minimize the error between the network's predicted output and the target distribution.

3. The all-optical vortex beam orbital angular momentum pattern recognition method as described in claim 1, characterized in that, In S1, random errors are introduced during each iteration of optimization of each diffraction modulation layer, so that error surface patterns are superimposed in a uniform distribution in the phase distribution after training. After the iteration optimization is completed, when each diffraction modulation layer converges the signal of the input target to the corresponding sub-region, the error surface patterns can overcome the mechanical alignment error between the diffraction modulation layers and the manufacturing error of the diffraction modulation layers. The random errors include: translational deviations of each modulation layer in space and phase shift errors of neurons.

4. The all-optical vortex beam orbital angular momentum pattern recognition method as described in claim 3, characterized in that, The random error is obtained by modeling the random phase scattering medium using the following formula: in, D ( x , y ) represents the random height of the medium in the optical path, and the randomness follows a uniform distribution; ∆n represents the refractive index of the medium; λ represents the wavelength of the coherent light source in air; t D [x,y] represents the transmission coefficient of the random scattering medium.

5. The all-optical vortex beam orbital angular momentum pattern recognition method as described in claim 1, characterized in that, In S2, the spatial regions are configured as 12, and the OAM modes corresponding to each number are as follows: arrive ,in, The topological charge number represents the orbital angular momentum of a vortex light.

6. An optical system employing the all-optical vortex beam orbital angular momentum pattern recognition method as described in any one of claims 1-5, characterized in that, include: A continuous-wave laser that provides a stable monochromatic coherent light source; A polarizer is placed at the exit of the continuous laser to construct the main optical path; Set up a beam splitter in the main optical path; A vortex phase loader that works in conjunction with one of the sub-optical paths of the beam splitter; An all-optical diffraction network that works in conjunction with another sub-optical path of the beam splitter; The detector is located on the output plane behind the all-optical diffraction network.

Images