Method and device for generating high-order light beam through physical model driven deep learning

Abstract

The invention provides a method and a device for generating a high-order light beam through physical model driven deep learning. The device comprises the following steps: firstly, establishing a non-training phase imaging physical model, loading phase modulation through a plane wave, carrying out Fourier transform by using a Fourier lens, and collecting light field intensity distribution on a rear focal plane through a detector; phase distribution and light field intensity distribution meet the Fourier transform relation, and the Fourier transform relation can be regarded as the process from phase forward coding to diffraction intensity diagram. And then, inputting the acquired light field intensity distribution I into the neural network, and outputting a phase result. And calculating the intensity distribution corresponding to the phase distribution by using the Fourier transform theory, and measuring the error between the input intensity and the calculated intensity. In the whole process, training of the neural network can be completed only by collecting a single light field propagation intensity graph, dependence on a large amount of training data is remarkably reduced, and reconstruction precision and generalization ability of phase imaging are improved.

Description

Technical Field

[0001] The present invention relates to a method and device for generating high-order light beams by driving deep learning with a physical model, and mainly relates to the field of deep learning. Background Art

[0002] In the field of optics and quantum physics, high-order Hermite-Gaussian modes, as an important form of light field distribution, have broad application prospects in quantum communication, precision measurement and other fields. However, the traditional method of generating high-order Hermite-Gaussian modes has many limitations, such as high cost, low efficiency, and poor flexibility. Therefore, seeking a more efficient and flexible generation method has become a research hotspot in this field.

[0003] As a data-driven method, deep learning has demonstrated powerful capabilities in image processing, pattern recognition and other fields. By building a deep neural network structure, it can automatically learn complex pattern features from a large amount of data and has strong generalization capabilities. However, deep learning models are often viewed as "black boxes" whose decision-making processes are difficult to explain and rely on a large amount of labeled data for training, which may face challenges in some physical problems due to data scarcity or high labeling costs.

[0004] At the same time, physical models solve problems by building precise mathematical models based on physical objectives, mechanisms, and prior knowledge. This approach has the advantages of predictable solution quality and deterministic solution methods, and is particularly suitable for physical problems that can be accurately modeled. In applications that generate high-order Hermite-Gaussian patterns, it is expected that accurate generation of patterns can be achieved by building precise physical models. However, the drawback of the model-driven approach is that it is difficult to accurately model in practical applications, which limits its widespread application. Summary of the invention

[0005] In view of the above shortcomings of the prior art, the present invention proposes a device for generating high-order Hermite-Gaussian beams based on physical models and deep learning technology. This solution achieves efficient and accurate phase recovery and beam generation. The core lies in combining the physical model of non-training phase imaging with the deep learning algorithm. The training of the neural network can be completed by collecting the light field intensity distribution once, thereby generating the required high-order Hermite-Gaussian beam.

[0006] To achieve the above object, the technical solution of the present invention is:

[0007] A method for generating a high-order Hermite-Gaussian beam based on a physical model and deep learning technology, characterized in that it comprises the following steps:

[0008] S1, constructing a non-training phase imaging physical model;

[0009] S2, deep learning processing;

[0010] S3, generating high-order Hermite-Gaussian beams.

[0011] Preferably, step S1 includes the following:

[0012] S1.1, prepare a plane wave source and ensure that its wavefront is uniform;

[0013] S1.2, using a phase modulator to modulate a plane wave to a specific phase;

[0014] S1.3, the modulated light wave is subjected to Fourier transform through a Fourier lens;

[0015] S1.4, place a detector on the back focal plane of the Fourier lens to collect the light field intensity distribution I.

[0016] Preferably, in step S1, a relationship between the phase distribution φ and the light field intensity I that satisfies Fourier transform is established, which is regarded as a process of forward encoding the phase information into the diffraction intensity map, and the equation is:

[0017] I=|FFT(A*exp(iΦ)) 2

[0018] Among them, FFT is fast Fourier transform, A is the incident plane wave light field, exp represents the e-index, and i represents the definition of the imaginary unit in complex numbers.

[0019] Preferably, step S2 includes the following:

[0020] S2.1, construct a neural network, whose input layer receives the light field intensity distribution I collected by the detector, and the output layer outputs the predicted phase result;

[0021] S2.2, using the collected light field intensity distribution I as the input of the neural network, performing forward propagation, and obtaining the predicted phase result;

[0022] S2.3, using Fourier transform theory, convert the predicted phase results into corresponding intensity distribution;

[0023] S2.4, calculating the error between the input intensity map I and the calculated intensity distribution as the loss function of the neural network;

[0024] S2.5, through the back propagation algorithm, the weights of the neural network are updated according to the loss function to reduce the error;

[0025] S2.6, repeating steps S2.1 to S2.5 until the loss function reaches a preset threshold or number of iterations, completing the training of the neural network;

[0026] It measures the mean square error between the input intensity distribution I and the standard light field intensity distribution ID, and its expression is:

[0027]

[0028] Wherein, j=1,2,...N,N≥2, is the sampling point sequence in the intensity distribution area, and ID is the standard light field intensity distribution.

[0029] Preferably, step S3 includes the following:

[0030] S3.1, using the trained neural network, perform phase recovery on the new light field intensity distribution to obtain the required phase distribution;

[0031] S3.2, loading the recovered phase distribution onto a phase modulator to phase modulate the plane wave;

[0032] S3.3, the modulated light wave passes through a Fourier lens to generate a high-order Hermite-Gaussian beam;

[0033] S3.4, perform performance testing and application verification on the generated high-order Hermite-Gaussian beams.

[0034] A device for generating a high-order Hermite-Gaussian beam based on a physical model and deep learning technology, based on the method for generating a high-order Hermite-Gaussian beam based on a physical model and deep learning technology as claimed in claim 1, comprising:

[0035] The non-training phase imaging physical model building module is used to set the plane wave loading phase modulation, and perform Fourier transform on the modulated light wave through the Fourier lens to generate the light field intensity distribution on the back focal plane;

[0036] A data acquisition module, used for connecting to the back focal plane of the non-training phase imaging physical model construction module to collect the light field intensity distribution;

[0037] A neural network processing module is used to receive the light field intensity distribution output by the data acquisition module as input, and output a phase result through a deep learning algorithm; the neural network processing module is internally configured with a back-propagation mechanism, and according to the error between the input intensity I and the intensity distribution calculated by Fourier transform theory, the network weight is adjusted to reduce the error;

[0038] A Fourier transform calculation module, connected to the neural network processing module, and calculating the phase result output by the neural network processing module through Fourier transform theory to obtain a corresponding intensity distribution;

[0039] The error measurement and feedback module is used to measure the error between the light field intensity distribution I collected by the data acquisition module and the intensity distribution calculated by the Fourier transform calculation module, and feed back the error signal to the neural network processing module for weight update until the error meets the preset threshold, thereby completing the training of the neural network.

[0040] Preferably, the non-training phase imaging physical model building module includes a phase modulator, a Fourier lens and a detector, the phase modulator is used to phase modulate the plane wave, the Fourier lens is used to perform Fourier transform on the modulated light wave, and the detector is arranged on the rear focal plane to collect the light field intensity distribution I.

[0041] Preferably, the neural network processing module adopts a convolutional neural network to extract features in the light field intensity distribution I and output a phase result.

[0042] Preferably, the error measurement and feedback module uses mean square error, cross entropy loss or other loss functions suitable for measuring intensity distribution differences to calculate the error.

[0043] Preferably, it also includes a beam generating module connected to the neural network processing module, for generating a high-order Hermite-Gaussian beam according to the phase result output by the trained neural network;

[0044] The beam generating module includes a spatial light modulator for modulating an incident light beam according to a phase result output by a neural network to generate a high-order Hermite-Gaussian beam.

[0045] The technical principles and beneficial effects of the present invention are as follows:

[0046] This solution solves the problem faced by existing data-driven deep learning phase imaging methods, namely, over-reliance on a large amount of labeled training data, and only when the distribution of the test data is similar to that of the training set can a relatively satisfactory reconstruction result be obtained. The significant advantage of the present invention is that the entire process only requires the collection of a single light field propagation intensity map to complete the training of the neural network, which greatly simplifies the training process and improves efficiency. Through this method, the present invention can accurately restore the required phase distribution and then generate a high-order Hermite-Gaussian beam, providing a new and powerful tool for research and application in the field of optics.

[0047] This device achieves efficient and accurate phase recovery and beam generation. Its core lies in combining the physical model of non-training phase imaging with the deep learning algorithm. By collecting the light field intensity distribution once, the neural network training can be completed, thereby generating the required high-order Hermite-Gaussian beam.

[0048] The method of combining physical models with deep learning in this solution came into being. This method effectively combines the advantages of physical models and deep learning technology, not only utilizing the prior knowledge and interpretability of physical models, but also relying on the automatic feature extraction and generalization capabilities of deep learning to achieve effective solutions to complex physical problems. In the application of generating high-order Hermite-Gaussian patterns, it is expected that efficient, flexible and interpretable pattern generation can be achieved by combining physical models with deep learning technology.

[0049] The present invention is used to generate high-order Hermite-Gaussian modes. The device is designed to overcome the limitations of traditional methods, achieve more efficient, flexible and interpretable mode generation, and provide new solutions for research and applications in the fields of optics and quantum physics. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only three of the drawings of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0051] Figure 1 The schematic diagram of the structure of a device for generating a high-order Hermite-Gaussian beam based on a physical model and deep learning technology includes an optical phase modulator 1, a Fourier lens 2, an output focal plane 3 and an intensity acquisition module 4;

[0052] Figure 2 This is a schematic diagram of deep learning combined with physical model processing;

[0053] Figure 3 This is a schematic diagram of the result of generating a high-order Hermite-Gaussian beam using the patented method of the present invention, wherein (a) represents a standard high-order Hermite-Gaussian beam, (b) represents the generated high-order Hermite-Gaussian beam, and (c) represents the phase hologram to be loaded. DETAILED DESCRIPTION

[0054] The technical solutions in the present invention will be described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are only preferred embodiments of the present invention, not all embodiments. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0055] Example

[0056] A method for generating a high-order Hermite-Gaussian beam based on a physical model and deep learning technology, characterized in that it comprises the following steps:

[0057] S1, constructing a non-training phase imaging physical model;

[0058] S2, deep learning processing;

[0059] S3, generating high-order Hermite-Gaussian beams.

[0060] Wherein step S1 comprises the following:

[0061] S1.1, prepare a plane wave source and ensure that its wavefront is uniform;

[0062] S1.2, using a phase modulator to modulate the plane wave in a specific phase, this step is the key to generate high-order Hermite-Gaussian beams;

[0063] S1.3, the modulated light wave is subjected to Fourier transform through a Fourier lens;

[0064] S1.4, place a detector on the back focal plane of the Fourier lens to collect the light field intensity distribution I.

[0065] In step S1, a relationship between the phase distribution φ and the light field intensity I that satisfies the Fourier transform is established, which is regarded as the process of forward encoding the phase information into the diffraction intensity map. The equation is:

[0066] I=|FFT(A*exp(iΦ)) 2

[0067] Among them, FFT is fast Fourier transform, A is the incident plane wave light field, exp represents the e-index, and i represents the definition of the imaginary unit in complex numbers.

[0068] Combination Figure 2 As shown, step S2 includes the following:

[0069] S2.1, construct a neural network, whose input layer receives the light field intensity distribution I collected by the detector, and the output layer outputs the predicted phase result;

[0070] S2.2, using the collected light field intensity distribution I as the input of the neural network, performing forward propagation, and obtaining the predicted phase result;

[0071] S2.3, using Fourier transform theory, convert the predicted phase results into corresponding intensity distribution;

[0072] S2.4, calculating the error between the input intensity map I and the calculated intensity distribution as the loss function of the neural network;

[0073] S2.5, through the back propagation algorithm, the weights of the neural network are updated according to the loss function to reduce the error;

[0074] S2.6, repeating steps S2.1 to S2.5 until the loss function reaches a preset threshold or number of iterations, completing the training of the neural network;

[0075] It measures the mean square error between the input intensity distribution I and the standard light field intensity distribution ID, and its expression is:

[0076]

[0077] Wherein, j=1,2,...N,N≥2, is the sampling point sequence in the intensity distribution area, and ID is the standard light field intensity distribution.

[0078] Combination Figure 3 As shown, step S3 includes the following:

[0079] S3.1, using the trained neural network, perform phase recovery on the new light field intensity distribution to obtain the required phase distribution;

[0080] S3.2, loading the recovered phase distribution onto a phase modulator to phase modulate the plane wave;

[0081] S3.3, the modulated light wave passes through a Fourier lens to generate a high-order Hermite-Gaussian beam;

[0082] S3.4, perform performance testing and application verification on the generated high-order Hermite-Gaussian beam to ensure that it meets the expected optical properties and application requirements.

[0083] A device for generating a high-order Hermite-Gaussian beam based on a physical model and deep learning technology, based on the method for generating a high-order Hermite-Gaussian beam based on a physical model and deep learning technology as described in claim 1, characterized in that the optimization of neural network parameters is completed by using a physical model to drive, including:

[0084] The non-training phase imaging physical model building module is used to set the plane wave loading phase modulation, and perform Fourier transform on the modulated light wave through the Fourier lens to generate the light field intensity distribution on the back focal plane;

[0085] A data acquisition module, used for connecting to the back focal plane of the non-training phase imaging physical model construction module to collect the light field intensity distribution;

[0086] A neural network processing module is used to receive the light field intensity distribution output by the data acquisition module as input, and output a phase result through a deep learning algorithm; the neural network processing module is internally configured with a back-propagation mechanism, and according to the error between the input intensity I and the intensity distribution calculated by Fourier transform theory, the network weight is adjusted to reduce the error;

[0087] A Fourier transform calculation module, connected to the neural network processing module, and calculating the phase result output by the neural network processing module through Fourier transform theory to obtain a corresponding intensity distribution;

[0088] The error measurement and feedback module is used to measure the error between the light field intensity distribution I collected by the data acquisition module and the intensity distribution calculated by the Fourier transform calculation module, and feed back the error signal to the neural network processing module for weight update until the error meets the preset threshold, thereby completing the training of the neural network.

[0089] Among them, the entire device can complete the training of the neural network by collecting the light field propagation intensity map once, without relying on a large amount of labeled training data, thereby improving the reconstruction accuracy and generalization ability of phase imaging.

[0090] Combination Figure 1 As shown, the device includes an optical phase modulator 1, a Fourier lens 2, an output focal plane 3 and an intensity acquisition module 4. The Fourier lens 2 is located between the optical phase plate 1 and the output focal plane 3, and the distances between the Fourier lens 2 and the optical phase plate 1 and the output focal plane 3 are all the lens focal length f. The optical phase modulator 1 performs specific phase modulation on the incident plane wave, which is a key step in achieving the generation of high-order Hermite-Gaussian beams, and the desired beam characteristics are achieved by precisely controlling the phase distribution. The Fourier lens 2 is a convex lens, which performs Fourier transform on the modulated light wave, simulates the propagation process of the light wave, and encodes the phase information into the light field intensity distribution. The intensity acquisition module 4 collects the light field intensity distribution at the output focal plane 3.

[0091] Combination Figure 2 As shown in the figure, the convolutional neural network is a multi-layer structure that takes the standard light field intensity distribution ID as input and outputs the predicted phase hologram φ through training. The weight of the network is updated by the RMSE between the input intensity distribution I and the standard light field intensity distribution ID. The structure and number of layers of the neural network can be designed according to actual needs to achieve efficient phase recovery.

[0092] The non-training phase imaging physical model building module includes a phase modulator, a Fourier lens and a detector. The phase modulator is used to phase modulate the plane wave, the Fourier lens is used to perform Fourier transform on the modulated light wave, and the detector is arranged on the rear focal plane to collect the light field intensity distribution I.

[0093] The neural network processing module uses a convolutional neural network to extract features in the light field intensity distribution I and output a phase result.

[0094] The error measurement and feedback module uses mean square error, cross entropy loss or other loss functions suitable for measuring intensity distribution differences to calculate the error.

[0095] It also includes a beam generation module, connected to the neural network processing module, for generating a high-order Hermite-Gaussian beam according to the phase result output by the trained neural network;

[0096] The beam generating module includes a spatial light modulator for modulating an incident light beam according to a phase result output by a neural network to generate a high-order Hermite-Gaussian beam.

[0097] The device first establishes a physical model of non-training phase imaging, loads phase modulation through plane waves, uses Fourier lenses to perform Fourier transform, and collects light field intensity distribution through detectors on the back focal plane. This physical model clearly defines the relationship between phase distribution and light field intensity distribution. The two satisfy the Fourier transform relationship and can be regarded as the process of forward encoding phase to diffraction intensity map. Subsequently, the collected light field intensity distribution I is input into the neural network, and a phase result is output. The intensity distribution corresponding to the phase distribution is calculated using Fourier transform theory, and the error between the input intensity and the calculated intensity is measured. The network weights are updated by back propagation to reduce the error. The entire process only requires the collection of a single light field propagation intensity map to complete the training of the neural network, which significantly reduces the dependence on a large amount of training data and improves the reconstruction accuracy and generalization ability of phase imaging. The present invention provides an efficient and accurate new method for the generation of high-order Hermite-Gaussian beams and the field of phase imaging.

[0098] This solution solves the problem faced by existing data-driven deep learning phase imaging methods, namely, over-reliance on a large amount of labeled training data, and only when the distribution of the test data is similar to that of the training set can a relatively satisfactory reconstruction result be obtained. The significant advantage of the present invention is that the entire process only requires the collection of a single light field propagation intensity map to complete the training of the neural network, which greatly simplifies the training process and improves efficiency. Through this method, the present invention can accurately restore the required phase distribution and then generate a high-order Hermite-Gaussian beam, providing a new and powerful tool for research and application in the field of optics.

[0099] This device achieves efficient and accurate phase recovery and beam generation. Its core lies in combining the physical model of non-training phase imaging with the deep learning algorithm. By collecting the light field intensity distribution once, the neural network training can be completed, thereby generating the required high-order Hermite-Gaussian beam.

[0100] The method of combining physical models with deep learning in this solution came into being. This method effectively combines the advantages of physical models and deep learning technology, not only utilizing the prior knowledge and interpretability of physical models, but also relying on the automatic feature extraction and generalization capabilities of deep learning to achieve effective solutions to complex physical problems. In the application of generating high-order Hermite-Gaussian patterns, it is expected that efficient, flexible and interpretable pattern generation can be achieved by combining physical models with deep learning technology.

[0101] The present invention is used to generate high-order Hermite-Gaussian modes. The device is designed to overcome the limitations of traditional methods, achieve more efficient, flexible and interpretable mode generation, and provide new solutions for research and applications in the fields of optics and quantum physics.

[0102] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for generating high-order Hermite-Gaussian beams based on physical models and deep learning technology, characterized in that: The steps include: S1, constructing a non-training phase imaging physical model; S2, deep learning processing; S3, generating high-order Hermite-Gaussian beams.

2. The method for generating a high-order Hermite-Gaussian beam based on a physical model and deep learning technology according to claim 1, characterized in that: Wherein step S1 comprises the following: S1.1, prepare a plane wave source and ensure that its wavefront is uniform; S1.2, using a phase modulator to modulate a plane wave to a specific phase; S1.3, the modulated light wave is subjected to Fourier transform through a Fourier lens; S1.4, place a detector on the back focal plane of the Fourier lens to collect the light field intensity distribution I.

3. The method for generating a high-order Hermite-Gaussian beam based on a physical model and deep learning technology according to claim 1, characterized in that: In step S1, a relationship between the phase distribution φ and the light field intensity I that satisfies the Fourier transform is established, which is regarded as the process of forward encoding the phase information into the diffraction intensity map. The equation is: I=|FFT(A*exp(iΦ)) 2 Among them, FFT is fast Fourier transform, A is the incident plane wave light field, exp represents the e-index, and i represents the definition of the imaginary unit in complex numbers.

4. The method for generating a high-order Hermite-Gaussian beam based on a physical model and deep learning technology according to claim 1, characterized in that: Wherein step S2 comprises the following: S2.1, construct a neural network, whose input layer receives the light field intensity distribution I collected by the detector, and the output layer outputs the predicted phase result; S2.2, using the collected light field intensity distribution I as the input of the neural network, performing forward propagation, and obtaining the predicted phase result; S2.3, using Fourier transform theory, convert the predicted phase results into corresponding intensity distribution; S2.4, calculating the error between the input intensity map I and the calculated intensity distribution as the loss function of the neural network; S2.5, through the back propagation algorithm, the weights of the neural network are updated according to the loss function to reduce the error; S2.6, repeating steps S2.1 to S2.5 until the loss function reaches a preset threshold or number of iterations, completing the training of the neural network; It measures the mean square error between the input intensity distribution I and the standard light field intensity distribution ID, and its expression is: Where j = 1, 2, ... N, N ≥ 2, is the sampling point sequence in the intensity distribution area, I D is the standard light field intensity distribution.

5. The method for generating a high-order Hermite-Gaussian beam based on a physical model and deep learning technology according to claim 1, characterized in that: Wherein step S3 comprises the following: S3.1, using the trained neural network, perform phase recovery on the new light field intensity distribution to obtain the required phase distribution; S3.2, loading the recovered phase distribution onto a phase modulator to phase modulate the plane wave; S3.3, the modulated light wave passes through a Fourier lens to generate a high-order Hermite-Gaussian beam; S3.4, perform performance testing and application verification on the generated high-order Hermite-Gaussian beams.

6. The device for generating high-order Hermite-Gaussian beams based on physical models and deep learning technology according to claim 1, characterized in that: A method for generating a high-order Hermite-Gaussian beam based on a physical model and deep learning technology according to claim 1, comprising: The non-training phase imaging physical model building module is used to set the plane wave loading phase modulation, and perform Fourier transform on the modulated light wave through the Fourier lens to generate the light field intensity distribution on the back focal plane; A data acquisition module, used for connecting to the back focal plane of the non-training phase imaging physical model construction module to collect the light field intensity distribution; A neural network processing module is used to receive the light field intensity distribution output by the data acquisition module as input, and output a phase result through a deep learning algorithm; the neural network processing module is internally configured with a back-propagation mechanism, and according to the error between the input intensity I and the intensity distribution calculated by Fourier transform theory, the network weight is adjusted to reduce the error; A Fourier transform calculation module, connected to the neural network processing module, and calculating the phase result output by the neural network processing module through Fourier transform theory to obtain a corresponding intensity distribution; The error measurement and feedback module is used to measure the error between the light field intensity distribution I collected by the data acquisition module and the intensity distribution calculated by the Fourier transform calculation module, and feed back the error signal to the neural network processing module for weight update until the error meets the preset threshold, thereby completing the training of the neural network.

7. The device for generating high-order Hermite-Gaussian beams based on physical models and deep learning technology according to claim 6, characterized in that: The non-training phase imaging physical model building module includes a phase modulator, a Fourier lens and a detector. The phase modulator is used to phase modulate the plane wave, the Fourier lens is used to perform Fourier transform on the modulated light wave, and the detector is arranged on the rear focal plane to collect the light field intensity distribution I.

8. The device for generating high-order Hermite-Gaussian beams based on physical models and deep learning technology according to claim 6, characterized in that: The neural network processing module uses a convolutional neural network to extract features in the light field intensity distribution I and output a phase result.

9. The device for generating high-order Hermite-Gaussian beams based on physical models and deep learning technology according to claim 6, characterized in that: The error measurement and feedback module uses mean square error, cross entropy loss or other loss functions suitable for measuring intensity distribution differences to calculate the error.

10. The device for generating high-order Hermite-Gaussian beams based on physical models and deep learning technology according to claim 6, characterized in that: It also includes a beam generation module, connected to the neural network processing module, for generating a high-order Hermite-Gaussian beam according to the phase result output by the trained neural network; The beam generating module includes a spatial light modulator for modulating an incident light beam according to a phase result output by a neural network to generate a high-order Hermite-Gaussian beam.

Images