A method for correcting co-phase error based on all-optical diffraction neural network

Abstract

This invention discloses a co-phase error correction method based on an all-optical diffraction neural network, which can be used for real-time correction of sub-aperture co-phase errors in optical synthetic aperture imaging systems. The method includes: generating an optical synthetic aperture imaging light field using a synthetic aperture imaging module and implementing co-phase error loading and compensation; extracting features and converting information from the input light field using a diffraction neural network module; and receiving the output light field from the all-optical diffraction neural network module using a co-phase error detection module, converting the output light field information into a light intensity distribution to characterize the predicted co-phase error value. Compared to previous co-phase detection techniques that rely on computer hardware, this method utilizes the free-space propagation characteristics of light and the principle of vector superposition to achieve matrix operations, possessing unique advantages in ultra-parallel processing and information transmission. It can perform co-phase detection tasks at the speed of light, which is particularly significant for high-frequency phase difference detection research in the field of aero-optics.

Description

Technical Field

[0001] This invention belongs to the field of optical synthetic aperture imaging, and specifically relates to a co-phase error correction method suitable for optical synthetic aperture imaging systems. Background Technology

[0002] One important way to improve the imaging resolution of a telescope is to increase its aperture. However, due to limitations in manufacturing technology, system assembly, and spacecraft payloads, the production of large-aperture telescopes has reached a bottleneck. Optical synthetic aperture imaging technology uses a synthetic aperture composed of a multi-path split-aperture array to replace the traditional single-aperture primary mirror, effectively achieving the high-resolution imaging performance of large-aperture telescope systems, and has broad application prospects.

[0003] In-phase superposition of sub-aperture beams on the image plane is crucial for achieving high-resolution imaging in synthetic aperture systems; otherwise, image quality will be severely degraded. How to quickly and effectively detect and correct co-phase errors has been a research hotspot in the field of optical synthetic aperture imaging. For atmospheric optics, co-phase errors caused by environmental turbulence have a low frequency, and existing co-phase detection methods can basically meet its real-time requirements, such as improved Shaker-Hartmann sensors, pyramidal detectors, and dispersive fringe detectors. However, for aerodynamic optics, co-phase errors caused by turbulence have a timescale of less than 10 microseconds, exceeding the limits of existing co-phase detection techniques.

[0004] Deep learning-based common-phase detection technology has garnered widespread attention both domestically and internationally in recent years. This technology can fit the functional mapping relationship between optical synthetic aperture imaging images and common-phase error values ​​based on data-driven approaches. A well-trained network only needs to perform one data forward transmission during the common-phase detection stage, theoretically achieving very high real-time performance. However, under the von Neumann architecture, the inherent limit of electronic computing speed has become a bottleneck for neural networks to achieve high-speed real-time computing, greatly restricting the application of deep learning technology in the field of common-phase detection. Summary of the Invention

[0005] To overcome the problems and limitations of existing methods, this invention provides a co-phase error correction method based on an all-optical diffraction neural network. The neural network is constructed using optical methods, and optical computing replaces electronic computing to obtain a faster (light speed) computing speed, thereby achieving strong real-time performance in co-phase detection.

[0006] The technical solution adopted in this invention is: a co-phase error correction method based on an all-optical diffraction neural network, comprising:

[0007] The optical synthetic aperture imaging light field is generated using a synthetic aperture imaging module, and the co-phase error is loaded and compensated.

[0008] The diffraction neural network module is used to extract features and convert information from the input light field.

[0009] The phase error detection module receives the output light field of the all-optical diffraction neural network module, and converts the output light field information into a detectable physical quantity to construct a mathematical model for the phase error characterization.

[0010] The synthetic aperture imaging module is based on point target imaging. The light wave is split by a multi-aperture array loaded with co-phase error, and then converged to the imaging surface by interference to obtain synthetic aperture imaging light field information.

[0011] The diffraction neural network module consists of multiple diffraction elements, and the physical parameters of the diffraction network include the light source wavelength, pixel size, diffraction layer size, and diffraction layer spacing.

[0012] The co-phase error detection module divides the co-phase error detection range into N sub-intervals, correspondingly dividing the detection image plane into N detection regions, and uses the light intensity distribution of the output light field of the diffraction network on the detection surface to characterize the co-phase error.

[0013] Among them, the diffraction neural network module uses the principle of superposition of light diffraction to realize the connection between two adjacent diffraction elements. Each pixel on the diffraction layer is equivalent to a neuron in the digital neural network. By changing the refractive index or transmittance of light when it passes through the pixel, the light field can be controlled in two dimensions: phase and amplitude. The modulation coefficient of the pixel on phase and amplitude is defined as the weight of the diffraction neural network.

[0014] The multi-aperture imaging light field information composed of phase information and amplitude information is used as the input of the diffraction network, and the light intensity distribution of the detection surface is used as the output of the diffraction network. The input data is transmitted forward in the diffraction network through the diffraction propagation of light in free space. The network weight is iteratively optimized through the gradient descent algorithm to ensure that the energy of the emitted light is concentrated in the ideal detection area, thereby realizing the mapping of the relationship between the imaging light field and the co-phase error.

[0015] In the actual detection optical path, the entire process from the input of the imaging light field to the output of the common phase error is an optical operation, which can realize the real-time detection of the common phase error at the speed of light.

[0016] The specific steps of a co-phase error correction method based on an all-optical diffraction neural network are as follows:

[0017] Step 1) Construct an optical synthetic aperture imaging numerical simulation platform, batch load the co-phase error to generate corresponding imaging light field data as input, and the corresponding ideal light intensity distribution as output, and construct simulation training set and simulation test set;

[0018] Step 2) Build a diffraction neural network and set the physical parameters of the diffraction neural network. The wavelength of the light source and the pixel size should be consistent with the imaging simulation platform. The diffraction layer size and diffraction layer spacing and other parameters should be adjusted according to the training effect.

[0019] Step 3) Optimize the weights of the diffraction network based on the simulation training set until the cost function converges to a very small value, so that the light intensity distribution of the network output light field on the detector surface is close to the ideal light intensity distribution.

[0020] Step 4) Evaluate the performance of the diffraction network based on the simulation test set, including the accuracy of cophase prediction, the contrast of light intensity in each detection area, the network's tolerance to diffraction layer spacing deviation and diffraction layer center alignment error, and its robustness to aberrations and noise, etc., and further optimize and adjust the network according to the cophase detection requirements.

[0021] Step 5) Train the diffraction network weights and then fabricate the diffraction elements using photolithography.

[0022] Step 6) The imaging light field of the optical synthetic aperture imaging system is transmitted to the detection surface through the diffraction network. The co-phase error detection result can be obtained according to the light intensity distribution of the image surface, and the co-phase error is compensated by the closed-loop control module.

[0023] Compared with the prior art, the present invention has the following advantages:

[0024] (1) This invention utilizes the free space propagation characteristics of light and the principle of vector superposition to realize complex matrix operations of neural networks. It has the unique advantages of ultra-parallel processing and information transmission, and can perform high-precision detection of co-phase error at the speed of light, which greatly improves the real-time performance of co-phase detection technology.

[0025] (2) The present invention uses simulation datasets to train and obtain the parameters of the diffraction network and then processes and manufactures it, thus avoiding the situation where the network performance is reduced due to the mismatch between experimental data and labels. Attached Figure Description

[0026] Figure 1 This is a schematic diagram illustrating the principle of a co-phase error correction method based on an all-optical diffraction neural network.

[0027] Figure 2 The flowchart illustrates the co-phase error correction method based on an all-optical diffraction neural network for a two-aperture array embodiment.

[0028] Figure 3This is a schematic diagram of a phase error detection system for a two-aperture array embodiment. Parallel light is reflected by a spliced ​​fast-reflecting mirror and passes through the pupil. The beam is converged by the imaging primary mirror and then input into a diffraction neural network composed of three diffraction layers. After phase modulation, the intensity distribution of the light emitted from the diffraction network is obtained on the detection plane. The prediction interval corresponding to the detection area where the light intensity is maximum is the phase error detection result.

[0029] Figure 4 This is a schematic diagram of the phase distribution of the three diffraction layers obtained from training the two-aperture array example.

[0030] Figure 5 The confusion matrix is ​​the simulation test result of the two-hole array embodiment.

[0031] Figure 6 This is a schematic diagram of the intensity distribution on the detection surface in the simulation test results of a two-aperture array embodiment. Detailed Implementation

[0032] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0033] This invention provides a co-phase error correction method based on an all-optical diffraction neural network, applicable to optical synthetic aperture imaging systems. The basic principle is as follows: Figure 1 As shown, the optical synthetic aperture imaging light field that varies with co-phase error is obtained by using a synthetic aperture imaging module; the imaging light field is input into a diffraction neural network, and is modulated by diffraction elements during diffraction propagation and converted into a specific light field distribution for output; the detection module converts the light field information output by the diffraction network into a light intensity distribution that can directly characterize the co-phase error, and realizes the detection of the co-phase error at the speed of light.

[0034] This invention is an embodiment of a two-aperture array, and the specific implementation steps are as follows:

[0035] Step 1) Implementation method and process as follows Figure 2 As shown, firstly, according to Figure 3 The synthetic aperture imaging module shown establishes a numerical simulation research platform for a two-aperture array. Based on single-wavelength point target imaging, with a light source wavelength λ = 600 nm, it batch-loads co-phase errors within a 1λ detection range to generate corresponding imaging light field data, where the imaging light field is the light field at the focal plane. For an optical synthetic aperture imaging system composed of N sub-apertures, its pupil function expression is:

[0036]

[0037] In equation (1), A sub (x0-x n y0-y n ) is the pupil function of the sub-aperture, (x n y n) is the center coordinate of the nth sub-aperture. It is the phase function of the nth sub-aperture. According to the Fourier imaging principle, the complex amplitude of the focal plane light field is obtained by performing a Fourier transform on the pupil function:

[0038] U(x f ,y f )=FT{A(x0,y0)} (2)

[0039] In equation (2), FT{} and (x f ,y f ) represent the spatial coordinates of the Fourier transform and the system focal plane, respectively.

[0040] Step 2) Divide the common phase error detection range 1λ into 10 intervals, namely:

[0041] Interval 0: (-0.5λ, -0.4λ), Interval 1: (-0.4λ, -0.3λ), Interval 2: (-0.3λ, -0.2λ),

[0042] Interval 3: (-0.2λ, -0.1λ), Interval 4: (-0.1λ, 0), Interval 5: (0λ, 0.1λ).

[0043] Interval 6: (0.1λ, 0.2λ), Interval 7: (0.2λ, 0.3λ), Interval 8: (0.3λ, 0.4λ)

[0044] Interval 9: (0.4λ, 0.5λ);

[0045] Correspondingly, the detection plane is divided into 10 detection areas, such as Figure 3 As shown in the phase error detection module, the ideal light intensity distribution of the detection plane is used as a label, and the imaging light field data generated in step 1) is used as input to construct a simulation training set and a simulation test set.

[0046] Step 3) Construct a diffraction neural network consisting of three diffraction layers, such as... Figure 3 As shown in the diffraction neural network module, ignoring the influence on amplitude, a pure phase-type diffraction layer is used, and the phase of each pixel is defined as the network weight.

[0047] Step 4) Train the network using the gradient descent algorithm: According to the Rayleigh-Sommerfeld principle, the free space light propagation formula can be expressed as:

[0048]

[0049] In equation (3), l represents the l-th layer of the network, and λ represents the wavelength. i represents the coordinate (x) of the l-th layer.i ,y i ,z i The i-th pixel at position (). For a pure phase-type diffraction network, the transmission coefficient of a pixel can be expressed as:

[0050]

[0051] In equation (4), Represents the coordinates (x) on the l-th layer. i ,y i ,z i The phase of the i-th pixel at position ). The output of the i-th pixel in the l-th layer network can be expressed as:

[0052]

[0053] in, This represents the superposition of light waves from the outputs of all neurons in the (l-1)th diffraction layer to the i-th neuron in the l-th diffraction layer.

[0054] Assuming the diffraction neural network consists of M diffraction layers, the intensity of the emitted light at the detector surface can be expressed as:

[0055]

[0056] Wherein, the (M+1)th layer represents the detection surface. This represents the superposition of light waves whose outputs from all neurons in layer M propagate to the i-th pixel on the detector surface.

[0057] If the detector surface contains Q pixels, the loss function is defined as the ratio of the intensity distribution of the network's outgoing light on the detector surface to the ideal intensity distribution. Mean square error between:

[0058]

[0059] The gradient descent algorithm is used to iteratively update the phase weights of the diffraction network, so that the intensity distribution of the detection plane continuously approaches the ideal intensity distribution.

[0060] Step 5) Evaluate the performance of the trained diffraction neural network using a simulation test set. Further optimize the detection performance by adjusting the network parameters, and finally determine the phase distribution of each diffraction layer in the network. The phase distribution of the three diffraction layers obtained after training is as follows: Figure 4 As shown. The confusion matrix of the common phase detection results of 1000 sets of simulation test data is as follows. Figure 5As shown, the interval detection accuracy is 97%. If cophase is performed based on the median value of the predicted interval, the maximum residual RMS after cophase is only λ / 20 when the interval prediction is accurate; even when the prediction is inaccurate, the prediction result deviates from the true label by only one interval, achieving very high detection accuracy. Figure 6 The diagram shows the intensity distribution of the detection surface in the simulation test results. The results show that the diffraction neural network can concentrate most of the energy into the correct detection area.

[0061] Step 6) Using the median value of the predicted interval as the result of the co-phase error detection, the closed-loop correction of the co-phase error is achieved by controlling the splicing fast reflection mirror based on all-optical detection.

[0062] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any method, apparatus, or system for correcting the co-phase error of an optical synthetic aperture imaging system based on a diffraction neural network falls within the scope of protection of the present invention.

Claims

1. A method for correcting phase error based on an all-optical diffraction neural network, characterized in that, include: The optical synthetic aperture imaging light field is generated using a synthetic aperture imaging module, and the co-phase error is loaded and compensated. The input light field information is extracted and converted using a diffraction neural network module. The diffraction neural network module is composed of multiple diffraction elements, and the physical parameters of the diffraction network include the light source wavelength, pixel size, diffraction layer size, and diffraction layer spacing. The co-phase error detection module divides the co-phase error detection range into N sub-intervals, corresponding to N detection areas on the detection image plane, and uses the light intensity distribution of the output light field of the diffraction network on the detection surface to characterize the co-phase error. The diffraction neural network module uses the principle of superposition of light diffraction to connect adjacent diffraction elements. Each pixel on the diffraction layer is equivalent to a neuron in a digital neural network. By changing the refractive index or transmittance of light when it passes through the pixel, the light field can be controlled in both phase and amplitude dimensions. The modulation coefficient of the pixel on phase and amplitude is defined as the weight of the diffraction neural network. The multi-aperture imaging light field information composed of phase information and amplitude information is used as the input of the diffraction network, and the light intensity distribution of the detection surface is used as the output of the diffraction network. The input data is transmitted forward in the diffraction network through the diffraction propagation of light in free space. The network weight is iteratively optimized through the gradient descent algorithm to ensure that the energy of the outgoing light is concentrated in the target detection area, thereby realizing the mapping of the relationship between the imaging light field and the co-phase error. The phase error detection module receives the output light field of the all-optical diffraction neural network module, and converts the output light field information into a detectable physical quantity to construct a mathematical model for the phase error characterization.

2. The co-phase error correction method based on an all-optical diffraction neural network according to claim 1, characterized in that, The synthetic aperture imaging module is based on point target imaging. The light wave is split by a multi-aperture array loaded with co-phase error, and then converged to the imaging surface by interference to obtain synthetic aperture imaging light field information.

3. The co-phase error correction method based on an all-optical diffraction neural network according to claim 1, characterized in that, In the actual detection optical path, the entire process from the input of the imaging light field to the output of the common phase error is an optical operation, which can realize the real-time detection of the common phase error at the speed of light.

4. The co-phase error correction method based on an all-optical diffraction neural network according to claim 1, characterized in that, The specific steps are as follows: Step 1) Construct an optical synthetic aperture imaging numerical simulation platform, batch load the co-phase error to generate corresponding imaging light field data as input, and the corresponding ideal light intensity distribution as output, and construct simulation training set and simulation test set; Step 2) Build a diffraction neural network and set the physical parameters of the diffraction neural network. The wavelength of the light source and the pixel size should be consistent with the imaging simulation platform. The diffraction layer size and diffraction layer spacing and other parameters should be adjusted according to the training effect. Step 3) Optimize the weights of the diffraction network based on the simulation training set until the cost function converges to a very small value, so that the light intensity distribution of the network output light field on the detector surface is close to the ideal light intensity distribution; Step 4) Evaluate the performance of the diffraction network based on the simulation test set, including the accuracy of cophase prediction, the contrast of light intensity in each detection area, the network's tolerance to diffraction layer spacing deviation and diffraction layer center alignment error, and its robustness to aberrations and noise, etc., and further optimize and adjust the network according to the cophase detection requirements. Step 5) Train the diffraction network weights and then fabricate the diffraction elements using photolithography. Step 6) The imaging light field of the optical synthetic aperture imaging system is transmitted to the detection surface through the diffraction network. The co-phase error detection result can be obtained according to the light intensity distribution of the image surface, and the co-phase error is compensated by the closed-loop control module.

Images