Anti-turbulence adaptive imaging method based on optical diffraction type diffusion model

Abstract

The invention relates to an anti-turbulence adaptive imaging method based on an optical diffraction type diffusion model. An optical diffraction transmission model and a diffusion model thought are combined, and an optical diffraction type diffusion model is provided to carry out low-beam-velocity optical calculation imaging on an image under atmospheric turbulence disturbance. Different from an existing anti-turbulence imaging method based on an electrical neural network and a traditional anti-turbulence imaging method, the method has the advantages that the high-quality image processing advantage of a light calculation normal form and a diffusion model is utilized, the idea that an intermediate layer is fully scattered through the diffusion model is adopted, and an optical diffraction type diffusion model architecture is constructed; meanwhile, the method has the advantages of light velocity calculation and efficient image generation. Compared with a traditional imaging method, the method is a novel optical calculation normal form, has potential to promote optical calculation to realize a more complex neural network and assist the calculation power and the speed of the neural network, is proved to have a remarkable effect in anti-turbulence imaging application, and has potential to promote landing construction of an anti-turbulence photoelectric calculation imaging system.

Description

Technical Field

[0001] The present invention relates to an anti-turbulence adaptive imaging method based on an optical diffraction diffusion model. An optical diffraction diffusion model is designed. By training an optical diffraction encoder and decoder, near-light-speed image restoration imaging is performed on images disturbed by atmospheric turbulence. A maturely trained diffraction encoder is used to further break up the distorted image light field under turbulence disturbance into a latent space in a near-Gaussian chaotic state. A diffraction decoder is then used to filter the information of non-target light fields in the latent space, thereby achieving light field correction and image restoration effects. Background Art

[0002] Adaptive optical imaging under atmospheric turbulence is an extremely important issue in many fields, such as deep space exploration, remote sensing imaging, astronomical observation, etc. With the development of adaptive optical technology, traditional iterative algorithms such as GS (Gerchberg-Saxton) phase recovery algorithm, SPGD (stochastic parallel gradient descent) random parallel gradient descent algorithm, and deep learning algorithms based on traditional convolutional electrical neural networks have been proposed for correction of optical distortion wavefronts to achieve image restoration. Among them, the adaptive optical imaging system based on traditional iterative algorithms includes wavefront detectors, wavefront controllers, and wavefront correctors. The hardware equipment wavefront sensors and wavefront controllers often bring response delays. At the same time, the algorithms often require long iterations and are prone to fall into local optimal solutions. There are defects such as long processing time and difficulty in convergence. Anti-turbulence imaging based on deep learning in the electrical field still has computational delays due to its reliance on electronic devices. It is not conducive to efficient response and energy conservation under the huge algorithm training execution, and the computing power is obviously insufficient in real-time low-power processing, which has gradually become a potential bottleneck for the development of the current intelligent imaging field.

[0003] In long-distance transmission imaging applications, the temperature and pressure of the atmosphere cause random fluctuations in the atmospheric refractive index, resulting in wavefront distortion and amplitude fluctuations of the light wave, causing beam jitter and light intensity flicker, which seriously affects the imaging quality. The accuracy of target parameter recognition will also be seriously affected. Therefore, it is very important to achieve high-quality near-light-speed calculation adaptive imaging through certain technologies, so that high-quality images can be obtained even in long-distance transmission, which is very important for applications in deep space exploration, remote sensing imaging, etc.

[0004] In recent years, with the development of optical artificial intelligence technology, neural network models based on physical optical transmission models have attracted the attention of a large number of researchers due to their significant advantages such as parallel processing capabilities close to the speed of light and extremely low energy consumption. Deep learning models based on physical optical free transmission have shown great potential in target classification, optical imaging, wavefront correction, inverse design and other fields. However, in the field of anti-turbulence adaptive optical imaging, there is no mature research work on optical diffraction neural network models that have the potential to achieve high-speed and high-resolution imaging effects.

[0005] Therefore, in the context of the urgent need for turbulence-resistant high-efficiency imaging, computational imaging of turbulence-resistant random phase disturbances at near-light speed has important research significance. The present invention proposes an optical diffractive diffusion model (ODDM) for near-real-time and efficient imaging of distorted complex amplitude light field targets under atmospheric turbulence disturbances, and designs an optical diffraction encoder and decoder to perform phase modulation on the complex amplitude light field disturbed by turbulence. Similar to the information exchange in the form of collisions between particles in the diffusion model, the diffraction transmission of light exchanges energy in the form of waves (manifested as a new particle distribution in the diffusion model and as a new wavefront distribution in the diffraction model of light). The image destroyed by the disturbance can be further destroyed through energy exchange to reach a near-Gaussian distribution chaotic state similar to the diffusion model, and then the light field is reversely diffused through specific diffraction transmission to achieve the effect of restoring the original state. Summary of the invention

[0006] The technical problem solved by the present invention is: in view of the high computational cost and difficulty of the current anti-turbulence imaging system, an optical diffraction diffusion model is designed and invented. By using a well-trained diffraction encoder to further break up the distorted image light field under turbulent disturbance into a latent space of a near-Gaussian chaotic state, and then using a diffraction decoder to filter the information of the non-target light field in the latent space, the correction of the light field and the restoration of the image are achieved. In the concept verification experiment, the model can be loaded onto a spatial light modulator for optical experiment construction. After training, the model can perform near-light-speed high-quality imaging of the MNIST (Mixed National Institute of Standards and Technology) test set and other unknown patterns under turbulent disturbance. This method has good flexibility, rapidity and robustness, is easy to operate, and has strong generalization ability.

[0007] The technical solution of the present invention is:

[0008] The present invention relates to an anti-turbulence adaptive imaging method based on an optical diffraction diffusion model, which mainly comprises the following steps:

[0009] (1) Design and construct the framework of the optical diffraction diffusion model and adjust its parameters. Use the MNIST (Mixed National Institute of Standards and Technology) handwritten digit dataset with a turbulence intensity of D / r0≈8.33 under the atmospheric turbulence phase screen to train the diffraction encoder and decoder, and implement the optical computing paradigm for the diffusion and inverse diffusion processes in the diffusion model.

[0010] (2) Numerical simulations were performed on the trained model. The unknown MNIST (Mixed National Institute of Standards and Technology) test set and other patterns were simulated using known (seen before) and unknown (unseen) turbulence phase screens of the model. The commonly used image quality evaluation functions PSNR (Peak signal-to-noise ratio) and SSIM (Structural Similarity) were used to evaluate the images.

[0011] (3) Calibrate the laser and the linearity of the spatial light modulator, design and build the experimental optical path according to the model training hyperparameters (diffraction layer distance, neuron size, etc.), load the diffraction encoder and decoder on the phase-type spatial light modulator, and perform high-quality restoration of the distorted image light field of the target loaded on the amplitude-type spatial light modulator after being disturbed by atmospheric turbulence.

[0012] (4) A camera is used to receive images at the detection layer of the model, and the experimental data is sorted and saved. An evaluation function is used to quantitatively evaluate the images collected with and without the model.

[0013] The principle of the present invention is:

[0014] In the Kolmogorov atmospheric turbulence model, a phase screen is often used to simulate the wavefront phase disturbance caused by atmospheric turbulence. In this paper, we use the Zernike polynomial method that obeys the Kolmogorov atmospheric turbulence distribution to simulate the turbulent disturbance. The turbulent phase can be linearly combined by a series of orthogonal Zernike polynomials, namely:

[0015]

[0016] Where q represents the number of Zernike polynomial modes, a j represents the j-th order Zernike polynomial Z j The coefficients of Zernike polynomial Z, ρ, θ represent the normalized polarization radial and angular directions. j is a set of orthogonal polynomials defined in the unit circle, which are defined as:

[0017]

[0018] Where m and n represent the angular frequency and radial frequency respectively, m and n are both integers and satisfy m≤n, n-|m| is an even number. The first 600 Zernike polynomials are selected to characterize atmospheric turbulence. When using Zernike polynomials to describe the phase of the vortex light wavefront, it is necessary to determine the coefficients a of each order of Zernike polynomials. j , but these coefficient variables are not independent of each other. According to the energy relationship, any two-order Zernike coefficients a i and a j The covariance between satisfies:

[0019]

[0020] Where n and m are Z i The radial frequency and angular frequency of Z are n' and m' respectively. j The radial frequency and angular frequency of ; Γ() represents the Gamma function; when i and j have the same parity, Take 1, otherwise take 0; D represents the beam diameter, r0 is the atmospheric coherence length, and D / r0 represents the atmospheric turbulence intensity; K z By Z i and Z j The radial and angular frequencies are determined by the following expression:

[0021]

[0022] From formula (6), we can get the covariance between the Zernike polynomial coefficients, and thus get the vector V consisting of the coefficients of each order of Zernike polynomials = [a1, a2, a3, ... a q-1 ,a q ] T The covariance matrix M of is as follows:

[0023]

[0024] From formula (7), we can see that M is a Hermitian matrix, so there must be a unitary moment U(U -1 =U T ), so that UMU T is a diagonal matrix. The unitary matrix U can be obtained by calculating the singular value decomposition (SVD) of the M matrix. The singular value decomposition of M is:

[0025] M=XSX T (8)

[0026] Among them, X is a unitary matrix and S is a diagonal matrix. Further, we get:

[0027] X T MX=X T (XSX T )X=X -1 XSX -1 X=S (9)

[0028] Therefore, we can get U = X T Then define a new vector B

[0029] B=UV (10)

[0030] The covariance matrix of vector B is:

[0031] E[BB T ]=E(UV(UV) T )=UE(VV T ) T =UMU T =S (11)

[0032] According to formula (11), the covariance matrix of B is a diagonal matrix, so the elements in B are independent and have a mean of 0. The variance is the Gaussian distribution of the diagonal elements of the diagonal matrix S. Therefore, the Zernike polynomial coefficient vector can be obtained:

[0033] V=U -1 B=U T B (12)

[0034] Therefore, by determining the number of Zernike polynomial terms q and the turbulence intensity D / r0, the Zernike polynomial coefficients can be obtained, and combined with formula (1), the atmospheric turbulence wave front phase can be obtained.

[0035] Different from the “invisible” connection between electronic neurons in electronic networks, the forward transmission of this optical neural network model is based on the Rayleigh-Somuffy principle:

[0036] The turbulent phase in formula (1) At the position z0, it is added to the optical path to perform phase perturbation on the amplitude-encoded image n (x, y, z = 0) d , then the complex amplitude light field after phase disturbance can be expressed as:

[0037] u0(x,y,z0)=t d ·[n(x,y,0)·w(x,y,z0)] (13)

[0038] in:

[0039]

[0040] w(x,y,z) is the transfer factor of the translation-invariant linear system following the Rayleigh-Somuffy formula, where λ is the wavelength of incident light, Then, formula (13) is used as the network input into the model, and the diffraction network at distance z l The transmission coefficient of the pure phase diffraction layer at can be expressed as:

[0041]

[0042] Therefore, the output light field u after the input light field is modulated by l diffraction layers is l (x,y,z l ) can be expressed as:

[0043] u l (x,y,z l ) = t l (x,y,z l )[u l-1 (x,y,z l-1 )·w(x,y,△z l )] (17)

[0044] Among them, △z l =z l -z l-1 is the axial distance between adjacent diffraction layers. The light field continues to transmit along the axial direction for a distance △z l+1 Arriving at the output layer, the light field intensity is calculated as the output o(x,y) of the model:

[0045] o(x,y)=|u l (x,y,z l )·w(x,y,△z l+1 )| 2 (18)

[0046] During the model training process, the mean square error function is used as the loss function to constrain the model training, and the network optimization goal is to minimize the mean square error between the network output layer image and the input target true value.

[0047]

[0048] Among them, B refers to the batch size, that is, the number of samples used each time the model weight parameters are updated. i represents the output light field of the i-th image in B, I i represents the actual light field of the real image of the i-th image in B.

[0049] In order to objectively evaluate the quality of restored images, this paper uses structural similarity (Structural

[0050] Similarity), SSIM) is used as an evaluation index to measure the quality of restored images, and its calculation method is:

[0051]

[0052] Among them, μ i and μ o are the means of the input original image i and the model output image o respectively; σ i and σ o are the variances of images i and o, σ io is the covariance of the two images, c1 and c2 are constants set to prevent the denominator from approaching 0. The SSIM range is (0,1), and the higher the similarity between the two images, the closer their SSIM is to 1.

[0053] In order to avoid the limitation of using one evaluation index, this paper also uses the peak signal-to-noise ratio (PSNR) as the evaluation index of the restored image quality. The calculation method is:

[0054]

[0055] Among them, MAX I is the maximum value of the image I under test. In this paper, MAX I =255, MSE is the mean square error between the original image and the restored image.

[0056] Compared with the existing solutions, the main advantages of the solution of the present invention are:

[0057] (1) The model combines the optical computing paradigm with the idea of ​​diffusion model to design and construct an optical diffraction diffusion model. After training, the model can perform high-quality imaging at near-light speed for distorted image light fields disturbed by turbulence.

[0058] (2) Model training The diffraction light encoder and decoder can be loaded onto the spatial light modulator and directly built onto the optical platform to control the light field. The image restoration effect can be achieved through light field control. At the same time, the real-time performance of the adaptive imaging system is greatly improved, and the computational cost is constrained.

[0059] (3) After the model anti-turbulence adaptive imaging, the PSNR value of the image can generally reach more than 14, and the highest can reach 17.0309. The SSIM can generally be increased to more than 65%, and the highest can reach 91.81%. The imaging effect is significantly improved, and the computational imaging process reaches the effect of near light speed. The stability and timeliness of imaging are significantly improved, providing an effective implementation plan model for the implementation of the anti-turbulence optoelectronic computational imaging system. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 Flow chart of the anti-turbulence adaptive imaging method

[0061] Figure 2 This is the principle diagram of anti-turbulence imaging based on the optical diffraction diffusion model;

[0062] Figure 3 This is a training framework diagram for the optical diffraction diffusion model;

[0063] Figure 4 Model training iteration curve (a) Iteration curve during the training of the model mean square error loss function; (b) Iteration curve during the training of the model PSNR accuracy rate;

[0064] Figure 5 Integration of simulation effects for MNIST test;

[0065] Figure 6 Simulation results for other pattern imaging test sets

[0066] Figure 7 Schematic diagram of the optical path of the model experiment

[0067] Figure 8 The experimental results for the MNIST test set are shown in the figure

[0068] Fig. 9 This is a display of other pattern imaging experimental effects

[0069] Fig.10 The average PSNR and average SSIM of the MNIST test set images before and after restoration using the model under known and unknown turbulence disturbances Specific implementation plan

[0070] The present invention uses the MNIST data set and other simple patterns as experimental objects, and the implementation object is a spatial light modulator. The specific implementation steps are as follows:

[0071] Based on Figure 2 The framework principle of the optical diffraction diffusion model shown in the figure is mainly composed of a diffraction encoder and a decoder. The initial light field u origin is the light field after amplitude encoding. After the light field is diffracted and transmitted for a certain distance, it is disturbed by the random phase of atmospheric turbulence. It continues to transmit and obtains the value shown in Figure u dis Here we define the disturbance caused by free transmission and turbulence phase as noise relative to the original light field. The distorted light field is further used as the input of the model. The model designs a diffraction encoder for the diffraction transmission diffusion process and trains a mature diffraction encoder to convert the incident distorted light field X dis Further perturbation obtains the chaotic light field u with a more chaotic near-Gaussian distribution in the middle of the modelchaos (Hidden Space), the chaotic light field includes turbulence information affected by diffraction transmission noise and images affected by turbulence disturbance and diffraction transmission noise. Next, the chaotic high-noise light field u chaos Reverse diffusion is performed in the form of diffraction transmission, and the light field modulated by the diffraction decoder is the original image u chaos The trained diffraction decoder can filter out the noise of the non-target light field like a filter to obtain the output light field u output , so as to achieve the purpose of correcting the light field and efficiently restoring the image.

[0072] Design and build Figure 3 As shown in the ODDM model framework diagram, it can be seen that the incident light is transmitted through the target layer amplitude encoding for a distance D0, is disturbed by several random atmospheric turbulence phase screens and transmitted for a distance D1 to obtain several distorted complex amplitude light field data sets. These complex amplitude light fields with preliminary turbulence noise are used as the input layer diffraction transmission D2 of the model. In this process, due to the influence of transmission and turbulence, the light field further reaches a more chaotic state. The diffraction encoding module of the model continues to transmit these semi-chaotic complex amplitude light fields for a distance D3, further diffusing them into the latent space of the chaotic state of near-Gaussian distribution. Then the diffraction decoding module of the model takes such a complex amplitude light field in the chaotic state of near-Gaussian distribution as input, and performs inverse diffusion transmission processing on the light field through transmission D4, thereby performing relatively excellent real-time optical calculation repair on the original light spot.

[0073] The model diffraction encoder and decoder are trained using the MNIST handwritten digit set, a public handwritten data set under the atmospheric turbulence phase screen disturbance with a turbulence intensity of D / r0≈8.33. Figure 4 The curve shows the gradual change of the model's loss function and accuracy with the number of iterations during the model training process. The horizontal axis represents the number of iterations of model training, and the vertical axis represents the logarithm of the target loss function. Figure 4 In the table in the upper right corner of (a), five points A (4, -5.08), B (8, -5.90), C (21, -6.61), D (43, -6.84), and E (263, -6.98) are selected to characterize the downward trend of the model loss function during network training. It can be clearly seen that the model has an obvious downward trend from point A to point D, and the downward speed from point D to point E tends to be gentle. Near point E, the model loss function no longer has an obvious downward trend, and the error between the model output and the target value has reached the minimum. Figure 4 (b) shows the iterative curve of the peak signal-to-noise ratio of the network accuracy evaluation function. As above, the horizontal axis represents the number of network iterations, and the vertical axis represents the peak signal-to-noise ratio PSNR between the image and the true value. Figure 4(b) In the table in the lower right corner, five points A (4, 14.62), B (11, 16.56), C (21, 17.74), D (61, 18.43), and E (210, 18.94) are selected on the curve, which more clearly represents the rising trend of the network accuracy. From A to C, the network accuracy shows a straight-line upward trend, which means that the update and adjustment of the network parameters in the early stage greatly improves the similarity between the network output and the target image; from C to E, the parameters of the model are fine-tuned, and the rate of increase of the model accuracy tends to be flat, and there is no large-scale floating trend near point E. The model training is mature, and the accuracy is stably maintained at a high level. It shows that the similarity between the model output and the real image has reached the highest level at this time, and the model has been trained to maturity. The maturely trained model can be directly loaded into Figure 6 The spatial light modulators Ph_SLM2 and Ph_SLM3 shown perform diffusion and de-diffusion processing on the input light field.

[0074] The model uses the Adam optimizer to update the pixel phase modulation value, and the learning rate Lr = 0.99 iteration *5*10 -2 Decrement, where iteration indicates the current iteration number. The model was trained for about 300 times using Python (v 3.8.18) and pytorch on an NVIDIAA100-SXM4-80GB x2, Intel(R) Xeon(R) Gold 6330 CPU@2.00GHz.

[0075] The trained model is simulated and verified. Figure 5 , 6 As shown, Figure 5 As shown in the first column (from left to right), we randomly selected the binary amplitudes of handwritten digits that the model has not seen in the MNIST dataset. Figure 2 , 5 , 8, 9 The distorted complex amplitude light field under known turbulence in the model training and unknown turbulence disturbance is used as the input of the model. After the known turbulence disturbance, the distorted complex amplitude light field obtained by free space transmission is as follows: Figure 4As shown on the left side of the second column, the PSNR and SSIM of the distorted image have been calculated and marked below the image. For example, the PSNR of the handwritten number 2 is as low as 12.1048 and the SSIM is as low as 0.5802. The image has been blurred by the known turbulence disturbance and the information cannot be identified. The image after ODDM imaging is shown on the right side of the second column. At this time, the distorted image has been well restored. For example, the PSNR and SSIM of the handwritten number 5 can be increased to 24.2554 and 0.9208 at most. At this time, the imaging effect of the image is highly similar to the original image. The distorted light field image presented by free space transmission after unknown turbulence disturbance is shown in the figure below. Figure 5 As shown on the left side of the third column, the PSNR of the disturbed handwritten digit 8 is as low as 11.6172, and the SSIM of the handwritten digit 2 is as low as 0.4769. The effective information is also unrecognizable. After the ODDM model, the imaging quality has been significantly improved. For example, the PSNR of the handwritten digit 5 ​​has been increased to 22.6687 and the SSIM has been increased to 0.9111. It can be seen from the image itself that the quality of the restored image has been greatly improved. From the simulation results, it can be seen that the model can perform good recovery of the handwritten digits in the test set under known turbulence and unknown turbulence disturbances.

[0076] In order to verify the generalization ability of the network to other types of images, four patterns not in the MNIST database were selected for simulation. The results are as follows: Figure 6 As shown in the first column, from top to bottom are the resolution target, the moon, the butterfly, and the English letter S. The input targets after known turbulence and unknown turbulence disturbances are blurred and unrecognizable, and their PSNR and SSIM generally do not exceed 15% and 65% respectively. The light field image imaged by the model has been significantly restored, and the PSNR (4.5-7dB) and SSIM (33-41%) have been improved to varying degrees. For example, the PSNR of the moon pattern under known turbulence disturbance is increased to 20.1918, and the SSIM is increased to 0.9182. Under unknown turbulence disturbance, its PSNR is increased to 18.3970, and SSIM is increased to 0.9174. This shows that for non-handwritten digital targets, the use of the model still has a stable restoration imaging effect.

[0077] In order to further verify the experimental effect of the network in the actual scenario, an experimental optical path was built. The experimental optical path mainly consists of a helium-neon laser, three phase-type SLMs (UPOLabs HDSLM80R) and one amplitude-type SLM (UPOLabs HDSLM80RA), a CMOS (Daheng MER2-2000-19U3M-L), a lens, and a polarization beam splitter PBS and a beam splitter BS. We designed that the plane wave encoded by the amplitude-type spatial light modulator is transmitted 150mm before reaching the phase-type spatial light modulator Ph_SLM1 loaded with turbulent phase disturbance. The distorted light field is transmitted 350mm after turbulent disturbance and enters the model input layer. After diffraction transmission of 170mm, it reaches the diffraction encoder loaded on Ph_SLM2. The encoder further diffuses and breaks up the light field into a chaotic state, and after diffraction of 170mm, it reaches the diffraction decoder loaded on Ph_SLM3. Finally, the output light field intensity is collected at a distance of 140mm from the diffraction decoder layer. We use z 0～4 They respectively represent the actual distances between Amp_SLM, Ph_SLM1, Ph_SLM2, Ph_SLM3 and CMOS in the experimental optical path, so z0=150mm, z1=350+170=520mm, z2=170mm, z3=170mm, z4=140mm.

[0078] The specific experimental optical path is as follows: Figure 7As shown, the optical path is divided into two parts. In the first part, 632.8nm light is emitted from a helium-neon laser, filtered by a spatial filtering system, and then expanded by a lens (Len2) into a 20mm circular beam (the light field incident on the SLM liquid crystal screen in the middle can be approximated as a flat-top light). After passing through the first polarization beam splitter prism 1 (PBS1) (in order to minimize the light intensity loss), the second polarization beam splitter prism 2 (PBS2) obtains (orthogonal) linearly polarized light with a diameter of about D = 5mm and is incident on an amplitude-type spatial light modulator Amp_SLM (loaded with a 680*680 pixel mnist handwritten digit set). The outgoing light (non-orthogonal) linearly polarized light passes through the polarization beam splitter prism 2 ( The light field of the laser is transmitted through the PBS2 and the beam splitter 1 (BS1) into the first Ph_SLM1 (loaded with a 1200*1200 pixel atmospheric turbulence phase screen), and then successively transmits 350mm of distorted handwritten digital light field through BS1 and BS2 as the input complex amplitude light field of the input model. In the second part, the input light field is first transmitted 170mm to Ph_SLM2 (loaded with a 1024*1024 diffraction encoder), and the outgoing scattered light field is transmitted 170mm through BS2 and BS3 to Ph_SLM3 (loaded with a 1024*1024 diffraction decoder). Finally, the outgoing light field continues to diffract freely for 140mm and is received by the camera CMOS (detection layer). Before the formal experiment begins, the laser, SLM and CMOS are calibrated to avoid the influence of large alignment errors on the experimental results.

[0079] After setting up the experiment, we load the model input on Amp_SLM, such as Figure 8 The first column (from left to right) shows the turbulence phase screen trained by the model loaded on Ph_SLM1, as follows Figure 8 In the first row of the second column, Ph_SLM2 and Ph_SLM3 are then left blank, and they are simply used as reflectors in the optical path, and the distorted light field spot is received at the CMOS, as shown in Figure 8 As shown in the intensity graph below the turbulence on the left side of the second column, we can see the imaging effect of the distorted light field of free space transmission after turbulence disturbance without passing through the ODDM model. The image quality evaluation indicators PSNR and SSIM are marked below each image. When we load the diffraction encoder and diffraction decoder on Ph_SLM2 and Ph_SLM3 respectively, the light field is instantly restored, and a clear image is presented at the receiving end CMOS, as shown in Figure 2. Figure 8 As shown on the right side of the second column, the acquisition of images and the evaluation of PSNR and SSIM evaluation indicators have been significantly improved. The best effect is to increase the PSNR of the distorted digital image 5 to 17.4047 and the SSIM to 0.8261. Then, the turbulence phase screen loaded on Ph_SLM1 is randomly changed to an unknown turbulence phase screen that has not participated in the model training, as shown in Figure 8As shown in the first row of the third column, when Ph_SLM2 and Ph_SLM3 are left blank, the distorted light field image received on the CMOS is as follows: Figure 8 As shown in the distortion intensity diagram of the first row of turbulence phase disturbance on the left side of the third column, the image distortion and blur at this time is also so blurred that its effective information cannot be recognized. The image collected at the CMOS is instantly restored to high quality after the diffraction encoder and diffraction decoder are loaded on Ph_SLM2 and Ph_SLM3 respectively, and its PSNR and SSIM are also significantly improved.

[0080] In order to further verify the imaging effect of the ODDM model on other types of patterns, we replace the binary amplitude map loaded on Amp_SLM to test it, such as Fig. 9 In the first column, when the ODDM model is not loaded, the distorted light fields of the free space transmission of the known turbulence and unknown turbulence patterns collected experimentally are shown as follows: Fig. 9 As shown in the distortion intensity diagrams on the left side of the second column and the left side of the third column, after the diffraction encoder and diffraction decoder are loaded on Ph_SLM2 and Ph_SLM3 respectively, the model performs computational imaging at near the speed of light and images the image efficiently. The calibration experimental results are consistent with the simulation effects. The model has good imaging effects for other unseen randomly selected simple patterns under known and unknown turbulence disturbances.

[0081] Fig.10 The average SSIM and average PSNR of the MNIST test set restored under known and unknown turbulent disturbances using the ODDM model are shown. It can be seen from the figure that the maturely trained model has a stable image quality improvement effect for distorted images under known and unknown turbulent disturbances, and the PSNR can be increased to about 18dB at most, and the SSIM can be increased to more than 80% at most. Therefore, the experiment proves that the ODDM model has an efficient anti-turbulence imaging effect, and the PSNR value of the experimental imaging image can generally reach more than 14, and the maximum can reach 17.0309, and the SSIM can generally be increased to more than 65%, and the maximum can reach 91.81%. The imaging effect is significantly improved, and the process of computational imaging reaches the effect of near light speed. The ODDM model and method proposed in the present invention is a new optical computing paradigm. This idea / method has the potential to promote the realization of more complex neural networks by optical computing and its computing power and speed, and has been proven to have significant results in anti-turbulence imaging applications, and has the potential to promote the implementation of anti-turbulence optoelectronic computing imaging systems.

[0082] In addition, the spatial light modulator has certain restrictions on the incident angle and power of the light beam, so the specific optical path design must be carried out according to the actual situation of the laboratory.

[0083] The contents not described in detail in this specification belong to the prior art known to the professionals in this field.

Claims

1. An anti-turbulence adaptive imaging method based on an optical diffraction diffusion model, characterized in that A new optical computing method is proposed, which combines the diffusion model with the optical diffraction computing paradigm. By adopting the idea of ​​fully dispersing the intermediate layer image information in the diffusion model, an optical diffraction diffusion model architecture is constructed. The imaging of this method relies on optical computing to achieve high efficiency and near-light-speed image restoration.

2. The anti-turbulence adaptive imaging method based on the optical diffraction diffusion model according to claim 1, characterized in that: An optical diffraction diffusion model architecture is proposed: the optical encoder further diffuses the distorted image disturbed by turbulence into a near-Gaussian chaotic state similar to the diffusion model. This state is a latent space (a multidimensional vector space used to represent the potential features or attributes of data. In this space, each vector represents a potential feature or attribute, but these features are usually difficult to explain directly). The optical decoder is then used to perform inverse diffusion transmission on the light field in the latent space. The inverse diffusion process can efficiently separate effective information and noise information in the latent space of the high-dimensional feature space like a "filter".

3. The anti-turbulence adaptive imaging method based on the optical diffraction diffusion model according to claim 2, characterized in that: The optical diffraction diffusion model is trained by inputting the complex amplitude light field of the distorted MNIST (Mixed National Institute of Standards and Technology) data set under turbulent random phase disturbance with turbulence intensity D / r0≈8.33 (where D is the light beam receiving aperture, r0=0.6 is the atmospheric coherence length, and D / r0 is used to characterize the turbulence intensity). After the training, the model diffraction encoder and decoder can be loaded onto the spatial light modulator for direct experiments, which has obvious improvement in the effect of the MNIST (Mixed National Institute of Standards and Technology) test set and other types of patterns with known and unknown turbulent phase screen disturbances at the same intensity.

Images