End-to-end turbulence suppression data set augmentation method based on diffusion model

Abstract

The invention provides an end-to-end turbulence suppression data set augmentation method based on a diffusion model, which is used for solving the problems of large calculation amount, long calculation time or insufficiency of generated phase screen high and low frequency information in the existing method for accurately representing space-time correlation of a turbulence degradation process. The technical problem that the visual effect is poor when the method is applied to a clear true-value image is solved. According to the end-to-end turbulence suppression data set augmentation method based on the diffusion model, a turbulence degradation image under a new turbulence degradation condition graph is directly generated from a clear image through the diffusion model, so that augmentation of an end-to-end turbulence suppression data set is realized; large-scale and high-quality image samples which are infinitely close to actual observation can be provided for a turbulence degraded image restoration algorithm based on data driving, simulation of complex atmospheric turbulence conditions is achieved, and then the observation capacity of a foundation large-aperture optical telescope is improved.

Description

Technical Field

[0001] The present invention relates to a data set augmentation method, and in particular to an end-to-end turbulence suppression data set augmentation method based on a diffusion model. Background Art

[0002] In the 400 years of telescope development, the aperture of ground-based telescopes has continued to increase, and their observation capabilities have become stronger and stronger, helping humans to continuously reveal the mysteries of the universe. However, in actual observations, the resolution of telescopes cannot reach the theoretically predicted diffraction limit due to the limitations of atmospheric turbulence above. Taking a 10m aperture telescope as an example, its corresponding theoretical diffraction limit is about 0.01 arc seconds, but even the 10m Keck telescope located on Mauna Kea in Hawaii, one of the two best observatory sites recognized in the world, can only achieve a best resolution of 0.25 arc seconds (corresponding to a wavelength of 500nm) under the seeing limit.

[0003] Due to the highly chaotic nature of atmospheric turbulence itself, traditional algorithms are difficult to accurately characterize the turbulence degradation process with strong spatiotemporal correlation affected by multiple factors, and thus serve the data-driven turbulence degradation image restoration task. In an era of greatly improved computing power, data-driven neural representation algorithms can rely on their powerful nonlinear feature representation capabilities to restore images. However, such algorithms rely on high-quality modeling of turbulence degradation processes. How to accurately characterize the spatiotemporal correlation of turbulence degradation processes has become the basic premise of data-driven neural representation algorithms.

[0004] At present, numerical simulation has been widely used because of its adjustable parameters and easy implementation. Among them, the classic step-by-step propagation algorithm, the power spectral density method and the Zernike polynomial method have become the common methods for accurately characterizing the spatiotemporal correlation of turbulent degradation processes. The classic step-by-step propagation algorithm divides the influence of atmospheric media on light waves into multiple phase screens, calculates the Fresnel diffraction integral of the image pixel by pixel, and models turbulence to cause the phase of light waves to fluctuate randomly in space and time, which is finally used to characterize dynamic aberrations. However, such methods require repeated Fourier transform and inverse Fourier transform, which are computationally intensive and time-consuming. Due to the constraints of Fourier transform itself, the existing power spectral density method generates a phase screen lacking low-frequency information. However, studies have shown that the distribution of turbulent power along the spatial frequency is nonlinear. As the frequency moves to the low-frequency region, the spectral power increases sharply, and the turbulent energy should be mainly concentrated in the low-frequency region. The turbulent phase screen generated based on the Zernike polynomial method has a good effect in the low-frequency part, but the high-frequency component is insufficiently undersampled. The above method lacks high and low frequency information of the phase screen, resulting in poor visual effect on the clear true value image. Summary of the invention

[0005] The object of the present invention is to solve the technical problems existing in the existing methods for accurately characterizing the spatio-temporal correlation of the turbulent degradation process, such as large computational amount, long computational time, or insufficient high and low frequency information of the generated phase screen, and poor visual effect when acting on a clear true-value image, and to provide an end-to-end turbulent suppression dataset augmentation method based on a diffusion model.

[0006] To achieve the above object, the technical solution provided by the present invention is as follows: An end-to-end turbulent suppression dataset augmentation method based on a diffusion model, which is characterized in that it includes the following steps: Step 1, construct a training set and a test set; the data in the training set includes the corresponding turbulent degradation image D, clear image V, and turbulent degradation condition map T, and the data in the test set includes the corresponding clear image V and turbulent degradation condition map T; Step 2, build a diffusion model, the diffusion model includes a forward noise addition unit, a noise prediction network, and a reverse denoising unit connected in sequence according to input and output; Step 3, generate a random Gaussian noise ε with the same size as the turbulent degradation image D, and add the random Gaussian noise ε to the turbulent degradation image D through the forward noise addition unit to obtain the noisy image D at different time steps i ; Step 4, splice the noisy images D at different time steps i with the clear image V and the turbulent degradation condition map T in the training set by channels respectively, and send the spliced images together with the corresponding time steps into the noise prediction network for iterative training to obtain a trained noise prediction network; when the noise prediction network is iteratively trained, the noise prediction network is optimized jointly by a noise prediction loss function, a structural loss function, and an edge loss function; Step 5, modify the tilt direction, tilt action radius, blur action radius, and blur degree of the turbulent degradation condition map T in the test set to obtain a new turbulent degradation condition map T new ; at the same time, define a new random Gaussian noise ε1 with the same size as the turbulent degradation image D, and splice the new random Gaussian noise ε1, the clear image V in the test set, and the new turbulent degradation condition map T new by channels, and send the spliced images into the trained noise prediction network, so that the trained noise prediction network outputs a predicted noise ; Step 6, perform reverse denoising on the predicted noise obtained in Step 5 through the reverse denoising unit to obtain the turbulent degradation image D under the new turbulent degradation condition map T new , completing the end-to-end turbulent suppression dataset augmentation based on the diffusion model. new

[0007] Further, in step 5, the modification of the tilt direction, tilt action radius, blur action radius, and blur degree of the turbulent degradation condition map T in the test set is specifically as follows: The turbulent degradation condition map T in the test set includes four channels. For each channel, different fields of view respectively use any real number within the range of 0-2 as the corresponding influence factor, which is multiplied by the pixel value of the corresponding field of view of the turbulent degradation condition map T to modify the turbulent degradation condition map T. Among them, the modification of the tilt direction is achieved by modifying the pixel values of each field of view in the first channel of the turbulent degradation condition map T, the modification of the tilt action radius is achieved by modifying the pixel values of each field of view in the second channel, the modification of the blur action radius is achieved by modifying the pixel values of each field of view in the third channel, and the modification of the blur degree is achieved by modifying the pixel values of each field of view in the fourth channel.

[0008] Further, in step 2, an array is set in the forward noise addition unit to store the noise scheduling parameters of different time steps.

[0009] Further, in step 2, the noise scheduling parameters of different time steps are obtained through the following steps: Step a1, define a sequence η including t numbers, η = [η1, η2, …, η t ; Step b1, calculate the scheduling sequence α according to the following formula i : α i = 1 - η i , i = 1, 2, …, t; Step c1, calculate the noise scheduling parameters of different time steps according to the scheduling sequence α i , and define and initialize the noise scheduling parameters of different time steps according to the specified rules and store them in the array.

[0010] Further, in step 3, the random Gaussian noise ε is added to the turbulent degradation image D through the forward noise addition unit to obtain the noisy images D at different time steps i Specifically: The random Gaussian noise ε is added to the turbulent degradation image D through the noise scheduling parameters of different time steps stored in the forward noise addition unit to obtain the noisy images D at different time steps i , and its specific expression is as follows: , i = 1, 2, …, t; Among them, is the noise scheduling parameter at time step i.

[0011] Further, in step 4, the noise prediction network is jointly optimized through the noise prediction loss function, the structure loss function, and the edge loss function specifically as follows: Step a2: Use the noise prediction loss function to calculate the output of the noise prediction network during iterative training. Minimize the distance between the random Gaussian noise ε; Step b2, calculate the clean turbulence-degraded image at different time steps :

[0012] Step c2, using the structural loss function, calculates the clean turbulence-degraded images at different time steps Minimize the distance between the image D and the turbulence degradation image; Step d2 uses the edge loss function to calculate the clean turbulence-degraded image at different time steps The distance between the image and the turbulence-degraded image D is minimized.

[0013] Further, in step 4, the noise prediction loss function is a Charbonnier loss function; The structural loss function is a multi-scale structural function; The edge loss function is a total variation TV function.

[0014] Furthermore, in step 6, the reverse denoising unit performs reverse denoising using a DDIM sampling method.

[0015] Furthermore, in step 6, the predicted noise obtained in step 5 is denoised by the reverse denoising unit. Perform reverse denoising to obtain a new turbulence degradation condition map T new The turbulence-degraded image D new Specifically: Step a3: Calculate the denoised image at time step t by the following formula: :

[0016] in, is the noise scheduling parameter at time step t; Step b3, calculate the turbulence degradation image at time step t-1 by the following formula: :

[0017] in, is the noise scheduling parameter at time step t-1; Step c3, using the method in step b to iteratively calculate the turbulence degradation image at all time steps, after the iterative calculation is completed, a new turbulence degradation condition map T is obtained. new The turbulence-degraded image D new。

[0018] Furthermore, in step 1, the turbulence degradation condition map T is obtained in the following way: Subtract the pixel values at the same positions of the corresponding clear image V and the turbulence-degraded image D respectively, and then take the absolute value of the result to obtain the turbulence degradation condition map T.

[0019] The beneficial effects of the present invention compared with the prior art are as follows: 1. An end-to-end turbulence suppression dataset augmentation method based on a diffusion model provided by the present invention directly generates a new turbulence-degraded image under a turbulence degradation condition map from a clear image through the diffusion model, realizing the augmentation of the end-to-end turbulence suppression dataset. It can provide large-scale, high-quality, and images that are infinitely close to actual observations for data-driven turbulence-degraded image restoration algorithms, realizing the simulation of complex atmospheric turbulence conditions, and further improving the observation ability of ground-based large-aperture optical telescopes. The present invention not only omits the complex calculation process of Fourier transform, but also compared with the existing power spectral density method and Zernike polynomial method, the present invention uses the powerful non-linear expression ability of the noise prediction network to solve a series of problems such as slow traditional simulation speed, low scene adaptability, and poor simulation accuracy.

[0020] 2. Through the end-to-end turbulence suppression dataset augmentation method based on the diffusion model of the present invention, turbulence-degraded images with specified tilt directions, tilt action radii, blur action radii, and blur degrees can be generated.

[0021] 3. An end-to-end turbulence suppression dataset augmentation method based on a diffusion model provided by the present invention can obtain a time-series related turbulence degradation sequence under the same new turbulence degradation condition map by processing a clear image for a target with short-time movement.

[0022] 4. The present invention jointly optimizes the noise prediction network through a noise prediction loss function, a structure loss function, and an edge loss function to retain high-frequency components, so as to realize the fast and high-precision generation of turbulence-degraded images under a new turbulence degradation condition map, and can effectively overcome the problem of loss of high-frequency details in turbulence-degraded images. Description of the Drawings

[0023] Figure 1 It is a schematic diagram showing the distribution difference of the image formed by the target passing through the optical system in step 5 of the embodiment of the present invention in different field-of-view spaces of the system after being affected by atmospheric turbulence; Figure 2Schematic diagram of the changes in the tilt action radius and tilt direction when the turbulence degradation condition map T in the test set is divided into 4×4 local blocks in step 5 of the embodiment of the present invention. Among them, (a) is a schematic diagram of the tilt action radius and tilt direction of block 1 and block 2 in the 4×4 local block at the initial moment, and (b) is a schematic diagram of the tilt action radius and tilt direction of block 1 and block 2 in the 4×4 local block at the next moment after being affected by the change of the atmospheric transmission medium. Detailed implementation manners

[0024] To make the advantages and features of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0025] An end-to-end turbulence suppression data set augmentation method based on a diffusion model specifically includes the following steps: Step 1, construct a training set and a test set.

[0026] The data in the training set includes a turbulent degradation image D, a clear image V, and a turbulence degradation condition map T, and the data in the test set includes a clear image V and a turbulence degradation condition map T.

[0027] Among them, the turbulent degradation image D and the clear image V correspond one by one. The turbulent degradation image D and the clear image V can be obtained in many ways. For example, in the morning in a hot and dry area, when the ground is gradually warming up, turbulence is likely to occur. At this time, the target can be photographed to obtain the turbulent degradation image D, and then the turbulent degradation image D can be restored by using a turbulence image restoration algorithm to obtain the clear image V. In addition, the turbulent degradation image D and the clear image V of the target can also be directly photographed under low-temperature and windless meteorological conditions, or the clear image V of the target can be photographed under low-temperature and windless meteorological conditions, and then the clear image V can be degraded by using a numerical simulation method to obtain the turbulent degradation image D.

[0028] Then, the obtained corresponding clear image V and turbulent degradation image D are subtracted pixel by pixel, that is, the pixel values at the same positions of the corresponding clear image V and turbulent degradation image D are subtracted respectively, and then the absolute value is taken to obtain the turbulence degradation condition map T.

[0029] Step 2, build a diffusion model.

[0030] The diffusion model includes a forward noise addition unit, a noise prediction network, and a reverse denoising unit connected in sequence according to input and output.

[0031] An array is set in the forward noise addition unit. First, define a sequence η including t numbers, η = [η1, η2,..., η t , and then calculate the scheduling sequence α according to the following formula i : α i = 1 - η i, i = 1, 2, …, t; where η1 and η t are usually taken as 10 -6 and 10 -2 , and the values between η1 and η t are obtained by adding the tolerance d in sequence. The calculation formula for the tolerance d is: d = .

[0032] Calculate the noise scheduling parameters at different time steps respectively through the following formula : ; Define and initialize the noise scheduling parameters at different time steps and store them in an array according to the specified rules for later call.

[0033] The reverse denoising unit performs reverse denoising using the DDIM sampling method. Define the total number of reverse denoising steps total_step as 2000 and the number of sampling steps num_step as 20. First, define a sampling time step array current_timestep. The length of the sampling time step array current_timestep is the same as the number of sampling steps num_step. All elements in the sampling time step array current_timestep form an arithmetic sequence with a common difference of total_step / num_step, which is 100. Then, define another sampling time step array prev_timestep. The sampling time step array prev_timestep is used to store the arithmetic sequence obtained by adding 0 at the beginning and discarding the last element of the sampling time step array current_timestep. For example, when the sampling time step array current_timestep is [1 101 201 301 1701 1801 1901], the sampling time step array prev_timestep is [0 1 101 201 1701 1801].

[0034] Then, reverse the sampling time step array current_timestep and the sampling time step array prev_timestep respectively. After reverse processing, the sampling time step array current_timestep is [1901 1801 1701 301 201 101 1], and after reverse processing, the sampling time step array prev_timestep is [1801 1701 201 101 1 0].

[0035] When performing reverse denoising, the number of loops is the sampling step number num_step, that is, loop 20 times. During the entire loop process, the elements in the sampled time step array current_timestep after reverse processing and the elements in the sampled time step array prev_timestep after reverse processing are taken in sequence. The element taken from the sampled time step array current_timestep is denoted as the sampled time step t_batch, and the element taken from the sampled time step array prev_timestep array is denoted as the sampled time step prev_t_batch. The sampled time step t_batch and the sampled time step prev_t_batch are used as access noise scheduling parameters to match the noise scheduling parameters and the index number of the corresponding time step t in

[0036] Step 3, forward noise addition

[0037] Generate a random Gaussian noise ε with the same size as the turbulent degradation image D .

[0038] Then, through the noise scheduling parameters at different time steps stored in the forward noise addition unit , add the random Gaussian noise ε to the turbulent degradation image D to obtain the noisy images D at different time steps i , and its specific expression is as follows , i = 1, 2, …, t

[0039] Step 4, concatenate the noisy images D at different time steps i with the clear images V and the turbulent degradation condition maps T in the training set channel by channel. After concatenation, send them into the noise prediction network together with the corresponding time steps for iterative training to obtain the trained noise prediction network. Among them, the clear images V and the turbulent degradation condition maps T are used as the control conditions of the noise prediction network. Denote the output of the noise prediction network during iterative training as .

[0040] When the noise prediction network is iteratively trained, the present invention optimizes the noise prediction network through the noise prediction loss function , the structure loss function and the edge loss function together, and the specific expression is ; where , , are the weighting coefficients of the noise prediction loss, the structure loss, and the edge structure loss, respectively, and can be adjusted according to the iterative training process of the noise prediction network.

[0041] In this embodiment, the noise prediction loss function adopts the Charbonnier loss function to minimize the distance between the output of the noise prediction network during training and the random Gaussian noise ε.

[0042] The clean turbulent degradation images at different time steps are deduced inversely by the following formula :

[0043] The structure loss function adopts the multi-scale structure function (MS-SSIM) to minimize the distance between the calculated clean turbulent degradation images at different time steps and the turbulent degradation image D; The edge loss function adopts the total variation TV function to minimize the distance between the calculated clean turbulent degradation images at different time steps and the turbulent degradation image D.

[0044] Step 5, obtain a new turbulent degradation condition map T new .

[0045] As Figure 1 shown, after the image formed by the target passing through the optical system is affected by atmospheric turbulence, it will be further degraded, and the degradation degree varies due to the distribution difference of atmospheric turbulence in different field-of-view spaces of the system. To illustrate the image degradation degree caused by the turbulent distribution difference, the object plane is divided into different field-of-view regions, and the centroid position of each field of view is used to represent the field-of-view region. Denote the coordinate values of the centroid positions of two field of views on the object plane as and , and denote the characteristics of the atmospheric transmission medium parameters corresponding to the respective fields of view as and . When the system observes and images the target, the target is affected by atmospheric turbulence when passing through the atmospheric transmission medium. The centroid coordinates and corresponding to the fields of view on the image plane and shift and become and . The light rays of the two fields of view and of the aforementioned target should normally correspond to the image plane and respectively, but after passing through different atmospheric transmission medium parameters and At this time, corresponding to the image plane At this point, its offset and It can be seen that the offsets of the two targets are different, that is, the spatial distribution difference of atmospheric turbulence causes different degrees of image degradation in different fields of view.

[0046] As Figure 2 shown, the turbulence degradation condition map T in the test set is divided into M×N local blocks. In this embodiment, M = N = 4, and the center of each local block is the center of the tilt action radius and the blur action radius. Taking the tilt action radius as an example, at the initial moment, blocks 1 and 2 in the 4×4 local block, the tilt action radii are r1 and r2 respectively. After being affected by the atmospheric transmission medium, the tilt action radii become r3 and r4 respectively at the next moment. The blur action radius is similar to the tilt action radius, and the change from r1, r2 to r3, r4 can be regarded as the change of the blur action radius on blocks 1 and 2 in the 4×4 local block. At the initial moment, the tilt directions of blocks 1 and 2 in the M×N local block are respectively and , after being affected by the change of the atmospheric transmission medium, the tilt directions become respectively and at the next moment.

[0047] The present invention uses the per-pixel turbulence degradation condition map T, specifies the per-pixel tilt direction and tilt action radius, and models high-order aberrations such as defocus, astigmatism, and coma as local blur to specify the blur action radius and blur degree. Specifically, the turbulence degradation condition map T is expressed as T = [F1, F2] ∈ , where F1 represents the tilt degradation parameter, F1 = [u; v] ∈ , u and v respectively represent the tilt direction and tilt action radius; F2 represents the blur degradation parameter, F2 = [m; n] ∈ , m and n respectively represent the blur action radius and blur degree of the blur aberration, represents the real number field, H represents the height of the clear image V, W represents the width of the clear image V, and 4 and 2 are the number of channels respectively.

[0048] The turbulence degradation condition diagram T includes four channels. When the present invention modifies the turbulence degradation condition diagram T in the test set, different fields of view in each channel of the turbulence degradation condition diagram T respectively use random decimals within the range of 0-2 as corresponding influence factors for modification. The influence factors of different fields of view in each channel are multiplied pixel by pixel with the turbulence degradation condition diagram T, that is, the influence factors of different fields of view in each channel are multiplied with the pixel values of the corresponding fields of view of the turbulence degradation condition diagram T, so as to achieve the purpose of modifying the turbulence degradation condition diagram T. Specifically, the modification of the tilt direction is achieved by modifying the pixel values of each field of view in the first channel of the turbulence degradation condition diagram T, the modification of the tilt action radius is achieved by modifying the pixel values of each field of view in the second channel, the modification of the blur action radius is achieved by modifying the pixel values of each field of view in the third channel, and the modification of the blur degree is achieved by modifying the pixel values of each field of view in the fourth channel. After the tilt direction, tilt action radius, blur action radius, and blur degree of the turbulence degradation condition diagram T in the test set are all completed, a new turbulence degradation condition diagram T is obtained new 。

[0049] Step 6, reverse denoising

[0050] Define a new random Gaussian noise ε1 with the same size as the turbulence degradation image D as the initial input for the entire reverse denoising. The new random Gaussian noise ε1, the clear image V in the test set, and the new turbulence degradation condition diagram T new After being spliced by channel, together with the sampling time step t_batch stored in the reverse denoising unit, are fed into the trained noise prediction network. The trained noise prediction network outputs the predicted noise 。

[0051] Use the following formula for reverse derivation calculation to obtain the denoised image at time step t :

[0052] Among them, is the noise scheduling parameter at time step t; Use the following formula to calculate the turbulence degradation image at time step t-1 :

[0053] Among them, is the noise scheduling parameter at time step t-1.

[0054] Iteratively calculate the turbulence degradation images at all time steps using the above method. After the iterative calculation is completed, that is, when the sampling time step t_batch takes 1 and the sampling time step prev_t_batch takes 0, the final output of the reverse denoising process is obtained, that is, a new turbulence degradation condition diagram T new The turbulence degradation image D undernew , the end-to-end turbulence suppression dataset augmentation based on the diffusion model is completed.

[0055] As described above, it is only used to illustrate the technical solution of the present invention, rather than to limit it. For those of ordinary professional skills in the art, the specific technical solution recorded in the above embodiments can be modified, or some of the technical features can be equivalently replaced. These modifications or replacements do not make the essence of the corresponding technical solution deviate from the scope of the technical solution protected by the present invention.

Claims

1. An end-to-end turbulence suppression dataset augmentation method based on diffusion models, characterized in that, It includes the following steps: Step 1: Construct a training set and a test set; the data in the training set includes corresponding turbulent degradation images D, clear images V, and turbulent degradation condition maps T, and the data in the test set includes corresponding clear images V and turbulent degradation condition maps T; Step 2: Build a diffusion model, which includes a forward noise-adding unit, a noise prediction network, and a reverse denoising unit connected in sequence according to input and output; Step 3: Generate a random Gaussian noise ε with the same size as the turbulent degradation image D, and add the random Gaussian noise ε to the turbulent degradation image D through the forward noise addition unit to obtain the noisy images D at different time steps i ; Step 4: For the noisy images D at different time steps i they are respectively concatenated with the clear images V and the turbulence degradation condition maps T in the training set channel by channel. After concatenation, they are sent into the noise prediction network together with the corresponding time steps for iterative training to obtain a trained noise prediction network. When the noise prediction network is iteratively trained, the noise prediction network is jointly optimized by a noise prediction loss function, a structure loss function, and an edge loss function; Step 5: Modify the tilt direction, tilt action radius, blur action radius, and blur degree of the turbulent degradation condition map T in the test set to obtain a new turbulent degradation condition map T new ; At the same time, define a new random Gaussian noise ε1 with the same size as the turbulent degradation image D, and combine the new random Gaussian noise ε1, the clear image V in the test set, and the new turbulent degradation condition map T new channel-wise, and after concatenation, feed it into the trained noise prediction network to make the trained noise prediction network output the predicted noise ; Step 6, perform reverse denoising on the predicted noise obtained in Step 5 to obtain a new turbulence degradation condition map T new and the turbulence degradation image D new under it, thus completing the augmentation of the end-to-end turbulence suppression dataset based on the diffusion model.

2. The end-to-end turbulent suppression dataset augmentation method based on a diffusion model according to claim 1, wherein: In step 5, the modification of the tilt direction, tilt action radius, blur action radius, and blur degree of the turbulent degradation condition map T in the test set is specifically as follows: The turbulent degradation condition map T in the test set includes four channels, and different fields of view of each channel respectively use any real number within the range of 0-2 as the corresponding influence factor, which is multiplied by the pixel value of the corresponding field of view of the turbulent degradation condition map T to modify the turbulent degradation condition map T; among them, the modification of the tilt direction is achieved by modifying the pixel values of each field of view of the first channel of the turbulent degradation condition map T, the modification of the tilt action radius is achieved by modifying the pixel values of each field of view of the second channel, the modification of the blur action radius is achieved by modifying the pixel values of each field of view of the third channel, and the modification of the blur degree is achieved by modifying the pixel values of each field of view of the fourth channel.

3. The end-to-end turbulent suppression dataset augmentation method based on a diffusion model according to claim 2, wherein: In step 2, an array is set in the forward noise-adding unit for storing noise scheduling parameters at different time steps.

4. The end-to-end turbulent suppression dataset augmentation method based on a diffusion model according to claim 3, wherein: In step 2, the noise scheduling parameters at different time steps are obtained through the following steps: Step a1, define a sequence η including t numbers, η = [η1, η2, …, η t ]; Step b1, calculate the scheduling sequence α according to the following formula i :[[]]END]] α i = 1 - η i , i = 1, 2, …, t; Step c1, according to the scheduling sequence α i Calculate the noise scheduling parameters for different time steps respectively, and store the noise scheduling parameters for different time steps in an array after defining and initializing them according to the specified rules.

5. The end-to-end turbulent suppression dataset augmentation method based on a diffusion model according to claim 4, wherein: In step 3, random Gaussian noise ε is added to the turbulent degradation image D through the forward noise addition unit to obtain the noisy image D at different time steps i Specifically: Add the random Gaussian noise ε to the turbulent degraded image D through the noise scheduling parameters at different time steps stored in the forward noise addition unit to obtain the noisy images D at different time steps. i The specific expression is as follows: , where \(i = 1, 2, \ldots, t\); Among them, is the noise schedule parameter at time step i.

6. The end-to-end turbulent suppression dataset augmentation method based on a diffusion model according to claim 5, wherein: In step 4, the noise prediction network is jointly optimized by a noise prediction loss function, a structure loss function, and an edge loss function, specifically as follows: Step a2, using a noise prediction loss function, minimizes and optimizes the distance between the output of the noise prediction network during iterative training and the random Gaussian noise ε; Step b2, calculate the clean turbulent degradation images at different time steps : ； Step c2, using a structural loss function, minimizes and optimizes the distance between the calculated clean turbulent degradation images at different time steps and the turbulent degradation image D; Step d2: Using an edge loss function, minimize the distance between the calculated clean turbulence-degraded images at different time steps and the turbulence-degraded image D.

7. The end-to-end turbulent suppression dataset augmentation method based on a diffusion model according to claim 6, wherein: In step 4, the noise prediction loss function is the Charbonnier loss function; The structure loss function is the multi-scale structure function; The edge loss function is the total variation TV function.

8. The end-to-end turbulent suppression dataset augmentation method based on a diffusion model according to claim 7, wherein: In step 6, the reverse denoising unit uses the DDIM sampling method for reverse denoising.

9. The end-to-end turbulent suppression dataset augmentation method based on a diffusion model according to claim 8, wherein: In step 6, the predicted noise obtained in step 5 is reversely denoised by the reverse denoising unit to obtain a new turbulence degradation condition map T new and a turbulence degradation image D under the new turbulence degradation condition map T new Specifically: Step a3, calculate the denoised image at time step t using the following formula : ； Among them, is the noise scheduling parameter at time step t; Step b3, calculate the turbulent degradation image at time step t-1 using the following formula :[[]]END]] ； Among them, is the noise scheduling parameter at time step t - 1; Step c3: Use the method in step b to iteratively calculate the turbulent degradation images at all time steps. After the iterative calculation is completed, a new turbulent degradation condition map T is obtained. new The turbulent degradation image D under new .

10. The end-to-end turbulent suppression dataset augmentation method based on a diffusion model according to claim 1, wherein: In step 1, the turbulent degradation condition map T is obtained through the following method: Subtract the pixel values at the same positions of the corresponding clear image V and the turbulence-degraded image D respectively, and then take the absolute value of the result to obtain the turbulence degradation condition map T.

Images