High-resolution reconstruction method of Fourier single-pixel imaging based on multi-domain diffusion model

Abstract

The present invention discloses a high-resolution reconstruction method for Fourier single-pixel imaging based on a multi-domain diffusion model, comprising: 1) acquiring the target object's light response signal through a single-pixel imaging system; 2) analyzing the photoelectric response value based on a phase modulation method to generate a Fourier sparse sampling spectrum of the target object and a corresponding low-resolution initial image; 3) performing data enhancement on the high-resolution image to obtain data samples. During the prior learning process, Gaussian noise is continuously added to the image domain and frequency domain data to perturb the data distribution, thereby obtaining a priori scores for the image domain and frequency domain data training samples; 4) achieving high-resolution image reconstruction through a serial or parallel collaborative strategy, in which Gaussian noise is gradually iterated from the learned prior scores to obtain the target image; and 5) introducing a PC sampler to correct errors in the inverse SDE process to obtain a corrected reconstructed image. The present method can significantly improve the reconstruction accuracy and high-frequency detail retention capability under ultra-sparse sampling.

Description

Technical Field

[0001] The present invention relates to the field of optical imaging technology, and in particular to a Fourier single-pixel imaging high-resolution reconstruction method based on a multi-domain diffusion model. Background Art

[0002] Fourier single-pixel imaging technology, by combining sparse sampling theory with computational imaging, enables high dynamic range imaging in low-photon flux, wide-spectral, and high-noise scenarios. In recent years, it has demonstrated significant application value in biological microscopy, industrial nondestructive testing, and remote sensing monitoring. The core challenge of this technology lies in reconstructing high-fidelity images from a very small amount of Fourier space (k-space) sampled data. Its performance is highly dependent on the reconstruction algorithm's efficient utilization of prior information. Existing methods fall into two main categories: one based on compressed sensing theory, which exploits k-space sparsity constraints to construct optimization models; the other based on deep learning, which learns image domain feature mappings through end-to-end networks. However, traditional compressed sensing methods suffer from severe underdetermination when sampling extremely sparsely (e.g., sampling rates below 5%), leading to a sharp decline in reconstruction quality. Purely data-driven deep learning models, lacking embedded physical imaging mechanisms, have limited generalization and are prone to introducing artifacts. To overcome the above-mentioned shortcomings, existing technologies have proposed a solution that integrates generative models (such as diffusion models) with physical imaging models. By constructing a diffusion process in the image domain and iteratively optimizing the reconstruction results, the imaging quality at low sampling rates is significantly improved. However, such methods still have obvious limitations: their optimization process relies solely on the image domain generative prior and fails to simultaneously utilize the inherent sparsity characteristics of k-space, resulting in insufficient ability to recover high-frequency details. Specifically, the single domain constraint forces the algorithm to make a trade-off between data consistency (k-space sampling matching) and image authenticity (generative prior). Especially when faced with complex textures or non-stationary noise interference, problems such as blurred edges and residual artifacts in the reconstructed image are prone to occur. In addition, the existing system lacks a dynamic collaborative optimization mechanism between the hardware control module and the reconstruction algorithm, making it difficult to adapt to the rapidly changing requirements of imaging scenarios. Summary of the Invention

[0003] The purpose of the present invention is to propose a high-resolution reconstruction method for Fourier single-pixel imaging based on a multi-domain diffusion model to meet actual needs. By generating prior collaborative constraints through the multi-domain diffusion model, the reconstruction accuracy and high-frequency detail retention capability of Fourier single-pixel imaging under ultra-sparse sampling are significantly improved.

[0004] In order to achieve the above-mentioned purpose, the present invention adopts the following technical solutions.

[0005] A high-resolution reconstruction method for Fourier single-pixel imaging based on a multi-domain diffusion model comprises the following steps:

[0006] Step S1, constructing a single-pixel imaging system and collecting the target object's light response signal;

[0007] The single-pixel imaging system includes a computer host, a helium-neon laser, a lens assembly, a digital micromirror device (DMD), and a photodetector. The helium-neon laser continuously emits red laser light, which is reflected by the lens assembly into the active area of the DMD. The computer host transmits the Fourier basis pattern to the DMD to modulate the light beam. The modulated light beam is focused to an aperture by a lens and then irradiated onto the target object. The photodetector detects the light response signal and synchronously transmits it to the computer host.

[0008] Step S2: analyzing the target object's light response signal;

[0009] The computer host obtains the Fourier coefficients of the Fourier basis pattern based on the collected light response signal of the target object through a four-step phase shift method, and reconstructs the Fourier sparse sampling spectrum of the target object and the corresponding low-resolution initial image;

[0010] Step S3: using a multi-domain diffusion model to perform data enhancement on the high-resolution image;

[0011] The multi-domain diffusion model includes an image domain diffusion model and a frequency domain diffusion model. The image domain diffusion model and the frequency domain diffusion model are trained simultaneously. During the forward diffusion process, data augmentation is performed on high-resolution image domain and frequency domain training images. Gaussian noise is continuously added to the training set to perturb the data distribution and obtain prior scores for the data samples in the image domain and frequency domain.

[0012] Step S4: multi-domain joint optimization based on serial or parallel collaboration;

[0013] The target object image is obtained by iterative steps based on the obtained prior scores: the serial collaborative approach generates the initial reconstructed image through the image domain diffusion model, and then optimizes the spectral characteristics through the frequency domain diffusion model; the parallel collaborative approach runs multiple domain diffusion models synchronously and fuses the output results;

[0014] Step S5: In the process of reconstructing the image based on the serial or parallel collaborative mode, a PC sampler is introduced to correct the error in the reverse SDE process to obtain a corrected reconstructed image.

[0015] Specifically, in step S1, the computer host transmits the binarized Fourier basis pattern to the digital micromirror device DMD to modulate the light beam; the photodetector is a single-pixel photodetector, which converts the detected light response signal into a voltage signal.

[0016] Specifically, the analysis of the target object light response signal in step S2 includes the following:

[0017] According to the collected light response signal of the target object, four Fourier basis patterns with different phases are obtained, and the Fourier coefficients are calculated by the image intensity difference. The Fourier coefficients are calculated using the following formula:

[0018] ;

[0019] In the above formula, are the Cartesian coordinates of the Fourier basis pattern; is an imaginary unit; The frequency is and the initial phase is The photoelectric response value of the Fourier basis pattern, or or or ; is the corresponding frequency The Fourier coefficients at ; is the system gain factor;

[0020] By recombining the obtained Fourier coefficients, the Fourier spectrum of the target object image is obtained. .

[0021] Specifically, in the forward diffusion process in step S3, the high-resolution image domain and frequency domain training images Gradually add Gaussian noise to get The random process of evolution is expressed as:

[0022] ;

[0023] In the above formula, for exist The change in Changes in the noise addition process; is the drift term, is the diffusion coefficient, is the standard Brownian motion noise;

[0024] The time schedule of Gaussian noise variance is expressed as:

[0025] ;

[0026] In the above formula, is Gaussian noise in time Variance when is the minimum value of Gaussian noise variance; is the minimum value of the Gaussian noise standard deviation; is the maximum value of the standard deviation of Gaussian noise;

[0027] Select and , then the forward diffusion process is expressed by the stochastic differential equation SDE as:

[0028] ;

[0029] Prior scores of data samples in image domain and frequency domain using score models To estimate, with parameters Score model Parameters are optimized by the following algorithm To train:

[0030] ;

[0031] In the above formula, is the optimal network parameter, To find the parameters that minimize the expected loss, For time variables expectations; is the time-dependent weighting function; For real data samples expectations; For a given Next, add noise data expectations; is the true score, that is, given hour The gradient of the log probability density, is the data after adding noise; is the transition probability distribution of the diffusion process; once the training is completed, , then use represents the prior score;

[0032] A weighting strategy is applied to the frequency domain spatial input data of the optimization process to reduce the dynamic range difference between high-frequency and low-frequency information so that the prior score Easier to obtain, the frequency domain spatial weighting strategy is expressed as:

[0033] ;

[0034] In the above formula, is the weighted frequency domain spatial input data; Input data in the frequency domain for training the model; is the weight matrix:

[0035] ;

[0036] In the above formula, and are the number of frequency encoding lines and phase encoding lines, respectively, and To adjust the weight, To set the cutoff value, is the smoothness of the weight boundary.

[0037] Specifically, the target object image is obtained by iteratively step by step based on the obtained prior scores in step S4. The target object image reconstruction process is expressed as:

[0038] ;

[0039] In the above formula, for exist The change in changes in the denoising process; To represent the noise in time Variance when represents the prior score, represents the standard Brownian process noise with time regression;

[0040] In each reconstruction iteration, in order to ensure that the generated image is consistent with the actual collected Fourier spectrum data, a data consistency update process is introduced:

[0041] Let the current reconstructed image be , whose Fourier transform is , the low-frequency Fourier data actually collected is , define the mask function as , mask function The value is 1 in the low frequency area and 0 in the rest of the area, then the image is reconstructed after the data consistency is updated. Expressed as:

[0042] .

[0043] Furthermore, in the serial collaborative mode described in step S4, the iterative processes of the image domain and frequency domain diffusion models are performed alternately, and after each iteration, the low-frequency spectrum components are constrained and corrected by updating the data consistency: first, the image domain diffusion model is used for preliminary reconstruction to obtain the intermediate result , and then the frequency domain diffusion model is used to Further processing is performed to obtain the final image .

[0044] Furthermore, in the parallel collaborative approach, the image domain diffusion model and the frequency domain diffusion model are independently used to obtain the intermediate results of preliminary reconstruction. and , perform weighted synthesis of the two to obtain the final reconstructed image , the synthesis formula is as follows:

[0045] ;

[0046] In the above formula, is the weight coefficient; is the inverse Fourier transform.

[0047] Specifically, in step S5, a PC sampler is introduced to correct the error in the reverse SDE process during the image reconstruction based on the serial or parallel collaborative mode to obtain a corrected reconstructed image; the PC sampler includes a predictor and a corrector, which is iteratively corrected by the Markov chain Monte Carlo MCMC method, and the prediction algorithm is based on the prior score. Providing directional updates, generating a preliminary sample update for iterative correction, the prediction algorithm is expressed as:

[0048] ;

[0049] In the above formula, To predict the previous time step The sample status; is the current time step The sample status; is the discretized time step; For about Prior scores of The mean is 0 and the covariance matrix is Gaussian noise, I is the identity matrix;

[0050] The correction algorithm is based on the same prior score , correct the prediction results along the gradient direction to reduce the deviation caused by random noise in the prediction step;

[0051] The correction algorithm is expressed as:

[0052] ;

[0053] In the above formula, is the sample after correction; is the gradient ascent step size.

[0054] Compared with the prior art, the present invention has the following beneficial effects:

[0055] 1. The method of the present invention breaks through the limitations of traditional diffusion models in a single domain by combining dual diffusion models in the image domain and frequency domain. It adopts serial and parallel collaborative training methods to effectively improve the accuracy of image reconstruction under extremely low sampling rates.

[0056] 2. The method of the present invention significantly improves the imaging quality of the system through multi-domain collaborative training, giving it a wider range of application prospects, especially in the fields of medical imaging, physical detection, and computer vision. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 This is a flow chart of the Fourier single-pixel imaging high-resolution reconstruction method based on the multi-domain diffusion model of the present invention;

[0058] Figure 2 It is a block diagram of the architecture of the single-pixel imaging system and multi-domain diffusion model of the present invention;

[0059] Figure 3 It is a parallel collaborative flow chart of high-resolution iterative reconstruction based on a multi-domain diffusion model of the present invention;

[0060] Figure 4 This is a partial flow chart of the reconstruction of the serial collaborative mode and the parallel collaborative mode of the present invention;

[0061] Figure 5 is a reconstruction result image of a cat as a target object in an embodiment of the present invention;

[0062] Figure 6 This is a comparison chart of the SSIM index of the reconstruction results of the method of the present invention and the traditional method. DETAILED DESCRIPTION

[0063] In order to facilitate those skilled in the art to understand and implement the present invention, each step of the method proposed in the present invention is described in detail below. It should be understood that these embodiments are only used to illustrate the present invention and are not intended to limit the scope of the present invention. In addition, it should be understood that after reading the content taught by the present invention, those skilled in the art can make various changes or modifications to the present invention, and these equivalent forms also fall within the scope defined by the claims appended hereto.

[0064] Example

[0065] like Figure 1 As shown, the present invention discloses a Fourier single-pixel imaging high-resolution reconstruction method based on a multi-domain diffusion model, comprising the following steps:

[0066] Step S1, constructing a single-pixel imaging system and collecting the target object's light response signal;

[0067] like Figure 2 As shown, Figure 2The single-pixel imaging system in the center left block diagram includes a computer host, a He-Ne laser, a lens assembly, a digital micromirror device (DMD), and a photodetector. The He-Ne laser continuously emits red laser light, which is reflected by the lens assembly into the active area of the DMD. The computer host transmits the Fourier basis pattern to the DMD to modulate the light beam. The modulated light beam is focused to an aperture by a lens and then irradiated onto the target object. The photodetector detects the light response signal and transmits it synchronously to the computer host.

[0068] Step S2: analyzing the target object's light response signal;

[0069] The computer host obtains the Fourier coefficients of the Fourier basis pattern based on the collected light response signal of the target object through a four-step phase shift method, and reconstructs the Fourier sparse sampling spectrum of the target object and the corresponding low-resolution initial image;

[0070] Step S3: using a multi-domain diffusion model to perform data enhancement on the high-resolution image;

[0071] like Figure 2 As shown, Figure 2 The multi-domain diffusion model in the middle right box includes an image domain diffusion model and a frequency domain diffusion model. The image domain diffusion model and the frequency domain diffusion model are trained simultaneously. During the forward diffusion process, data enhancement is performed on the high-resolution image domain and frequency domain training images. Gaussian noise is continuously added to the training set to perturb the data distribution and obtain a priori scores for the data samples in the image domain and frequency domain.

[0072] Step S4: multi-domain joint optimization based on serial or parallel collaboration;

[0073] The target object image is obtained by iterative steps based on the obtained prior scores: the serial collaborative approach generates the initial reconstructed image through the image domain diffusion model, and then optimizes the spectral characteristics through the frequency domain diffusion model; the parallel collaborative approach runs multiple domain diffusion models synchronously and fuses the output results;

[0074] Step S5: In the process of reconstructing the image based on the serial or parallel collaborative mode, a PC sampler is introduced to correct the error in the reverse SDE process to obtain a corrected reconstructed image.

[0075] Specifically, in step S1, the computer host transmits the binarized Fourier basis pattern to the digital micromirror device DMD to modulate the light beam; the photodetector is a single-pixel photodetector, which converts the detected light response signal into a voltage signal.

[0076] Specifically, the analysis of the target object light response signal in step S2 includes the following:

[0077] According to the collected light response signal of the target object, four Fourier basis patterns with different phases are obtained, and the Fourier coefficients are calculated by the image intensity difference. The Fourier coefficients are calculated using the following formula:

[0078] ;

[0079] In the above formula, are the Cartesian coordinates of the Fourier basis pattern; is an imaginary unit; The frequency is and the initial phase is The photoelectric response value of the Fourier basis pattern, or or or ; is the corresponding frequency The Fourier coefficients at ; is the system gain factor;

[0080] By recombining the obtained Fourier coefficients, the Fourier spectrum of the target object image is obtained. .

[0081] like Figure 3 As shown, Figure 3 In the above figure, FFT stands for Fast Fourier Transform, and IFFT stands for Inverse Fast Fourier Transform. In the forward diffusion process in step S3, the high-resolution image domain and frequency domain training images Gradually add Gaussian noise to get The random process of evolution is expressed as:

[0082] ;

[0083] In the above formula, for exist The change in Changes in the noise addition process; is the drift term, is the diffusion coefficient, is the standard Brownian motion noise;

[0084] The time schedule of Gaussian noise variance is expressed as:

[0085] ;

[0086] In the above formula, is Gaussian noise in time Variance when is the minimum value of Gaussian noise variance; is the minimum value of the Gaussian noise standard deviation; is the maximum value of the standard deviation of Gaussian noise;

[0087] Select and , that is, the forward diffusion process is expressed by the stochastic differential equation SDE as:

[0088] ;

[0089] Prior scores of data samples in image domain and frequency domain using score models To estimate, with parameters Score model Parameters are optimized by the following algorithm To train:

[0090] ;

[0091] In the above formula, is the optimal network parameter, To find the parameters that minimize the expected loss, For time variables expectations; is the time-dependent weighting function; For real data samples expectations; For a given Next, add noise data expectations; is the true score, that is, given hour The gradient of the log probability density, is the data after adding noise; is the transition probability distribution of the diffusion process; once the training is completed, , then use represents the prior score;

[0092] Since the low-frequency and high-frequency components are located in the central area and the peripheral area respectively in the frequency domain, the frequency domain data has a large dynamic range change between low spatial frequencies and high spatial frequencies. Therefore, a weighting strategy is applied to the frequency domain input data of the optimization process to reduce the dynamic range difference between high-frequency and low-frequency information, so that the prior score Easier to obtain, the frequency domain spatial weighting strategy is expressed as:

[0093] ;

[0094] In the above formula, is the weighted frequency domain spatial input data; Input data in the frequency domain for training the model; is the weight matrix:

[0095] ;

[0096] In the above formula, and are the number of frequency encoding lines and phase encoding lines, respectively, and To adjust the weight, To set the cutoff value, is the smoothness of the weight boundary.

[0097] Specifically, the target object image is obtained by iteratively step by step based on the obtained prior scores in step S4. The target object image reconstruction process is expressed as:

[0098] ;

[0099] In the above formula, for exist The change in changes in the denoising process; is Gaussian noise in time Variance when represents the prior score, represents the standard Brownian process noise with time regression;

[0100] In each reconstruction iteration, in order to ensure that the generated image is consistent with the actual collected Fourier spectrum data, a data consistency update process is introduced:

[0101] Let the current reconstructed image be , whose Fourier transform is , the low-frequency Fourier data actually collected is , define the mask function as , mask function The value is 1 in the low frequency area and 0 in the rest of the area. The image is reconstructed after the data consistency is updated. Expressed as:

[0102] .

[0103] like Figure 4 As shown, Figure 4 Where FFT stands for Fast Fourier Transform, IFFT stands for Inverse Fast Fourier Transform, and Mean stands for Average. In the serial collaborative approach, the iterative processes of the image domain and frequency domain diffusion models are performed alternately. After each iteration, the low-frequency spectrum components are constrained and corrected by updating the data consistency: First, the image domain diffusion model is used for preliminary reconstruction to obtain the intermediate result. , and then the frequency domain diffusion model is used to Further processing is performed to obtain the final image ;

[0104] In the parallel collaborative approach, the image domain diffusion model and the frequency domain diffusion model are independently used to obtain the intermediate results of preliminary reconstruction. and , perform weighted synthesis of the two to obtain the final reconstructed image , the synthesis formula is as follows:

[0105] ;

[0106] In the above formula, is the weight coefficient; is the inverse Fourier transform.

[0107] Specifically, in step S5, a PC sampler is introduced to correct the error in the reverse SDE process during the image reconstruction based on the serial or parallel collaborative mode to obtain a corrected reconstructed image; the PC sampler includes a predictor and a corrector, which is iteratively corrected by the Markov chain Monte Carlo MCMC method, and the prediction algorithm is based on the prior score. Providing directional updates, generating a preliminary sample update for iterative correction, the prediction algorithm is expressed as:

[0108] ;

[0109] In the above formula, To predict the previous time step The sample status; is the current time step The sample status; is the discretized time step; For about Prior scores of The mean is 0 and the covariance matrix is Gaussian noise, I is the identity matrix;

[0110] The correction algorithm is based on the same prior score , correct the prediction results along the gradient direction to reduce the deviation caused by random noise in the prediction step;

[0111] The correction algorithm is expressed as:

[0112] ;

[0113] In the above formula, is the sample after correction; is the gradient ascent step size.

[0114] The technical effect of the method of the present invention is further illustrated below through an example.

[0115] Experimental equipment configuration and preparation:

[0116] A HeNe laser (model: JDSU-1137, manufacturer: Power Technology, wavelength: 632.8 nm, beam diameter: 0.84 mm) emits a beam; the beam passes through a beam expansion system consisting of plano-convex lenses L1 (focal length: 18 mm) and L2 (focal length: 150 mm), with a magnification of 8.3 times; the beam is then reflected by a mirror M1 to a digital micromirror device DMD (model: V-7001VIS, manufacturer: Aunion Tech, resolution: 1024×768); a computer transmits the binarized Fourier basis pattern to a digital micromirror device (DMD) to modulate the light beam; the modulated Fourier basis pattern is amplified by a plano-convex lens L3 (focal length: 150 mm); its first-order diffraction light passes through an aperture and illuminates the target, where the diffuse reflected light signal of the reconstructed target is collected by a plano-convex lens L4 and uniformly focused on an A4 sheet of paper; a single-pixel photodetector (model: DET025A / M, manufacturer: Thorlabs, wavelength range: 400-1700 nm) converts the detected light response signal into a voltage signal, which is synchronously transmitted to the computer host by a data acquisition card DAQ (model: PicoScope 3405D, manufacturer: Pico Technology, bandwidth: 100 MHz, sampling rate: 1 GS / s).

[0117] like Figure 5 The figure shows the reconstruction result of the target object being a cat in this embodiment; Figure 5 The first column in represents the sparse sampling rate (5%, 3% and 1%); Figure 5 The second column is the reconstruction result using the traditional FSPI method, followed by two columns in a group. The reconstruction model is marked above each group. The left column of each group is the reconstructed image. The number above the reconstructed image indicates the number of iterations for the best reconstruction result. The right column is the spectrum diagram. The last group is the true value GT of the image; the SSIM value and PSNR value are marked in the upper left corner of the reconstructed image.

[0118] Through the reconstructed images and spectrograms, it can be clearly seen that the reconstruction results of the multi-domain diffusion model proposed by the method of the present invention are better than those of the single-domain diffusion model, whether in terms of the SSIM evaluation index or the PSNR evaluation index. In addition, due to the loss of high-frequency information during the reconstruction process of the traditional FSPI method, the reconstruction results obtained are relatively poor, while the image reconstructed by the method of the present invention has richer high-frequency information.

[0119] Furthermore, to verify the effectiveness of the multi-domain diffusion model proposed in this invention, a series of ablation experiments were conducted in this embodiment. Based on the same data, the serial and parallel collaborative methods of the single-frequency domain diffusion model, the single-image domain diffusion model, and the multi-domain diffusion model were compared. The structural similarity index (SSIM) between the reconstructed image and the real image was statistically calculated. At the same time, to evaluate the robustness of the model, eight different animal (cat) images were selected for experimental analysis.

[0120] like Figure 6 As shown in the figure, the experimental results are presented in the form of a bar graph, where the horizontal axis represents the sparse sampling rate (5%, 3%, 1% and average value), and the vertical axis represents the SSIM value. At each sampling rate, the reconstruction effects of different methods such as unprocessed (DC), single frequency domain diffusion model (single frequency domain), single image domain diffusion model (single image domain), parallel collaboration (parallel) and serial collaboration (series) of multi-domain diffusion models are shown. Figure 6 It can be clearly concluded that compared with the traditional FSPI method, the image reconstruction results of the multi-domain diffusion model proposed in this invention are significantly better than those of the traditional method, and the reconstruction quality of the multi-domain diffusion model is better than that of the single-domain diffusion model, which further verifies the effectiveness and robustness of the multi-domain diffusion model in image reconstruction tasks.

[0121] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other manner. Any person skilled in the art may utilize the above-disclosed technical content to modify or modify the present invention into equivalent embodiments. However, any simple modifications, equivalent variations, and modifications to the above embodiments that do not depart from the technical content of the present invention and are based on the technical essence of the present invention remain within the scope of protection of the present invention.

Claims

1. A high-resolution reconstruction method for Fourier single-pixel imaging based on a multi-domain diffusion model, characterized in that: The following steps are involved: Step S1, constructing a single-pixel imaging system and collecting the target object's light response signal; The single-pixel imaging system includes a computer host, a helium-neon laser, a lens assembly, a digital micromirror device (DMD), and a photodetector. The helium-neon laser continuously emits red laser light, which is reflected by the lens assembly into the active area of the DMD. The computer host transmits the Fourier basis pattern to the DMD to modulate the light beam. The modulated light beam is focused to an aperture by a lens and then irradiated onto the target object. The photodetector detects the light response signal and synchronously transmits it to the computer host. Step S2: analyzing the target object's light response signal; The computer host obtains the Fourier coefficients of the Fourier basis pattern based on the collected light response signal of the target object through a four-step phase shift method, and reconstructs the Fourier sparse sampling spectrum of the target object and the corresponding low-resolution initial image; Step S3: using a multi-domain diffusion model to perform data enhancement on the high-resolution image; The multi-domain diffusion model includes an image domain diffusion model and a frequency domain diffusion model. The image domain diffusion model and the frequency domain diffusion model are trained simultaneously. During the forward diffusion process, data augmentation is performed on high-resolution image domain and frequency domain training images. Gaussian noise is continuously added to the training set to perturb the data distribution and obtain prior scores for the data samples in the image domain and frequency domain. Step S4: multi-domain joint optimization based on serial or parallel collaboration; The target object image is obtained by iterative steps based on the obtained prior scores: the serial collaborative approach generates the initial reconstructed image through the image domain diffusion model, and then optimizes the spectral characteristics through the frequency domain diffusion model; the parallel collaborative approach runs multiple domain diffusion models synchronously and fuses the output results; Step S5: In the process of reconstructing the image based on the serial or parallel collaborative mode, a PC sampler is introduced to correct the error in the reverse SDE process to obtain a corrected reconstructed image.

2. The high-resolution Fourier single-pixel imaging reconstruction method based on a multi-domain diffusion model according to claim 1, characterized in that: In step S1, the computer host transmits the binarized Fourier basis pattern to the digital micromirror device (DMD) to modulate the light beam; the photodetector is a single-pixel photodetector, which converts the detected light response signal into a voltage signal.

3. The high-resolution Fourier single-pixel imaging reconstruction method based on a multi-domain diffusion model according to claim 1, characterized in that: The analysis of the target object light response signal in step S2 includes the following: According to the collected light response signal of the target object, four Fourier basis patterns with different phases are obtained, and the Fourier coefficients are calculated by the image intensity difference. The Fourier coefficients are calculated using the following formula: ； In the above formula, are the Cartesian coordinates of the Fourier basis pattern; is an imaginary unit; The frequency is and the initial phase is The photoelectric response value of the Fourier basis pattern, or or or ; The corresponding frequency The Fourier coefficients at ; is the system gain factor; By recombining the obtained Fourier coefficients, the Fourier spectrum of the target object image is obtained. .

4. The method for high-resolution Fourier single-pixel imaging reconstruction based on a multi-domain diffusion model according to claim 1, characterized in that: In the forward diffusion process in step S3, the high-resolution image domain and frequency domain training images Gradually add Gaussian noise to get The random process of evolution is expressed as: ； In the above formula, for exist The change in Changes in the noise addition process; is the drift term, is the diffusion coefficient, is the standard Brownian motion noise; The time schedule of Gaussian noise variance is expressed as: ； In the above formula, is Gaussian noise in time Variance when is the minimum value of Gaussian noise variance; is the minimum value of the Gaussian noise standard deviation; is the maximum value of the standard deviation of Gaussian noise; Select and , then the forward diffusion process is expressed by the stochastic differential equation SDE as: ； Prior scores of data samples in image domain and frequency domain using score models To estimate, with parameters Score model Parameters are optimized by the following algorithm To train: ； In the above formula, is the optimal network parameter, To find the parameters that minimize the expected loss, For time variables expectations; is the time-dependent weighting function; For real data samples expectations; For a given Next, add noise data expectations; is the true score, that is, given hour The gradient of the log probability density, is the data after adding noise; is the transition probability distribution of the diffusion process; once the training is completed, , then use represents the prior score; A weighting strategy is applied to the frequency domain spatial input data of the optimization process to reduce the dynamic range difference between high-frequency and low-frequency information so that the prior score Easier to obtain, the frequency domain spatial weighting strategy is expressed as: ； In the above formula, is the weighted frequency domain spatial input data; Input data in the frequency domain for training the model; is the weight matrix: ； In the above formula, and are the number of frequency encoding lines and phase encoding lines, respectively, and To adjust the weight, To set the cutoff value, is the smoothness of the weight boundary.

5. The method for high-resolution Fourier single-pixel imaging reconstruction based on a multi-domain diffusion model according to claim 1, characterized in that: In step S4, the target object image is obtained by iterative stepwise based on the obtained prior scores. The target object image reconstruction process is expressed as: ； In the above formula, for exist The change in changes in the denoising process; is Gaussian noise in time Variance when represents the prior score, represents the standard Brownian process noise with time regression; In each reconstruction iteration, in order to ensure that the generated image is consistent with the actual collected Fourier spectrum data, a data consistency update process is introduced: Let the current reconstructed image be , whose Fourier transform is , the low-frequency Fourier data actually collected is , define the mask function as , mask function The value is 1 in the low frequency area and 0 in the rest of the area. The image is reconstructed after the data consistency is updated. Expressed as: 。 6. The method for high-resolution Fourier single-pixel imaging reconstruction based on a multi-domain diffusion model according to claim 5, characterized in that: In the serial collaborative mode described in step S4, the iterative process of the image domain and frequency domain diffusion models is performed alternately. After each iteration, the low-frequency spectrum components are constrained and corrected by updating the data consistency: First, the image domain diffusion model is used for preliminary reconstruction to obtain the intermediate result. , and then the frequency domain diffusion model is used to Further processing is performed to obtain the final image .

7. The method for high-resolution Fourier single-pixel imaging reconstruction based on a multi-domain diffusion model according to claim 5, characterized in that: In the parallel collaborative mode described in step S4, the image domain diffusion model and the frequency domain diffusion model are independently used to obtain the intermediate results of preliminary reconstruction. and , perform weighted synthesis of the two to obtain the final reconstructed image , the synthesis formula is as follows: ； In the above formula, is the weight coefficient; is the inverse Fourier transform.

8. The method for high-resolution Fourier single-pixel imaging reconstruction based on a multi-domain diffusion model according to claim 1, characterized in that: In step S5, a PC sampler is introduced to correct the error in the reverse SDE process during the image reconstruction based on the serial or parallel collaborative mode to obtain a corrected reconstructed image; the PC sampler includes a predictor and a corrector, which is iteratively corrected by the Markov chain Monte Carlo MCMC method, and the prediction algorithm is based on the prior score. Providing directional updates, generating a preliminary sample update for iterative correction, the prediction algorithm is expressed as: ； In the above formula, To predict the previous time step The sample status; is the current time step The sample status; is the discretized time step; For about Prior scores of The mean is 0 and the covariance matrix is Gaussian noise, I is the identity matrix; The correction algorithm is based on the same prior score , correct the prediction results along the gradient direction to reduce the deviation caused by random noise in the prediction step; The correction algorithm is expressed as: ； In the above formula, is the sample after correction; is the gradient ascent step size.

Images