Lensless imaging method for combined generation of twin image and original image

Abstract

A lensless imaging method for combined generation of a twin image and an original image comprises two parts of prior learning of a high-dimensional combined solution space and combined generation of the twin image and the original image, and in the prior learning stage of the high-dimensional combined solution space, a twin image domain scoring network and an original image domain scoring network are trained respectively, and a high-dimensional combined solution space is obtained; wherein the twin image domain scoring network specially learns artifact distribution characteristics formed by diffraction effect and conjugate wave interference, and the original image domain scoring network focuses on intrinsic structure characteristics of a modeling target object; by constructing a high-dimensional joint solution space formed by an original image and a twin image, an artifact removal problem in traditional Fresnel zone plate lens-free imaging is converted into an independent generation problem of the original image and the twin image, so that mutual interference in an iterative optimization process is effectively avoided.

Description

Technical Field

[0001] This invention relates to the field of computational optical imaging technology, and more particularly to Fresnel zone plate lensless imaging technology. Background Technology

[0002] Lensless imaging technology utilizes encoders to replace traditional lenses, offering advantages such as lightweight systems and low cost, and has broad application prospects in many key fields such as biomedicine and mobile imaging. However, because image sensors are only sensitive to light intensity, phase information is lost during acquisition, leading to twin images in the reconstructed images. Twin images are an inherent problem in coaxial holography, manifesting as the superposition of the object's conjugate image on the original image, significantly reducing image quality. Traditional reconstruction algorithms can remove artifacts to some extent, but image details are significantly damaged, limiting the widespread application of lensless imaging. Therefore, developing a high-quality, artifact-free lensless imaging method has become a critical problem that urgently needs to be solved.

[0003] The existing technology has the following technical problems:

[0004] (1) Hardware technology solution: Traditional Fresnel zone plate-based coding masks often use single-phase zone plates or multi-phase combinations of non-orthogonal phases, which are difficult to effectively suppress twin artifacts at the hardware level and require subsequent algorithm optimization, thus limiting the imaging quality.

[0005] (2) Algorithm technology: Traditional reconstruction methods, such as backpropagation algorithm, compressed sensing algorithm and deep learning solutions developed in recent years, generally treat twin images and original images as a whole for suppression or removal. They fail to fully explore the separable characteristics of the two at the information level, resulting in incomplete removal of artifacts and easy loss of image details, making it difficult to achieve high-quality reconstruction.

[0006] In summary, current lensless imaging technology focuses on forcibly removing twin artifacts, but the removal effect is limited and it is difficult to balance image quality and detail preservation. This is far from meeting the requirements of high-precision lensless imaging applications in fields such as biomedical imaging and industrial inspection. Summary of the Invention

[0007] This invention provides a lensless imaging method that jointly generates twin and original images. By constructing a high-dimensional joint solution space composed of the original and twin images, the artifact removal problem in traditional Fresnel zone plate lensless imaging is transformed into an independent generation problem of the original and twin images, thus effectively avoiding mutual interference during the iterative optimization process. This invention employs an encoder with orthogonal phase to construct the lensless imaging system, further eliminating twin interference. In the prior learning of the high-dimensional joint solution space, a dual-diffusion model is constructed to learn the prior distribution information of the twin and original image domains, respectively. During the joint generation stage of the twin and original images, an iterative strategy of alternating twin and original image generation is used to simultaneously reconstruct the original and twin images. The twin prior information is used to separate the corresponding twin interference signal, while the original prior information guides the reconstruction result to approximate the structure and texture of the real scene. A data consistency process based on a Fresnel zone plate lensless imaging propagation model is used to ensure fidelity. Through an alternating iterative process based on the Fresnel zone propagation model, artifacts are gradually transferred to the twin image, ultimately yielding the original image without artifacts.

[0008] This invention provides a lensless imaging method for jointly generating twin images and original images. The method includes two parts: prior learning of a high-dimensional joint solution space and joint generation of twin and original images.

[0009] In the prior learning phase of the high-dimensional joint solution space, twin image domain scoring networks and original image domain scoring networks are trained separately. The twin image domain scoring network is specifically trained to learn the artifact distribution characteristics formed by diffraction effects and conjugate wave interference, while the original image domain scoring network focuses on modeling the intrinsic structural features of the target object.

[0010] In the generation stage of the high-dimensional joint solution space, a rotating iterative strategy is used to simultaneously reconstruct the twin image and the original image. The original image and the twin image are respectively subjected to regularization constraints based on prior information and data consistency constraints based on the Fresnel zone plate lensless imaging physical propagation model, and a dual-domain transformation is performed through the superposition property of Fresnel zone plate lensless imaging. The reconstruction process includes:

[0011] The first step is to randomly generate a two-dimensional Gaussian noise image as the inverse starting point of the diffusion model;

[0012] The second step is to input the Gaussian noise image into the original image domain scoring network and solve the stochastic differential equation of the inverse-time variance explosion of the inverse-time original image training model.

[0013] The third step involves introducing prior information from the original image domain to constrain the prediction of the target information for the next time step using a predictor, and using the annealing Langevin equation as a corrector to input the predicted original image into the inner loop for correction, so as to obtain the initial reconstructed original image.

[0014] The fourth step is to use the fidelity term of the physical propagation model of lensless imaging based on Fresnel zone plates to constrain data consistency, and then subtract the fidelity term from the original image in the previous step to obtain a twin image.

[0015] The fifth step is to input the twin images into the twin image domain scoring network and solve the stochastic differential equation of the inverse-time variance explosion of the inverse-time original image training model.

[0016] The sixth step involves introducing prior information constraints from the twin image domain. The predictor forecasts the target information for the next time step, and the predicted twin image is input into the internal loop of the annealing Langevin corrector for correction to obtain the optimized twin image.

[0017] Step 7: Subtract the fidelity value from the twin image from the previous step to obtain the original image, and use it as input again;

[0018] Step 8: Repeat steps 2 through 7 a preset number of times, then end the loop and use the original image output after the last loop as the reconstructed image that is similar to the target object.

[0019] One or more technical solutions provided by this invention have at least the following technical effects or advantages:

[0020] This invention employs an encoder with orthogonal phase to construct a lensless imaging system. It builds a high-dimensional joint solution space of the original image and twin image, using this space as the overall optimization objective to simultaneously generate high-quality original and twin images. A fractional diffusion model is used to model the prior information of the original and twin images separately, constraining the joint generation process of the two domains. This gradually transfers artifact interference information from the reconstructed image to the twin image, thereby separating the original image with clear details and no artifacts. Compared to existing single-phase or non-orthogonal phase encoders, this invention uses orthogonal phase, eliminating most twin images and significantly improving image quality. Compared to solution methods based on constraints of a single original image domain, this invention can almost completely eliminate twin artifacts in the reconstructed image, accurately reproducing the detailed texture features and color fidelity of the target.

[0021] This invention effectively solves the fundamental problem of twin image artifact suppression in lensless imaging. It not only provides reliable technical support for the widespread application of lensless imaging, but also provides a general solution for high-quality image reconstruction. It has broad application prospects in computational optical imaging systems affected by twin images, such as digital holography.

[0022] This invention achieves a significant breakthrough in the reconstruction effect of Fresnel lensless imaging. Experimental results show that the peak signal-to-noise ratio of the reconstructed image is improved by nearly 2.86 dB compared with the latest proposed diffusion model method, and the structural similarity value reaches 0.92, which significantly improves the overall image quality. Attached Figure Description

[0023] Figure 1 This is a flowchart of the algorithm reconstruction in one embodiment of the present invention;

[0024] Figure 2 This is a simulation experiment result diagram in one embodiment of the present invention. Detailed Implementation

[0025] This invention provides a lensless imaging method that jointly generates twin images and original images. It employs orthogonal phase to construct a lensless imaging system and achieves high-quality Fresnel zone plate lensless imaging by jointly generating original and twin images. Image generation is performed in the high-dimensional solution space of the original-twin image, progressively transferring artifacts to the twin image to generate an artifact-free and structurally clear original image. This invention effectively suppresses twin artifacts in the reconstructed image while preserving texture details and color fidelity, solving the problem of twin artifact interference in lensless imaging and significantly improving imaging quality and accuracy.

[0026] First, the technical terms used in this invention will be explained.

[0027] An orthogonal phase encoder is a key optical component used for wavefront modulation in the lensless imaging system of this invention. It consists of a pair of encoders with a specific phase relationship. The complex amplitude transfer functions of this encoder pair are mathematically orthogonal, causing the twin images generated by the two orthogonal phases in the imaging result to exhibit complementary characteristics and cancel each other out when superimposed. This encoding design significantly reduces the correlation between the original image and the twin image in the observation data at the physical level, thereby effectively suppressing the interference of the twin image on the original image at the source of the system.

[0028] The high-dimensional joint solution space refers to the modeling of the original image and the twin image as coupled variables in the lensless imaging process of this invention within a high-dimensional parameter space. This space not only contains the structural information of the original image but also covers the artifact components of the twin image. Modeling this space using a diffusion model allows for the learning of prior distribution information in both the twin image domain and the original image domain. The artifact distribution characteristics formed by diffraction effects and conjugate wave artifacts are modeled by the scoring network of the twin image domain, while the structural characteristics of the target are modeled by the scoring network of the original image domain.

[0029] The iterative rotation strategy is a method for jointly generating twin and original images as proposed in this invention. This strategy treats the original and twin images as two independent optimization variables, updating them alternately within a unified iterative framework. In each iteration, the original image is first updated using a prior model and data consistency constraints. Then, the twin image is obtained by subtracting the original image from the fidelity term. The twin image is then updated in the same way, returning to the original image domain. During the iterative reconstruction process, the model gradually guides artifact energy to migrate from the original image to the twin image, thereby effectively separating and suppressing twin artifacts while preserving the texture details and color fidelity of the original image, achieving high-quality, structurally clear image reconstruction.

[0030] To better understand the lensless imaging method based on the joint generation of the original image and twin image described above, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods. Obviously, the embodiments described in this invention are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without inventive effort are within the scope of protection of this invention.

[0031] The lensless imaging method for jointly generating twin images and original images of the present invention mainly includes two stages: a priori learning stage of high-dimensional joint solution space and a joint generation stage of twin images and original images.

[0032] (a) Prior learning stage of high-dimensional joint solution space;

[0033] like Figure 1 In the prior learning phase of the high-dimensional joint solution space, the original image and twin image datasets are first constructed separately: a color three-channel dataset. Magnification factor is Pipe imaging to construct raw image dataset The original image after system scaling. Forward propagation is performed with transfer functions of phases 0 and 0.5π respectively to generate corresponding coded images. Then, backward propagation is performed on the two sets of coded images with different phases, and the results are superimposed to obtain a two-phase backward propagation reconstructed image. The difference between the backward propagation reconstructed image and the original image is calculated to construct a twin image dataset. Its expression is as follows:

[0034] (1)

[0035] In the formula, Represents a color three-channel dataset. Represents the original image dataset, Represents a twin image dataset, This indicates that a pinhole camera is being performed on the input image. This represents the forward propagation operator for lensless imaging of Fresnel zone plates. This represents the backpropagation operator for lensless imaging of Fresnel zone plates.

[0036] It should be noted that this embodiment uses an orthogonal phase composed of 0 and 0.5π for lensless imaging. In practical applications, the selected phase is not limited in any way. Single-phase or multi-phase can be selected according to actual needs, and this invention does not impose any restrictions.

[0037] The construction process for the original image dataset and the twin image dataset involves, for example, selecting the LSUN-Church dataset as the original dataset for the target objects, and preprocessing each selected image to a size of 256×256 pixels through cropping and scaling to complete the preprocessing of the original dataset. Next, 8000 images are selected from the preprocessed dataset, of which 4000 images are processed into the original image dataset. Another 4,000 images were processed into a twin dataset. .

[0038] After constructing the original image dataset and twin datasets Then, the prior learning process begins, which includes steps S11 to S13.

[0039] Step S11: Transfer the original image dataset The input is placed into the original image domain scoring network to learn prior information about the intrinsic structural features of the target object;

[0040] The process of obtaining the original image domain scoring network includes steps S111 to S112.

[0041] Step S111: Configure the stochastic differential equation for variance explosion used in the model training phase, as shown in the following expression:

[0042] (2)

[0043] In the formula Represents the original image dataset, This represents a monotonically increasing function in the original image domain. The original image domain diffusion coefficient is... It is uniform sampling. Indicates in Brownian motion in 3D real space Represent real numbers, Represents the vector dimension.

[0044] The stochastic differential equations of variance explosion can effectively distribute data across a wide range of noise levels, thereby enabling more effective learning of the underlying data structure and improving the quality of generated samples.

[0045] Step S112: Train a time-varying score network To estimate Gradient of all logarithmic data distributions This process can be modeled as solving the core objective function within a scoring-based stochastic differential equation framework:

[0046] (3)

[0047] In the formula, These represent the optimal parameters for training the original image domain neural network. This represents the training parameters of the original image domain neural network. Expressing expectations, Represents a positive weighting function. Represents the training samples of the original image domain. by A Gaussian perturbation kernel centered on the core. This represents the continuous-time correlation fractional function of the original image domain network. for The gradient of the logarithmic data distribution.

[0048] Step S12: Transfer the twin image dataset Input into the twin image domain scoring network to learn prior information about the artifact distribution characteristics formed by diffraction effects and conjugate wave interference.

[0049] The process of obtaining the twin image domain scoring network includes steps S121 to S122.

[0050] Step S121: Configure the stochastic differential equation for variance explosion used in the model training phase, as shown in the following expression:

[0051] (4)

[0052] In the formula, Represents a twin image dataset, This represents a monotonically increasing function in the twin image domain. The diffusion coefficient of the twin image domain. It is uniform sampling. Indicates in Brownian motion in 3D real space Represent real numbers, Represents the vector dimension.

[0053] The stochastic differential equations of variance explosion can effectively distribute data across a wide range of noise levels, thereby enabling more effective learning of the underlying data structure and improving the quality of generated samples.

[0054] Step S122: Train a time-varying score network To estimate Gradient of all logarithmic data distributions This process can be modeled as solving the core objective function within a scoring-based stochastic differential equation framework:

[0055] (5)

[0056] In the formula, The optimal parameters for training a twin-image domain neural network are represented. This represents the training parameters of the twin-image domain neural network. Expressing expectations, Represents a positive weighting function. The training samples represent the twin image domain. by A Gaussian perturbation kernel centered on the core. This represents the continuous-time correlation fractional function of the twin image domain network. for The gradient of the logarithmic data distribution.

[0057] Step S13: Configure training parameters. Radius of the innermost region of the Fresnel zone mask. The distance was set to 0.23 mm, and two phases, 0 and 0.5π, were used for encoding. The distance between the mask and the sensor was set to 3 mm, and the pixel pitch was 0.014 mm. The target size was set to 180 × 180 mm, and the distance between the target and the sensor was set to 300 mm. Gaussian noise with values ​​ranging from 0.01 to 380 was added to perturb the data distribution. The Adam optimizer was used for network optimization, with a learning rate set to 0.0002 and 500,000 training iterations. All training and experiments were performed on an NVIDIA GeForce RTX 4090D 24GB graphics card.

[0058] (ii) The combined generation of twin images and original images;

[0059] like Figure 1 As shown, in the joint generation stage of the twin image and the original image, an iterative rotation strategy is used to simultaneously generate the twin image and the original image. The iteration starts from the pure Gaussian noise image in the original image domain. In the... In the next iteration, the original image After applying the regularization term constraint shown in equation (6a) to the original image domain diffusion model, and... Twin images obtained by subtraction In the twin domain, twins After applying the regularization term constraint shown in equation (6c) to the twin image domain diffusion model, and... The original image is obtained by subtraction. The final output is a high-quality reconstructed image without artifacts. The process expression is:

[0060]

[0061] In the formula, Indicates the first A primitive image, far from the length of a walk. This represents an external index with a total time step of . This represents a regularization term based on prior information from the original image domain. Indicates the first A pair of twins, one long and one short, For data fidelity items, The regularization term is represented by the prior information of the twin image domain, and formulas (6b) and (6d) correspond to the data consistency term.

[0062] The solution process for the rotational iteration strategy includes steps one through eight.

[0063] The first step is to randomly generate a two-dimensional Gaussian noise image as the inverse starting point for the diffusion model.

[0064] The second step involves inputting the Gaussian noise image into the original image domain scoring network to solve the stochastic differential equation of the inverse-time variance explosion of the inverse-time original image training model.

[0065] The second step specifically involves utilizing the trained original image scoring network. Solve the stochastic differential equations for the inverse-time variance explosion of the inverse-time original image training model:

[0066] (7)

[0067] In the formula, Represents the original image dataset, This represents a monotonically increasing function in the original image domain. The original image domain diffusion coefficient is... It is uniform sampling. This represents the continuous-time correlation fractional function of the original image domain network. Indicates in Brownian motion in 3D real space Represent real numbers, Represents the vector dimension.

[0068] The third step is to introduce a regularization constraint based on prior information of the original image domain in formula (6a). This is achieved through a predictor corrector in the diffusion model. The predictor in the diffusion model refers to the numerical solver of the inverse-time stochastic differential equation, which predicts the ... Original image of the target at each time step The corrector uses the Langevin Markov chain correction algorithm to correct the predicted original image. Correction is performed in the gradient ascent direction. Its expression is:

[0069] (8)

[0070] In the formula, Indicates the first The original image predicted from the distance of the walk, Indicates the total time step as External indexes, Indicates the first A primitive image, far from the length of a walk. Indicates the first The noise intensity during a long walk express The square of, Indicates the first The noise intensity during a long walk express The square of, Indicates Gaussian noise. Indicates in As the input of the inner loop, the first A primitive image that is far from the distance of a long walk. Indicates in As the input of the inner loop, the first A primitive image that is far from the distance of a long walk. Indicates the first A corrective stride length for long walks. The total step size is The internal iterative index.

[0071] Once the predictive correction process is complete, the first... The original image at a discrete time step .

[0072] The fourth step, as shown in formula (6b), utilizes the fidelity term based on the lensless imaging physical propagation model of Fresnel wave zone plates. Perform data consistency constraints, comparing the fidelity items with those from the previous step. Twin images are obtained by subtracting the original images. Among them, the fidelity term... The acquisition includes:

[0073] Encoded images with phases of 0 and 0.5π were acquired respectively. Perform independent reconstructions, then superimpose the two reconstruction results to obtain the final biphase reconstruction result, which is the deconvolutioned image. Its expression is as follows:

[0074] (9)

[0075] In the formula, This is the result of backpropagation of orthogonal phases. For the acquired coded image intensity information, Let be the transfer function in the physical propagation process in the frequency domain. Indicates Fourier transform, Indicates the inverse Fourier transform. Indicates the number of phases.

[0076] The fifth step is to input the twin images into the twin image domain scoring network and utilize the trained prior distribution. Solve the stochastic differential equations for the inverse-time variance explosion of the inverse-time original image training model;

[0077] (10)

[0078] In the formula, Represents a twin image dataset, This represents a monotonically increasing function in the original image domain. The original image domain diffusion coefficient is... It is uniform sampling. This represents the continuous-time correlation fractional function of the twin image domain network. Indicates in Brownian motion in 3D real space Represent real numbers, Represents the vector dimension.

[0079] Step 6, the regularization term in formula (6c) based on prior information from the twin image domain. This is achieved through a predictive corrector. The predictor predicts the first... A target twin at a time step The predicted twin image The input annealing process is performed within the internal loop of the Langevin orthodontic system for correction; its expression is:

[0080] (11)

[0081] In the formula, Indicates the first A predicted twin image of a walking distance, This represents an external index with a total time step of . Indicates the first A pair of twins, one long and one short, Indicates the first The noise intensity during a long walk express The square of, Indicates the first The noise intensity during a long walk express The square of, Indicates Gaussian noise. Indicates in As the input of the inner loop, the first A pair of twins who have been walking for a long time. Indicates in As the input of the inner loop, the first A pair of twins who have been walking for a long time. Indicates the first A corrective stride length for long walks. The total step size is The internal iterative index.

[0082] Step 7: Ensure the authenticity of the item. With the A pair of twins walking together Difference to obtain the next time step The original image And serve as input for the next iteration.

[0083] Step 8: Repeat steps 2 through 6 a preset number of times, then end the loop and output the original image after the last loop. As a reconstructed image similar to the target object. In specific implementation, for example, when the preset number of iterations is 1000, the seventh step outputs a reconstructed image with very few artifacts and high definition.

[0084] To evaluate the performance of this invention, 100 additional images were randomly selected from the LSUN-church dataset as a test dataset. To simulate actual captured images, each image in the test set was convolved with a Fresnel zone mask projection to obtain the image acquired by the image sensor. Backpropagation, compressed sensing, and diffusion model algorithms were selected for comparison. Peak signal-to-noise ratio and structural similarity were chosen as evaluation metrics. The results are shown in Table 1.

[0085] Table 1. Average peak signal-to-noise ratio and structural similarity.

[0086] method Peak signal-to-noise ratio (in dB) Structural similarity Backpropagation 20.54 0.80 Compressed sensing 20.78 0.80 diffusion model 21.51 0.75 This invention 24.37 0.92

[0087] Reconstruction results as follows Figure 2As shown. To better demonstrate the quality of the reconstructed image, the sub-image below the reconstruction result is a magnified view of the details within the red box. For complex targets, the backpropagation algorithm and the compressed sensing algorithm produce similar reconstruction results. While both can restore the basic outline of the image, they both exhibit obvious mesh artifacts and severe color distortion. Although the diffusion model algorithm can remove mesh artifacts to some extent, it leads to an overall blurring of the reconstructed image, specifically manifested as loss of texture details and significant degradation of edge sharpness. Furthermore, twin image interference remains in the surrounding black borders of its reconstruction result. In contrast, this invention demonstrates significant advantages in twin image elimination, complex outline reconstruction, and color fidelity. Its reconstructed image not only effectively removes conjugate wave interference from the internal patterns and black borders, accurately restoring the geometric features of the internal structure, but also achieves high-precision color restoration, resulting in a more realistic overall visual effect and detail presentation. The results show that the image quality reconstructed by this invention is superior to other technical solutions.

[0088] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A lensless imaging method for jointly generating a twin image and a original image, characterized in that, It includes two parts: prior learning of the high-dimensional joint solution space and joint generation of twin images and original images. In the prior learning stage of the high-dimensional joint solution space, the twin image domain scoring network and the original image domain scoring network are trained separately. The twin image domain scoring network is specifically trained to learn the artifact distribution characteristics formed by diffraction effect and conjugate wave interference, while the original image domain scoring network focuses on modeling the intrinsic structural characteristics of the target object. In the joint generation stage of twin and original images, a rotating iterative strategy is used to simultaneously reconstruct the twin and original images. The original and twin images are respectively subjected to regularization constraints based on prior information and data consistency constraints based on the Fresnel zone plate lensless imaging physical propagation model, and a dual-domain transformation is performed through the superposition property of Fresnel zone plate lensless imaging. The reconstruction process includes: The first step is to randomly generate a two-dimensional Gaussian noise image as the inverse starting point of the diffusion model; The second step is to input the Gaussian noise image into the original image domain scoring network and solve the stochastic differential equation of the inverse-time variance explosion of the inverse-time original image training model. The third step involves introducing prior information from the original image domain to constrain the prediction of the target information for the next time step using a predictor, and using the annealing Langevin equation as a corrector to input the predicted original image into the inner loop for correction, so as to obtain the initial reconstructed original image. The fourth step is to use the fidelity term of the physical propagation model of lensless imaging based on Fresnel zone plates to constrain data consistency, and then subtract the fidelity term from the original image in the previous step to obtain a twin image. The fifth step is to input the twin images into the twin image domain scoring network and solve the stochastic differential equation of the inverse-time variance explosion of the inverse-time original image training model. The sixth step involves introducing prior information constraints from the twin image domain. The predictor forecasts the target information for the next time step and inputs the predicted twin image into the internal loop of the annealing Langevin corrector for correction, in order to obtain the optimized twin image. Step 7: Subtract the fidelity value from the twin image from the previous step to obtain the original image, and use it as input again; Step 8: Repeat steps 2 through 7 a preset number of times, then end the loop and use the original image output after the last loop as the reconstructed image that is similar to the target object.

2. The lensless imaging method for jointly generating a twin image and a original image as described in claim 1, characterized in that, The prior learning process for the original image domain scoring network and the twin image domain scoring network of the high-dimensional joint solution space includes: Construct original image and twin image datasets separately: Color three-channel dataset Magnification factor is Pipe imaging to construct raw image dataset The original image after system scaling Forward propagation is performed with transfer functions of phases 0 and 0.5π respectively to generate corresponding coded images. Subsequently, backward propagation is performed on the coded images of the two phases respectively, and the results are superimposed to obtain a two-phase backward propagation reconstructed image. The difference between the backward propagation reconstructed image and the original image is calculated to construct a twin image dataset. ; (1) In the formula, Represents a color three-channel dataset. Represents the original image dataset, Represents a twin image dataset, This indicates that a pinhole camera is being performed on the input image. This represents the forward propagation operator for lensless imaging of Fresnel zone plates. This represents the backpropagation operator for lensless imaging of Fresnel zone plates; By training a score network that varies over time and Estimate separately Gradient of all logarithmic data distributions and This process can be modeled as solving the core objective function within a scoring-based stochastic differential equation framework: (2) In the formula, These represent the optimal parameters for training the original image domain neural network. This represents the training parameters of the original image domain neural network. Expressing expectations, Represents a positive weighting function. Represents the training samples of the original image domain. by A Gaussian perturbation kernel centered on the core. This represents the continuous-time correlation fractional function of the original image domain network. for The gradient of the logarithmic data distribution. The optimal parameters for training a twin-image domain neural network are represented. This represents the training parameters of the twin-image domain neural network. The training samples represent the twin image domain. by A Gaussian perturbation kernel centered on the core. This represents the continuous-time correlation fractional function of the twin image domain network. for The gradient of the logarithmic data distribution.

3. The lensless imaging method for jointly generating a twin image and a original image as described in claim 1, characterized in that, The second to seventh steps of the joint generation stage of the twin image and the original image can be described using the following expression: In the formula, Indicates the first A primitive image, far from the length of a walk. This represents an external index with a total time step of . This represents a regularization term based on prior information from the original image domain. Indicates the first A pair of twins, one long and one short, walking together. For data fidelity items, The regularization term is based on prior information from the twin image domain. Formulas (3b) and (3d) correspond to the data consistency term.

4. The lensless imaging method for jointly generating a twin image and a original image as described in claim 3, characterized in that, The regularization term in formula (3a) based on prior information of the original image domain This is achieved through a predictive corrector based on a diffusion model, and its expression is as follows: (4) In the formula, Indicates the first The original image predicted from the distance of the walk, Indicates the total time step as External indexes, Indicates the first A primitive image, far from the length of a walk. Indicates the first The noise intensity during a long walk express The square of, Indicates the first The noise intensity during a long walk express The square of, Indicates Gaussian noise. Indicates in As the input of the inner loop, the first A primitive image that is far removed from a long walk. Indicates in As the input of the inner loop, the first A primitive image that is far removed from a long walk. Indicates the first A corrective stride length for long walks. The total step size is The internal iterative index.

5. The lensless imaging method for jointly generating a twin image and a original image as described in claim 3, characterized in that, To further improve the underlying physical consistency of the reconstructed image, equations (3b) and (3d) introduce fidelity terms using a lensless imaging physical propagation model based on Fresnel zone plates. Its expression is: (5) In the formula, This is the result of backpropagation of orthogonal phases. For the acquired coded image intensity information, Let be the transfer function in the physical propagation process in the frequency domain. Indicates Fourier transform, Indicates the inverse Fourier transform. Indicates the number of phases.

6. The lensless imaging method for jointly generating a twin image and a original image as described in claim 3, characterized in that, The regularization term based on prior information from the twin image domain in formula (3c) This is achieved through a predictive corrector based on a diffusion model, and its expression is as follows: (6) In the formula, Indicates the first A predicted twin image of a walking distance, Indicates the total time step as External indexes, Indicates the first A pair of twins, one long and one short, walking together. Indicates the first The noise intensity during a long walk express The square of, Indicates the first The noise intensity during a long walk express The square of, Indicates Gaussian noise. Indicates in As the input of the inner loop, the first A pair of twins who have been walking for a long time. Indicates in As the input of the inner loop, the first A pair of twins who have been walking for a long time. Indicates the first A corrective stride length for long walks. The total step size is The internal iterative index.

7. The lensless imaging method for jointly generating a twin image and a original image as described in claim 3, characterized in that, Fidelity item With the A pair of twins walking together Difference to obtain the next time step The original image This image serves as the input for the next iteration. The loop continues until the iteration ends, and the original image output from the last iteration is used as the input for the next iteration. As output.

Images