CN-121995623-A - Lens-free imaging method for combined generation of twin image and original image

CN121995623ACN 121995623 ACN121995623 ACN 121995623ACN-121995623-A

Abstract

A lens-free imaging method for combined generation of a twin image and an original image comprises a priori learning of a high-dimensional combined solution space and combined generation of the twin image and the original image, wherein a twin image domain scoring network and an original image domain scoring network are trained respectively in the priori learning stage of the high-dimensional combined solution space, the twin image domain scoring network is used for specially learning artifact distribution characteristics formed by diffraction effects and conjugate wave interference, the original image domain scoring network is focused on intrinsic structural characteristics of a modeling target object, and artifact removal problems in the traditional Fresnel zone plate lens-free imaging are converted into independent generation problems of the original image and the twin image by constructing the high-dimensional combined solution space formed by the original image and the twin image, so that mutual interference in an iterative optimization process is effectively avoided.

Inventors

WAN WENBO
YU QI
YU TIANSHUI
HUANG MINGCHUN
NIE ZONGJUN
LIU QIEGEN

Assignees

南昌大学
南昌市一境信息技术有限公司

Dates

Publication Date: 20260508
Application Date: 20260129

Claims (7)

1. A lens-free imaging method for combined generation of a twin image and an original image is characterized by comprising prior learning of a high-dimensional combined solution space and combined generation of the twin image and the original image, In the prior learning stage of the high-dimensional joint solution space, training a twin image domain scoring network and an original image domain scoring network respectively, wherein the twin image domain scoring network is used for specially learning artifact distribution characteristics formed by diffraction effects and conjugate wave interference, and the original image domain scoring network is focused on the intrinsic structural characteristics of a modeling target object; In the joint generation stage of the twin image and the original image, the twin image and the original image are synchronously reconstructed by utilizing a rotation iteration strategy, wherein the original image and the twin image respectively pass through regularization constraint of prior information and data consistency constraint based on a Fresnel zone plate lens-free imaging physical propagation model, and double-domain conversion is carried out through superposition of the Fresnel zone plate lens-free imaging, and the reconstruction process comprises the following steps: firstly, randomly generating a two-dimensional Gaussian noise image serving as a reverse starting point of a diffusion model; Secondly, inputting the Gaussian noise image into an original image domain scoring network, and solving a random differential equation of inverse time variance explosion of an inverse time original image training model; Thirdly, introducing prior information constraint of an original image domain, predicting target information of the next time step through a predictor, and inputting a predicted original image into an internal circulation for correction by using an annealing Lanshould equation as an corrector so as to obtain a primary reconstructed original image; Fourthly, performing data consistency constraint by using a fidelity item based on a Fresnel zone plate lens-free imaging physical propagation model, and differencing the fidelity item and the original image in the last step to obtain a twin image; Fifthly, inputting the twin image into a twin image domain score network, and solving a random differential equation of inverse time variance explosion of an inverse time original image training model; Step six, introducing a twin image domain prior information constraint, predicting target information of the next time step through a predictor, and inputting the predicted twin image into the internal circulation of the annealing Langmuir corrector to correct so as to obtain an optimized twin image; Seventhly, performing difference between the fidelity item and the twin image in the previous step to obtain an original image, and re-using the original image as input; and eighth, repeating the second step to the seventh step for preset times, ending the cycle, and taking the original image output after the last cycle as a reconstructed image with similar target objects.
2. A lens-free imaging method for combined generation of a twin image and an original image as claimed in claim 1, wherein the prior learned original image domain scoring network and twin image domain scoring network of the high-dimensional combined solution space comprises: respectively constructing original image and twin image data set Undergo magnification of Construction of raw image data sets by pinhole imaging Original image after system scaling Forward propagating with transfer functions with phases of 0 and 0.5 pi respectively to generate corresponding coded images, then performing backward propagation on the coded images with two groups of phases respectively, superposing the results to obtain a double-phase backward propagation reconstructed image, and differencing the backward propagation reconstructed image and the original image to construct a twin image data set ; (1) In the formula, A color three-channel dataset is represented, Representing the data set of the original image, A twin image dataset is represented and, Representing the imaging of a small hole in an input image, A forward propagation operator representing fresnel zone plate lens-free imaging, A back propagation operator representing fresnel zone plate lens-less imaging; Training a time-varying score network And Separate estimation Gradients of all logarithmic data distributions of (a) And This process can be modeled as solving a core objective function in a scoring-based stochastic differential equation framework: (2) in the formula, Representing optimal parameters for training of the original image domain neural network, Representing training parameters of the original image domain neural network, It is indicated that the desire is to be met, Representing the positive weight function of the vehicle, A training sample representing the original image field, To be used for Is the gaussian disturbance kernel of the center, Representing an original image domain network continuous time correlation score function, Is that Is a logarithmic data distribution gradient of (c), Representing optimal parameters for twin domain neural network training, Representing training parameters of the twin domain neural network, A training sample representing a twin image domain, To be used for Is the gaussian disturbance kernel of the center, Representing a continuous time dependent score function of the twin domain network, Is that Is a logarithmic data distribution gradient of (c).
3. A lens-free imaging method for combined generation of a twin image and an original image as claimed in claim 1, wherein the second to seventh steps of the combined generation stage of the twin image and the original image can be described by the following expression: in the formula, Represent the first The original image of the individual discrete steps is taken, Representing an external index of the total time step, Representing regularized terms based on original image domain prior information, Represent the first Twin images of the individual discrete steps are taken, In the case of a data-fidelity item, Representing regularization terms based on twin domain prior information, equations (3 b) and (3 d) correspond to data consistency terms.
4. A lens-free imaging method for combined generation of a twin image and an original image as claimed in claim 3, wherein regularization term based on prior information of original image field in formula (3 a) The predictive corrector realized by the diffusion model has the following expression: (4) in the formula, Represent the first The predicted raw image of a discrete step size, Representing a total time step as Is used for the external indexing of (a), Represent the first The original image of the individual discrete steps is taken, Represent the first The noise strength at discrete steps, Representation of Is defined by the square of (a), Represent the first The noise strength at discrete steps, Representation of Is defined by the square of (a), Representing the gaussian noise and the noise level of the signal, Is shown in As the first under the input of the internal circulation The original image at the time of the discrete steps, Is shown in As the first under the input of the internal circulation The original image at the time of the discrete steps, Represent the first Correction step sizes at discrete step sizes, Is the total step length of Is included.
5. A lens-free imaging method as defined in claim 3, wherein to further enhance the underlying physical consistency of the reconstructed image, equations (3 b) and (3 d) introduce fidelity terms using a fresnel zone plate-based lens-free imaging physical propagation model The expression is: (5) in the formula, As a result of the quadrature-phase back-propagation, For the acquired intensity information of the encoded image, As a transfer function during physical propagation in the frequency domain, Representing the fourier transform of the signal, Representing the inverse fourier transform of the signal, Representing the number of phases.
6. A lens-free imaging method for combined generation of a twin image and an original image as claimed in claim 3, wherein the regularization term based on the prior information of the twin image field in the formula (3 c) The predictive corrector realized by the diffusion model has the following expression: (6) in the formula, Represent the first A predicted twin image of a discrete step size, Representing a total time step as Is used for the external indexing of (a), Represent the first Twin images of the individual discrete steps are taken, Represent the first The noise strength at discrete steps, Representation of Is defined by the square of (a), Represent the first The noise strength at discrete steps, Representation of Is defined by the square of (a), Representing the gaussian noise and the noise level of the signal, Is shown in As the first under the input of the internal circulation Twin images at discrete steps are obtained, Is shown in As the first under the input of the internal circulation Twin images at discrete steps are obtained, Represent the first Correction step sizes at discrete step sizes, Is the total step length of Is included.
7. A lens-free imaging method for combined generation of a twin image and an original image as claimed in claim 3, wherein the fidelity term is And the first Twin image of discrete steps Taking the difference to obtain the next time step Is the original image of (a) And as the input of a new iteration, repeatedly executing the iterative strategy, ending the loop, and outputting the original image after the last loop As an output.

Description

Lens-free imaging method for combined generation of twin image and original image Technical Field The invention relates to the technical field of computational optical imaging, in particular to a Fresnel zone plate lens-free imaging technology. Background The lens-free imaging technology utilizes the encoder to replace the traditional lens, has the advantages of light system, low cost and the like, and has wide application prospect in a plurality of key fields such as biomedicine, mobile imaging and the like. However, since the image sensor is only sensitive to light intensity, phase information is lost during acquisition, resulting in the appearance of twin images in the reconstructed image. Twin images are an inherent problem in-line holography, and are represented by the superposition of conjugate images of objects on the original image, which significantly reduces the imaging quality. Traditional reconstruction algorithms can remove artifacts to a certain extent, but image details are obviously damaged, and the wide application of lens-free imaging is limited. Therefore, developing a lens-free imaging method with high quality and no artifacts has become a key problem to be solved. The prior art has the following technical problems: (1) The hardware technical scheme is that a traditional encoding mask based on a Fresnel zone plate mostly adopts a single-phase zone plate or a multi-phase combination of non-orthogonal phases, so that twin image artifacts are difficult to effectively inhibit from a hardware level, follow-up algorithm optimization is needed, and imaging quality is limited. (2) The algorithm technical scheme is that a traditional reconstruction method, such as a back propagation algorithm, a compressed sensing algorithm and a recently developed deep learning scheme, generally takes a twin image and an original image as a whole to inhibit or remove, and the separable characteristics of the twin image and the original image in an information layer can not be fully mined, so that artifact removal is incomplete, image details are easy to lose, and high-quality reconstruction is difficult to realize. In conclusion, the current lens-free imaging technology is focused on forcedly removing twin image artifacts, the removing effect is limited, the imaging quality and detail retention are difficult to be considered, and the gap between the requirements of high-precision lens-free imaging in the fields of biomedical imaging, industrial detection and the like is large. Disclosure of Invention According to the lens-free imaging method for combined generation of the twin image and the original image, which is provided by the invention, the artifact removal problem in the traditional Fresnel zone plate lens-free imaging is converted into the independent generation problem of the original image and the twin image by constructing the high-dimensional combined solution space formed by the original image and the twin image, so that mutual interference in the iterative optimization process is effectively avoided. The invention adopts the encoder with the quadrature phase to construct the lens-free imaging system, thereby further eliminating the interference of twin images. And a double diffusion model is constructed aiming at Gao Weijie space in prior learning of high-dimensional joint solution space, and prior distribution information of twin image domain data and original image domain data is respectively learned. And in the joint generation stage of the twin image and the original image, synchronously reconstructing the original image and the twin image by adopting a rotation iteration strategy of the twin image and the original image. The twin image priori information is used for separating corresponding twin image interference signals, and the original image priori information guides the reconstruction result to approach to the structure and texture of the real scene. And a data consistency process based on a Fresnel zone plate lens-free imaging propagation model is adopted for fidelity. Through an alternate iterative process based on the Fresnel zone plate propagation model, the artifacts are gradually transferred into the twin images, and finally the original image without the artifacts is obtained. The invention provides a lens-free imaging method scheme for jointly generating a twin image and an original image, which comprises a priori learning of a high-dimensional joint solution space and jointly generating the twin image and the original image, In the prior learning stage of the high-dimensional joint solution space, a twin image domain scoring network and an original image domain scoring network are trained respectively, wherein the twin image domain scoring network is specially used for learning artifact distribution characteristics formed by diffraction effects and conjugate wave interference, and the original image domain scoring network is focused on the intrinsic structural characteristics of a