Digital holographic reconstruction optimization model and method based on cycle consistency adaptive optimization

Abstract

The invention relates to the technical field of holographic reconstruction, in particular to a digital holographic reconstruction optimization model and method based on cycle consistency adaptive optimization. The invention provides a digital holographic reconstruction optimization model which comprises a first optimization network, a first reconstruction unit and a second optimization network. The calibration image PSF is optimized by the first optimization network, the hologram Ori is optimized by the second optimization network, the optimized calibration image PSF1 output by the first optimization network and the optimized hologram Or1 output by the second optimization network are combined by the first reconstruction unit for reconstruction, and a holographic reconstruction image R is obtained. According to the method, the PSF is optimized through the neural network, the features are extracted, meanwhile, the features of the incoherent hologram are extracted through another optimization network, reconstruction is carried out in combination with the features of the optimized PSF and the features of the incoherent hologram, and therefore a high-quality reconstructed image is achieved.

Description

Technical Field

[0001] The present invention relates to the field of holographic reconstruction technology, and in particular to a digital holographic reconstruction optimization model and method based on cycle consistency adaptive optimization. Background Art

[0002] As a core characteristic function of an optical system, the point spread function (PSF) plays a key role in digital holography. It not only fully describes the system's response to a point light source, reflecting the optical system's diffraction behavior and imaging quality, but also serves as a crucial bridge between object space and image space. Optimizing the PSF is crucial for improving the quality of digital holographic reconstruction.

[0003] In the development of PSF applications, coded aperture correlation holography (COACH) is an important milestone. In 2011, Choi et al. proposed the turbid lens imaging technology, which increased the effective NA (numerical aperture) of the system by scattering high-frequency waves, and achieved super-resolution reproduction of random patterns. In 2016, Anand and Rosen expanded the secondary phase aperture of Fresnel correlation holography to a chaotic phase aperture, creating a new method that combines coded phase aperture imaging with incoherent digital holography. In 2017, Rosen et al. developed the interference-free COACH (I-COACH) technology, which extracts three-dimensional object information by generating speckle intensity through a single wavefront, significantly simplifying the operation process. Subsequently, researchers made a series of improvements to I-COACH, including an improved COACH system for 3D imaging and an annular coded aperture endoscopic imaging technology.

[0004] As can be seen, existing technologies utilize engineered PSFs to significantly improve imaging performance, including spatial resolution and volumetric imaging throughput. However, these methods require optimization of the PSF acquisition method and improvements to the laboratory environment. PSFs obtained in existing laboratory environments are difficult to optimize, and these PSFs may contain aberrations, distortion, and noise, which can affect reconstruction quality.

[0005] In recent years, significant progress has been made in the field of digital holographic reconstruction. End-to-end deep learning-based methods have enabled off-axis digital holography for imaging scattering media. Imaging networks combining Fourier transforms with CNNs have achieved breakthroughs in generalization and inference speed. Physically embedded untrained neural networks and self-supervised learning methods based on angular spectrum forward diffraction models have further advanced research in unsupervised learning and physical consistency constraints. However, existing methods generally rely on large amounts of sample data for training; even self-supervised models require extensive simulated data.

[0006] However, due to the strict requirements of experimental conditions and the differences in optical system configurations, obtaining high-quality holographic datasets remains a challenge. In addition, the inherent characteristics of digital holography, such as phase jumps, speckle noise, and phase differences, limit the direct application of traditional neural networks. Summary of the Invention

[0007] In order to overcome the problem in the above-mentioned prior art that the PSF itself defects affect the quality of holographic reconstruction, the present invention proposes a training method for a digital holographic reconstruction optimization model based on cycle consistency adaptive optimization. The PSF is optimized and features are extracted through a neural network, and the features of the incoherent hologram are extracted through another optimization network. The features of the optimized PSF and the features of the incoherent hologram are combined for reconstruction, thereby achieving high-quality reconstructed images.

[0008] The present invention proposes a training method for a digital holographic reconstruction optimization model based on cycle-consistent adaptive optimization. First, a basic model and a data set {(PSF, Pri); Ori} are constructed. PSF represents the calibration image before spatial modulation, Pri represents the test object image before spatial modulation, and Ori represents the incoherent hologram after Pri is spatially modulated.

[0009] The basic model includes: a first optimization network, a first reconstruction unit, a second optimization network, an angular spectrum propagation unit, and a second reconstruction unit; the first optimization network optimizes the calibration image PSF, and the second optimization network optimizes the hologram Ori. The first reconstruction unit reconstructs the optimized calibration image PSF1 output by the first optimization network and the optimized hologram Ori1 output by the second optimization network to obtain a holographic reconstructed image R; the holographic reconstructed image R is subjected to angular spectrum propagation to obtain a cyclic hologram Cyc, and the cyclic hologram Cyc and the optimized calibration image PSF1 are reconstructed by the second reconstruction unit to obtain a cyclic reconstructed image CR;

[0010] The second optimized network has the same structure as the first optimized network, and its parameters are randomly initialized and fixed. The basic model is trained on the dataset, and the first optimized network is updated during the training process with the goal of minimizing the loss function.

[0011] After the basic model training is completed, the connection structure of the first optimization network, the first reconstruction unit and the second optimization network is extracted to form a holographic reconstruction optimization model.

[0012] Preferably, the loss function L = λ1L f +λ2L b +λ3L c ; or, L = λ1L f +λ2L b ;L f represents the mean square error loss between the reconstructed image R and the test object image Pri; Lb represents the mean square error loss of hologram Ori and cyclic hologram Cyc, L c represents the mean square error loss between the cyclic reconstruction image CR and the test object image Pri; λ1, λ2, and λ3 all represent weight coefficients with values in the interval (0,1).

[0013] Preferably, the first optimization network and the second optimization network adopt fully connected networks.

[0014] Preferably, the first reconstruction unit and the second reconstruction unit adopt convolutional networks.

[0015] Preferably, the construction of the data set includes the following steps:

[0016] A holographic acquisition system is constructed, which modulates the light wave carrying the test object information through a second spatial modulator. The phase mask loaded on the second spatial modulator adopts a phase-constrained amplitude of a concentric ring structure, so that the output light of the second spatial modulator forms multiple focal points on the optical axis; the electronic coupler moves on the output optical axis of the second spatial modulator, and outputs holograms Ori at different focal points to form a sample by combining the calibration image PSF used to modulate the holographic acquisition system, the test object image Pri and the hologram Ori.

[0017] Preferably, the holographic acquisition system comprises: a white light source, a first polarizer, a first beam splitter, a first spatial modulator, a second polarizer, a collimating lens, a second beam splitter, a second spatial modulator, and an electronic coupler;

[0018] The light emitted by the white light source is polarized by the first polarizer and then redirected by the first beam splitter to be projected onto the first spatial modulator. The first spatial modulator is loaded with the test object image Pri. The reflected light of the first spatial modulator carries the test object information. The reflected light is polarized by the second polarizer and calibrated by the collimating lens. It is then redirected by the beam splitter and projected onto the second spatial modulator. The second spatial modulator reflects the light onto the electronic coupler, which outputs a hologram.

[0019] A digital holographic reconstruction optimization method proposed in the present invention is characterized in that a holographic reconstruction optimization model, a holographic acquisition system and a calibration image PSF are first obtained, the holographic acquisition system is calibrated with the calibration image, and then the test object image Pri is input into the holographic acquisition system. After the holographic acquisition system spatially modulates the test object image Pri, the holographic acquisition system outputs a hologram Ori through an electronic coupler CCD; the PSF, Pri and Ori are input into the holographic reconstruction optimization model to obtain a holographic reconstructed image R output by the holographic reconstruction optimization model.

[0020] Preferably, holograms Ori at different focal points are collected by a holographic system, each hologram Ori is reconstructed by a holographic reconstruction optimization model, and then a three-dimensional hologram of the test object is constructed by combining the holographic reconstructed images R corresponding to each focal point.

[0021] The present invention proposes a digital holographic reconstruction optimization system, which includes a memory and a processor. The memory stores a computer program, and the processor is connected to the memory. The processor is used to execute the computer program to implement the digital holographic reconstruction optimization method.

[0022] The present invention proposes a storage medium storing a computer program, which is used to implement the digital holographic reconstruction optimization method when executed.

[0023] The advantages of the present invention are:

[0024] (1) This paper proposes a digital holographic reconstruction optimization method that constructs a PSF adaptive optimization network and achieves dynamic optimization of the PSF through deep learning methods. This method provides a new technical path for achieving high-quality two-dimensional (axial slice) super-resolution reconstruction in multi-depth digital holography. This breakthrough not only addresses the limitations of traditional methods but also opens up new directions for the further development of digital holography technology.

[0025] (2) In the present invention, the incoherent hologram Ori is subjected to a second optimization network with fixed and randomized weights to optimize the angles of feature extraction and transformation. This performs specific feature extraction and transformation on the input hologram, allowing for better learning of the relationship between features and facilitating subsequent reconstruction. Furthermore, the second optimization network, acting as a stable nonlinear transformer, helps reduce data redundancy, provides a consistent feature representation, and enables high-quality data preprocessing.

[0026] (3) The optimized network uses a fully connected layer network, which can better learn the nonlinear relationship between features.

[0027] (4) PSF (point spread function), as the core point response function of the holographic system, determines the diffraction behavior and imaging clarity of the hologram; PSF is the key feature of the imaging system, and the quality of PSF directly determines the quality of the reconstructed image. The present invention focuses on optimizing PSF to solve the root cause of the problem of holographic reconstruction effect. PSF contains the optical characteristic information of the imaging system. Optimizing PSF is equivalent to optimizing the entire imaging system, and PSF can be used for different hologram reconstructions. This method is more in line with the nature of the physical system. The present invention is applicable to holographic reconstruction in different scenarios, which is conducive to a new breakthrough in the field of holographic reconstruction.

[0028] (5) The present invention proposes a training method for a digital holographic reconstruction optimization model based on cycle consistency adaptive optimization. First, a data set is generated according to different imaging requirements through a holographic acquisition system, and its axial response characteristics at different imaging depths are analyzed. In particular, in three-dimensional holographic imaging, corresponding focus PSFs are generated for target points at different axial positions to ensure accurate recovery of depth information. In the data generation stage, diverse PSFs are generated for different types of three-dimensional targets and different environmental interference factors to ensure the diversity and adaptability of subsequent training data, laying a data foundation for the training of deep learning models.

[0029] (6) The present invention adopts a comprehensive optimization objective using a triple loss function, covering the entire image reconstruction process, while considering both forward and reverse reconstruction quality and adding optimization constraints. In particular, the application of the cycle consistency loss function ensures that the PSF remains consistent in bidirectional transformations, reducing information loss and improving reconstruction reliability.

[0030] (7) The present invention only trains one optimization network, which reduces computational complexity, training difficulty and memory requirements, and can concentrate computing resources on the most critical parts, making the training convergence faster and more stable.

[0031] (8) In terms of network architecture design, the present invention innovatively adopts a single-input fully connected network (FCN) structure to achieve adaptive optimization of feature constraints, and proposes a new cycle consistency constraint loss function to effectively improve the reconstruction quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 Optimizing models for digital holographic reconstruction;

[0033] Figure 2 As the basic model;

[0034] Figure 3 Schematic diagram of loss function;

[0035] Figure 4 Optimizing methods for digital holographic reconstruction;

[0036] Figure 5 It is a holographic acquisition system;

[0037] Figure 6(a) shows the phase-constrained amplitude;

[0038] Figure 6(b) shows the phase mask;

[0039] Figure 7 This is a physical picture of the holographic acquisition system;

[0040] Figure 8 The trend of the marginal correlation coefficient using different loss functions in the embodiment;

[0041] Figure 9 The trend of structural similarity changes using different loss functions in the embodiment;

[0042] Figure 10 The experimental results are shown, where (a) is the incoherent hologram, (b) is the PSF, (c) is the reconstruction of the comparison model, (d)-(f) are the reconstructions of the model of the present invention after different training rounds, and (g) is the intensity distribution of the dotted lines in (c) and (f);

[0043] FIG11( a ) is an incoherent hologram according to an embodiment;

[0044] Figure 11(b) is the reconstructed hologram of Figure 11(a) output by the comparison model;

[0045] FIG11( c ) is the reconstructed hologram of FIG11( a ) output by the reconstructed holographic optimization model proposed in the present invention. DETAILED DESCRIPTION

[0046] The following will provide a clear and complete description of the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0047] Reference Figure 1 This embodiment provides a digital holographic reconstruction optimization model, including: a first optimization network, a first reconstruction unit, and a second optimization network; the first optimization network optimizes the calibration image before spatial modulation, namely the point spread function PSF, and the second optimization network optimizes the incoherent hologram Ori of the test object image after spatial modulation; the first reconstruction unit reconstructs the optimized calibration image PSF1 output by the first optimization network and the optimized hologram Ori1 output by the second optimization network to obtain a holographic reconstructed image.

[0048] Reference Figure 2 、 Figure 3 This embodiment provides a training method for a digital holographic reconstruction optimization model based on cycle consistency adaptive optimization, comprising the following steps:

[0049] St1. Construct a basic model and data set {(PSF, Pri); Ori}; PSF represents the calibration image before spatial modulation, i.e., point spread function, which can be a binary pinhole image; Pri represents the test object image before spatial modulation; Ori represents the hologram after Pri is spatially modulated;

[0050] The basic model includes: a first optimization network, a first reconstruction unit, a second optimization network, an angular spectrum propagation unit and a second reconstruction unit;

[0051] The first optimization network optimizes the calibration image PSF, and the second optimization network optimizes the hologram Ori after the test object image is spatially modulated. The first reconstruction unit combines the optimized calibration image PSF1 output by the first optimization network and the optimized hologram Ori1 output by the second optimization network to reconstruct a holographic reconstructed image R. The holographic reconstructed image R is propagated through the angular spectrum to obtain a cyclic hologram Cyc. The cyclic hologram Cyc and the optimized calibration image PSF1 are reconstructed by the second reconstruction unit to obtain a cyclic reconstructed image CR.

[0052] St2. Let the basic model be trained on the dataset {(PSF, Pri); Ori}. During the training process, the first optimization network is updated with the goal of minimizing the loss function. The loss function is calculated as follows:

[0053] L=λ1L f +λ2L b +λ3L c

[0054] Among them, L f represents the mean square error loss between the reconstructed image R and the test object image Pri; L b represents the mean square error loss of hologram Ori and cyclic hologram Cyc, L c represents the mean square error loss between the cyclic reconstruction image CR and the test object image Pri; λ1, λ2, and λ3 all represent weight coefficients taking values in the interval (0, 1). In the embodiment, λ1 = λ2 = 0.3; λ3 = 0.2 can be specifically set.

[0055] In specific implementation, during the construction of the data set {(PSF, Pri); Ori}, the spatial modulator that modulates the test object image Pri in the holographic acquisition system that generates Ori can be loaded with a multi-focal length mask, thereby obtaining holograms Ori corresponding to different focal lengths, enriching the data set, and allowing the first optimization network to adapt to the PSF requirements of holographic acquisition systems with different focal lengths.

[0056] Reference Figure 5 6(a), a holographic acquisition system proposed in the present invention includes: a white light source, a first polarizer P1, a first beam splitter BS1, a first spatial modulator SLM1, a second polarizer P2, a collimating lens CL, a second beam splitter BS2, a second spatial modulator SLM2 and an electronic coupler CCD.

[0057] The white light source uses a xenon lamp. Light emitted by the xenon lamp is polarized by the first polarizer P1, redirected by the first beam splitter BS1, and projected onto the first spatial modulator SLM1, which is loaded with the test object image Pri. The reflected light from the first spatial modulator SLM1 carries the test object information. This reflected light is polarized by the second polarizer P2, collimated by the collimating lens CL, and then redirected by the beam splitter BS2 to the second spatial modulator SLM2, which is loaded with a phase mask. SLM2 reflects the light onto the electronic coupler CCD, which outputs the hologram.

[0058] The spatial modulator SLM can only modulate polarized light. In this optical path, P1 cooperates with SLM1, and P2 cooperates with SLM2 to ensure the smooth implementation of spatial modulation.

[0059] In this system, the phase mask loaded on the second spatial modulator SLM2 adopts a phase-constrained amplitude with a concentric ring structure so that the reflected light of the second spatial modulator SLM2 forms multiple focal points on the optical axis; the mobile electronic coupler CCD can collect the hologram Ori at each focus, thereby constructing samples {(PSF, Pri); Ori} corresponding to different focal points.

[0060] Reference Figure 4 , a digital holographic reconstruction optimization method proposed in this embodiment includes the following steps:

[0061] S1. Obtain the holographic acquisition system and the calibration image PSF. After calibrating the holographic acquisition system with the calibration image, input the test object image Pri into the holographic acquisition system. After the holographic acquisition system spatially modulates the test object image Pri, it outputs the hologram Ori through the electronic coupler CCD.

[0062] S2. Input PSF, Pri and Ori into the holographic reconstruction optimization model to obtain the holographic reconstructed image R output by the holographic reconstruction optimization model.

[0063] When the second spatial modulator SLM2 in the holographic acquisition system adopts the phase mask with phase-constrained amplitude as shown in FIG6( a ), each time the electronic coupler CCD switches the focus, step S1 needs to be re-executed to obtain the hologram Ori corresponding to the different focus.

[0064] In this embodiment, the mobile CCD collects the hologram Ori at each focus, and the Ori at each focus is optimized using the holographic reconstruction optimization model to obtain the corresponding holographic reconstructed image R; the holographic reconstructed images R corresponding to each focus are collected to construct a three-dimensional holographic image of the test object.

[0065] The above holographic reconstruction optimization model is verified below in conjunction with specific embodiments.

[0066] In this embodiment, the Figure 5 The holographic acquisition system shown generates a data set, and the actual picture is as follows Figure 7 As shown, the mask loaded by the second spatial modulator SLM2 satisfies the constraints shown in FIG6( a ), and the mask parameters can be iterated by the GSA algorithm.

[0067] In this embodiment, the GSA algorithm is iterated 100 times under the constraints shown in FIG6(a) to obtain the mask shown in FIG6(b). Its basic parameters are as follows:

[0068] Size: 512×512 pixels;

[0069] Pixel size: 1 micron / pixel;

[0070] Ring parameters:

[0071] Quantity: 10 concentric rings;

[0072] Radius settings: [20,40,60,80,100,120,140,160,180,200] pixels;

[0073] Ring thickness: uniformly set to 8 pixels;

[0074] Adjacent ring spacing: 12 pixels;

[0075] Scattering parameters: gradually increase from the inner ring to the outer ring, respectively [0.05, 0.06, 0.07, 0.08, 0.09, 0.1, 0.11, 0.12, 0.13, 0.14];

[0076] Initial phase: randomly distributed in the interval (0-2π).

[0077] In the mask, each ring has its own scattering degree, so when the incident light is reflected by the mask and focused, a small amount of light will be scattered, constructing a self-interference digital hologram.

[0078] In this implementation, the dataset is divided into a training set and a test set. The model is trained on the training set and then validated on the test set after training.

[0079] In this embodiment, a resolution plate is used as a test object, and both the first reconstruction unit and the second reconstruction unit adopt convolutional neural networks.

[0080] In this embodiment, the model training process is first demonstrated.

[0081] In this embodiment, for the training method of the digital holographic reconstruction optimization model based on cycle consistency adaptive optimization, four loss functions are used for model training. The four loss functions are:

[0082] L1=0.3L f +0.3L b +0.2L c

[0083] L2=0.3L b +0.2L c

[0084] L3=0.3L f +0.2L c

[0085] L4=0.3L f +0.3L b

[0086] The model training processes involved in the above four loss functions are as follows: Figure 8 、 Figure 9 As shown. It can be seen that in terms of convergence, the edge correlation coefficient (ECC) tends to be stable after 1000 iterations, while the structural similarity (SSIM) continues to improve until 2000 iterations, reflecting the optimization characteristics of the model for different frequency components. f ), reverse consistency loss (L b ) and cycle consistency constraints (L c ) loss function L1 training method achieved the best performance, combined with the forward reconstruction loss (L f ) and reverse consistency loss (L b ) The training effect of loss function L4 is slightly worse than that of loss function L1, and the training effects of loss functions L2 and L3 are completely incomparable to those of loss functions L1 and L4.

[0087] As can be seen, the key to achieving excellent model performance in this invention is the synergistic effect of the forward reconstruction loss (Lf), the backward consistency loss (Lb), and the cycle consistency constraint (Lc). The model structure proposed in this invention, using the proposed loss function, achieves high performance without relying on a large-scale training set, opening up a new approach for digital holographic reconstruction under data-scarce conditions.

[0088] In this embodiment, the first optimization network, the first reconstruction unit and the second optimization network are extracted from the basic model trained using the loss function L1 to form the Figure 1 The holographic reconstruction optimization model is shown.

[0089] In this embodiment, the first optimization network and the first reconstruction unit constitute a comparison model. In the comparison model, the first optimization network optimizes the PSF and outputs PSF1, and the first reconstruction unit combines PSF1 and the test object image Ori to perform image reconstruction to output a reconstructed hologram R'.

[0090] Figure 10 The experimental results are shown in Figure 1. The resolution plate is used as the test object and the binary pinhole image is used as the PSF.

[0091] Sub-image (a) is the test object image Ori, specifically an incoherent hologram, and sub-image (b) is the PSF;

[0092] Sub-figure (c) shows the reconstructed hologram R' output by the comparison model combined with Ori and PSF; and the first optimized network in the comparison model is extracted from the base model trained for 2300 rounds;

[0093] Sub-figure (d) is the reconstructed hologram R output by the holographic reconstruction optimization model combined with Ori and PSF after 200 rounds of training; Sub-figure (e) is the reconstructed hologram R output by the holographic reconstruction optimization model combined with Ori and PSF after 500 rounds of training; Sub-figure (f) is the reconstructed hologram R output by the holographic reconstruction optimization model combined with Ori and PSF after 2300 rounds of training;

[0094] The red curve in sub-image (g) is the pixel intensity change of the red dotted line in sub-image (c), and the blue curve is the pixel intensity change of the blue dotted line in sub-image (f).

[0095] The holographic reconstruction optimization model trained for 200 rounds refers to a holographic reconstruction optimization model consisting of the first optimization network, the first reconstruction unit, and the second optimization network extracted from the basic model trained for 200 rounds; and so on.

[0096] from Figure 10 It can be seen that when the holographic reconstruction optimization model was trained to 2300 rounds, it had achieved an excellent reconstruction effect. It can be seen from sub-graph (g) that the reconstructed hologram obtained by optimizing the hologram Ori through the second optimization network and then combining it with PSF1 is better than that without optimizing Ori, and the strength of the reconstructed hologram is significantly increased. This is because, although the present invention does not train the second optimization network, the second optimization network uses the same network structure as the first optimization network. Even if the second optimization network uses random initialization parameters, the second optimization network can also map the image features of the hologram Ori and the features of the optimized PSF1 to the same space, thereby facilitating the first reconstruction unit to better step the features of PSF and Ori and perform fusion reconstruction.

[0097] In this example, after 2300 rounds of training, the incoherent hologram shown in Figure 11(a) was reconstructed using the comparison model and the holographic reconstruction optimization model. The reconstructed hologram output by the comparison model is shown in Figure 11(b), and the reconstructed hologram output by the holographic reconstruction optimization model proposed in the present invention is shown in Figure 11(c). It can be seen that the reconstruction effect of the model proposed in the present invention is better, and it can more completely preserve the original image information.

[0098] Of course, it will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, but also encompasses the same or similar structures that can be implemented in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be considered in all respects as illustrative and non-restrictive, and the scope of the invention is defined by the appended claims, not the foregoing description, and it is intended that all variations that fall within the meaning and range of equivalents of the claims be encompassed within the present invention. Any reference signs in the claims should not be construed as limiting the claim to which they relate.

[0099] In addition, it should be understood that although this specification is described in terms of implementation methods, not every implementation method contains only one independent technical solution. This narrative method of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment can also be appropriately combined to form other implementation methods that can be understood by those skilled in the art.

[0100] The technology, shape, and structure not described in detail in the present invention are all well-known technologies.

Claims

1. A training method for a digital holographic reconstruction optimization model based on cycle consistency adaptive optimization, characterized in that: First, the basic model and data set {(PSF, Pri); Ori} are constructed; PSF represents the calibration image before spatial modulation, Pri represents the test object image before spatial modulation; Ori represents the incoherent hologram after Pri is spatially modulated; The basic model includes: a first optimization network, a first reconstruction unit, a second optimization network, an angular spectrum propagation unit, and a second reconstruction unit; the first optimization network optimizes the calibration image PSF, and the second optimization network optimizes the hologram Ori. The first reconstruction unit reconstructs the optimized calibration image PSF1 output by the first optimization network and the optimized hologram Ori1 output by the second optimization network to obtain a holographic reconstructed image R; the holographic reconstructed image R is subjected to angular spectrum propagation to obtain a cyclic hologram Cyc, and the cyclic hologram Cyc and the optimized calibration image PSF1 are reconstructed by the second reconstruction unit to obtain a cyclic reconstructed image CR; The second optimized network has the same structure as the first optimized network, and its parameters are randomly initialized and fixed. The basic model is trained on the dataset, and the first optimized network is updated during the training process with the goal of minimizing the loss function. After the basic model training is completed, the connection structure of the first optimization network, the first reconstruction unit and the second optimization network is extracted to form a holographic reconstruction optimization model.

2. The training method for a digital holographic reconstruction optimization model based on cycle consistency adaptive optimization according to claim 1, characterized in that: Loss function L = λ1L f +λ2L b +λ3L c ; or, L = λ1L f +λ2L b ;L f represents the mean square error loss between the reconstructed image R and the test object image Pri; L b represents the mean square error loss of hologram Ori and cyclic hologram Cyc, L c represents the mean square error loss between the cyclic reconstruction image CR and the test object image Pri; λ1, λ2, and λ3 all represent weight coefficients with values in the interval (0,1).

3. The training method for a digital holographic reconstruction optimization model based on cycle consistency adaptive optimization according to claim 1, characterized in that: The first optimization network and the second optimization network adopt fully connected networks.

4. The training method for a digital holographic reconstruction optimization model based on cycle consistency adaptive optimization according to any one of claims 1 to 3, characterized in that: The first reconstruction unit and the second reconstruction unit adopt convolutional networks.

5. The training method of the digital holographic reconstruction optimization model based on cycle consistency adaptive optimization according to claim 1, characterized in that: The construction of the dataset includes the following steps: A holographic acquisition system is constructed, which modulates the light wave carrying the test object information through a second spatial modulator. The phase mask loaded on the second spatial modulator adopts a phase-constrained amplitude of a concentric ring structure, so that the output light of the second spatial modulator forms multiple focal points on the optical axis; the electronic coupler moves on the output optical axis of the second spatial modulator, and outputs holograms Ori at different focal points to form a sample by combining the calibration image PSF used to modulate the holographic acquisition system, the test object image Pri and the hologram Ori.

6. The training method for a digital holographic reconstruction optimization model based on cycle consistency adaptive optimization according to claim 5, characterized in that: The holographic acquisition system includes: a white light source, a first polarizer, a first beam splitter, a first spatial modulator, a second polarizer, a collimating lens, a second beam splitter, a second spatial modulator and an electronic coupler; The light emitted by the white light source is polarized by the first polarizer and then redirected by the first beam splitter to be projected onto the first spatial modulator. The first spatial modulator is loaded with the test object image Pri. The reflected light of the first spatial modulator carries the test object information. The reflected light is polarized by the second polarizer and calibrated by the collimating lens. It is then redirected by the beam splitter and projected onto the second spatial modulator. The second spatial modulator reflects the light onto the electronic coupler, which outputs a hologram.

7. A digital holographic reconstruction optimization method using the training method of the digital holographic reconstruction optimization model based on cycle consistency adaptive optimization according to any one of claims 1 to 6, characterized in that: First, the holographic reconstruction optimization model, holographic acquisition system and calibration image PSF are obtained. After the calibration image is used to calibrate the holographic acquisition system, the test object image Pri is input into the holographic acquisition system. After the holographic acquisition system spatially modulates the test object image Pri, the holographic acquisition system outputs the hologram Ori through the electronic coupler CCD; PSF, Pri and Ori are input into the holographic reconstruction optimization model to obtain the holographic reconstructed image R output by the holographic reconstruction optimization model.

8. The digital holographic reconstruction optimization method according to claim 7, wherein: Holograms Ori at different focal points are collected through the holographic system, and each hologram Ori is reconstructed through the holographic reconstruction optimization model. Then, a three-dimensional hologram of the test object is constructed by combining the holographic reconstructed image R corresponding to each focal point.

9. A digital holographic reconstruction optimization system, characterized in that: It includes a memory and a processor, the memory stores a computer program, the processor is connected to the memory, and the processor is used to execute the computer program to implement the digital holographic reconstruction optimization method as described in claim 7 or 8.

10. A storage medium, characterized in that: A computer program is stored, and when the computer program is executed, it is used to implement the digital holographic reconstruction optimization method according to claim 7 or 8.

Images