Holographic reconstruction method and system for frequency-sensing physical guidance neural network

Abstract

The invention discloses a holographic reconstruction method and system for a frequency-sensing physical guidance neural network, and relates to the technical field of holographic reconstruction of pictures, and the method comprises the steps: collecting a label data pair set; performing frequency sensing data driven network training by adopting the training set to obtain a pre-trained network model; building a holographic imaging system of an actual environment according to the imaging light path, carrying out collimation operation on the holographic imaging system, and measuring a real object by using the built holographic imaging system to obtain a hologram; and inputting the obtained hologram into the trained network model, adding an imaged forward model on the basis of the network model to carry out iterative optimization, and finally outputting a holographic reconstruction image after enhancement optimization. The method has universality and imaging quality at the same time, the physical forward model and Fourier domain learning are incorporated into network reasoning, the discomfort of digital holographic reconstruction is reduced to a great extent, and high-throughput 3D microscopic imaging is assisted.

Description

Technical Field

[0001] The present invention relates to the technical field of image holographic reconstruction, and in particular to a holographic reconstruction method and system of a frequency-aware physically guided neural network. Background Art

[0002] Lensless digital holographic imaging technology intentionally bypasses expensive and complex optical component systems. As a low-cost and compact alternative, it can provide another observation dimension in addition to amplitude: phase. This capability is widely used in microscopic quantitative analysis such as biological samples, nanomanufacturing, and optical detection and metrology. However, since existing photodetectors can only record light intensity information, phase information is lost. At the same time, solving phase information from two-dimensional intensity information is an ill-posed problem. There are holographic imaging systems that provide prior information for effective constraints based on multi-frame encoding (see reference patent: Cao Liangcai, Zhang Wenhui, Zhang Hao, etc. Quasi-common optical path multi-angle illumination digital holographic imaging system and method [P]. Beijing: CN107741691B, 2019-07-26.), but additional scanning devices are usually required, which makes the optical path system more complicated. Therefore, digital holographic systems and solutions based on single-frame imaging began to be designed, of which there are two main categories: methods based on compressed sensing and methods based on deep learning (see references: Zhang W, Cao L, Brady DJ, et al. Twin-image-free holography: a compressive sensing approach [J]. Physical Review Letters, 2018, 121 (9): 093902. Reference patent: Di Jianglei, Wang Kaiqiang, Li Ying, et al. A holographic reconstruction algorithm based on deep learning [P]. Shaanxi Province: CN109459923A, 2019-03-12.). So far, the two existing holographic reconstruction methods always need to make a trade-off between universality and imaging quality.

[0003] To solve the above technical problems, by exploring the differences between Fourier domain networks and model-driven neural networks in frequency recovery mechanisms, the present invention proposes a holographic reconstruction system and method that combines frequency perception learning and physical guidance. This method can be used for high-resolution holographic reconstruction, and its idea can select and construct appropriate deep learning models based on different imaging tasks. Summary of the invention

[0004] In view of the above-mentioned problems, the present invention is proposed.

[0005] Therefore, the problem to be solved by the present invention is: how to solve the problem that the existing holographic reconstruction methods always need to make a trade-off between universality and imaging quality.

[0006] To solve the above technical problems, the present invention provides the following technical solutions: a frequency-aware physically guided neural network holographic reconstruction method, comprising: collecting a set of labeled data pairs; using a training set to perform frequency-aware data-driven network training to obtain a pre-trained network model; building a holographic imaging system of an actual environment according to the imaging optical path, and aligning the holographic imaging system, using the built holographic imaging system to measure a real object to obtain a hologram; inputting the acquired hologram into the trained network model, and on the basis of the network model, adding an imaging forward model for iterative optimization, and finally outputting an enhanced and optimized holographic reconstruction image.

[0007] As a preferred solution of the holographic reconstruction method of a frequency-aware physically guided neural network described in the present invention, the set of label data pairs includes: the network input image and the label are regarded as a pair of data, and the open source data set is cropped into a 128x128 image according to the input size of the network.

[0008] As a preferred solution of the holographic reconstruction method of a frequency-aware physically guided neural network described in the present invention, the pre-trained network model includes: the network input image is first subjected to preliminary feature extraction in the spatial domain through two 3×3 convolution layers, ReLU activation function, and batch normalization operations; sampling is performed based on the extracted preliminary features, and four down-sampling and four up-sampling are performed successively, and a frequency perception module is passed after each down-sampling and up-sampling; finally, the output global features are passed through a 1×1 convolution layer to obtain a holographic reconstruction image output by the entire network, and the output reconstruction image is Compare with label y by minimizing the loss function to train the first phase network model

[0009] As a preferred solution of the holographic reconstruction method of a frequency-aware physically guided neural network described in the present invention, the frequency perception module includes: the extracted features are converted from the spatial domain to the frequency domain through fast Fourier transform, global features are extracted through two 3×3 convolution layers, LeakyReLU activation function, and batch normalization operations, and finally returned to the spatial domain through inverse Fourier transform.

[0010] As a preferred solution of the holographic reconstruction method of a frequency-aware physical guided neural network described in the present invention, the pre-trained network model also includes using the proposed frequency-aware neural network for supervised learning to perform denoising and high-frequency information enhancement, and continuing to guide the training of the existing model through the diffraction model. The object of the two-step training is the same neural network. The entire network is based on the U-Net model. The hologram is used as input and first passes through two 3x3 convolution layers. The fine distribution of frequency is learned in the Fourier domain using a two-dimensional real Fourier transform, and then returns to the spatial domain through a two-dimensional real inverse Fourier transform. After passing through the output layer of the 1x1 convolution, a reconstructed phase map is obtained, which is expressed as:

[0011]

[0012]

[0013] in, is the reconstructed phase object output by the network, f x is the spatial frequency along the x-axis, f y is the spatial frequency along the y-axis, F(·) is the Fourier transform, the absolute value of u(x,y,d) is the reconstructed image, and |u(x,y,d)| 2 is the forward physical process of the system, i is an imaginary number, H(f x ,f y ,λ,d) is the propagation factor of light, d is the propagation distance between the phase object and the detector plane, and λ is the wavelength.

[0014] As a preferred embodiment of the holographic reconstruction method of a frequency-aware physically guided neural network described in the present invention, the loss function includes the loss function L for the first and second stages of the optimization training. stage1 and L stage2 , which combines the mean squared error, mean absolute error, and total variation regularizer, is expressed as:

[0015]

[0016]

[0017] Among them, F(y) and Represents the reconstruction of the first stage And the target truth value y performs Fourier transform operation, F(j) and represents the measured diffraction hologram j and the predicted diffraction hologram in the second stage Perform Fourier transform operation, ||·||1 and ||·||2 represent 1-norm and 2-norm respectively, α is the weight coefficient corresponding to mean square error, β is the weight coefficient corresponding to mean absolute error, and γ is the penalty coefficient of TV regularizer; for the first stage of network training, the reconstructed image output by the network is Calculate L with the target true value y stage1 The network is then optimized by the back propagation algorithm. The second stage is to compare the measured diffraction hologram j input by the network with the predicted diffraction hologram Calculate L stage2 The first-stage network model is fine-tuned through the back-propagation algorithm.

[0018] As a preferred solution of the holographic reconstruction method of a frequency-aware physical guided neural network described in the present invention, the holographic imaging system includes: building a holographic system in a real environment, and performing optical path collimation, beam expansion, and filtering preprocessing operations. After the preprocessing operations of the optical path experiment are completed, the real sample is placed in the above-built holographic system, and a diffraction hologram is obtained by collecting it with a camera after the coherent light source illuminates the propagation distance d; the enhanced and optimized holographic reconstruction image includes: inputting the obtained diffraction hologram into the trained network model R θ An accurate initial phase value is obtained, and the forward model of imaging is added to iteratively optimize the initial phase. The optimization goal is to minimize the error between the predicted hologram and the camera-captured hologram. After the iterative optimization is completed, the enhanced holographic reconstruction image is output.

[0019] Another object of the present invention is to provide a system for a holographic reconstruction method of a frequency-aware physically-guided neural network, which can solve the problem of holographic reconstruction of a frequency-aware physically-guided neural network by constructing a picture holographic reconstruction system.

[0020] To solve the above technical problems, the present invention provides the following technical solutions: a frequency-aware physically guided neural network holographic reconstruction system, comprising a data acquisition module, a model training module, a system building module and an imaging module; the data acquisition module is used to collect a set of labeled data pairs; the model training module uses the training set to perform frequency-aware data-driven network training to obtain a pre-trained network model; the system building module builds a holographic imaging system of an actual environment according to the imaging optical path, and aligns the holographic imaging system, and uses the built holographic imaging system to measure a real object to obtain a hologram; the imaging module is used to input the acquired hologram into the trained network model, and on the basis of the network model, add an imaging forward model for iterative optimization, and finally output an enhanced and optimized holographic reconstruction image.

[0021] A computer device includes a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the steps of the holographic reconstruction method of a frequency-aware physically guided neural network are implemented as described above.

[0022] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of a holographic reconstruction method of a frequency-aware physically guided neural network as described above.

[0023] The beneficial effects of the present invention are as follows: Compared with end-to-end deep learning and traditional physics-based methods, the frequency-aware physics-guided neural network holographic reconstruction method provided by the present invention can have both universality and imaging quality. By incorporating physical forward models and Fourier domain learning into network reasoning, the instability of digital holographic reconstruction is greatly reduced. The combined ideas of the present invention are particularly suitable for biological microscopic imaging with rich fine features, and will help high-throughput 3D microscopic imaging. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative labor.

[0025] Figure 1 A flow chart of a frequency-aware physically guided neural network holographic reconstruction method provided for the first embodiment of the present invention.

[0026] Figure 2 A schematic diagram of network training for a frequency-aware physically guided neural network holographic reconstruction method provided in the first embodiment of the present invention.

[0027] Figure 3 A schematic diagram of the optical path of a holographic imaging system for a holographic reconstruction method of a frequency-aware physically guided neural network provided in the first embodiment of the present invention.

[0028] Figure 4 A structural diagram of a frequency-aware physically guided neural network holographic reconstruction system provided for the second embodiment of the present invention.

[0029] Figure 5 A comparison diagram of the visualization results of the distributed intra-data of a holographic imaging system of a holographic reconstruction method of a frequency-aware physically guided neural network provided in the third embodiment of the present invention.

[0030] Figure 6A comparison diagram of visualization results of out-of-distribution data of a holographic imaging system of a holographic reconstruction method of a frequency-aware physically guided neural network provided in the third embodiment of the present invention.

[0031] Figure 7 This is a diagram of the reconstruction results of skeletal muscle cells of a holographic imaging system of a frequency-aware physically-guided neural network holographic reconstruction method provided in the third embodiment of the present invention. DETAILED DESCRIPTION

[0032] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the accompanying drawings.

[0033] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0034] Example 1

[0035] Reference Figure 1 to Figure 3 , which is the first embodiment of the present invention, and provides a holographic reconstruction method of a frequency-aware physically guided neural network, comprising: collecting a set of label data pairs; using a training set to perform frequency-aware data-driven network training to obtain a pre-trained network model; building a holographic imaging system of an actual environment according to an imaging optical path, and aligning the holographic imaging system, and using the built holographic imaging system to measure a real object to obtain a hologram; inputting the obtained hologram into the trained network model, and on the basis of the network model, adding an imaging forward model for iterative optimization, and finally outputting an enhanced and optimized holographic reconstruction image.

[0036] S1. Collect a set of label data pairs.

[0037] Prepare a set of labeled data pairs, where the network input image j k With label y k Considered as a pair of data, that is, (i k ,v k )∈S T , k = 1, 2, ... K, K is the total sample size of the training set. And the open source dataset is cropped into 128x128 images according to the input size of the network.

[0038] S2. Use the training set to perform frequency-aware data-driven network training to obtain a pre-trained network model. Figure 2 As shown in a, it is the first stage of network training optimization.

[0039] S2.1. The image input to the network first undergoes two 3×3 convolutional layers in the spatial domain, followed by a ReLU activation function and a batch normalization operation for preliminary feature extraction.

[0040] S2.2, then according to the preliminary features extracted above, four down-sampling and four up-sampling are performed successively, and after each down (up) sampling, a frequency perception module is passed.

[0041] S2.2.1. The extracted features are first converted from the spatial domain to the frequency domain by fast Fourier transform.

[0042] S2.2.2.2, then through two 3×3 convolutional layers - LeakyReLU activation function - batch normalization operation to extract global features, and finally through inverse Fourier transform back to the spatial domain.

[0043] S2.3. Finally, the output global features are passed through a 1×1 convolutional layer to obtain the holographic reconstruction image output by the entire network. Compare it with the label y by minimizing the loss function to train the first stage network model

[0044] S3. Build a holographic imaging system of the actual environment according to the imaging optical path, align the holographic imaging system, and use the built holographic imaging system to measure the real object to obtain a hologram.

[0045] according to Figure 3 The optical path schematic diagram shown in the figure builds a holographic system in a real environment, and performs pre-processing operations such as optical path collimation, beam expansion, and filtering. After the pre-processing operation of the optical path experiment is completed, a real sample (such as a graticule, a biological tissue specimen, etc.) is placed in the holographic system built above, and a diffraction hologram is obtained by the camera after being illuminated by a coherent light source with a propagation distance d.

[0046] S4. Input the acquired hologram into the trained network model, and based on the network model, add the imaging forward model for iterative optimization, and finally output the enhanced and optimized holographic reconstruction image.

[0047] Input the hologram obtained by S3 into the network model R trained by S2 θ A relatively accurate initial phase value can be obtained from the hologram. Then, on this basis, the forward model of imaging is added to iteratively optimize the initial phase, with the goal of minimizing the error between the predicted hologram and the camera-captured hologram, such as Figure 2 (b) shows the second stage of network training optimization. Finally, after the iterative optimization is completed, the enhanced holographic reconstruction image is output.

[0048] The network architecture designed by the present invention is different from the existing end-to-end reconstruction mode, but adopts a two-step reconstruction strategy. First, the proposed frequency-aware neural network is used for supervised learning to perform denoising and high-frequency information enhancement; then the existing model is guided and trained through the diffraction model. It is worth noting that the object of the two-step training is the same neural network. The entire network is based on the U-Net model. The hologram is used as input and first passes through two 3x3 convolution layers. Then, the two-dimensional real Fourier transform is used to learn the fine distribution of its frequency in the Fourier domain, and then it is returned to the spatial domain through a two-dimensional real inverse Fourier transform. Finally, the reconstructed phase map is obtained after the output layer of the 1x1 convolution. At the same time, the physical guidance stage uses the angular spectrum method to further enhance its high-frequency components after passing through the above-mentioned Fourier network. The specific process is as follows:

[0049]

[0050]

[0051] in, is the reconstructed phase object output by the network. x and f y are the spatial frequencies along the x-axis and y-axis, respectively. F(·) is the Fourier transform. The absolute value of u(x,y,d) is the reconstructed image, and |u(x,y,d)| 2 is the forward physical process of the system, i is an imaginary number, H(f x ,f y ,λ,d) is the propagation factor of light, d is the propagation distance between the phase object and the detector plane, and λ is the wavelength.

[0052] The network training of the present invention adopts two stages for training. For the first stage and the second stage of the optimization training, the loss function L stage1 and L stage2 , which combines the mean square error (MSE), mean absolute error (MAE) and total variation regularizer (TV), is expressed as:

[0053]

[0054]

[0055] For the first stage of network training, the reconstructed image output by the network is Calculate L with the target true value y stage1 The network is then optimized by the back propagation algorithm; the second stage is to compare the measured diffraction hologram j input by the network with the predicted diffraction hologram Calculate L stage2 The first-stage network model is fine-tuned through the back-propagation algorithm. Among them: F(y) and Represents the reconstruction of the first stage and the target true value y to perform Fourier transform operation; F(j) and represents the measured diffraction hologram j and the predicted diffraction hologram in the second stage Perform Fourier transform operation. ||·||1 and ||·||2 represent 1-norm and 2-norm respectively. α and β are the corresponding weight coefficients, set to 0.5 and 1, and γ is the penalty coefficient of TV regularizer, set to 10 -9 .

[0056] In step S2, for the first phase of network training, 30,000 images from the CelebAMask-HQ open source dataset were selected as the training set. The training batch was 1, the initial learning rate was 0.0001, and the training cycle was 64. In step S4, the wavelength of the light source was 632.8 nm, and the pixel pitch of the camera was 4.8 μm.

[0057] S5. If it is a new sample to be tested, repeat steps S3 to S4.

[0058] Example 2

[0059] Reference Figure 4 , which is the second embodiment of the present invention, is different from the previous embodiment in that it provides a frequency-aware physically guided neural network holographic reconstruction system, including: a data acquisition module, a model training module, a system building module and an imaging module.

[0060] The data collection module is used to collect a set of label data pairs.

[0061] The model training module uses the training set to perform frequency-aware data-driven network training to obtain a pre-trained network model.

[0062] The system building module builds a holographic imaging system of the actual environment according to the imaging optical path, aligns the holographic imaging system, and uses the built holographic imaging system to measure the real object to obtain a hologram.

[0063] The imaging module is used to input the acquired hologram into the trained network model, and on the basis of the network model, add the imaging forward model for iterative optimization, and finally output the enhanced and optimized holographic reconstruction image.

[0064] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the methods described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc., which can store program codes.

[0065] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as an ordered list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by an instruction execution system, device or apparatus (such as a computer-based system, a system including a processor, or other system that can fetch instructions from an instruction execution system, device or apparatus and execute instructions), or in conjunction with such instruction execution systems, devices or apparatuses. For the purposes of this specification, "computer-readable medium" can be any device that can contain, store, communicate, propagate or transmit a program for use by an instruction execution system, device or apparatus, or in conjunction with such instruction execution systems, devices or apparatuses.

[0066] More specific examples of computer-readable media (a non-exhaustive list) include the following: an electrical connection with one or more wires (electronic device), a portable computer disk case (magnetic device), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), an optical fiber device, and a portable compact disk read-only memory (CDROM). In addition, the computer-readable medium may even be a paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, deciphering or, if necessary, processing in another suitable manner, and then stored in a computer memory.

[0067] It should be understood that the various parts of the present invention can be implemented by hardware, software, firmware or a combination thereof. In the above-mentioned embodiments, a plurality of steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, it can be implemented by any one of the following technologies known in the art or their combination: a discrete logic circuit having a logic gate circuit for implementing a logic function for a data signal, a dedicated integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0068] Example 3

[0069] Reference Figure 5 to Figure 7 , which is the third embodiment of the present invention, and which is different from the first two embodiments in that it is used to verify and illustrate the technical effects adopted in the present invention, so as to verify the real effects of this method.

[0070] The reconstruction results of this embodiment use structural similarity (SSIM) and peak signal-to-noise ratio (PSNR) to quantitatively evaluate the performance of the present invention and the comparison method. As shown in Table 1, in the numerical simulation experiment, the average structural similarity and peak signal-to-noise ratio of the samples distributed in the data of the present invention are 0.89 and 26.48 dB, respectively, which are 0.08 and 5.67 dB higher than the latest physical guidance network PGNet, and 0.13 and 8.34 dB higher than the frequency-aware UNet. Although not as fast as the data-driven UNet reconstruction speed, the present invention has better universality and high-quality reconstruction in holographic reconstruction compared to the UNet method. The visualization effect diagram is as follows Figure 5 and Figure 6 , indicating the reconstruction effect of the data inside the distribution and outside the distribution. The first row is the reconstruction result of the diffraction hologram by various methods; the second row is the error result between the first row result and the true value; the third row is the spectrum of the first row result; the fourth row is the error result between the third row result and the true spectrum. The reconstructed image obtained by the method of the present invention has higher resolution and smaller error in frequency components, which is better than other reconstruction methods.

[0071] Table 1 Quantitative metrics for 100 images randomly selected from the test set

[0072]

[0073] Build a coaxial digital holographic imaging system for optical experiments. The schematic diagram of the experimental setup is shown in the figure. Figure 3As shown. A beam of coherent light with a wavelength of 632.8nm is emitted by a helium-neon laser and first filtered and expanded by a spatial filter. Then it passes through a collimating lens with a focal length of f=100mm and irradiates the transparent phase object. After diffraction at a distance of d, the camera obtains its intensity information as network input. Under this system, skeletal muscle cell slices were reconstructed in 2D and 3D. The present invention is superior to the compared deep learning scheme and can obtain more delicate biological structures, indicating that the technical means of Fourier network, model drive and deep integration can significantly reconstruct the microscopic features of biological cells. The visualization effect can be seen in Figure 7 As shown in the figure, the reconstruction results of each method include 2D reconstruction (left) and 3D reconstruction (right).

[0074] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A frequency-aware physically guided neural network holographic reconstruction method, characterized in that: include, Collect a set of labeled data pairs; The training set is used to perform frequency-aware data-driven network training to obtain a pre-trained network model; Build a holographic imaging system in the actual environment according to the imaging optical path, align the holographic imaging system, and use the built holographic imaging system to measure the real object to obtain a hologram; The acquired hologram is input into the trained network model, and based on the network model, the imaging forward model is added for iterative optimization, and finally the enhanced and optimized holographic reconstruction image is output.

2. The frequency-aware physically guided neural network holographic reconstruction method according to claim 1, characterized in that: The label data pair set includes a network input image and a label regarded as a pair of data, and the open source data set is cropped into a 128x128 image according to the input size of the network.

3. The holographic reconstruction method of a frequency-aware physically guided neural network according to claim 2, characterized in that: The pre-trained network model includes: the network input picture is firstly subjected to preliminary feature extraction in the spatial domain through two 3×3 convolution layers, ReLU activation function, and batch normalization operation; According to the extracted preliminary features, sampling is performed, and four down-sampling and four up-sampling are performed successively. After each down-sampling and up-sampling, a frequency perception module is passed; Finally, the output global features are passed through a 1×1 convolutional layer to obtain the holographic reconstruction image output by the entire network. Compare with label y by minimizing the loss function to train the first phase network model 4. The holographic reconstruction method of a frequency-aware physically guided neural network according to claim 3, characterized in that: The frequency perception module includes: performing fast Fourier transform to convert the extracted features from the spatial domain to the frequency domain, extracting global features through two 3×3 convolution layers, LeakyReLU activation function, and batch normalization operations, and finally returning to the spatial domain through inverse Fourier transform.

5. The holographic reconstruction method of a frequency-aware physically guided neural network according to claim 4, characterized in that: The pre-trained network model also includes using the proposed frequency-aware neural network for supervised learning to perform denoising and high-frequency information enhancement, and continuing to guide the training of the existing model through the diffraction model. The object of the two-step training is the same neural network. The entire network is based on the U-Net model. The hologram is used as input and first passes through two 3x3 convolution layers. The fine distribution of frequency is learned in the Fourier domain using a two-dimensional real Fourier transform, and then returns to the spatial domain through a two-dimensional real inverse Fourier transform. After passing through the output layer of the 1x1 convolution, a reconstructed phase map is obtained, which is expressed as, in, is the reconstructed phase object output by the network, f x is the spatial frequency along the x-axis, f y is the spatial frequency along the y-axis, F(·) is the Fourier transform, the absolute value of u(x,y,d) is the reconstructed image, and |u(x,y,d)| 2 is the forward physical process of the system, i is an imaginary number, H(f x ,f y ,λ,d) is the propagation factor of light, d is the propagation distance between the phase object and the detector plane, and λ is the wavelength.

6. The frequency-aware physically guided neural network holographic reconstruction method according to claim 5, characterized in that: The loss function includes the loss function L for the first and second stages of the optimization training. stage1 and L stage2 , which combines the mean square error, mean absolute error, and total variation regularizer, is expressed as, Among them, F(y) and Represents the reconstruction of the first stage And the target truth value y performs Fourier transform operation, F(j) and represents the measured diffraction hologram j and the predicted diffraction hologram in the second stage Perform Fourier transform operation, ||·||1 and ||·||2 represent 1-norm and 2-norm respectively, α is the weight coefficient corresponding to mean square error, β is the weight coefficient corresponding to mean absolute error, and γ is the penalty coefficient of TV regularizer; For the first stage of network training, the reconstructed image output by the network is Calculate L with the target true value y stage1 The network is then optimized by the back propagation algorithm. The second stage is to compare the measured diffraction hologram j input by the network with the predicted diffraction hologram Calculate L stage2 The first-stage network model is fine-tuned through the back-propagation algorithm.

7. The frequency-aware physically guided neural network holographic reconstruction method according to claim 6, characterized in that: The holographic imaging system includes building a holographic system in a real environment, and performing optical path collimation, beam expansion, and filtering preprocessing operations. After the preprocessing operations of the optical path experiment are completed, a real sample is placed in the above-built holographic system, and a diffraction hologram is obtained by collecting it with a camera after being illuminated by a coherent light source with a propagation distance d; The enhanced optimized holographic reconstruction image includes: inputting the acquired diffraction hologram into the trained network model R θ An accurate initial phase value is obtained, and the forward model of imaging is added to iteratively optimize the initial phase. The optimization goal is to minimize the error between the predicted hologram and the camera-captured hologram. After the iterative optimization is completed, the enhanced holographic reconstruction image is output.

8. A system using a holographic reconstruction method of a frequency-aware physically guided neural network as claimed in any one of claims 1 to 7, characterized in that: Including data acquisition module, model training module, system building module and imaging module; The data acquisition module is used to collect a set of label data pairs; The model training module uses the training set to perform frequency-aware data-driven network training to obtain a pre-trained network model; The system building module builds a holographic imaging system of the actual environment according to the imaging optical path, performs a collimation operation on the holographic imaging system, and uses the built holographic imaging system to measure the real object to obtain a hologram; The imaging module is used to input the acquired hologram into the trained network model, and on the basis of the network model, add the imaging forward model for iterative optimization, and finally output the enhanced and optimized holographic reconstruction image.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, it implements the steps of a frequency-aware physically guided neural network holographic reconstruction method as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of a holographic reconstruction method of a frequency-aware physically guided neural network described in any one of claims 1 to 7 are implemented.

Images