Method for realizing high-quality holographic imaging through self-calibration complex valued network and random compensation

Abstract

The invention provides a method for realizing high-quality holographic imaging by combining a self-calibration complex valued network with random compensation training, and relates to color holographic display. Comprising the following steps: S1, randomly extracting an image from a COCO2017 training data set, and performing super-resolution on the image by using a residual local feature super-resolution module; s2, predicting complex amplitude through a complex amplitude generator; s3, predicting a phase hologram through a phase hologram generator; s4, using a mean square error of the target amplitude and the predicted amplitude as a loss function optimization network; and S5, verifying the method through simulation and optical experiments. According to the algorithm, the robustness of the neural network is improved through randomized data compensation, the performance of the neural network is brought into full play, and then relevant information from a long-range region is adaptively coded through self-calibration convolution, so that the reconstruction quality of a holographic image is improved.

Description

Technical Field

[0001] The present invention relates to the fields of three-dimensional display and fast high-quality holographic imaging, and in particular to a method for achieving real-time high-quality holographic imaging based on a self-calibration complex-valued network and random compensation training. Background Art

[0002] Holographic display technology, a core solution in the field of three-dimensional imaging, achieves realistic three-dimensional representation of multiple objects and perspectives through spatial light field reconstruction. It holds irreplaceable application value in fields such as virtual reality, augmented reality, and medical imaging. In education, holographic classrooms enable one-to-one, one-to-many, and many-to-many holographic remote teaching, supporting access to multiple teaching platforms, effectively expanding classroom capacity and facilitating the shared development of school-based teaching resources. In the cultural tourism sector, Nanjing Dasiqi Digital Technology Co., Ltd.'s "Application of Holographic Display Equipment Technology in Traditional Cultural Scenic Spots" project uses air levitation and glasses-free 3D display effects to tailor content for scenic spots, create internet-famous landmarks, and present stunning light and shadow shows for tourists. In the commercial sector, holographic projection technology is widely used in exhibitions and advertising media, creating stunning three-dimensional effects through creative design. In the medical field, holographic technology can achieve more realistic medical imaging and surgical simulations, improving the accuracy and efficiency of medical diagnosis and treatment.

[0003] Over the past few decades, researchers have developed various computer-generated holographic algorithms in an attempt to achieve better display quality within a limited generation time. These conventional methods can be roughly divided into two categories: iterative methods and non-iterative methods, as well as some other computer-generated holographic methods, which aim to reduce computational time and improve display quality. In recent years, deep learning technology has brought breakthroughs to computational holography, especially the real-valued neural network architecture has significantly improved the efficiency of hologram generation. However, the existing schemes have two key technical defects: (1) the real-valued network is mismatched with the physical nature of light wave propagation in the complex domain, resulting in insufficient phase prediction accuracy; (2) the lack of a dynamic error compensation mechanism, which easily generates speckle noise and reconstruction distortion in complex scenes. The present invention proposes a self-calibrating complex-valued network architecture, which constructs an end-to-end wavefront propagation model through a complex convolution kernel, achieving the coordinated optimization of imaging quality and computational efficiency. Summary of the Invention

[0004] In response to the above-mentioned defects, the present invention proposes a method for achieving high-quality holographic imaging by combining a self-calibrated complex-valued network with random compensation training. Through a randomized data compensation strategy, the model can be exposed to more input data during training, increasing the diversity of weight updates during training, reducing training instability caused by small data volume, and improving network performance. At the same time, the addition of a self-calibration mechanism helps the network better understand the global structure and long-distance dependencies of the image, enabling the network to process a wider range of information, effectively improving the quality of reconstructed holographic images.

[0005] To achieve the above object, the technical solution adopted by the present invention is as follows:

[0006] A method for achieving real-time high-quality holographic imaging based on a self-calibrated complex-valued network and random compensation training includes the following steps:

[0007] S1. First, randomly extract images from the COCO2017 training dataset and super-resolve them using the residual local feature super-resolution module;

[0008] S2, predicting the complex amplitude through the complex amplitude generator;

[0009] S3, predicting a phase hologram through a phase hologram generator;

[0010] S4, the mean square error between the target amplitude and the predicted amplitude is used as the loss function to optimize the network;

[0011] S5. Verified through simulation and optical experiments.

[0012] Preferably, the method of step S1 is as follows: randomly extracting images from the COCO2017 training data set, satisfying the independent and identically distributed conditions, and the probability of each image being selected satisfies the uniform distribution condition, and then improving the resolution of the extracted images through the residual local feature super-resolution network.

[0013] Preferably, the method of step S2 is as follows: the target complex amplitude is predicted by the complex amplitude generator of the network. At this time, the predicted complex amplitude is in the target plane, and the predicted complex amplitude is forward propagated to the spatial light modulator (SLM) plane through the angular spectrum method to facilitate final imaging.

[0014] Preferably, the specific method of step S3 is: predicting the phase hologram through the phase hologram generator of the network U-Net architecture, the self-calibration module in the generator expands the receptive field during convolution, and improves the generation quality of the phase hologram, and the phase hologram is used for verification of subsequent optical experiments.

[0015] Preferably, the specific method in step S4 is: comparing the reconstructed amplitude with the target amplitude to obtain a mean square error and using it as a loss function to optimize the network.

[0016] Preferably, the method of randomly extracting images from the COCO2017 training dataset in step S1 is specifically as follows:

[0017] Step S11, the COCO2017 training dataset is mathematically expressed as a set Each element I k Corresponding to an image, M is the number of images, D COCO Denotes the COCO2017 training set, from D COCO Randomly extract N images from , which can be expressed as: Among them D sto-ext represents an image randomly sampled from the COCO2017 training dataset; i j ~Unifor(1,I M ), where i j are independent and identically distributed random variables, and ensure that i j ≠i k ; The probability of each image being selected satisfies the uniform distribution condition:

[0018] Step S12: Super-resolution processing is performed on the extracted image. The mathematical expression is:

[0019]

[0020]

[0021]

[0022] in represents the convolution operation for feature extraction, is the extracted feature map. Then, multiple RLFBs are used in a cascaded manner for deep feature extraction. represents the nth RLFB function, H smooth Indicates the use of a 3×3 convolution layer to smooth the gradually refined depth feature map, I SR is the image output after super-resolution.

[0023] Preferably, the method for complex amplitude prediction in step S2 is specifically as follows:

[0024] The prediction of complex amplitude is a complex-valued self-calibration network based on U-Net. In downsampling, the convolution kernel size is 3×3 and the stride is 2; in upsampling, the convolution kernel size is 4×4 and the stride is 2; four downsampling and four upsampling are performed respectively.

[0025] Preferably, the forward propagation method of the angular spectrum method in step S2 is specifically as follows:

[0026] The forward propagation of the angular spectrum method propagates the complex amplitude from the target plane to the SLM plane. The calculation of the forward propagation can be expressed as:

[0027]

[0028]

[0029] Among them, CA SLM-plane represents the complex amplitude propagating to the SLM plane, represents the two-dimensional inverse fast Fourier transform, represents the two-dimensional fast Fourier transform, CA Tar-plane represents the complex amplitude of the target plane, H is the transfer function in the angular spectrum method, where the training set image is used as the amplitude, the zero matrix is used as the initial phase, k is the wave number, z is the propagation distance, λ is the wavelength, f x , f y are the spatial frequencies in the x and y directions, respectively.

[0030] Preferably, step S3 predicts a phase hologram through a phase hologram generator of a network U-Net architecture. The self-calibration module in the generator expands the receptive field during convolution and improves the generation quality of the phase hologram. The phase hologram is used for verification in subsequent optical experiments.

[0031] Preferably, the complex amplitude and phase holographic prediction method in steps S2-S3 is specifically as follows:

[0032] Combining self-calibration convolution with complex-valued convolutional neural network to achieve high-quality holographic imaging; in the complex-valued convolutional neural network, the calculation of the complex convolution layer can be expressed as:

[0033] Y0=X*K=(X r *K r -X i *K i )+j(X r *K r +j(X r *K r )

[0034] CRelu(X)=Relu(X r )+jRelu(X i )

[0035] Where X represents the input complex-valued matrix, K represents the convolution kernel, subscripts r and i are the real and imaginary parts respectively; CRelu is the complex-valued activation function; the calculation of the self-calibrated complex convolution layer is expressed as:

[0036]

[0037] F cro-dim=σ{Y0+Up[Γ2(Y0)]}⊙Y0

[0038] Y′0=Complex_Avgpools(Y0)

[0039] where Y self-calibrated is the feature after the self-calibration complex convolution layer operation, Y0 represents the input complex value matrix, Γ1(·)~Γ3(·) represent the convolution operation, the convolution kernel sizes are 3, 5, and 1 respectively, the step size is 1, the padding is 1, 2, and 0 respectively, and F cro-dim represents cross-dimensional features generated through downsampling and low-dimensional spatial transformation, used to capture long-range dependencies. ⊙ represents element-wise multiplication, and σ(·) is the sigmoid function, which maps features to attention weights in the range [0, 1] and dynamically adjusts the contribution of the real and imaginary convolution results. Up(·) represents bilinear interpolation with a scale factor of 2. Y'0 represents processing in the self-calibrated scale space, using a convolution kernel size of ε×ε=2×2, a stride of s=2, and complex-valued average pooling with zero padding. The final output after the self-calibrated convolution is:

[0040]

[0041] Y=Y self-calibrated +Y1

[0042] Y1 represents the output feature after a convolution with a kernel size of 3, a step size of 1, and a padding of 1; Y self-calibrated represents the output features after the self-calibration operation, and Y is the final output feature.

[0043] The activation functions of the self-calibrated complex convolutional layer include complex-valued Sigmoid (CSigmoid) and complex-valued Relu (CRelu), which can be expressed as:

[0044] CSigmoid=Sigmoid(X r )+jSigmoid(X i )

[0045] Preferably, the specific method for calculating the mean square error loss function between the target amplitude and the predicted amplitude in step S4 is:

[0046]

[0047] Among them A Rec,i and A Tar,i Represent the amplitude values of the reconstructed image and the target image at the i-th pixel respectively, and N is the total number of pixels.

[0048] The beneficial effects of the present invention are:

[0049] (1) The present invention uses a randomized data compensation strategy to allow the model to be exposed to more input data during training, which can increase the diversity of weight updates during training, reduce training instability caused by small data volume, and improve network performance.

[0050] (2) Through self-calibration convolution, the network can better understand the global structure and long-distance dependencies of the image, enabling the network to process a wider range of information and effectively improving the quality of reconstructed holographic images.

[0051] (3) The present invention introduces a multi-scale feature fusion module that can effectively integrate feature information at different scales. In image data, features at different scales contain information at different levels. This fusion method can provide richer and more comprehensive feature information for reconstructing holographic images, making the reconstructed images more vivid and realistic, further improving the reconstruction quality of holographic images. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments.

[0053] Figure 1 This is a flowchart of the steps of the method for realizing real-time high-quality holographic imaging based on a self-calibrated complex-valued network and random compensation training according to the present invention;

[0054] Figure 2 1 is an overall structural diagram of a self-calibration complex-valued network and a randomized compensation training network in an embodiment of the present invention;

[0055] Figure 3 : is a diagram of the overall structure of a super-resolution network in an embodiment of the present invention;

[0056] Figure 4 is a structural diagram of a complex amplitude and phase holographic prediction network in an embodiment of the present invention;

[0057] Figure 5 1 is a diagram showing the results of a single-channel numerical simulation experiment in an embodiment of the present invention;

[0058] Figure 6 Graph showing the results of numerical simulation experiments of generalization experiments in an embodiment of the present invention;

[0059] Figure 7 A color numerical simulation experiment result diagram in an embodiment of the present invention;

[0060] Figure 8 1 is a diagram showing the results of a single-channel optical experiment in an embodiment of the present invention;

[0061] Figure 9 Graph showing the results of color optical experiments in an embodiment of the present invention;

[0062] Figure 10 This is an optical experimental platform in an embodiment of the present invention. DETAILED DESCRIPTION

[0063] The following will clearly and completely describe 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. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0064] The embodiment of the present invention provides a method for achieving real-time high-quality holographic imaging based on a self-calibrated complex-valued network and random compensation training, such as Figure 1 As shown, the solution of the present invention includes six aspects:

[0065] In a first aspect, the present invention provides a method for randomly extracting image data, which is applied to an electronic device. The method includes: the electronic device performs extraction algorithm operation, calls a random extraction algorithm in Python, and randomly extracts a corresponding number of images from the COCO2017 training data set.

[0066] In a second aspect, the present invention provides an image super-resolution method, which is applied to an electronic device. The method includes: the electronic device calls the super-resolution algorithm in Python through a super-resolution algorithm to achieve resolution improvement of low-resolution images extracted from the COCO2017 training dataset.

[0067] In the third aspect, the present invention provides a complex amplitude generation method based on a self-calibration complex-valued convolutional neural network, which is applied to electronic devices. The method includes: the electronic device calls the self-calibration complex-valued convolutional neural network algorithm in Python, trains the network through self-calibration adjustment, and then uses the mean square error between the target amplitude and the reconstructed amplitude as the loss function to tune the network parameters, and finally realizes the prediction of complex amplitude.

[0068] In the fourth aspect, the present invention provides a phase hologram generation method based on a self-calibration complex-valued convolutional neural network, which improves the generation quality and generation speed of the hologram and is applied to electronic devices. The method includes: the electronic device calls the self-calibration complex-valued convolutional neural network algorithm in Python, trains the network through self-calibration adjustment, and then uses the mean square error between the target amplitude and the reconstructed amplitude as the loss function to tune the network parameters, and finally realizes the prediction of the phase hologram.

[0069] In the fifth aspect, the present invention provides a fast and high-quality holographic generation solution based on the above method, so as to quickly realize the generation and imaging of holograms. The method includes: the electronic device implements the operation of the self-calibration complex-valued convolutional neural network algorithm in Python, and uses the mean square error between the target amplitude and the reconstructed amplitude as the loss function to tune the parameters of the entire self-calibration complex-valued convolutional neural network. At the same time, the randomized data compensation strategy allows the model to be exposed to more input data during training, which can increase the diversity of weight updates during training, reduce training instability caused by small data volume, and improve the robustness of the network; the self-calibration convolution operation helps the network better understand the global structure and long-distance dependencies of the image, so that the network can process a wider range of information, effectively improving the quality of the reconstructed holographic image.

[0070] In a sixth aspect, the present invention provides a hologram optical reconstruction platform that can reproduce the three-dimensional form of an object. The method includes: a personal computer is connected to an SLM, the laser emitted by an RGB laser passes through an attenuator, a beam expander and an aperture to reach the SLM, the RGB hologram on the personal computer is loaded onto an SLM adapter board, and finally displayed on a camera.

[0071] Example:

[0072] In the process of super-resolution improvement through the residual local feature network after image extraction of the COCO2017 training dataset, the specific steps are as follows: first, shallow feature extraction is performed, the function call: Conv2D(kernel_size=3×3,stride=1,padding=1), after inputting the low-resolution image, the initial feature map is extracted through a single 3×3 convolution layer, retaining the original texture and edge information; local feature extraction, function call: stacking multiple residual local feature blocks, Conv(Relu(Conv(Relu(Conv()))) gradually extracts local features through 3 layers of convolution and nonlinear activation, reduces network fragmentation, and enhances spatial attention, calls the function: Conv2D(kernel_size=3×3,stride=1,padding=1) to smooth the stacked deep features, reduce noise interference, accurately extract features, call the function PixelShuffle(scale_factor=4), achieve 4x upsampling through channel rearrangement, and generate high-resolution image output, such as Figure 3 shown.

[0073] In the process of complex amplitude prediction and phase hologram generation, you first need to call torch, and then call nn and optim in torch to build and train neural networks. It provides many commonly used neural network layers, loss functions, optimizers and other functions.

[0074] Call the OpenCV library, which is a widely used computer vision library commonly used for tasks such as image processing, video analysis, object detection, and a series of functions for building neural networks and image processing.

[0075] Called numpy, numpy is a powerful python library mainly used for numerical computing, especially when dealing with large, multidimensional arrays and matrices. It provides many functions, including mathematical operations, random number generation, linear algebra, Fourier transform, etc.

[0076] The torch.FFT module provides support for the Fast Fourier Transform (FFT). FFT is an efficient algorithm for computing the Discrete Fourier Transform (DFT).

[0077] Call ComplexPytorch, call ComplexConv2D and ComplexConvTranspose2D in the ComplexPytorch.ComplexLayers library, and provide convolution layers for processing complex data. They are very similar to ordinary convolution layers of Pytorch (such as Conv2d and ConvTranspose2d), but can handle complex input and output; call ComplexRelu and ComplexSigmoid in ComplexPyTorch.ComplexFunctions. ComplexRelu is a complex version of the Relu activation function. The traditional Relu processes the input value. If the input is greater than 0, the input value is output; if the input is less than 0, 0 is output. For the case of complex numbers, ComplexRelu acts on the modulus of the complex number and processes the real and imaginary parts of each complex number, usually by keeping the real part non-negative and setting the negative real part to 0. The complex tensor output by ComplexSigmoid, each element of which is transformed by the Sigmoid function, will be in the interval [0,1]. In tqdm, call trange. trange is a very convenient function in the tqdm library for quickly generating a progress bar. It is equivalent to range, but it automatically creates a progress bar.

[0078] The specific steps are as follows: the U-net architecture maintains the resolution consistency between the input and output images, and the jump connection retains the extracted detail information; in downsampling, the function is called: ComplexConv2D(kernel_size=3×3,stride=2,padding=1), and the input image or feature map is convolved to extract local features, while changing the spatial dimension of the input. In upsampling, the function is called: ComplexConvTranspose2D(kernel_size=4×4,stride=2,padding=1), and the transposed convolution uses a suitable convolution kernel to enlarge the size of the input data while retaining or reconstructing the feature information in the input; the input image first undergoes an ordinary complex-valued convolution to widen the feature channel, and then the feature is self-calibrated and convolved. The self-calibrated convolution performs feature processing at each layer of downsampling. The specific steps are as follows: in the Γ1 part, the function is called: ComplexConv2D(kernel_size=3×3,stride=1,padding=1,dilation=1,groups=1,bias s=False); in the Γ2 part, first call the function: ComplexAvgPool2d(kernel_size=2×2,stride=2), a two-dimensional complex average pooling layer, which performs a pooling operation on the two-dimensional input, reduces the size of the data by averaging the elements in the region, thereby achieving dimensionality reduction and feature extraction, and then calls the function: ComplexConv2D(kernel_size=5×5,stride=1,padding=2,dilation=1,groups=1,bias=False) to further process the features, and then call the bilinear interpolation function to enlarge the feature size with a size factor of 2, and then map the features to [0,1] by calling the function: ComplexSigmoid(input); in the Γ3 part, call the function: ComplexConv2D(kernel_size=1×1,stride=1,padding=1,dilation=1,groups=1,bias=False) to process the features, and finally merge the features Y self-calibrated And Y1 to get the output feature Y, such as Figure 4 shown.

[0079] In the process of implementing numerical simulation, images of the DIV2K validation dataset were selected and two different images not included in the training or validation dataset were used to perform numerical reconstruction in order to evaluate the generalization ability of the model. The results of the monochrome numerical simulation are shown in Figure 2. Figure 5 As shown in the figure, the numerical simulation results of the single-color generalization experiment are as follows Figure 6The colored numerical simulation results are shown in Figure 7 shown.

[0080] In the process of optical reconstruction of the hologram, a personal computer (PC) is connected to the SLM. The laser light emitted by the RGB laser passes through the attenuator (A1), beam expander (BE) and aperture (A2) to reach the SLM. The RGB hologram on the PC is loaded onto the SLM adapter board, and the three-dimensional image is finally displayed on the camera. The monochrome hologram reconstruction result is as follows Figure 8 As shown in the figure, the color hologram reconstruction results are as follows Figure 9 As shown in Figure 9 shown.

[0081] The experimental results show that the numerically reconstructed color image of this embodiment has high contrast, clear image details, and good color feature restoration. The optical reconstruction result has weak speckle and clear imaging details. Figure 10 shown.

[0082] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, it is still possible for those skilled in the art to modify the technical solutions described in the aforementioned embodiments or to substitute equivalents for some of the technical features. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A method for achieving high-quality holographic imaging using a self-calibrated complex-valued network and random compensation, characterized in that: The holographic imaging method comprises the following steps: S1, randomly extract images from the COCO2017 training dataset and super-resolve them using the residual local feature super-resolution module; S2, predicting the complex amplitude through the complex amplitude generator; S3, predicting a phase hologram through a phase hologram generator; S4, the mean square error between the target amplitude and the predicted amplitude is used as the loss function to optimize the network; S5. Verified through simulation and optical experiments.

2. The method according to claim 1, characterized in that The step S1 includes randomly extracting images from the COCO2017 training dataset, satisfying the independent and identically distributed condition, and the probability of each image being selected satisfies the uniform distribution condition, and then improving the resolution of the extracted images through the residual local feature super-resolution network.

3. The method according to claim 1, characterized in that The step S2 predicts the target complex amplitude through the complex amplitude generator of the network U-Net architecture. At this time, the predicted complex amplitude is in the target plane, and the predicted complex amplitude is forward propagated through the angular spectrum method to propagate the predicted complex amplitude to the spatial light modulator plane to facilitate the final imaging.

4. The method according to claim 1, wherein The step S3 predicts the phase hologram through the phase hologram generator of the network U-Net architecture. The self-calibration module in the generator expands the receptive field during convolution and improves the generation quality of the phase hologram. The phase hologram is used for verification of subsequent optical experiments.

5. The method according to claim 1, wherein In step S4, the reconstructed amplitude is compared with the target amplitude, and the mean square error is obtained as a loss function to optimize the network. The calculation method of the loss function is: Among them A Rec,i and A Tar,i Represent the amplitude values of the reconstructed image and the target image at the i-th pixel respectively, and N is the total number of pixels.

6. The method according to any one of claims 1 or 2, characterized in that The method of randomly extracting images from the COCO2017 training dataset in step S1 is specifically as follows: Step S11, the COCO2017 training dataset is mathematically expressed as a set Each element I k Corresponding to an image, M is the number of images, Represents the COCO2017 training set; from Randomly extract N images from , which can be expressed as: in represents an image randomly selected from the COCO2017 training set; i j ~Unifor(1,I M ), where i j are independent and identically distributed random variables, and ensure that i j ≠i k ; The probability of each image being selected satisfies the uniform distribution condition: Step S12: Super-resolution processing is performed on the extracted image. The mathematical expression is: in represents the convolution operation for feature extraction, is the extracted feature map; then, multiple residual local feature blocks RLFB are used in a cascade manner to perform deep feature extraction, represents the nth RLFB function, H smooth Indicates the use of a 3×3 convolution layer to smooth the gradually refined depth feature map, I SR is the image output after super-resolution.

7. The method according to claim 1, characterized in that The method for complex amplitude prediction in step S2 is specifically as follows: based on a self-calibrated complex-valued network of U-Net, in downsampling, the size of the convolution kernel is 3×3, and the stride is 2; in upsampling, the size of the convolution kernel is 4×4, and the stride is 2; four downsampling and four upsampling are performed respectively.

8. The method according to claim 1, characterized in that The method for predicting phase holography in step S3 is specifically as follows: based on a self-calibrated complex-valued network of U-net, in downsampling, the size of the convolution kernel is 3×3, and the stride is 2; in upsampling, the size of the convolution kernel is 4×4, and the stride is 2; three downsampling and three upsampling are performed respectively.

9. The method according to any one of claims 7 or 8, characterized in that Combining self-calibration convolution with complex-valued convolutional neural network to achieve high-quality holographic imaging; in the complex-valued convolutional neural network, the calculation of the complex convolution layer can be expressed as: Y0=X*K=(X r *K r -X i *K i )+j(X r *K i +X i *K r ) CRelu(X)=ReRelu(X r )+jRelu(X i ) Where X represents the input complex-valued matrix, K represents the convolution kernel, subscripts r and i are the real and imaginary parts respectively; CRelu is the complex-valued activation function; the calculation of the self-calibrated complex convolution layer is expressed as: F cro-dim =σ{Y0+Up[Γ2(Y′0)]}⊙Y0 Y′0=Complex_Avgpools(Y0) where Y self-calibbrated is the feature after the self-calibration complex convolution layer operation, Y0 represents the input complex value matrix, Γ1(·)~Γ3(·) represent the convolution operation, the convolution kernel size is 3, 5, 1, the step size is 1, and the padding is 1, 2, 0 respectively; F cro-dim represents the cross-dimensional features generated by downsampling and low-dimensional space transformation, which is used to capture long-distance dependencies. ⊙ represents the multiplication of each element. σ(·) is the Sigmoid function, which maps the features to attention weights in [0, 1] and dynamically adjusts the contribution of the real and imaginary part convolution results. Up(·) represents bilinear interpolation with a scale factor of 2. Y'0 represents processing in the self-calibrated scale space, using a convolution kernel size of ε×ε=2×2, a step size of s=2, and complex-valued average pooling with zero padding. The final output after the self-calibrated convolution is: Y=Y self-calibrated +Y1 Y1 represents the output feature after a convolution with a kernel size of 3, a stride of 1, and a padding of 1; Y self-calibrated represents the output feature after the self-calibration operation, and Y is the final output feature; The activation functions of the self-calibrated complex convolutional layer include complex-valued Sigmoid (CSigmoid) and complex-valued Relu (CRelu) activation functions, which can be expressed as: CSigmoid=Sigmoid(X r )+jSigmoid(X i )。 10. The method according to claim 3, characterized in that The angular spectrum forward propagation propagates the complex amplitude from the target plane to the spatial light modulator plane. The angular spectrum propagation calculation can be expressed as: Among them, CA SLM-plane represents the complex amplitude propagating to the SLM plane, represents the two-dimensional inverse fast Fourier transform, represents the two-dimensional fast Fourier transform, CA Tar-plane represents the complex amplitude of the target plane, where the training set image is used as the amplitude, the zero matrix is used as the initial phase, k is the wave number, z is the propagation distance, λ is the wavelength, and f x , f y are the spatial frequencies in the x and y directions, respectively.

Images