Non-interference, non-iterative complex amplitude reading method and device

Abstract

The present invention discloses a non-interference, non-iterative complex amplitude reading method and device. The reading method comprises the following steps: diffracting a light beam containing amplitude and phase information to obtain an intensity image as a diffraction pattern with varying light intensity; constructing and training a diffraction intensity-complex amplitude model based on the correlation between the diffraction pattern and the amplitude and phase information; and directly applying the trained model to new diffraction images to obtain amplitude and phase information; and being able to detect complex amplitude information including amplitude and phase from an intensity image based on a single diffraction image. This method can improve the stability and accuracy of phase reading results, increase calculation speed, and simplify the optical system. The method is suitable for fields such as holographic storage, biomedical image processing, and microscopic imaging.

Description

Technical Field

[0001] The present invention relates to the field of imaging technology, and in particular to a non-interference, non-iterative complex amplitude reading method and device. Background Art

[0002] Phase and amplitude are two fundamental properties of light. In fields such as information storage, biomedicine, and computational imaging, it is often necessary to simultaneously obtain both amplitude and phase information.

[0003] Current detectors can only detect the intensity of light, and the phase information needs to be obtained through indirect calculations using interferometry or non-interferometric iterative methods.

[0004] The interferometry method requires the introduction of a reference light beam to convert phase information into an interference pattern, and then calculates to obtain the phase. This results in a relatively complex optical system, unstable phase reading results, and low accuracy.

[0005] While non-interferometric methods don't require reference light and are relatively simple, they often require multiple iterations or multiple captures to obtain relatively accurate phase information. Therefore, the disadvantage of non-interferometric iterative methods is their slow computational speed.

[0006] However, in some specialized fields, such as information storage and access, and real-time image processing for computational imaging, there are high demands for speed and accuracy in reading information. Accurate and rapid reading of amplitude and phase information is required using simple optical systems. Therefore, both the interferometric and non-interferometric methods mentioned above in the prior art cannot meet these requirements. Summary of the Invention

[0007] This invention proposes a non-interferometric, non-iterative complex amplitude reading method and device, which detects complex amplitude information from an intensity image based on a single diffraction pattern. This method aims to overcome the drawbacks of existing interferometric and non-interferometric methods for acquiring light phase information, such as complex optical systems, unstable phase reading results, low accuracy, and slow computation speed. This method improves the stability and accuracy of phase reading results, increases computation speed, and simplifies the optical system, making it suitable for applications in holographic storage, biomedical image processing, microscopy, and other fields.

[0008] In a first aspect, the present invention provides a non-interferometric, non-iterative complex amplitude reading method, comprising the following steps:

[0009] Step S01, diffracting a light beam containing amplitude information and phase information to obtain an intensity image as a diffraction pattern with light intensity variation;

[0010] Step S02 : constructing and training a diffraction intensity-complex amplitude model based on the correlation between the diffraction pattern and the amplitude information and the phase information, and applying the diffraction intensity-complex amplitude model to a new diffraction pattern to directly obtain the amplitude information and the phase information.

[0011] Preferably, in step S02, learning and training are performed through the correspondence between multiple inputs and multiple outputs to establish the neural network model parameters of the diffraction intensity-complex amplitude model.

[0012] Preferably, in step S01, the following steps are included:

[0013] Step S11, generating experimental images: generating n amplitude A images and phase P images of different modes, where n≥1 and n is a positive integer;

[0014] Step S12, experimental image capture: performing amplitude and phase modulation on the light beam using the complex amplitude image dataset C including the amplitude A image and the phase P image, and capturing a diffraction intensity image I corresponding to the amplitude and phase modulation;

[0015] In step S02, the following steps are included:

[0016] Step S21, data set preparation: the diffraction intensity image I and the complex amplitude image data set C are combined into a data set D IC , the dataset D IC Divide into mutually exclusive neural network training data sets T IC and validation dataset V IC , respectively used for training and verification of the neural network model CNN;

[0017] Step S22, model building: establishing the diffraction intensity-complex amplitude model consistent with the neural network model CNN.

[0018] Preferably, after step S22, the following steps are further included:

[0019] Step S221, model optimization: setting the loss function L of the neural network model CNN, and using the neural network training data set T IC The parameters of the neural network model CNN are trained until the loss function L converges.

[0020] Preferably, after step S221, the following steps are further included:

[0021] Step S222, model verification: through the verification data set V IC Verify the model generalization performance of the neural network model CNN and obtain the generalized neural network model CNN.

[0022] Preferably, in step S02, after step S22, when obtaining the amplitude information and the phase information, the following steps are included:

[0023] Step S23, model application: input the new diffraction intensity image into the trained and verified neural network model CNN, and output the amplitude A image and the phase P image.

[0024] Preferably, in step S11, the amplitude A image and the phase P image are randomly coded amplitude images (a) and phase images (b), or the amplitude A image and the phase P image are natural visual images.

[0025] Preferably, in step S22, the neural network model CNN adopts an unsupervised neural network model structure based on a combination with a physical optics diffraction model, or the neural network model CNN adopts an end-to-end neural network model structure based on data-driven, or the neural network model CNN includes an intensity-amplitude neural network model CNN1 and an intensity-phase neural network model CNN2, which are respectively used to restore the amplitude A image and the phase P image of the complex amplitude image, or the neural network model CNN is set to an intensity-amplitude-phase neural network model CNN_12 with a single input of the diffraction intensity image and dual outputs of the amplitude A image and the phase P image.

[0026] In a second aspect, the present invention provides a non-interference, non-iterative complex amplitude reading device for implementing the non-interference, non-iterative complex amplitude reading method described in any one of the first aspects of the present invention, wherein the reading device comprises an optical system and an electronic device, wherein the optical system comprises a laser, a beam paralleling component, a first 1 / 2 wave plate, an aperture, a first imaging component, a first polarizer, and a first transmission-reflection beam splitter arranged in sequence along the propagation direction of the incident light beam; an amplitude spatial light modulator is arranged in the propagation direction of the transmitted light beam of the first beam splitter, and a second polarizer, a second imaging component, a second 1 / 2 wave plate, and a second transmission-reflection beam splitter are arranged in sequence in the propagation direction of the reflected light beam of the first beam splitter; a phase spatial light modulator is arranged in the propagation direction of the transmitted light beam of the second beam splitter, and a third imaging component and a photodetector are arranged in sequence in the propagation direction of the reflected light beam of the second beam splitter; the electronic device comprises one or more processors and a memory, wherein one or more computer programs are stored in the memory, and when the one or more processors receive the diffraction pattern captured by the photodetector and execute the one or more computer programs, the steps of the non-interference, non-iterative complex amplitude reading method described in any one of the first aspects of the present invention are implemented.

[0027] Preferably, the beam parallelizing component includes a pinhole filter and a collimating lens sequentially arranged along the propagation direction of the incident light beam.

[0028] The present invention provides a non-interference, non-iterative complex amplitude reading method, which can be implemented by the non-interference, non-iterative complex amplitude reading device provided by the present invention.

[0029] The non-interference, non-iterative complex amplitude reading method and device of the present invention can achieve at least the following beneficial effects:

[0030] 1. The non-interference, non-iterative complex amplitude reading method of the present invention can detect complex amplitude information including amplitude and phase from an intensity image based on a single diffraction image, overcoming the defects of the interference method used in the prior art for obtaining the phase information of light, such as a complex optical system and unstable phase reading results, and the relatively low accuracy and slow calculation speed of the non-interference iterative method. As a non-interference method, there is no need for interferometric reference light to read the phase, and the amplitude and phase are read by directly reconstructing the model without iteration. Only a single diffraction image needs to be captured, making the complex amplitude information reading operation simple. It can realize simultaneous reading of amplitude and phase, and can realize real-time reading of multi-grayscale amplitude and multi-grayscale phase, which can improve the stability and accuracy of the phase reading results. It can realize direct reading of complex amplitude without iterative calculation, thereby improving the calculation speed. It can also simplify the optical system and is suitable for fields such as holographic storage, biomedical image processing, and microscopic imaging.

[0031] 2. The non-interference, non-iterative complex amplitude reading method of the present invention can adopt the neural network model of the existing technology, or construct a neural network model with a different structure. After training and verification, it uses the intensity image with diffraction information as input to accurately and directly predict the corresponding amplitude and phase information, thereby improving the reading accuracy and efficiency of the amplitude and phase information including the complex amplitude, and meeting the needs of modern and intelligent optical applications.

[0032] 3. The non-interferometric, non-iterative complex amplitude reading device of the present invention is a non-interferometric reading system that only requires the light beam to propagate a preset distance in the air. No reference light, additional lenses, or other optical systems are required for reading. The device has a reasonable and compact structure, fast amplitude and phase reading speed, and accurate, stable, and reliable reading results. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0034] Figure 1 This is a flow chart of a non-interference, non-iterative complex amplitude reading method according to an embodiment of the present invention, which realizes rapid complex amplitude reading using a single diffraction intensity image;

[0035] Figure 2This is a flow chart of a non-interference, non-iterative complex amplitude reading method for realizing rapid complex amplitude reading using a single diffraction intensity image in another embodiment of the present invention;

[0036] Figure 3 Schematic diagram of the complex amplitude reading principle in the non-interference, non-iterative complex amplitude reading method according to one embodiment of the present invention;

[0037] Figure 4 Schematic diagram of a non-interference, non-iterative complex amplitude reading method according to an embodiment of the present invention, in which both the phase image and the amplitude image are four-grayscale random coding images, and the output of the optical system is a diffraction intensity image;

[0038] Figure 5 Schematic diagram of a U-net neural network model in a non-interference, non-iterative complex amplitude reading method according to an embodiment of the present invention;

[0039] Figure 6 Schematic diagram of the training and verification process of the amplitude neural network model and the phase neural network model in the non-interference, non-iterative complex amplitude reading method according to one embodiment of the present invention;

[0040] Figure 7 Schematic diagram of a process for directly reading amplitude images and phase images in a non-interference, non-iterative complex amplitude reading method according to an embodiment of the present invention;

[0041] Figure 8 Schematic diagram of a single-input, dual-output neural network model in a non-interference, non-iterative complex amplitude reading method according to an embodiment of the present invention;

[0042] Figure 9 Schematic diagram of the training and verification process of the amplitude neural network model and the phase neural network model in the non-interference, non-iterative complex amplitude reading method according to another embodiment of the present invention;

[0043] Figure 10 Schematic diagram of a process for directly reading amplitude images and phase images in a non-interference, non-iterative complex amplitude reading method according to another embodiment of the present invention;

[0044] Figure 11 FIG. 1 is a schematic diagram of the optical system structure of a non-interference, non-iterative complex amplitude reading device according to an embodiment of the present invention.

[0045] In the figure, 1 is a laser, 2 is a pinhole filter, 3 is a collimating lens, 4 is a first 1 / 2 wave plate, 5 is an aperture, 6 is a first relay lens, 7 is a second relay lens, 8 is a first polarizer, 9 is a first beam splitter, 10 is an amplitude spatial light modulator, 11 is a second polarizer, 12 is a third relay lens, 13 is a fourth relay lens, 14 is a second 1 / 2 wave plate, 15 is a second beam splitter, 16 is a phase spatial modulator, 17 is a fifth relay lens, 18 is a sixth relay lens, and 19 is a photodetector. DETAILED DESCRIPTION

[0046] To make the objectives, technical solutions, and advantages of the present invention more clear, the technical solutions of the present invention will be clearly and completely described below in conjunction with specific embodiments of the present invention and corresponding drawings. Obviously, the embodiments described are only some embodiments of the present invention, not all 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.

[0047] The technical solutions provided by various embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0048] Example 1

[0049] Please refer to Figure 1 This embodiment provides a non-interferometric, non-iterative complex amplitude reading method, the purpose of which is to use a single diffraction intensity image to realize a method for quickly reading amplitude information and phase information by directly reconstructing the complex amplitude. The method may include the following steps:

[0050] Step S01, diffracting a light beam containing amplitude information and phase information to obtain an intensity image as a diffraction pattern with light intensity variation;

[0051] Step S02: constructing and training a diffraction intensity-complex amplitude model based on the correlation between the diffraction pattern and the amplitude information and phase information, and applying the model to directly obtain the amplitude information and phase information of a new diffraction pattern.

[0052] The non-interferometric, non-iterative complex amplitude reading method of this embodiment is based on the following relationship between the diffraction pattern of light and amplitude and phase information: the light field contains both amplitude and phase information. For example, when amplitude and phase information are loaded onto an object surface, a certain light field distribution or light field pattern is generated. At this time, z = 0, where z is the propagation direction of the light field. When light propagates forward a preset distance, such as z = 2 mm, a diffraction pattern of the light field is generated at this location. This diffraction pattern is an intensity image determined by both the amplitude and phase of the object surface. The distribution of this diffraction pattern is determined by the diffraction behavior of the light field in air and is objectively unique. When calculating this diffraction distribution, many models constructed using approximate diffraction formulas can be proposed, such as the angular spectrum transmission model and the point spread function model. These models can all approximately describe the distribution of this diffraction pattern. By using a neural network to construct a diffraction intensity-complex amplitude model, the inverse correlation between the diffraction pattern and the amplitude and phase information is expressed. By training the neural network model with specific data, a neural network model that can accurately describe the intensity-complex amplitude relationship is obtained. For the new diffraction intensity image, the corresponding amplitude and phase information can be obtained by directly inputting it into the neural network model.

[0053] Among them, the angular spectrum transmission model describes the forward propagation process of light, while the neural network model describes the reverse propagation process, that is, the inverse process.

[0054] This embodiment can be specific. In step S01, a light beam containing amplitude information and phase information is propagated for a preset distance d and then incident on the photodetector 19. Through the diffraction effect of light, the intensity image received by the photodetector 19 is a diffraction pattern that produces light intensity changes.

[0055] Please refer to Figure 2 Specifically, in step S02, the diffraction intensity-complex amplitude model is constructed as a neural network model.

[0056] Among them, the neural network model can be learned and trained through the correspondence between multiple inputs and multiple outputs to establish the neural network model parameters of the diffraction intensity-complex amplitude model, so that the intensity map of the new diffraction pattern can be used as input to accurately predict the corresponding amplitude and phase information.

[0057] It can be seen that by training the neural network model, it is possible to restore complex amplitude information including amplitude information and phase information using a single captured diffraction intensity image I. The captured diffraction intensity image is a non-interference image.

[0058] Please refer to Figure 2 This embodiment may further include the following steps in step S01:

[0059] Step S11, experimental image generation: generate n amplitude A images and phase P images of different modes, where n≥1, and n is a positive integer; generally speaking, the larger n is, the more data there is, that is, the more images there are, and the greater the difference between the data, that is, the greater the difference between the images, the better the reading result.

[0060] Step S12, experimental image capture: amplitude and phase modulate the light beam using a complex amplitude image dataset C containing an amplitude A image and a phase P image, and capture a diffraction intensity image I corresponding to the amplitude and phase modulation;

[0061] In step S02, the following steps are included:

[0062] Step S21, data set preparation: combine the diffraction intensity image I and the complex amplitude image data set C into a data set D IC , the dataset D IC The neural network training data set T can be divided into mutually exclusive sets according to the preset ratio IC and validation dataset V IC , respectively used for training and verification of the neural network model CNN;

[0063] Step S22, model building: establishing the diffraction intensity-complex amplitude model consistent with the neural network model CNN. That is, establishing the neural network model CNN to describe the diffraction intensity-complex amplitude relationship.

[0064] This embodiment may further include, in step S11 , the amplitude A image and the phase P image being randomly coded amplitude images (a) and phase images (b).

[0065] The randomly coded amplitude image (a) and the phase image (b) may be, for example, multi-grayscale randomly coded amplitude image (a) and phase image (b) within 16 levels, for data storage.

[0066] Alternatively, in step S11 , the amplitude A image and the phase P image are natural visual images.

[0067] Among them, natural visual images are used for complex amplitude image reconstruction.

[0068] Please refer to Figure 3 This embodiment can further, in step S12, load the complex amplitude image data set C containing the amplitude A image and the phase P image into the spatial light modulator to perform amplitude and phase modulation on the laser beam, and capture the diffraction intensity image I under the corresponding amplitude and phase modulation through the photodetector 19.

[0069] The laser beam may be a single-wavelength laser beam.

[0070] Specifically, two spatial light modulators may be used to upload the amplitude A image and the phase P image respectively, so as to modulate the amplitude and phase of the laser beam.

[0071] At this time, if the complex amplitude modulated laser beam is directly detected according to the existing technology, only the amplitude information can be detected, but the phase information cannot be directly detected.

[0072] In an embodiment of the present invention, a laser beam containing amplitude information and phase information propagates a preset distance d in free space, and then is incident on an image capturing device such as a photodetector 19. Due to the diffraction of light, the photodetector 19 receives a diffraction intensity image I that produces a change in light intensity. The diffraction pattern of the diffraction image is related to the amplitude image information and phase image information of the laser beam. Then, a diffraction intensity-complex amplitude neural network model is established and trained using the correlation between the diffraction image and the amplitude image information and the phase image information. For a new diffraction image, the hidden amplitude information and phase information can be directly obtained by inputting the trained neural network model.

[0073] Please refer to Figure 4 , the amplitude image (a) and the phase image (b) are combined to form a complex amplitude distribution, which is transmitted in the direction of the optical axis, or diffracted a preset distance in the air to obtain a diffraction intensity image (c).

[0074] This embodiment may further include the following steps after step S22:

[0075] Step S221, model optimization: set the loss function L of the neural network model CNN, and train the data set T through the neural network. IC Train the parameters of the neural network model CNN until the loss function L converges.

[0076] It should be noted that step S221 trains the internal parameters of the neural network model CNN.

[0077] This embodiment may further include the following steps after step S221:

[0078] Step S222, model verification: through the verification data set V IC Verify the model generalization performance of the neural network model CNN and obtain the generalized neural network model CNN.

[0079] Here, generalization means that the neural network can be applied to unknown diffraction intensity-complex amplitude correlations.

[0080] This embodiment may further include the following steps in step S02, after step S22, when obtaining the amplitude information and the phase information:

[0081] Step S23, model application: input the new diffraction intensity image into the trained and verified neural network model CNN, and output the amplitude A image and the phase P image.

[0082] The diffraction intensity image may be any new image.

[0083] It can be seen that the above embodiment can realize a single diffraction intensity image, input it into the trained deep neural network model CNN, directly obtain the corresponding amplitude and phase information, and realize the direct reading of amplitude information and phase information in the complex amplitude image.

[0084] Figure 5 In the U-net neural network model shown, the numbers below each layer, such as 384*384, 192*192, etc., represent the pixel size of the image in that neural network layer.

[0085] The numbers above each layer, such as 64, 128, etc., indicate the number of channels in the output image after convolution (which corresponds to the number of convolution kernels in the previous step).

[0086] 3*3 convolution ReLU means: the convolution kernel size is 3*3, and after convolution of the input, the output is sent to the ReLU activation function for activation.

[0087] 1*1 convolution Sigmoid means: the convolution kernel size is 1*1, and after convolution of the input, the output is activated by the Sigmoid activation function.

[0088] Sigmoid and ReLU are both activation functions in neural networks. Their function is to increase the nonlinear factors in the neural network and solve the defect of insufficient expression ability of linear models. The activation function ultimately determines the content to be transmitted to the next neuron.

[0089] Max pooling 2*2 means that for a 2×2 input, each element of the output is the maximum value of the element. After 2*2 max pooling, the height and width of the feature map are halved, while the number of channels remains unchanged.

[0090] Upsampling 2*2 means: enlarging the feature map, doubling the height and width of the feature map, and keeping the number of channels unchanged.

[0091] This embodiment can further include that in step S22, the neural network model can be established on the basis of an existing neural network model, for example, Figure 5 The U-net neural network model shown in FIG, and the intensity-amplitude model CNN1 and the intensity-phase model CNN2 have the same model structure.

[0092] At this point, the model optimization process of the U-net neural network model is as follows Figure 6As shown, that is, Figure 6 It shows the training and verification process of the intensity-amplitude model CNN1 and the intensity-phase model CNN2 of the U-net neural network model.

[0093] At this time, the model application process of the U-net neural network model is as follows Figure 7 As shown, that is, Figure 7 It shows the direct reading process of the amplitude A image and phase P image of the U-net neural network model.

[0094] Alternatively, the neural network model CNN adopts an unsupervised neural network model structure based on a combination with a physical optics diffraction model.

[0095] Alternatively, in step S22, the neural network model CNN adopts a data-driven end-to-end neural network model structure.

[0096] Since the U-net neural network model is a data-driven end-to-end neural network model structure, the neural network model CNN can adopt other data-driven end-to-end neural network model structures.

[0097] Alternatively, in step S22, the neural network model CNN includes an intensity-amplitude neural network model CNN1 and an intensity-phase neural network model CNN2, which are respectively used to restore the amplitude A image and the phase P image of the complex amplitude image.

[0098] Alternatively, the network structures of the intensity-amplitude neural network model CNN1 and the intensity-phase neural network model CNN2 may be the same or different.

[0099] Or we can go further and, in step S221, the hyperparameters of the intensity-amplitude neural network model CNN1 and the intensity-phase neural network model CNN2 are set to be the same or different.

[0100] Figure 8 In the single-input, dual-output neural network model shown, the numbers below each layer, such as 384*384, 192*192, etc., indicate the pixel size of the image at that neural network layer.

[0101] The numbers above each layer, such as 64, 128, etc., indicate the number of channels in the output image after convolution (which corresponds to the number of convolution kernels in the previous step).

[0102] 3*3 convolution ReLU means: the convolution kernel size is 3*3, and after convolution of the input, the output is sent to the ReLU activation function for activation.

[0103] 1*1 convolution Sigmoid means: the convolution kernel size is 1*1, and after convolution of the input, the output is activated by the Sigmoid activation function.

[0104] Sigmoid and ReLU are both activation functions in neural networks. Their function is to increase the nonlinear factors in the neural network and solve the defect of insufficient expression ability of linear models. The activation function ultimately determines the content to be transmitted to the next neuron.

[0105] Max pooling 2*2 means: for a 2×2 input, each element of the output is the maximum value of the element. After 2*2 max pooling, the height and width of the feature map are halved, while the number of channels remains unchanged.

[0106] Upsampling 2*2 means: enlarging the feature map, doubling the height and width of the feature map, and keeping the number of channels unchanged.

[0107] Alternatively, in step S22, the neural network model CNN is set as follows: Figure 8 The intensity-amplitude-phase neural network model CNN_12 shown has a single input of diffraction intensity image and dual outputs of amplitude A image and phase P image.

[0108] At this time, the intensity-amplitude-phase neural network model CNN_12, which has a single input of the diffraction intensity image and a dual output of the amplitude A image and the phase P image, has a model optimization process as follows: Figure 9 As shown, that is, Figure 9 It shows the training and verification process of the single-input, dual-output intensity-amplitude-phase neural network model CNN_12.

[0109] At this time, the model application process of the single-input, dual-output intensity-amplitude-phase neural network model CNN_12 is as follows Figure 10 As shown, that is, Figure 10 It shows the direct reading process of the amplitude A image and phase P image of the single-input, dual-output intensity-amplitude-phase neural network model CNN_12.

[0110] Example 2

[0111] Figure 4 The phase image and amplitude image shown are both four-grayscale randomly coded images, and the output of the optical system is a diffraction intensity image.

[0112] Figure 4 In the figure, 384*384 represents the image pixel size.

[0113] [0, 0.1, 0.4, 0.7, 1] represents the amplitude size after normalization, where 0 represents the amplitude of 0, that is, no light; 0.1, 0.4, 0.7, 1 represent the four-gray amplitude level encoding size.

[0114] [π / 6, 2π / 3, π, 3π / 2] represents four grayscale phase encoding values.

[0115] 10*10 represents pixels / data point, that is, each data point is represented by 10*10 pixels.

[0116] Non-interferometric, non-iterative complex amplitude readout methods, such as Figure 4 As shown, both the phase image and the amplitude image are four-grayscale randomly encoded images, and the system outputs a single diffraction intensity image. Two neural network models CNN1 and CNN2 are used to establish the intensity-amplitude neural network model and the intensity-phase neural network model respectively. The two neural network models are trained using the intensity-amplitude dataset and the intensity-phase dataset respectively. For the new diffraction intensity image, it is input into the trained intensity-amplitude neural network model CNN1 and the intensity-phase neural network model CNN2 respectively to obtain the corresponding amplitude information and phase information, thereby realizing complex amplitude decoding. The specific steps are as follows:

[0117] Step S11, generating experimental images: Figure 4 As shown, 10,000 randomly encoded four-grayscale amplitude images (a) and four-grayscale phase images (b) are generated.

[0118] Step S12, experimental image capture: 10,000 amplitude images (a) and phase images (b) are loaded into a spatial light modulator to perform amplitude modulation and phase modulation on the laser beam, and then the corresponding 10,000 diffraction intensity images (c) are captured using a photodetector 19; each diffraction intensity image (c) and its corresponding amplitude image (a) constitute an intensity-amplitude image pair, and each diffraction intensity image (c) and its corresponding phase image (b) constitute an intensity-phase image pair.

[0119] Step S21, data set preparation: 10,000 diffraction intensity images (c) and the corresponding 10,000 amplitude images (a) are used as the amplitude training data set D IA ; Among them, 9000 intensity-amplitude images are randomly selected as the amplitude training dataset T IA The remaining 1000 intensity-amplitude images are used as the amplitude verification dataset V IA ; Using T IA Train the amplitude neural network model CNN1 and use V IA Verify the generalization ability of the amplitude neural network model CNN1; use 10,000 diffraction intensity images (c) and the corresponding 10,000 phase images (b) as the phase training dataset D IP . Among them, 9000 intensity-phase images are randomly selected as the phase training dataset T IPThe remaining 1000 intensity-phase images are used as the phase verification dataset V IP ; Using T IP Train the phase neural network model CNN2 and use V IP The generalization ability of the phase neural network model CNN2 is verified.

[0120] Step S22, model building: select Figure 5 The U-net neural network model shown is used as the network model. The intensity-amplitude neural network model CNN1 and the intensity-phase neural network model CNN2 have the same model hyperparameters;

[0121] Step S221, model optimization: set the hyperparameters of the intensity-amplitude neural network model CNN1 and the intensity-phase neural network model CNN2 according to Table 1.

[0122] Table 1: Hyperparameter settings of neural network models CNN1 and CNN2

[0123] Loss Function MSE Number of training rounds 50 Learning rate <![CDATA[10 -4 ]]> Batch size 4 Optimizer Adam

[0124] The mean square error (MSE) is selected as the loss function of the intensity-amplitude neural network model CNN1 and the intensity-phase neural network model CNN2:

[0125]

[0126] Where W is the width of the input image and H is the height of the input image.

[0127] J is the training batch size, which refers to the batch size in deep learning training. The training batch size determines the number of samples used in a single training session. For example, if the training data size is the total number of samples, such as 9,000 training images, then the training batch size could be 4, meaning that 4 samples are taken from the 9,000 training images for each training session. The samples in the training batch can be randomly selected, for example, 4 samples are randomly selected from the 9,000 training images for each training session.

[0128] Output prediction value for the neural network model, C j (u,v) is the corresponding true value. We use the amplitude training data set T IA and phase training dataset T IP The amplitude neural network model CNN1 and the phase neural network model CNN2 are trained until the loss function MSE converges. IA and phase validation dataset V IPThe generalization ability of the amplitude neural network model CNN1 and the phase neural network model CNN2 is verified. Under the conditions of the amplitude neural network model CNN1 and the phase neural network model CNN2 and the hyperparameter settings, the training process of the amplitude neural network model CNN1 and the phase neural network model CNN2 is as follows: Figure 6 shown.

[0129] Step S23, model application: input any new diffraction intensity image I into the trained amplitude neural network model CNN1 and phase neural network model CNN2 respectively, and directly output the amplitude A image and phase P image to realize direct reading of complex amplitude. The complex amplitude reading process is as follows: Figure 7 shown.

[0130] Example 3

[0131] A non-interferential, non-iterative complex amplitude reading method, in which both the phase image and the amplitude image are four-grayscale randomly encoded images, and the system output is a single diffraction intensity image. A neural network model (CNN) is used to establish an intensity-amplitude-phase neural network model. The single-input, dual-output neural network model is trained using the intensity-amplitude-phase dataset. New diffraction intensity images are input into the trained intensity-amplitude-phase neural network model (CNN) to obtain the corresponding amplitude and phase information, thereby achieving complex amplitude decoding. The specific steps are as follows:

[0132] Step S11, generating experimental images: Figure 4 As shown, 10,000 randomly encoded four-grayscale amplitude images (a) and four-grayscale phase images (b) are generated.

[0133] Step S12, experimental image capture: 10,000 amplitude images (a) and phase images (b) are loaded into a spatial light modulator to perform amplitude and phase modulation on the light beam, and 10,000 corresponding diffraction intensity images (c) are captured using a photodetector 19; each diffraction intensity image (c) and its corresponding amplitude image (a) constitute an intensity-amplitude image pair, and each diffraction intensity image (c) and its corresponding phase image (b) constitute an intensity-phase image pair.

[0134] Step S21, data set preparation: 10,000 intensity images (c) are combined with the corresponding 10,000 amplitude images (a) and 10,000 phase images (b) to form 10,000 intensity-amplitude-phase images. 9,000 sets of intensity-amplitude-phase images are randomly selected as amplitude and phase training data T. IC The remaining 1000 sets of intensity-amplitude-phase images are used as the amplitude and phase verification dataset V IC ; Using T IC Train the single-input dual-output neural network model CNN using VIC The generalization ability of the single-input dual-output neural network model CNN is verified.

[0135] Step S22, model building: select Figure 8 The intensity-amplitude-phase neural network model CNN_12 shown with a single input of the diffraction intensity image and dual outputs of the amplitude A image and the phase P image is used as a single-input dual-output neural network model CNN.

[0136] Step S221, model optimization: set the single-input dual-output neural network model CNN hyperparameters according to Table 2.

[0137] Table 2: Neural Network Model CNN Hyperparameter Settings

[0138] Loss Function MSE Number of training rounds 50 Learning rate <![CDATA[10 -4 ]]> Batch size 4 Optimizer Adam

[0139] The mean square error (MSE) is selected as the loss function of the single-input dual-output neural network model CNN:

[0140] MSE=MSE A +λMSE P

[0141]

[0142] Where W is the width of the input image, H is the height of the input image, and J is the training batch size; Output prediction value for the neural network, A j (u,v) is the true value; using the intensity-amplitude-phase training set T IC The single-input dual-output neural network model CNN is trained until the loss function MSE converges, and the intensity-amplitude-phase validation dataset V is used. IC The generalization ability of the single-input dual-output neural network model CNN was verified; under the conditions of this model and hyperparameter settings, the model training process is as follows Figure 9 shown.

[0143] Among them, the loss function is a combination of amplitude and phase, so the true value is neither a simple phase nor a simple amplitude, but amplitude + λ × phase.

[0144] Step S23, model application: input any new diffraction intensity image I into the trained amplitude neural network model CNN, directly output the amplitude A image and the phase P image, and realize the direct reading of the complex amplitude. The complex amplitude reading process is as follows: Figure 10 shown.

[0145] Example 4

[0146] Please refer to Figure 11This embodiment provides a non-interference, non-iterative complex amplitude reading device for implementing the non-interference, non-iterative complex amplitude reading method described in any one of Embodiments 1 to 3 of the present invention.

[0147] The non-interference, non-iterative complex amplitude reading device includes an optical system and an electronic device. The optical system includes a laser 1, a beam paralleling component, a first 1 / 2 wave plate 4, an aperture 5, a first imaging component, a first polarizer 8 and a transmissive-reflective first beam splitter 9 arranged in sequence along the propagation direction of the incident light beam; an amplitude spatial light modulator 10 is arranged in the propagation direction of the transmitted light beam of the first beam splitter 9, and a second polarizer 11, a second imaging component, a second 1 / 2 wave plate 14 and a transmissive-reflective second beam splitter 15 are arranged in sequence in the propagation direction of the reflected light beam of the first beam splitter 9; a phase spatial light modulator 16 is arranged in the propagation direction of the transmitted light beam of the second beam splitter 15, and a third imaging component and a photodetector 19 are arranged in sequence in the propagation direction of the reflected light beam of the second beam splitter 15; the electronic device includes one or more processors and a memory, and one or more computer programs are stored in the memory. When the one or more processors receive the diffraction pattern captured by the photodetector 19 and execute the one or more computer programs, the non-interference, non-iterative complex amplitude reading method steps of any one of embodiments 1 to 3 of the present invention are implemented.

[0148] The electronic device in the non-interferometric, non-iterative complex amplitude reading device can be a computer. The computer generates amplitude and phase images of the complex amplitude, which are then captured by the optical system to obtain the corresponding experimental diffraction intensity image, thereby completing step S01 of the non-interferometric complex amplitude reading method. The computer then completes step S02 of the non-interferometric, non-iterative complex amplitude reading method.

[0149] In actual applications, in the neural network model application stage after the neural network model training and verification are completed, the diffraction intensity map is still captured by the optical system and then input into the neural network model for complex amplitude reconstruction; that is, both computers and optical systems are needed.

[0150] Specifically, the beam parallelization component includes a pinhole filter 2 and a collimating lens 3 arranged in sequence along the propagation direction of the incident light beam. The function of the aperture 5 is to control the size of the beam diameter. The first imaging component includes a first relay lens 6 and a second relay lens 7 arranged in sequence along the propagation direction of the light beam. The first relay lens 6 and the second relay lens 7 constitute a 4f system. The first polarizer 8 is a horizontal polarizer. The first beam splitter 9 is a polarizing beam splitter and a non-polarizing stereo beam splitter. The second polarizer 11 is a vertical polarizer. The second imaging component includes a third relay lens 12 and a fourth relay lens 13 arranged in sequence along the propagation direction of the light beam. The third relay lens 12 and the fourth relay lens 13 constitute a 4f system. The second beam splitter 15 is a non-polarizing stereo beam splitter. The third imaging component includes a fifth relay lens 17 and a sixth relay lens 18 arranged in sequence along the propagation direction of the light beam. The fifth relay lens 17 and the sixth relay lens 18 constitute a 4f system.

[0151] The specific working process of the non-interference, non-iterative complex amplitude reading device may be:

[0152] The laser 1 emits a laser, for example, a green laser with a wavelength of 532 nm. After passing through the pinhole filter 2 and the collimating lens 3, the laser is converted into parallel light with good beam quality. After the parallel light passes through the first 1 / 2 wave plate 4 and the aperture 5, the beam cross-section of the parallel light is converted from a circular shape to the aperture shape of the aperture 5. The first relay lens 6 and the second relay lens 7 form a 4f system, the function of which is to image the aperture 5 onto the plane where the spatial light modulator 10 is located. After the above-mentioned laser beam passes through the first polarizer 8 and the first beam splitter 9 and is incident on the spatial light modulator 10, its reflected light passes through the first beam splitter 9 again and is reflected in a direction perpendicular to the original optical path. The spatial light modulator 10 can be an amplitude-type spatial light modulator, or it can be a phase-type spatial light modulator combined with the first polarizer 8 and the second polarizer 11 whose polarization directions are perpendicular to each other, so as to achieve amplitude modulation of the laser beam. Specifically, the spatial light modulator 10 uploads a specific amplitude image A. After the laser beam is incident on the spatial light modulator 10, it is reflected and passes through the second polarizer 11, where it acquires accurate amplitude information. The laser beam carrying amplitude information then passes through the third relay lens 12 and the fourth relay lens 13. These lenses also form a 4f system, which images the amplitude spatial light modulator 10 onto the plane of the second spatial light modulator 16. After passing through the 4f system formed by the third and fourth relay lenses 12 and 13, the laser beam continues through the second half-wave plate 14 and the second beam splitter 15, and is incident on the second beam splitter 15. The second half-wave plate 14 adjusts the polarization state of the laser beam to meet the polarization state requirements of the phase-type spatial light modulator 16. The phase spatial light modulator 16 uploads a specific phase image P to achieve phase modulation of the laser beam. The laser beam with amplitude and phase information reflected from the phase-type spatial light modulator 16 is reflected in a direction perpendicular to the original optical path after passing through the second beam splitter 15, and then enters the photodetector 19 through the fifth relay lens 17 and the sixth relay lens 18. The fifth relay lens 17 and the sixth relay lens 18 form a 4f system, and their function is to accurately image the plane where the second beam splitter 15 is located onto the back focal plane of the sixth relay lens 18. Along the propagation direction of the laser beam, the photodetector 19 is located on the plane behind the back focal plane of the sixth relay lens 18. Therefore, the light beam located at the back focal plane of the sixth relay lens 18 has precise amplitude and phase information. After the laser beam continues to propagate for a preset distance, the diffracted light enters the photodetector 19 and is received by the photodetector 19 to obtain a diffraction pattern with varying light intensity, namely the diffraction intensity image I, or Figure 4 and Figure 5 Diffraction intensity image in (c).

[0153] The non-interference, non-iterative complex amplitude reading method and device of the embodiment of the present invention can further improve the speed and accuracy of amplitude and phase reading and simplify the device by: building an optical system for non-interference lensless complex amplitude diffraction reconstruction, using an electronic device such as a computer to generate amplitude and phase images, uploading the amplitude and phase to the optical system to obtain a diffraction intensity image to establish a neural network data set, building a neural network model structure and setting corresponding parameters, using the data set to train the neural network model and verify its generalization, and inputting any diffraction image into the neural network model to directly output amplitude and phase images.

[0154] The above are merely embodiments of the present invention and are not intended to limit the present invention. It will be apparent to those skilled in the art that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention are intended to be included within the scope of the claims of the present invention.

Claims

1. A non-interferometric, non-iterative complex amplitude reading device, characterized in that: The reading device includes an optical system and an electronic device, wherein the optical system includes a laser, a beam paralleling component, a first 1 / 2 wave plate, an aperture, a first imaging component, a first polarizer, and a first transmissive-reflective beam splitter arranged in sequence along the propagation direction of the incident light beam; an amplitude spatial light modulator is arranged in the propagation direction of the transmitted light beam of the first beam splitter, and a second polarizer, a second imaging component, a second 1 / 2 wave plate, and a second transmissive-reflective beam splitter are arranged in sequence in the propagation direction of the reflected light beam of the first beam splitter; a phase spatial light modulator is arranged in the propagation direction of the transmitted light beam of the second beam splitter, and a third imaging component and a photodetector are arranged in sequence in the propagation direction of the reflected light beam of the second beam splitter; the electronic device includes one or more processors and a memory, wherein one or more computer programs are stored on the memory, and when the one or more processors receive the diffraction pattern captured by the photodetector and execute the one or more computer programs, the steps in the non-interference, non-iterative complex amplitude reading method are implemented: Step S01, diffracting a light beam containing amplitude information and phase information to obtain an intensity image as a diffraction pattern with light intensity variation; Step S02: constructing and training a diffraction intensity-complex amplitude model based on the correlation between the diffraction pattern and the amplitude information and the phase information, and applying the diffraction intensity-complex amplitude model to a new diffraction pattern to directly obtain the amplitude information and the phase information.

2. The non-interferometric, non-iterative complex amplitude reading device according to claim 1, wherein: In step S02, learning and training are performed through the correspondence between multiple inputs and multiple outputs to establish the neural network model parameters of the diffraction intensity-complex amplitude model.

3. The non-interferometric, non-iterative complex amplitude reading device according to claim 1, wherein: In step S01, the following steps are included: Step S11, generating experimental images: generating n amplitude A images and phase P images of different modes, where n≥1, and n is a positive integer; Step S12, experimental image capture: performing amplitude and phase modulation on the light beam using the complex amplitude image dataset C including the amplitude A image and the phase P image, and capturing a diffraction intensity image I corresponding to the amplitude and phase modulation; In step S02, the following steps are included: Step S21, data set preparation: the diffraction intensity image I and the complex amplitude image data set C are combined into a data set DIC, and the data set DIC is divided into a mutually exclusive neural network training data set TIC and a verification data set VIC, which are used for training and verification of a neural network model CNN, respectively; Step S22, model building: establishing the diffraction intensity-complex amplitude model consistent with the neural network model CNN.

4. The non-interference, non-iterative complex amplitude reading device according to claim 3, characterized in that: After step S22, the following steps are also included: Step S221, model optimization: setting the loss function L of the neural network model CNN, and training the parameters of the neural network model CNN using the neural network training data set TIC until the loss function L converges.

5. The non-interference, non-iterative complex amplitude reading device according to claim 4, characterized in that: After step S221, the following steps are further included: Step S222, model verification: verifying the model generalization performance of the neural network model CNN through the verification data set VIC to obtain the generalized neural network model CNN.

6. The non-interference, non-iterative complex amplitude reading device according to any one of claims 3 to 5, characterized in that: In step S02, after step S22, when obtaining the amplitude information and the phase information, the following steps are included: Step S23, model application: input the new diffraction intensity image into the trained and verified neural network model CNN, and output the amplitude A image and the phase P image.

7. The non-interference, non-iterative complex amplitude reading device according to any one of claims 3 to 5, characterized in that: In step S11 , the amplitude A image and the phase P image are randomly coded amplitude images and phase images, or the amplitude A image and the phase P image are natural visual images.

8. The non-interference, non-iterative complex amplitude reading device according to any one of claims 3 to 5, characterized in that: In step S22, the neural network model CNN adopts an unsupervised neural network model structure based on a combination with a physical optics diffraction model, or the neural network model CNN adopts an end-to-end neural network model structure based on data-driven, or the neural network model CNN includes an intensity-amplitude neural network model CNN1 and an intensity-phase neural network model CNN2, which are respectively used to restore the amplitude A image and the phase P image of the complex amplitude image, or the neural network model CNN is set to an intensity-amplitude-phase neural network model CNN_12 with a single input of the diffraction intensity image and dual outputs of the amplitude A image and the phase P image.

9. The non-interferometric, non-iterative complex amplitude reading device according to claim 1, wherein: The beam parallelization component comprises a pinhole filter and a collimating lens which are sequentially arranged along the propagation direction of the incident beam.

Images