U-NET network for orthogonal phase gradient prediction and training method thereof

Abstract

The invention relates to a training method of a U-NET network for orthogonal phase gradient prediction. The training method comprises the following steps: S21, collecting phase diagrams of different samples; s22, expanding the number of phase diagrams; s23, respectively extracting the phase gradient of each phase diagram along the x direction and the phase gradient of each phase diagram along the y direction; and S24, inputting the phase gradient sheared along the x direction into a U-NET network, and learning a nonlinear mapping relation between orthogonal phase gradients. The U-NET network for orthogonal phase gradient prediction has the advantage that the phase gradient in the orthogonal direction can be predicted quickly and precisely through the phase gradient in one direction.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of microscopy, in particular to a U-NET network for orthogonal phase gradient prediction and a training method thereof. BACKGROUND

[0002] Differential interference contrast (DIC) microscopy is an optical microscopy technique based on the principle of partial coherent light shearing interference, which can encode the phase gradient information of a transparent sample into intensity information to realize high-contrast imaging of the transparent sample. DIC microscopy uses a pair of Wollaston prisms or Nomarski prisms to encode the phase gradient information of a sample into intensity information. The incident polarized light is split into two beams of orthogonal polarized light that are slightly offset from each other by a prism, and the offset direction is called the shearing direction, and the offset distance is called the shearing amount. The two beams of polarized light carry different position information of the sample due to their different spatial positions. The two beams of light carrying different position information are recombined by a second prism, and after passing through a polarizer, interference occurs to form a differential interference image with light and dark changes, which represent changes in interference intensity. The phase gradient of the sample along the shearing direction can be obtained by the differential interference image, reflecting the phase change of the sample along the direction.

[0003] Common DIC microscopy techniques can only have a single shearing direction, and cannot simultaneously obtain phase gradients in different shearing directions, limiting the observation of specific information of the sample along other shearing directions. In order to obtain phase gradients in multiple shearing directions, the sample or the shearing prism is usually rotated. However, the method of rotating the sample or the shearing prism introduces alignment problems of sample information on the pixels or causes image defocusing due to mechanical vibration, which seriously affects high-precision phase gradient imaging.

[0004] In order to be able to obtain orthogonal phase gradients in one imaging, a chessboard grating, a spatial light modulator and an off-axis digital holography method have been further designed. However, the imaging resolution of the orthogonal phase gradient obtained by the chessboard grating is reduced by half compared to the image in a single shearing direction, affecting the high-precision imaging of the phase gradient; the method of realizing orthogonal phase gradient acquisition using a spatial light modulator has a complex optical structure, and sacrifices the time resolution, which requires re-setting the image generated by the spatial light modulator or re-adjusting the control voltage of the liquid crystal, affecting the rapid imaging of the phase gradient; and the off-axis digital holography system is complex, has high requirements for environmental stability, and is complex to operate, and the imaging speed and data processing speed are slow. SUMMARY

[0005] Based on this, the purpose of the present application is to obtain orthogonal phase gradients by imaging only once, and to provide a training method of a U-NET network for orthogonal phase gradient prediction.

[0006] The training method of the U-NET network for orthogonal phase gradient prediction comprises the following steps:

[0007] S1: Collecting phase maps of different samples; S2: Expanding the number of phase maps; S3: Extracting the phase gradient along the x direction and the phase gradient along the y direction of each phase map, respectively; S4: Inputting the phase gradient along the x direction into the U-NET network to learn the nonlinear mapping relationship between the orthogonal phase gradients; The training target of the U-NET network for orthogonal phase gradient prediction is: , Wherein, is a loss function, is a weight parameter of the network; N represents the total number of training samples; represents a phase gradient Fourier domain loss function, represents the y direction phase gradient output by the network, represents the true value label of the y direction phase gradient, represents a TV regularization term, which is used to accelerate network convergence and alleviate network overfitting problems, represents a weight parameter of the regularization term.

[0008] Compared with the prior art, the present application continuously trains and adjusts according to the phase gradient along the x direction and the phase gradient along the y direction, and the set training target, and finally obtains a U-NET network model capable of predicting the phase gradient along the y direction according to the phase gradient along the x direction; the trained U-NET network can predict the phase gradient along the y direction perpendicular to the x direction according to the phase gradient along the x direction, and has high prediction accuracy, that is, even in conventional differential interference phase contrast microscopy, orthogonal images can be obtained by imaging only once, without the need to increase additional components, thereby avoiding the reduction of precision and resolution caused by additional components; at the same time, the prediction process consumes less resources and is fast, and combined with differential interference phase contrast microscopy, it can simultaneously have the advantages of simplicity, speed and high imaging accuracy.

[0009] Further, the phase map is collected by a digital holographic microscope; and the step S3 specifically comprises: S3a1: Fourier shifting each of the expanded phase maps to obtain phase images shifted along the x direction in the spatial domain.

[0010] S3a2: obtaining the phase gradient along x direction by translating the two phase patterns along x direction.

[0011] S3a3: adding random noise in the phase gradient along x direction.

[0012] S3b1: obtaining the phase image translated along y direction by performing Fourier shift on each of the phase patterns.

[0013] S3b2: obtaining the phase gradient along y direction by translating the two phase patterns along y direction.

[0014] Further, the phase patterns are collected by digital holographic microscope. The phase patterns obtained by the off-axis digital holographic interference optical path contain more complete phase information, and high-precision phase gradient along x direction and phase gradient along y direction can be separated by processing, which is beneficial to the training of U-NET network. Further, the sample region of the phase gradient along x direction is segmented, and the random noise is added to the background region.

[0015] Further, the random noise added in step S23a3 is Gaussian random noise with a standard deviation of 0.05.

[0016] Further, in step S23, the phase gradient along x direction and the corresponding phase gradient along y direction are taken as a group of data sets, and each group of data sets is in a ratio of 8:1:1.

[0017] Further, the U-NET network comprises a down-sampling convolution module processing the phase gradient along the x direction as an input image, and an up-sampling convolution module taking the result of the down-sampling convolution module as an input image; the down-sampling convolution module comprises 6 down-sampling convolution units successively performing down-sampling convolution, and each down-sampling convolution unit performs down-sampling convolution taking the sampling result of the previous down-sampling unit as an input, except the down-sampling convolution unit taking the phase gradient along the x direction as an input image; the 6 down-sampling convolution units perform 6 times of down-sampling convolution, respectively obtaining a first down-sampling convolution image to a sixth down-sampling convolution image; the down-sampling convolution reduces the spatial resolution of the input image and increases the number of channels; the up-sampling convolution module comprises 6 up-sampling convolution units successively performing up-sampling convolution, each up-sampling convolution unit comprising a transposed convolution layer and a skip connection layer, the transposed convolution layer performing transposed convolution to increase the spatial resolution of the input image and reduce the number of channels, and the skip connection layer splicing the image of the transposed convolution layer with a down-sampling convolution image to obtain an up-convolution image of this time of up-sampling convolution; wherein the transposed convolution image of the first up-sampling convolution unit is spliced with the sixth down-sampling convolution image, the transposed convolution image of the second up-sampling convolution unit is spliced with the fifth down-sampling convolution image, and the transposed convolution images of the third up-sampling convolution unit to the sixth up-sampling convolution unit are sequentially spliced with one of the fourth down-sampling convolution image to the first down-sampling convolution image, respectively.

[0018] Further, the down-sampling convolution performs down-sampling convolution with a stride of 2 and a convolution kernel size of 4*4, and performs batch normalization processing and Relu activation function processing after each convolution; the convolution layer of the up-sampling convolution module performs inverted convolution with a stride of 2 and a convolution kernel size of 4*4.

[0019] Based on the same inventive concept, the application further comprises a U-NET network-based orthogonal phase gradient imaging method, which comprises: acquiring a phase shift differential interference image of a sample to be measured; extracting a phase gradient along the x direction in the phase shift differential interference image; inputting the phase gradient along the x direction into a U-NET network to obtain a phase gradient along the y direction.

[0020] Compared with the prior art, the U-NET network-based orthogonal phase gradient imaging method of the application can use the U-NET network to predict the phase gradient in the orthogonal direction after imaging the phase gradient in one direction, without the need to adjust the cell sample or the microscopic imaging device and then re-shoot, which can effectively reduce the operation complexity and the required time, and avoid reducing the accuracy and clarity of the orthogonal image due to adjustment of the device.

[0021] Based on the same inventive concept, the present invention also includes an electronic device comprising a processor; a memory for storing a computer program executed by the processor; wherein, when the processor executes the computer program, it implements the U-NET network for orthogonal phase gradient prediction and / or the training method for the U-NET network for orthogonal phase gradient prediction as described above.

[0022] To better understand and implement this invention, the following detailed description is provided in conjunction with the accompanying drawings. Attached Figure Description

[0023] Figure 1 The flowchart shows the orthogonal phase gradient imaging method based on U-Net network. Figure 2 This is a schematic diagram of the structure of the U-NET network used for orthogonal phase gradient prediction; Figure 3 This is a flowchart illustrating the training process of the U-NET network used for orthogonal phase gradient prediction. Figure 4 The input is the U-NET network, and the phase gradient images are obtained and compared. Detailed Implementation

[0024] The inventors discovered that in orthogonal phase gradient imaging, the orthogonal phase gradients share similarities in image features and that the images are correlated in phase. Addressing the problems of low resolution, slow imaging, and defocusing in current orthogonal phase gradient imaging, this invention designs a U-NET network for orthogonal phase gradient prediction. In orthogonal phase gradient imaging, it predicts the phase gradient in another orthogonal direction based on the phase gradient in one direction, thereby obtaining the phase gradient along the x-direction and the phase gradient along the y-direction in a single differential interferometric phase-contrast imaging, achieving fast, high-precision, and high-resolution orthogonal phase gradient imaging.

[0025] Please see Figure 1 , Figure 1 This is a flowchart of an orthogonal phase gradient imaging method based on a U-Net network. An orthogonal phase gradient imaging method based on a U-Net network includes: S11: Acquire the phase-shift differential interferometry (PII) pattern of the sample under test. A four-step phase-shift differential interferometry pattern with phase shifts of 0, π / 2, π, and 3π / 2 is acquired using differential interferometry phase-contrast microscopy. In differential interferometry phase-contrast microscopy, two orthogonally polarized beams are introduced into a small spatial displacement by passing through the first polarizing beam splitter. This displacement is typically smaller than the Airy disk diameter, producing two image field replicas. and These two image fields, after passing through a second polarizing beam splitter and a polarizer, coherently superimpose to form a differential interference pattern, the intensity of which can be expressed as: , wherein, and are the amplitudes of the two image fields, and are the phases of the two image fields, denotes the phase gradient along the x direction, is the n-th DIC interferogram captured, represents the corresponding phase delay introduced by the phase shifter or prism, .

[0026] S12: Extracting the phase gradient along the x direction in the phase-shifted differential interference figure. The phase gradient along the x direction is extracted by using a four-step phase shift algorithm on the differential interference figure images of 0, π / 2, π, 3π / 2, and the formula is: , wherein, is the small displacement introduced in space by the two orthogonally polarized lights, is the n-th DIC interferogram captured, represents the corresponding phase delay introduced by the phase shifter or prism, .

[0027] S13: Inputting the phase gradient along the x direction into the U-NET network to obtain the phase gradient along the y direction. The U-NET network predicts the phase gradient along the y direction according to the phase gradient along the x direction.

[0028] Please refer to Figure 2 , Figure 2 is a structural diagram of the U-NET network for orthogonal phase gradient prediction. The U-NET network for orthogonal phase gradient prediction includes a down-sampling convolution module and an up-sampling convolution module. The down-sampling convolution module processes the phase gradient along the x direction as an input image to extract high-dimensional image feature information. The down-sampling convolution module processes the convolution result of the down-sampling convolution module as an input to decode the feature information of the high-dimensional image.

[0029] The down-sampling convolution module adopts multiple convolution units with the same structure to perform continuous multiple times of down-sampling convolution on the input image; in this embodiment, a total of 6 times of down-sampling convolution are performed continuously. The down-sampling convolution module includes a first down-sampling convolution unit 11, a second down-sampling convolution unit 12, a third down-sampling convolution unit 13, a fourth down-sampling convolution unit 14, a fifth down-sampling convolution unit 15, and a sixth down-sampling convolution unit 16, and the six down-sampling convolution units have the same structure; each of the down-sampling convolution units includes two convolution layers, wherein the first layer is a compression expansion convolution layer, which performs convolution on the input image with a stride of 2 and a convolution kernel size of 4x4 to obtain a first layer convolution image, realizes halving of the spatial resolution and doubling of the channel number of the input image, reduces spatial redundancy information, expands the receptive field, and better captures global semantic information; the second layer is a normalization activation layer, which performs batch normalization processing and Relu activation function processing on the first layer convolution result to obtain a down-sampling image. The first down-sampling convolution unit 11a is the first processing stage of the down-sampling convolution module, which performs down-sampling convolution, batch normalization processing and Relu activation function processing on the input image respectively to obtain and save a first down-sampling convolution image; the second down-sampling convolution unit 12a is the second processing stage of the down-sampling convolution module, which takes the first down-sampling image as input to obtain a second down-sampling convolution image; the third down-sampling convolution unit 13a to the sixth down-sampling unit 16a take the down-sampling convolution image obtained in the previous stage as input to obtain a down-sampling convolution image in the current stage as a third down-sampling convolution image to a sixth down-sampling convolution image; the down-sampling convolution images obtained in each processing stage are saved in the order of the first down-sampling convolution image to the sixth down-sampling convolution image. The batch normalization is to calculate the mean and variance of each feature map, and adjust the input data distribution to a standard normal distribution with a mean of 0 and a variance of 1. The rectified linear unit is a kind of activation function, which is used to introduce nonlinearity and avoid overfitting. The processing of 6 stages in this embodiment can balance the training effect and the power consumption, because fewer stages will result in insufficient feature information extracted by the network, leading to poor training effect, and more stages will result in overfitting of the network and waste of computing resources.

[0030] The up-sampling convolution module adopts multiple convolution units with the same structure to perform continuous multiple times of up-sampling convolution on the input image; in this embodiment, a total of 6 times of up-sampling convolution are performed continuously. The up-sampling convolution module includes 6 identical up-sampling convolution units, which are a first up-sampling convolution unit 21, a second up-sampling convolution unit 22, a third up-sampling convolution unit 23, a fourth up-sampling convolution unit 24, a fifth up-sampling convolution unit 25, and a sixth up-sampling convolution unit 26. As Figure 2As shown, the first up-sampling convolution unit 21 includes a transpose convolution layer 21a and a skip connection layer 21b. The transpose convolution layer 21a performs transpose convolution on the input image with a stride of 2 and a kernel size of 4x4 to obtain a first transpose image, which doubles the spatial resolution of the input image while reducing the number of channels by half, thereby improving the image resolution and reducing resource consumption and processing time. The skip connection layer 21b splices the first transpose image with the down-sampled convolution image of the sixth down-sampling convolution unit to obtain a first fusion image. The second up-sampling convolution unit 22 to the sixth up-sampling convolution unit 26 have the same structure as the first up-sampling convolution unit and each include a transpose convolution layer and a skip connection layer. The second up-sampling convolution unit processes the first fusion image as input and splices the fifth down-sampling convolution image obtained by the fifth down-sampling convolution to obtain a second fusion image. The third up-sampling convolution unit processes the second fusion image as input and splices the fourth down-sampling convolution image obtained by the fourth down-sampling convolution to obtain a third fusion image. The fourth up-sampling convolution unit processes the third fusion image as input and splices the fourth down-sampling convolution image obtained by the fourth down-sampling convolution to obtain a fourth fusion image. The fifth up-sampling convolution unit processes the fourth fusion image as input and splices the second down-sampling convolution image obtained by the second down-sampling convolution to obtain a fifth fusion image. The sixth up-sampling convolution unit processes the fifth fusion image as input and splices the first down-sampling convolution image obtained by the first down-sampling convolution to obtain a sixth fusion image. The sixth fusion image obtained by the sixth up-sampling convolution unit is the predicted gradient.

[0031] Please refer to Figure 3 , Figure 3 is a flowchart of the training process of the U-NET network for orthogonal phase gradient prediction. The training process of the U-NET network includes: S21: Collecting phase maps of different samples. The phase map of the cell sample is obtained by using an off-axis digital holographic interference imaging device. The off-axis digital holographic interference imaging device records the interference fringes formed by the object light modulated by the cell sample and the reference light not irradiated on the cell sample, extracts the phase information of the cell by combining the digital reconstruction algorithm, and obtains the phase map of the cell sample. The phase map obtained by the off-axis digital holographic interference optical path contains more complete phase information, and high-precision phase gradients along the x direction and phase gradients along the y direction can be separated by processing. High-precision phase gradients are beneficial to the training of the U-NET network. In this embodiment, the off-axis digital holographic interference optical path with a Mach-Zehnder structure is adopted, the reference cell sample is a HeLa cell, and 1700 phase maps of different HeLa cell samples are collected. In the training process, the off-axis digital holographic interference optical path can obtain a phase map containing more complete phase information and higher precision, which is beneficial to improve the training precision in the U-NET network training process and ensure the prediction accuracy of the network. At the same time, the U-NET network trained based on the phase map with higher precision and more complete information is applied to differential interference phase contrast microscopy, which can obtain clear orthogonal phase gradients without reducing the resolution on the basis of the simple operation characteristics of the imaging, without the need for complicated operations of off-axis digital holographic interference imaging.

[0032] S22: Expanding the number of phase maps. New phase maps are generated by rotating the phase maps, and the reference data is expanded to reduce the number of samples to be collected. In this embodiment, the 1700 phase maps are respectively copied and rotated by 90 degrees, 180 degrees and 270 degrees, and 6800 phase maps are obtained by expansion.

[0033] S23: Extracting the phase gradient along the x direction and the phase gradient along the y direction of each phase map, respectively. Specifically, step S3 specifically includes: S23a1: Fourier shifting each of the expanded phase maps to obtain a phase image shifted along the x direction in the spatial domain.

[0034] S23a2: Obtaining the phase gradient along the x direction by shifting two phase maps along the x direction. The formula for obtaining the phase gradient along the x direction is: , wherein, is the phase gradient, is the original phase map, is the phase map shifted along the x direction, is the distance shifted along the x direction.

[0035] S23a3: Adding random noise to the phase gradient sheared along the x direction. The noise level of digital holographic microscope imaging is inconsistent with that of differential interference contrast microscope. In order to enable the subsequently trained model to effectively predict the interference pattern obtained under the differential interference contrast microscope, random noise is added to the phase gradient to simulate the noise level of the interference pattern of the differential interference contrast microscope. In this embodiment, the sample area of the phase gradient along the x direction is segmented, and Gaussian random noise with a standard deviation of 0.05 is added to the background area.

[0036] S23b1: Fourier shift is performed on each of the expanded phase patterns to obtain phase images translated along the y direction in the spatial domain.

[0037] S23b2: The phase gradient sheared along the y direction is obtained by translating the two phase patterns before and after along the y direction. The phase gradient sheared along the y direction obtained by the phase pattern is the true value gradient. The formula for obtaining the phase gradient along the y direction is: is the phase gradient, is the original phase pattern, is the phase pattern translated along the y direction, is the distance of translation along the y direction.

[0038] S24: Inputting the phase gradient sheared along the x direction into the U-NET network to learn the nonlinear mapping relationship between the orthogonal phase gradients. The U-NET network adopts a minimum network output, predicts the phase gradient sheared along the y direction from the phase gradient sheared along the x direction as the predicted gradient, and calculates the difference between the predicted gradient and the true value gradient in the Fourier domain through a loss function. According to the difference, the network parameters are optimized through back propagation, and the loss function is minimized to improve the prediction accuracy. The training target of the U-NET network is: is the loss function, is the weight parameter of the network; N represents the total number of training samples; represents the phase gradient Fourier domain loss function, represents the y direction phase gradient output by the network, represents the true value label of the y direction phase gradient, represents the TV regularization term, which is used to accelerate network convergence and relieve network overfitting problem, represents the weight parameter of the regularization term. In this embodiment, the is set to e-7.

[0039] ​​​​Since the phase gradient reflects the degree of phase change of the sample and is mainly high-frequency information, the difference in the Fourier domain can better constrain the network output. In this embodiment, the phase gradient along the x-direction and the phase gradient along the y-direction obtained from the same augmented phase map are used as a dataset. The dataset is divided into a training set, a validation set, and a test set in a ratio of 8:1:1.

[0040] To verify the prediction accuracy of the trained U-NET network, validation set data was input into the U-NET network for prediction. The prediction result had a root mean square error of 1.9 mrad / μm compared with the official result, indicating that accurate prediction was possible.

[0041] Furthermore, its predictive capabilities were validated by inputting validation set data into the U-NET network. See also... Figure 4 , Figure 4 The input is the U-NET network and the obtained and compared phase gradient images, wherein, Figure 4 (a) is the input phase gradient image along the x-direction. Figure 4 (b) is the predicted gradient image along the y-direction. Figure 4 (c) is the ground truth gradient image along the y-direction obtained by orthogonal imaging. Figure 4 (d) is Figure 4 (a) is a magnified view of a portion of the image. Figure 4 (e) is Figure 4 (b) is a magnified view of a portion of the image. Figure 4 (f) is Figure 4 (c) is a magnified view of a section, where the arrows indicate the shearing direction. According to... Figure 4 As can be seen from the figures, the predicted phase gradient along the y-direction is highly consistent with the actual phase gradient obtained from the imaging along the y-direction. During the verification process, the root mean square error (RMSE), peak signal-to-noise ratio (PSNR), and structural similarity were used to evaluate the results of the training and test sets. Specifically, the training set had an RMSE of 0.591 mrad / μm, a PSNR of 64.98 dB, and a structural similarity of 0.9622; the test set had an RMSE of 1.174 mrad / μm, a PSNR of 60.73 dB, and a structural similarity of 0.9580. Compared to existing technologies, the U-NET network for orthogonal phase gradient prediction of this invention can predict the phase gradient in the y-direction, which is perpendicular to the x-direction, based on the phase gradient along the x-direction, and has high prediction accuracy. Even in conventional differential interferometric phase-contrast microscopy, orthogonal images can be obtained in a single imaging process without the need for additional components, thus avoiding the reduction in accuracy and resolution caused by additional components. At the same time, the prediction process consumes fewer resources and the imaging is faster. When combined with differential interferometric phase-contrast microscopy, it can simultaneously achieve the advantages of simplicity, speed, and high imaging accuracy.

[0042] The terminology used in the description of the application herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. As used in the description of the application and the appended claims, the singular forms "a", "an" and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise. It also will be understood that the term "or" as used herein encompasses the

[0043] The embodiments of the application described above are merely intended to express the several embodiments of the application, and the description is more specific and detailed, but it should not be understood as a limitation to the scope of the patent. It should be pointed out that for those skilled in the art, some modifications and improvements can be made without departing from the concept of the application, and the application also intends to include these modifications and improvements.

Claims

1. A training method of a U-NET network for quadrature phase gradient prediction, characterized in that, The method comprises the following steps: S21: collecting phase maps of different samples; S22: expanding the number of phase maps; S23: extracting phase gradients in the x direction and phase gradients in the y direction of each phase map, respectively; S24: inputting the phase gradients in the x direction into a U-NET network to learn a nonlinear mapping relationship between the orthogonal phase gradients; The training target of the U-NET network is: ， wherein, is a loss function, is a weight parameter of the network; N represents the total number of training samples; represents a phase gradient Fourier domain loss function, represents a y-direction phase gradient output by the network, represents a true value label of the y-direction phase gradient, represents a TV regularization term, which is used to accelerate network convergence and relieve network overfitting problems, represents a weight parameter of the regularization term.

2. The training method of the U-NET network for quadrature phase gradient prediction according to claim 1, characterized in that, The phase maps are collected by a digital holographic microscope; and the step S3 specifically comprises: S23a1: Fourier shifting each of the expanded phase maps to obtain phase images translated in the x direction in the spatial domain; S23a2: obtaining phase gradients in the x direction by using two phase maps before and after translation in the x direction; S23a3: adding random noise to the phase gradients in the x direction; S23b1: Fourier shifting each of the expanded phase maps to obtain phase images translated in the y direction in the spatial domain; S23b2: obtaining phase gradients in the y direction by using two phase maps before and after translation in the y direction.

3. The training method of the U-NET network for quadrature phase gradient prediction according to claim 2, characterized in that, The phase maps are collected by a digital holographic microscope.

4. The training method of the U-NET network for quadrature phase gradient prediction according to claim 2, characterized in that, In step S23a3, the random noise is added in the following manner: the sample region of the phase gradient in the x direction is segmented, and the random noise is added to the background region.

5. The training method of the U-NET network for quadrature phase gradient prediction according to claim 4, characterized in that, In step S23a3, the random noise added is Gaussian random noise with a standard deviation of 0.

05.

6. The training method of the U-NET network for quadrature phase gradient prediction according to claim 5, characterized in that, In step S23, the phase gradient in the x direction and the corresponding phase gradient in the y direction are taken as a group of data sets, and each group of data sets is in a ratio of 8:1:

1.

7. The training method of a U-NET network for quadrature phase gradient prediction according to claim 1, characterized in that, The U-NET network comprises a down-sampling convolution module for processing the phase gradient in the x direction as an input image, and an up-sampling convolution module for taking the result of the down-sampling convolution module as an input image; The down-sampling convolution module comprises six down-sampling convolution units that are sequentially and continuously subjected to down-sampling convolution, and each of the six down-sampling convolution units is subjected to convolution processing with the result of the previous down-sampling convolution unit as an input, except for the down-sampling convolution unit that takes the phase gradient in the x direction as an input image; the six down-sampling convolution units are subjected to six times of down-sampling convolution, and six images of a first down-sampling convolution image to a sixth down-sampling convolution image are obtained; the down-sampling convolution reduces the spatial resolution of the input image and increases the number of channels; The up-sampling convolution module comprises six up-sampling convolution units that successively perform up-sampling convolution, each of the up-sampling convolution units comprises a transposed convolution layer and a skip connection layer, the transposed convolution layer performs transposed convolution to increase the spatial resolution of the input image and reduce the number of channels, and the skip connection layer splices the image of the transposed convolution layer with a set of the down-sampling convolution images to obtain an up-convolution image of the up-sampling convolution; wherein the transposed convolution image of the first up-sampling convolution unit is spliced with the sixth down-sampling convolution image, the transposed convolution image of the second up-sampling convolution unit is spliced with the fifth down-sampling convolution image, and the transposed convolution images of the third up-sampling convolution unit to the sixth up-sampling convolution unit are sequentially spliced with one of the fourth down-sampling convolution image to the first down-sampling convolution image.

8. The training method of the U-NET network for quadrature phase gradient prediction according to claim 4, characterized in that, The down-sampling convolution performs down-sampling convolution with a stride of 2 and a convolution kernel size of 4*4, and performs batch normalization processing and Relu activation function processing after each convolution; the convolution layer of the up-sampling convolution module performs inverted convolution with a stride of 2 and a convolution kernel size of 4*4. 9.A method for orthogonal phase gradient imaging based on a U-NET network, characterized in that, The method comprises collecting a phase shift differential interference figure of a sample to be measured; extracting a phase gradient in an x direction from the phase shift differential interference figure; and inputting the phase gradient in the x direction into a U-NET network to obtain a phase gradient in a y direction.

10. An electronic device, comprising: a processor; a memory for storing a computer program executed by the processor; wherein the processor implements the training method of the U-NET network for orthogonal phase gradient prediction according to any one of claims 1-8 when executing the computer program.

Images