Single-shot coherent diffraction imaging method under low photon counting condition based on multi-constraint non-training neural network

Abstract

The present application relates to a single coherent diffraction imaging method, in particular to a single coherent diffraction imaging method under low photon counting conditions based on a multi-constraint non-training neural network, which solves the technical problems of insufficient precision and poor stability of the existing single coherent diffraction imaging method under low photon counting conditions. The present application is based on a non-training neural network architecture, belongs to the category of unsupervised learning, and can restore the amplitude and phase information of the sample using a single intensity diffraction image. Compared with supervised training neural networks, it does not rely on large-scale labeled data sets, avoids the generalization problem caused by data distribution bias, and has better physical consistency and interpretability. At the same time, the physical constraint, the structure constraint and the support constraint are jointly used as a joint loss function to guide the parameter optimization of the multi-constraint non-training neural network, realizing multi-constraint collaborative optimization, significantly improving the precision and stability of single coherent diffraction imaging, and still having good imaging performance under low photon counting conditions.

Description

Technical Field

[0001] This invention relates to a single-shot coherent diffraction imaging method, specifically to a single-shot coherent diffraction imaging method based on a multi-constrained untrained neural network under low photon counting conditions. Background Technology

[0002] Coherent diffraction imaging is a lensless imaging technique based on highly coherent light sources (such as X-rays or electron beams). It obtains high spatial resolution structural images by measuring the far-field or Fraunhofer diffraction intensity distribution of a sample and reconstructing the complex amplitude information of the sample using phase retrieval algorithms. Because it eliminates the need for imaging lenses, coherent diffraction imaging overcomes the limitations of numerical aperture and aberrations in traditional microscopes, achieving spatial resolution at the nanoscale or even near the atomic scale. It is currently widely used in materials science, biomedicine, semiconductor devices, nanostructure characterization, and ultrafast dynamics.

[0003] Coherent diffraction imaging techniques can be mainly divided into two categories based on the method of acquiring diffraction data: single-shot imaging and scanning imaging. Single-shot coherent diffraction imaging acquires complete diffraction intensity information through a single exposure and combines it with phase retrieval algorithms to reconstruct the sample structure. It is suitable for studying fast dynamic processes or non-reproducible samples. Scanning coherent diffraction imaging, on the other hand, moves the sample or beam to acquire overlapping diffraction patterns at multiple locations and performs joint reconstruction. A typical method is stacked coherent diffraction imaging, which can significantly improve imaging resolution and reconstruction stability, but usually requires a longer data acquisition time and cannot achieve high temporal resolution. Therefore, single-shot coherent diffraction imaging has irreplaceable advantages in ultrafast dynamics research.

[0004] In certain sample cases, such as biomaterials, polymers, or organic semiconductors, it is necessary to obtain diffraction patterns under low-dose radiation to prevent sample damage. Simultaneously, under certain light source conditions (such as water-window attosecond sources), it is difficult to provide sufficient photon flux due to limitations in output power. In this context, low photon counts (102) are crucial. 6 Single-shot coherent diffraction imaging under conditions of magnitude 6.5 (on the order of magnitude of a given scale) presents a significant challenge, introducing substantial shot noise into the acquired diffraction pattern, severely reducing the signal-to-noise ratio and increasing the difficulty of subsequent image reconstruction. Traditional iterative phase retrieval algorithms, such as Error Reduction (ER) and Hybrid Input-Output (HIO), are prone to getting trapped in local minima or stagnation, leading to decreased convergence performance or even reconstruction failure. While end-to-end neural network methods offer fast reconstruction capabilities and a degree of tolerance for noisy data, they are highly dependent on training data and prone to insufficient generalization ability. Summary of the Invention

[0005] The purpose of this invention is to solve the technical problems of insufficient accuracy and poor stability of existing single-shot coherent diffraction imaging methods under low photon counting conditions, and to provide a single-shot coherent diffraction imaging method based on a multi-constraint untrained neural network under low photon counting conditions.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] A single-shot coherent diffraction imaging method based on a multi-constraint untrained neural network under low photon counting conditions is characterized by the following steps:

[0008] S1. Construct a single-shot coherent diffraction imaging system with a pinhole, and then use it to collect coherent diffraction light carrying information about the sample in the Fraunhofer diffraction region of the sample to obtain the original intensity diffraction pattern.

[0009] S2. Preprocess the original intensity diffraction pattern to obtain a preprocessed intensity diffraction pattern;

[0010] S3. Define physical constraints, structural constraints, and support constraints; construct a physical model for Fraunhofer diffraction between the sample to be tested and the original intensity diffraction pattern acquisition location. The physical constraints are based on the physical model, the structural constraints are based on the structural characteristics of the sample to be tested, and the support constraints are based on the spatial region limitations of the sample to be tested.

[0011] S4. Generate a random vector based on the size of the preprocessed intensity diffraction pattern. Then, using the random vector as input, the predicted amplitude and phase distribution of the sample to be tested as output, and physical constraints, structural constraints, and support constraints as joint loss functions, construct a multi-constraint untrained neural network and reset the number of iterations to zero.

[0012] S5. Input the random vector into the multi-constraint untrained neural network. The multi-constraint untrained neural network outputs the predicted amplitude and phase distribution of the sample to be tested. Then, combined with the physical model, the predicted intensity diffraction pattern is obtained.

[0013] S6. Calculate the joint loss function based on the predicted amplitude and phase distribution of the sample to be tested, the preprocessed intensity diffraction pattern, and the predicted intensity diffraction pattern;

[0014] S7. If the joint loss function converges, or the number of iterations reaches the preset number of iterations, the predicted amplitude and phase distribution of the sample to be tested are used as the amplitude and phase distribution of the sample to be tested, respectively, to complete a single coherent diffraction imaging; otherwise, the parameters of the multi-constraint untrained neural network are optimized and updated according to the joint loss function, the number of iterations is incremented by 1 and the process returns to S5.

[0015] Furthermore, in S3, the physical constraint is Λ phy (I pred ;I meas), where Λ phy The physical loss function is set as the mean absolute error function, and then compared. and Their respective first-order differentials in the horizontal and vertical directions; I pred To predict the intensity diffraction pattern, I meas This is the preprocessed intensity diffraction pattern;

[0016] The structural constraint is Λ str (A*;φ*), where Λ str Let A* be the structural loss function, set as the total variational loss. We calculate the first derivatives of A* and φ* in the horizontal and vertical directions, respectively, and then calculate the average of the sum of the absolute values ​​of the first derivatives. A* and φ* are the predicted amplitude and phase distributions of the sample to be tested, respectively.

[0017] The support constraint is Λ sup (A*;0), where 0 represents an all-zero matrix, Λ sup To support the loss function, the mean absolute error function is set, and the electron density of the test sample outside the pinhole region is compared with the all-zero matrix. The electron density reflects the predicted amplitude distribution of the test sample.

[0018] Furthermore, in S4, the joint loss function is: β1Λ phy +β2Λ str +β3Λ sup , where β1, β2, and β3 are all weight parameters.

[0019] Furthermore, in S4, the multi-constraint untrained neural network adopts a convolutional neural network architecture.

[0020] Furthermore, in S4, the multi-constraint untrained neural network adopts a symmetric encoder-decoder architecture, including five downsampling layers, five upsampling layers, a 1×1 convolutional layer, and a sigmoid activation function layer connected in sequence.

[0021] Furthermore, in S7, the Adam optimizer is used to optimize the parameters of the multi-constrained untrained neural network.

[0022] Furthermore, in S2, the preprocessing includes noise reduction, pixel merging, and zeroing out negative numbers.

[0023] Furthermore, in S4, the random vector satisfies a normal distribution with a mean of 0 and a standard deviation of 0.1.

[0024] Furthermore, in S1, the single-shot coherent diffraction imaging system includes a coherent light source, and a beam expander, a pinhole, and an image sensor arranged sequentially along the output light path of the coherent light source. The light-emitting surface of the pinhole is used to set the sample to be tested. The image sensor is located in the Fraunhofer diffraction region of the sample to be tested. The image sensor is connected to a computer, which is used to preprocess the original intensity diffraction pattern and deploy a multi-constraint untrained neural network.

[0025] Furthermore, in S1, the output of the coherent light source is visible light, X-rays, or an electron beam;

[0026] The pinhole is square, round, or triangular;

[0027] The image sensor is either a CCD camera or a CMOS camera.

[0028] Compared with the prior art, the present invention has the following beneficial effects:

[0029] 1. The single coherent diffraction imaging method under low photon counting conditions provided by the present invention is based on a multi-constraint untrained neural network architecture and belongs to the category of unsupervised learning. It can recover the amplitude and phase information of the sample using a single intensity diffraction pattern. Compared with supervised training neural networks, it does not rely on large-scale labeled datasets, avoids the generalization problem caused by data distribution bias, and has better physical consistency and interpretability.

[0030] 2. The single-shot coherent diffraction imaging method based on a multi-constraint untrained neural network provided by this invention uses physical constraints, structural constraints and support constraints as a joint loss function to guide the parameter optimization of the multi-constraint untrained neural network, realizing multi-constraint collaborative optimization, significantly improving the accuracy and stability of single-shot coherent diffraction imaging, and still having good imaging performance under low photon count conditions.

[0031] 3. The single-shot coherent diffraction imaging method based on multi-constraint untrained neural networks under low photon counting conditions provided by this invention has strong robustness to low signal-to-noise ratio environments, can effectively suppress noise interference, improve the stability of single-shot coherent diffraction imaging, and reduce the requirements for experimental conditions while improving imaging performance. It can still achieve high-fidelity imaging under limited light source power and limited detection conditions, and has good engineering applicability and promotion value.

[0032] 4. The single-shot coherent diffraction imaging method based on a multi-constraint untrained neural network under low photon counting conditions provided by this invention can simultaneously recover the amplitude and phase information of the sample, and is suitable for imaging complex value samples.

[0033] 5. The single-shot coherent diffraction imaging method based on a multi-constrained untrained neural network under low photon counting conditions provided by this invention has a simpler imaging system compared to scanning imaging. This not only greatly reduces the requirements for system stability, but also provides a feasible path for dynamic coherent diffraction imaging. Attached Figure Description

[0034] Figure 1 This is a flowchart of a method according to an embodiment of the present invention;

[0035] Figure 2 This is a schematic diagram of the single-shot coherent diffraction imaging system in embodiment S1 of the present invention;

[0036] Figure 3 This is a flowchart of the preprocessing in embodiment S2 of the present invention;

[0037] Figure 4 This is an architecture diagram of the multi-constraint untrained neural network in embodiment S4 of the present invention;

[0038] Figure 5 This is a schematic diagram of the joint loss function in embodiment S4 of the present invention;

[0039] Figure 6 This is a flowchart illustrating the execution of the multi-constraint untrained neural network in embodiments S5-S7 of the present invention.

[0040] Figure 7 The diagram shows the coherent diffraction imaging results of pure amplitude and pure phase samples obtained by using the embodiments of the present invention. In the diagram, the test samples in (a), (b), and (c) are pure amplitude USAF1951 positive resolution target, pure phase USAF1951 resolution target, and pure amplitude USAF1951 negative resolution target, respectively.

[0041] Figure 8 To implement the embodiments of the present invention in 10 6 A schematic diagram of coherent diffraction imaging results of complex-valued samples obtained under photon counting of the order of magnitude, wherein (a) is a schematic diagram of amplitude distribution and (b) is a schematic diagram of phase distribution;

[0042] The annotations in the attached figures are explained as follows:

[0043] 1-Coherent light source, 2-Beam expander system, 3-Pinhole, 4-Sample to be tested, 5-Image sensor, 6-Computer. Detailed Implementation

[0044] The present invention provides a detailed description of a single-shot coherent diffraction imaging method based on a multi-constraint untrained neural network under low photon counting conditions, in conjunction with the accompanying drawings and specific embodiments. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the invention and are not intended to limit the scope of protection of the invention.

[0045] A single-shot coherent diffraction imaging method based on a multi-constraint untrained neural network under low photon counting conditions, such as Figure 1 As shown, it includes the following steps:

[0046] S1. Construct a single-shot coherent diffraction imaging system with a pinhole, and then use it to collect coherent diffraction light carrying information about the sample in the Fraunhofer diffraction region of the sample to obtain the original intensity diffraction pattern.

[0047] Single-shot coherent diffraction imaging system, such as Figure 2 As shown, the system includes a coherent light source 1, and a beam expander system 2, a pinhole 3, and an image sensor 5 arranged sequentially along the output light path of the coherent light source 1. The light-emitting surface of the pinhole 3 is used to hold the sample 4 to be tested, and the image sensor 5 is located in the Fraunhofer diffraction region of the sample 4. The image sensor 5 is connected to a computer 6, which is used to preprocess the original intensity diffraction pattern and deploy a multi-constraint untrained neural network. The output of the coherent light source 1 is visible light, X-rays, or an electron beam. The pinhole 3 is square, circular, or triangular. The image sensor 5 uses a CCD camera or a CMOS camera. In this embodiment, the output of the coherent light source 1 is a helium-neon laser with a wavelength of 632.8 nanometers. The pinhole 3 is triangular, and the image sensor 5 uses a CMOS camera with a pixel count of 2048×2028 and a pixel size of 6.5 micrometers.

[0048] Specifically, the coherent light source 1, after passing through the beam expander system 2, forms a converging spherical wave, which is then projected onto the sample 4 under test through the pinhole 3, creating a finite-sized illumination field. This field propagates backward through the sample 4, undergoing coherent diffraction. The original intensity diffraction pattern is then acquired using an image sensor 5 in the Fraunhofer diffraction region. The image sensor 5 is placed on the focusing plane of the converging spherical wave, with its center coinciding with the center of the intensity diffraction pattern. The transmitted light field of the sample 4 and the planar light field of the image sensor 5 satisfy a Fourier transform relationship, allowing the recording of a scaled-down Fraunhofer intensity diffraction image of the sample 4.

[0049] S2. The original intensity diffraction pattern is preprocessed using digital image processing techniques to obtain a preprocessed intensity diffraction pattern. For example... Figure 3As shown, the preprocessing includes denoising, pixel merging, and zeroing out negative values. Specifically, the image sensor 5 is set with an appropriate exposure time, and the sample 4 under test records an original intensity diffraction image and a background image under both illuminated and dark conditions, respectively, and stores them in the computer 6. The background image is subtracted from the original intensity diffraction image to complete the background noise removal operation. To improve the signal-to-noise ratio and computational efficiency, pixel merging is performed on the denoised intensity diffraction image. In this embodiment, 4×4 pixel merging is performed, and the size of the merged diffraction image is 512×512. Finally, pixel values ​​less than 0 in the diffraction image are set to 0 to obtain the preprocessed intensity diffraction image.

[0050] S3. Define physical constraints, structural constraints, and support constraints. Specifically, a physical model for Fraunhofer diffraction is constructed between the sample 4 and the image sensor 5. The physical constraints, based on this model, ensure physical consistency between the network output and the experimentally measured diffraction pattern, and are defined as Λ. phy (I pred ;I meas ), where Λ phy The physical loss function is set as the mean absolute error function, and then compared. and Their respective first-order differentials in the horizontal and vertical directions; I pred The predicted intensity diffraction pattern of sample 4 is obtained by using a physical model based on the predicted amplitude and phase distribution. meas This is the preprocessed intensity diffraction pattern. The structural constraint, based on the structural characteristics of the sample under test, is defined as Λ. str (A*;φ*), where Λ str The structural loss function is set as the total variational loss. First-order derivatives of A* and φ* are calculated in the horizontal and vertical directions, respectively, and then the average of the sum of the absolute values ​​of these first-order derivatives is calculated. A* and φ* represent the predicted amplitude and phase distributions of the sample, respectively. Structural constraints utilize the total variational prior to constrain the spatial continuity of the sample's amplitude and phase, suppressing non-physical oscillations caused by noise. Support constraints are based on spatial region limitations of the sample and are defined as Λ. sup (A*;0), where 0 represents an all-zero matrix, Λ sup To support the loss function, a mean absolute error function is used. The electron density of sample 4 outside the pinhole region is compared with the all-zero matrix. The electron density reflects the predicted amplitude distribution of the sample. The support constraint requires that the electron density of the sample outside the pinhole region be as small as possible. The support constraint uses the known spatial support region formed by the pinhole to restrict the spatial distribution of the sample and improve the uniqueness of the solution.

[0051] S4. Generate a random vector based on the size of the preprocessed intensity diffraction pattern. Then, using the random vector as input and the predicted amplitude and phase distribution of the sample 4 to be tested as output, construct a multi-constraint untrained neural network using physical constraints, structural constraints, and support constraints as joint loss functions, and reset the number of iterations to zero.

[0052] In this embodiment, the random vectors follow a normal distribution with a mean of 0 and a standard deviation of 0.1, and have a size of 64×512×512. The joint loss function combines the various constraints as a common constraint, summing the physical loss function, structural loss function, and support loss function, and assigning certain weights, specifically: β1Λ phy +β2Λ str +β3Λ sup , where β1, β2, and β3 are all weight parameters.

[0053] like Figure 5 As shown, the multi-constraint untrained neural network is a convolutional neural network, employing a symmetric encoder-decoder architecture, comprising five downsampling layers, five upsampling layers, a 1×1 convolutional layer, and a sigmoid activation function layer connected sequentially. The input random vector is first compressed to 128×16×16 by the encoder. Each downsampling layer contains two sequential operations: ① a 3×3 convolution with a stride of 1 + a 2×2 max pooling with a stride of 2 + batch normalization + LeakyReLU activation function; ② a 3×3 convolution with a stride of 1 + batch normalization + LeakyReLU activation function. Subsequently, the decoder upsamples the features to 128×512×512. Each upsampling layer contains two sequential operations: ① a 2× nearest neighbor upsampling operation + a 3×3 convolution with a stride of 1 + batch normalization + LeakyReLU activation function; ② a 1×1 convolution with a stride of 1 + batch normalization + LeakyReLU activation function. Then, a 1×1 convolution operation with a stride of 1 maps the features to two output channels, and a sigmoid activation function is used to constrain the output values ​​to the range of 0 to 1. Both output channels are 512×512 in size, representing the predicted amplitude and phase distribution of the sample to be tested, respectively.

[0054] S5. Input the random vector into the multi-constraint untrained neural network. The multi-constraint untrained neural network outputs the predicted amplitude and phase distribution of the sample to be tested. Then, combined with the physical model, the predicted intensity diffraction pattern is obtained.

[0055] S6. Calculate the joint loss function based on the amplitude and phase information of the sample to be tested, the preprocessed intensity diffraction pattern, and the predicted intensity diffraction pattern.

[0056] S7. If the joint loss function converges, or the number of iterations reaches the preset number of iterations, the predicted amplitude and phase distribution of the sample to be tested are used as the amplitude and phase distribution of the sample to be tested, respectively, to complete a single coherent diffraction imaging; otherwise, according to the joint loss function, the Adam optimizer is used to optimize and update the parameters of the multi-constrained untrained neural network, and the number of iterations is incremented by 1 before returning to S5.

[0057] S5-S7 describes the execution process of a multi-constraint untrained neural network, such as... Figure 6 As shown, the input to the multi-constraint untrained neural network is a fixed random vector. During the optimization process, it does not participate in parameter iteration updates, and its output is the predicted amplitude and phase distribution of the sample under test. The multi-constraint untrained neural network does not have a pre-training process. During the optimization process, multiple constraints are used as a joint loss function to guide the optimization direction of the multi-constraint untrained neural network. The network parameters are iteratively updated through backpropagation, and the Adam optimizer is used to optimize the parameters, thereby gradually approaching the optimal solution and achieving accurate reconstruction of the amplitude and phase distribution of the sample under test.

[0058] The coherent diffraction imaging results of the pure amplitude and pure phase samples obtained using this embodiment are as follows: Figure 7 As shown, the coherent diffraction imaging results of the complex-valued sample are as follows: Figure 8 As shown, the single-shot coherent diffraction imaging method based on a multi-constrained untrained neural network under low photon counting conditions provided in this embodiment has high accuracy and stability, and still has good imaging performance under low photon counting conditions.

Claims

1. A single-shot coherent diffraction imaging method based on a multi-constraint untrained neural network under low photon counting conditions, characterized in that, Includes the following steps: S1. Construct a single-shot coherent diffraction imaging system with a pinhole, and then use it to collect coherent diffraction light carrying information about the sample in the Fraunhofer diffraction region of the sample to obtain the original intensity diffraction pattern. S2. Preprocess the original intensity diffraction pattern to obtain a preprocessed intensity diffraction pattern; S3. Define physical constraints, structural constraints, and support constraints; construct a physical model for Fraunhofer diffraction between the sample to be tested and the original intensity diffraction pattern acquisition location. The physical constraints are based on the physical model, the structural constraints are based on the structural characteristics of the sample to be tested, and the support constraints are based on the spatial region limitations of the sample to be tested. S4. Generate a random vector based on the size of the preprocessed intensity diffraction pattern. Then, using the random vector as input, the predicted amplitude and phase distribution of the sample to be tested as output, and physical constraints, structural constraints, and support constraints as joint loss functions, construct a multi-constraint untrained neural network and reset the number of iterations to zero. S5. Input the random vector into the multi-constraint untrained neural network. The multi-constraint untrained neural network outputs the predicted amplitude and phase distribution of the sample to be tested. Then, combined with the physical model, the predicted intensity diffraction pattern is obtained. S6. Calculate the joint loss function based on the predicted amplitude and phase distribution of the sample to be tested, the preprocessed intensity diffraction pattern, and the predicted intensity diffraction pattern; S7. If the joint loss function converges, or the number of iterations reaches the preset number of iterations, the predicted amplitude and phase distribution of the sample to be tested will be used as the amplitude and phase distribution of the sample to be tested, respectively, to complete a single coherent diffraction imaging. Otherwise, based on the joint loss function, the parameters of the multi-constrained untrained neural network are optimized and updated, the iteration count is incremented by 1, and the process returns to S5.

2. The single-shot coherent diffraction imaging method based on a multi-constraint untrained neural network under low photon counting conditions according to claim 1, characterized in that: In S3, the physical constraint is Λ phy (I pred ;I meas ), where Λ phy The physical loss function is set as the mean absolute error function, and then compared. and Their respective first-order differentials in the horizontal and vertical directions; I pred To predict the intensity diffraction pattern, I meas This is the preprocessed intensity diffraction pattern; The structural constraint is Λ str (A*;φ*), where Λ str Let A* be the structural loss function, set as the total variational loss. We calculate the first derivatives of A* and φ* in the horizontal and vertical directions, respectively, and then calculate the average of the sum of the absolute values ​​of the first derivatives. A* and φ* are the predicted amplitude and phase distributions of the sample to be tested, respectively. The support constraint is Λ sup (A*;0), where 0 represents an all-zero matrix, Λ sup To support the loss function, the mean absolute error function is set, and the electron density of the test sample outside the pinhole region is compared with the all-zero matrix. The electron density reflects the predicted amplitude distribution of the test sample.

3. The single-shot coherent diffraction imaging method based on a multi-constraint untrained neural network under low photon counting conditions according to claim 2, characterized in that, In S4, the joint loss function is: β1Λ phy +β2Λ str +β3Λ sup , where β1, β2, and β3 are all weight parameters.

4. The single-shot coherent diffraction imaging method based on a multi-constraint untrained neural network under low photon counting conditions according to any one of claims 1-3, characterized in that: In S4, the multi-constraint untrained neural network adopts a convolutional neural network architecture.

5. The single-shot coherent diffraction imaging method based on a multi-constraint untrained neural network under low photon counting conditions according to claim 4, characterized in that: In S4, the multi-constraint untrained neural network adopts a symmetric encoder-decoder architecture, which includes five downsampling layers, five upsampling layers, a 1×1 convolutional layer, and a sigmoid activation function layer connected in sequence.

6. The single-shot coherent diffraction imaging method based on a multi-constraint untrained neural network under low photon counting conditions according to claim 5, characterized in that: In S7, the Adam optimizer is used to optimize the parameters of multi-constrained untrained neural networks.

7. The single-shot coherent diffraction imaging method based on a multi-constraint untrained neural network under low photon counting conditions according to claim 6, characterized in that: In S2, the preprocessing includes noise reduction, pixel merging, and zeroing out negative numbers.

8. The single-shot coherent diffraction imaging method based on a multi-constraint untrained neural network under low photon counting conditions according to claim 7, characterized in that: In S4, the random vector satisfies a normal distribution with a mean of 0 and a standard deviation of 0.

1.

9. The single-shot coherent diffraction imaging method based on a multi-constraint untrained neural network under low photon counting conditions according to claim 8, characterized in that: In S1, the single coherent diffraction imaging system includes a coherent light source (1), and a beam expander (2), a pinhole (3), and an image sensor (5) arranged sequentially along the output light path of the coherent light source (1). The light-emitting surface of the pinhole (3) is used to set the sample to be tested (4). The image sensor is located in the Fraunhofer diffraction region of the sample to be tested (4). The output end of the image sensor (5) is connected to a computer (6). The computer (6) is used to preprocess the original intensity diffraction pattern and deploy a multi-constraint untrained neural network.

10. The single-shot coherent diffraction imaging method based on a multi-constraint untrained neural network under low photon counting conditions according to claim 9, characterized in that: In S1, the output of the coherent light source (1) is visible light, X-rays or an electron beam; The pinhole (3) is square, circular or triangular; The image sensor (5) is a CCD camera or a CMOS camera.

Images