Phase distribution measurement device, phase distribution measurement method, and machine learning method

Abstract

The present invention provides a phase distribution measurement method that utilizes AI and that makes it possible to perform high-speed processing with high accuracy. The present invention comprises: a diffusion plate that diffuses incident light; a light-receiving device that receives light which has passed through the diffusion plate; and a reconstruction processing unit that estimates the phase distribution of the incident light on the incident surface of the diffusion plate on the basis of the intensity distribution of the received light.

Description

Phase distribution measurement device, phase distribution measurement method, and machine learning method CROSS-REFERENCE TO RELATED APPLICATIONS

[0001] This application is based on U.S. Provisional Patent Application No. 63 / 557,709, filed February 26, 2024, the contents of which are incorporated herein by reference.

[0002] The present disclosure relates to a phase distribution measurement device, a phase distribution measurement method, and a machine learning method.

[0003] In optical measurements, the amplitude and phase of light each contain different information about the object being measured. However, image sensors such as CCDs acquire an intensity distribution proportional to the square of the amplitude, so while they can acquire information contained in the amplitude, it has been difficult to acquire information contained in the phase distribution. For this reason, phase distributions have traditionally been measured using techniques such as Shack-Hartmann sensors, interferometry, and iterative Fourier transform methods. In recent years, measurement methods utilizing AI (artificial intelligence) have also been developed (e.g., Non-Patent Document 1).

[0004] A. Sinha et al, “Lensless computational imaging through deep learning,” Optica, vol. 4, No. 9, September 2017, pp. 1117-1125, https: / / doi.org / 10.1364 / OPTICA.4.001117

[0005] However, the Shack-Hartmann sensor has the problem of low spatial resolution due to the large size of the microlens array used. Furthermore, the interferometry method is vulnerable to air fluctuations and vibrations because it uses interference with a reference light, and it also has the problem of taking time because it repeatedly captures the interference light. Furthermore, the iterative Fourier transform method has the problem of long measurement times. Furthermore, the AI-based method described in Non-Patent Document 1 is limited to measuring symbolic objects such as images of animals or people, and is not intended for physical measurement or observation applications.

[0006] Therefore, an object of the present disclosure is to provide a phase distribution measurement method that utilizes AI to achieve high accuracy and high-speed processing.

[0007] The phase distribution measuring device according to the present disclosure includes a diffuser plate that diffuses incident light, a light receiving device that receives light that has passed through the diffuser plate, and a reconstruction processing unit that estimates the phase distribution of the incident light on the incident surface of the diffuser plate based on the intensity distribution of the received light.

[0008] The phase distribution measurement method according to the present disclosure includes the steps of diffusing incident light using a diffuser plate, receiving the light that has passed through the diffuser plate, inputting the intensity distribution of the received light into a reconstruction function, and estimating the phase distribution of the incident light at the incident surface of the diffuser plate from the intensity distribution using the reconstruction function.

[0009] The machine learning method according to the present disclosure is a machine learning method for a reconstruction model implemented as the reconstruction processing unit of the phase distribution measurement device according to the present invention, which forms a phase distribution that serves as correct data and causes it to enter the diffuser plate, estimates the phase distribution of the incident light at the incident surface of the diffuser plate based on the intensity distribution of the light received by the light receiving device, and evaluates the estimated phase distribution based on the phase distribution that serves as correct data, thereby performing machine learning of the reconstruction model.

[0010] According to the present disclosure, it is possible to provide a phase distribution measurement method that utilizes AI and is highly accurate and capable of high-speed processing.

[0011] FIG. 1 is a diagram illustrating an overview of a phase distribution measurement apparatus 1 according to the present invention. FIG. 2 is a diagram illustrating a measurement process of a phase distribution in the phase distribution measurement apparatus 1 according to the present invention. FIG. 3 is a diagram illustrating machine learning of a reconstructed model of a phase distribution in the phase distribution measurement apparatus 1 according to the present invention. FIG. 4 is a diagram illustrating the device configuration of a phase distribution measurement apparatus 1 according to an embodiment of the present invention. FIG. 5 is a diagram illustrating a flow of phase measurement in a learning stage according to an embodiment of the present invention. FIG. 6 is a diagram illustrating the device configuration of a phase distribution measurement apparatus 1 according to an embodiment of the present invention. FIG. 7 is a diagram illustrating a reconstructed model according to an embodiment of the present invention.

[0012] FIG. 1 is a diagram illustrating an overview of a phase distribution measurement device 1 according to the present invention. As shown in FIG. 1, the phase distribution measurement device 1 diffuses optical data (incident light) L, such as image data of an object to be observed, using a diffuser 11, and then receives the light using two cameras (light receiving devices) 12 and 13. The light received by the cameras 12 and 13 may be light that has passed through different phase modulators. The phase modulators may be, for example, lenses or SLMs (Spatial Light Modulators). The cameras 12 and 13 are complementary metal oxide semiconductor (CMOS) image sensors that acquire an intensity distribution proportional to the square of the amplitude of the incident light.

[0013] By diffusing the incident light L with the diffuser 11, it is possible to prevent the local information of the incident light L from being rounded by the discrete pixels of the cameras 12 and 13, and it is possible to acquire high in-plane frequency components of the phase distribution without missing them. Furthermore, by modulating the phase with a lens or an SLM, it is possible to acquire different frequency spatial images of the light after passing through the diffuser 11. In the example of FIG. 1, in order to acquire sufficient information on both phase and amplitude, the cameras 12 and 13 acquire intensity distributions of light phase-modulated by different methods, but it is also possible to phase-modulate only one of the lights, or to use only one camera to acquire the intensity distribution.

[0014] The intensity distribution data acquired by the cameras 12 and 13 is input to a phase distribution reconstruction model F (y = F(x)). The reconstruction model F is a model that has undergone machine learning using a method described later, and estimates and outputs the amplitude and phase distribution of light on the measurement plane 14 based on the input intensity distribution.

[0015] 2 is a diagram illustrating an outline of the phase distribution measurement process in the phase distribution measurement device 1 according to the present invention. As shown in FIG. 2, in this embodiment, the phase distribution on the measurement surface 14 is measured. Incident light passing through the diffuser 11 is split by a beam splitter (BS), one light is directly acquired by the camera 12, and the other light passes through a phase modulator (SLM) 15 and is then acquired by the camera 13. The intensity distributions of the light acquired by the cameras 12 and 13 are input to a reconstruction function F, and the phase distribution on the measurement surface 14 is reconstructed using the function F and output.

[0016] 3 is a diagram illustrating an outline of machine learning of a phase distribution reconstruction model in the phase distribution measurement device 1 according to the present invention. In Fig. 3, a learning phase distribution forming system 16 projects the phase distribution of an input image P1 onto a measurement surface 14. Machine learning is performed on a function F using the light intensity distributions acquired by the cameras 12 and 13 as input data and the phase distribution projected onto the measurement surface 14 by the learning phase distribution forming system 16 as correct data.

[0017] FIG. 4 illustrates the configuration of a phase distribution measurement device 1 for performing machine learning of a reconstruction model (function F). As shown in FIG. 4, the device includes a learning phase distribution formation system 16 and a phase distribution measurement system 17. The learning phase distribution formation system 16 projects phase distribution data, which serves as ground truth data, onto the measurement plane 14 of the phase distribution measurement system 17. The phase distribution measurement system 17 acquires a real space image (a real space image can be acquired at a conjugate plane of the lens) using a camera 12 and acquires a frequency space image using a camera 13. A mask 18 is provided on the light-receiving surface of the diffuser 11. The mask 18 is provided for the purpose of limiting the light-receiving area of ​​the diffuser 11, thereby limiting the incidence of ambient light and stray light on the diffuser 11 and controlling the passage of only light from areas phase-modulated by the SLM 15. This prevents ambient light other than light emitted from the laser 19 and light from areas where the phase distribution phase-modulated by the SLM 15 is not displayed from being measured as stray light.

[0018] Fig. 5 is a diagram illustrating the flow of phase measurement and reconstruction model learning in the learning stage using the device shown in Fig. 4. As shown in Fig. 5, optical data obtained by projecting a phase distribution D1 formed by a learning phase distribution forming system 16 is incident on the diffuser 11 of the phase distribution measurement system 17, and after being diffused by the diffuser 11, the cameras 12 and 13 acquire intensity distributions of light that have been phase-modulated using different methods. Each intensity distribution data is input to the reconstruction model F, which outputs an estimated phase distribution D2 (an estimated phase distribution at the measurement plane 14) reconstructed by the reconstruction model F. Meanwhile, the phase distribution D1 is input to the reconstruction model F as ground truth data, and the estimation accuracy is evaluated by comparing the ground truth data with the estimated phase distribution D2 using a loss function.

[0019] It is desirable that the range of the intensity distribution data acquired by the cameras 12 and 13 used as input data to the reconstruction model F be as wide as possible so as to include the entire light diffused by the diffuser 11. This makes it possible to reproduce in detail information, particularly in the high-frequency region of the frequency space image.

[0020] Fig. 6 shows another example of the configuration of the phase distribution measurement device 1 for performing machine learning of the reconstruction model. In the example of Fig. 4, the phase distribution measurement system 17 uses two cameras to acquire a real space image and a frequency space image, but in the example of Fig. 6, the phase distribution measurement system 22 includes only one camera 23. As shown in Fig. 6, the phase distribution measurement system 22 uses the camera 23 to acquire the intensity distribution of light that has passed through the diffuser 11 and propagated through space a distance d. The acquired intensity distribution is input to the reconstruction model F, and the phase distribution on the measurement plane 14 is estimated. In the learning phase distribution formation system 21, the phase distribution data input to the SLM 15 is projected onto the measurement plane 14 of the phase distribution measurement system 22, as in the example of Fig. 4.

[0021] (Reconstruction Model) The reconstruction model according to this embodiment may employ a CNN (Convolutional Neural Network)-based model such as U-Net. Furthermore, a reconstruction model based on a Vision Transformer, as illustrated in FIG. 7 , may be employed as a model capable of efficiently representing the relationship between distant points, such as light diffused by a diffuser. Reconstruction models based on other learning methods may also be employed. Specific examples include, but are not limited to, a Recurrent Neural Network (RNN), a Long Short Term Memory (LSTN), an Echo State Network (ESN), a Transformer, a Bidirectional Encoder Representations from Transformers (BERT), a Generative Adversarial Network (GAN), a Variational Autoencoder (VAE), and models including a combination of these. Furthermore, the reconstruction model may be applied to moving images if reconstruction is performed on multiple images input in time series.

[0022] As described above, according to this embodiment, the phase distribution of incident light is estimated by a reconstruction model based on the intensity distribution of incident light diffused by the diffuser plate 11. In this way, by diffusing the incident light by the diffuser plate 11, it is possible to acquire the high in-plane frequency components of the phase distribution without missing any of them, and therefore it is possible to estimate a phase distribution with high reproducibility.

[0023] Furthermore, by using two cameras to acquire the intensity distribution of the real space image and the frequency space image, respectively, it has become possible to more accurately estimate the phase distribution of the incident light.

[0024] Furthermore, by providing a mask that limits the light receiving area of ​​the diffuser 11, ambient light and light in areas that do not contain a phase distribution are not measured, thereby enabling more accurate estimation of the phase distribution. Furthermore, in the learning stage, a reference phase area may be placed in a part of the input image, or a part of the input image may be designated as the reference phase area, and learning may be performed using the reference phase area as input data. When measuring the phase distribution using a learned reconstruction model, the reference phase area may be placed together with the input image of the object to be observed, and information on the reference phase area may be incident on the diffuser 11.

[0025] It should be noted that the present invention is not limited to the above-described embodiment, and can be embodied in various other forms without departing from the spirit of the present invention. Therefore, the above-described embodiment is merely an example in all respects and should not be interpreted as being limiting.

[0026] The present invention can be applied to amplitude measurement as well as phase measurement with a similar configuration. Furthermore, the present invention can also measure phase and amplitude simultaneously (measure complex amplitude).

[0027] 1... Phase distribution measurement device, 11... Diffuser, 12, 13, 23... Camera, 14... Measurement surface, 15... SLM, 16, 21... Learning phase distribution formation system, 17, 22... Phase distribution measurement system, 18... Mask, 19... Laser

Claims

1. A phase distribution measurement device comprising: a diffuser plate that diffuses incident light; a light receiving device that receives light that has passed through the diffuser plate; and a reconstruction processing unit that estimates the phase distribution of the incident light on the incident surface of the diffuser plate based on the intensity distribution of the received light.

2. The phase distribution measuring device according to claim 1, wherein the light receiving device comprises a first light receiving device that acquires the intensity distribution of a real space image and a second light receiving device that acquires the intensity distribution of a frequency space image, and the reconstruction processing unit estimates the phase distribution of the incident light on the incident surface of the diffuser based on the first intensity distribution acquired by the first light receiving device and the second intensity distribution acquired by the second light receiving device.

3. A phase distribution measuring device according to claim 2, further comprising a phase modulator for modulating the phase of light incident on said second light receiving device.

4. The phase distribution measuring device according to claim 1, further comprising a mask for limiting the light receiving area of ​​said diffusing plate.

5. A phase distribution measurement method comprising the steps of: diffusing incident light using a diffuser; receiving the light that has passed through the diffuser; inputting the intensity distribution of the received light into a reconstruction function; and estimating the phase distribution of the incident light on the incident surface of the diffuser from the intensity distribution using the reconstruction function.

6. A machine learning method for a reconstruction model implemented as the reconstruction processing unit of the phase distribution measurement device described in claim 1, comprising: forming a phase distribution that serves as ground truth data and allowing it to enter the diffuser; estimating the phase distribution of incident light on the incident surface of the diffuser based on the intensity distribution of light received by the light receiving device; and evaluating the estimated phase distribution based on the phase distribution that serves as ground truth data, thereby performing machine learning of the reconstruction model.