Single-exposure compressive ultrafast coherent modulation imaging device

Abstract

The application discloses a single-exposure compressed ultrafast coherent modulation imaging device. By combining coherent modulation imaging and compressed ultrafast imaging technology, a sample to be measured is illuminated by coherent light, after the sample diffracts the light wave, the phase modulation of a random phase modulation plate is carried out, the integral intensity distribution of the modulated wave front is recorded by deflection on a detection plane by using a slit fully opened stripe camera, a single compressed diffraction pattern is obtained, a spatial two-dimensional to space-time three-dimensional image reconstruction algorithm based on a plug-and-play framework and a depth image priori is developed, and data recovery is carried out on the collected image by using the algorithm. Compared with traditional ultrafast optical imaging and phase imaging technology, the application can quantitatively obtain complete intensity and phase information of a transient scene in a single measurement process, has the characteristics of picosecond time resolution and real-time quantitative intensity-phase imaging capacity in single-exposure phase imaging, and can be widely applied to various ultrafast phenomena requiring phase observation.

Description

Technical Field

[0001] This invention relates to the field of ultrafast optical imaging technology, and in particular to a single-exposure compressed ultrafast coherent modulation imaging device (CS-CMUI). Background Technology

[0002] Single-exposure ultrafast imaging is an important branch of ultrafast optical imaging. Compressed ultrafast imaging (CUP) is suitable for recording non-repeatable or irreversible transient events and recovering the original dynamic scene from a two-dimensional image using image reconstruction algorithms. Phase, as a crucial parameter, describes the wavefront shape and optical path difference of light waves in optical research and has a significant impact on optical imaging. Phase can provide information about an object's refractive index, thickness, etc., that light intensity cannot provide, which is crucial for understanding the physical phenomena under test. Coherent modulation imaging (CMI) is a lensless diffraction imaging technique within phase imaging. By introducing a random phase or amplitude modulator between the sample and the image sensor, multiple modulations of the object wave increase constraints and improve the convergence speed of iterative algorithms, enabling the reconstruction of complex optical fields from a single diffraction intensity map.

[0003] However, ultrafast optical imaging is an intensity recording method and cannot acquire the phase information of an object; phase imaging technology can only achieve static recording and cannot acquire time-resolved information. Currently, no technology can quantitatively acquire complete intensity and phase information of a transient scene in a single measurement. The challenge of ultrafast phase imaging lies in the fact that current sensors or other recording media can only sense the intensity information of the sample and cannot directly acquire phase information. Therefore, optical phase imaging requires special optical design and signal processing methods. Furthermore, phase information is easily affected by distortions such as refraction, scattering, and absorption in the optical path, resulting in poor image quality and application performance. Summary of the Invention

[0004] To overcome the difficulty of traditional optical devices in detecting phase information and to break through the bottleneck of traditional phase imaging technology in acquiring time-resolved information, this invention proposes a spatiotemporal three-dimensional imaging device that simultaneously possesses picosecond-level time resolution and quantitative intensity-phase imaging capabilities in a single exposure. It can overcome the defect of current ultrafast optical imaging technology in being unable to quantitatively detect phase changes in real time, and can simultaneously acquire the intensity and phase changes of ultrafast transient scenes over time, thereby improving the performance of ultrafast optical imaging and expanding the application fields of optical imaging technology. This invention is a single-exposure compressed ultrafast coherent modulation imaging (CS-CMUI) that combines coherent modulation imaging (CMI) and compressed ultrafast imaging (CUP). It uses a lensless design to reduce the impact of optical path structure distortion, and achieves quantitative detection of phase and intensity information based on CMI. It uses compressed sensing principle and streak camera to achieve compressed recording of spatiotemporal holographic information under single exposure. Finally, it uses the iterative reconstruction algorithm PnP-DIP based on plug and play (PnP) framework and deep image prior (DIP) to restore the measurement data, and finally achieves quantitative ultrafast imaging of phase and intensity with picosecond-level temporal resolution.

[0005] The specific technical solution for achieving the objective of this invention is as follows:

[0006] A single-exposure compressed ultrafast coherent modulation imaging device, characterized in that the device includes a probe light phase modulation optical path, a streak camera and a data processing system;

[0007] The probe light phase modulation optical path includes a coherent illumination source, a test object containing phase information (which can be a dynamic scene), and a random phase modulation board connected in sequence.

[0008] The probe light phase modulation optical path adopts a lensless design, which reduces the impact of phase distortion of the optical path structure caused by lenses.

[0009] The stripe camera internally includes an entrance slit, a photocathode, a microchannel plate, a scanning electrode, a fluorescent screen, and a built-in camera, which are connected in sequence by optical paths.

[0010] The data processing system consists of a computer equipped with a graphics card and a CUDA general-purpose parallel computing architecture; wherein:

[0011] The lens center of the coherent illumination source, the geometric center of the object under test, the geometric center of the random phase modulation plate, and the lens center of the streak camera are on the same straight line, which is the direction of light propagation.

[0012] The coherent illumination source, the object under test containing phase information, the random phase modulation plate, and the streak camera correspond to four planes in the optical path, namely the outgoing plane, the incident plane, the modulation plane, and the detection plane, respectively; all four planes are perpendicular to the direction of light propagation, and the intersection of the optical path and the plane is at the geometric center of the plane.

[0013] The CUDA general-purpose parallel computing architecture includes an image reconstruction algorithm, namely the PnP-DIP algorithm, a spatial two-dimensional to spatiotemporal three-dimensional image reconstruction algorithm based on a plug-and-play framework and depth image priors. After signal acquisition, the data processing system uses the PnP-DIP algorithm to process the recorded compressed diffraction holographic pattern to reconstruct the phase and amplitude information of the measured object. The specific data processing procedure of PnP-DIP is as follows:

[0014] Let the complex amplitude of the object to be measured be:

[0015] U0(x,y,t;0)=I(x,y,t;0)exp[iφ(x,y,t;0)] (1)

[0016] Where i represents the imaginary unit, x and y are the two-dimensional spatial coordinates of the object, t represents time, and the 0 after the semicolon ";" represents the current propagation distance; I(x,y,t;0) represents the amplitude information of the object, and φ(x,y,t;0) represents the phase information of the object; after propagation over a distance d1, the complex amplitude of the diffracted light is expressed as:

[0017]

[0018] in The transfer function represents the distance d1, k represents the wave number, and λ represents the wavelength of the coherent illumination source; It is the two-dimensional Fourier transform of U0, f x and f y These are the spatial frequencies of the image in the x and y directions, respectively; the complex amplitude after passing through the random phase modulation plate is expressed as:

[0019]

[0020] Where m(x,y) represents the spatial phase distribution of the phase modulation plate; U M After propagating for another d2 distance, the complex amplitude of the diffracted light is expressed as:

[0021]

[0022] At this moment, the light intensity information received by the camera is:

[0023]

[0024] Inside the streak camera, the time-varying light intensity information is deflected and then integrated and superimposed, resulting in the following compressed result:

[0025]

[0026] Where v represents the deflection speed of the streak camera. All the above processes can be represented in matrix form as follows:

[0027] I(x,y)=TS|H d2 MH d1 U(x,y,t)| 2 (7)

[0028] Where U(x,y,t) represents the complex amplitude of the object under test, and H d1 and H d2 For spatial transfer operators, M represents the phase modulation operator, S represents the spatiotemporal deflection operator of the streak camera, and T represents the integration operator of the streak camera CMOS during a single exposure.

[0029] The above forward acquisition process can be further simplified as follows:

[0030] y = Ax (8)

[0031] in This represents the complex amplitude distribution of the object under test. Represents the measurement matrix. This represents the final compressed data result; here, variables W, H, and B represent the number of pixels in the x, y, and t directions of the dynamic scene, respectively, and variable N represents the number of pixels in the y direction in the compressed snapshot coordinate system. Typically, the value of N is greater than the value of H, but much smaller than the value of H×B. The data recovery process is the process of calculating the test data x from the known signal y. From a mathematical perspective, solving for the original signal x given y and A is clearly an underdetermined problem because the amount of known data is much smaller than the amount of unknown data. Therefore, to accurately recover the signal, other regularization constraints must be added to reduce the size of the solution space. Simultaneously, because the matrix form of A is too large, solving A... -1 The process of finding the inverse is quite computationally and time-consuming. Here, we introduce DIP to replace the inverse process, and rewrite the inverse problem as an optimization problem in the following form:

[0032]

[0033] Where R(x) is a prior term, λ is a regularization parameter used to balance the fidelity term and the prior term; G(θ,z) represents a deep convolutional neural network with updatable parameters, whose input is... It is a standard Gaussian white noise of the same size as the original data x, which remains unchanged during the optimization process. θ is an updatable weight parameter in the network, and G(θ,z) is abbreviated as G(θ). Furthermore, by introducing an updatable auxiliary parameter u and a manually set penalty parameter ρ, an augmented Lagrangian function L is constructed. ρ The constrained optimization problem described above can be rewritten as an unconstrained optimization problem:

[0034]

[0035] Using the ADMM framework, The minimization problem can be broken down into the following three subproblems:

[0036] 1)

[0037] 2)

[0038] 3)u k+1 =u k +(x k+1 -G(θ k+1 (3)

[0039] For the update of x, it is observed that it is the input G(θ) k )-u k about The form of the proximal operator is used here, and a plug-and-play design is used to replace the proximal operator with a denoising algorithm to prevent overfitting of the depth image prior; for the update of θ, a first-order gradient descent step is used in each iteration;

[0040] It is easy to see from equations (10), (11), (12), and (13) that as the number of iterations k increases, the solution object x... k+1 It will continuously approach the complex amplitude distribution x of the real object under test, and finally converge to x when k is infinitely large. According to equation (8), x is the complex amplitude information of the object under test, U0(x,y,t;0)=I(x,y,t;0)exp[iφ(x,y,t;0)]; by manually selecting an appropriate number of iterations k, the phase and amplitude information of the object under test can be reconstructed from the obtained compressed data results.

[0041] The coherent illumination source of this invention is a 532 nm Q-switched laser with a pulse width of 10 nanoseconds (full width at half maximum) and a maximum pulse energy of 5 millijoules.

[0042] The random phase modulation plate is formed by irregular etching at a depth of 577 nm on a fused silica substrate, with a unit size of 50 μm × 50 μm, and exhibits a binary random distribution of 0 and π phase modulation at a wavelength of 532 nm.

[0043] The stripe camera is a time-deflection imaging device. The photocathode converts incident light into electrons that are proportional to the light intensity. The electrons undergo different vertical deflections under the action of the scanning electrodes according to the different light intensities, thereby converting the light intensity into electrical intensity.

[0044] In this invention, the light originates from a coherent illumination source, passes through the object under test and a random phase modulation plate, and is finally received by a streak camera, forming a complete light path.

[0045] The stripe camera receives electrical signals and inputs them into the computer in the data processing system.

[0046] The PnP-DIP algorithm of this invention utilizes high-performance computing graphics cards and the CUDA general-purpose parallel computing architecture to accelerate the training and inference processes. The large-scale parallel computing capability of graphics cards can effectively improve the algorithm's running speed and reduce computation time.

[0047] Except for the data processing system, all other components of this invention must be built on an optical platform, and the ambient light requirement should be minimized to reduce the impact on the system's detection results.

[0048] The probe light phase modulation optical path adopts a lensless design, which reduces the impact of phase distortion of the optical path structure caused by lenses.

[0049] The beneficial effects of this invention are that the optical imaging technology, possessing both ultra-sensitive phase detection capability and ultra-high temporal resolution capability, can realize the imaging and diagnosis of dynamic processes of samples on ultra-short timescales. This invention combines coherent modulation imaging technology and compressed ultrafast imaging technology to propose a novel single-exposure compressed ultrafast coherent modulation imaging technique, enabling quantitative intensity and phase imaging of dynamic processes on ultra-short timescales. This provides strong technical support for exploring the dynamic processes of phase-type ultrafast transient events and is expected to promote the progress of many related studies. Attached Figure Description

[0050] Figure 1 This is a schematic diagram of the structure of the present invention;

[0051] Figure 2 A schematic diagram showing how an illumination laser pulse is modulated through an E-shaped pattern to obtain an E-shaped laser pulse.

[0052] Figure 3 The intensity and phase diagrams obtained by capturing E-shaped nanosecond laser pulses in this invention;

[0053] Figure 4 This is a schematic diagram illustrating the real-time detection of the laser ablation process of ITO thin film (indium tin oxide semiconductor transparent conductive film) using the present invention.

[0054] Figure 5The intensity map and phase map obtained by detecting laser ablation of ITO thin films in this invention are shown. Detailed Implementation

[0055] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0056] See Figure 1 The present invention comprises three parts: a probe light phase modulation optical path 100, a streak camera 200, and a data processing system 300. The probe light phase modulation optical path 100 comprises three parts: a coherent illumination source 1, a test object 2 containing phase information (which can be a dynamic scene), and a random phase modulation plate 3. The streak camera 200 is internally composed of six parts: an entrance slit 4, a photocathode 5, a microchannel plate 6, a scanning electrode 7, a fluorescent screen 8, and a built-in camera 9. The data processing system 300 is composed of a computer 10.

[0057] After the system is set up, the object under test (AUT) 2 is first illuminated by ultrafast coherent pulsed light from coherent illumination source 1. Based on the CMI principle, due to the diffraction of coherent light by AUT 2, the beam spreads in the propagation direction. After propagating 70mm, it is modulated by random phase modulation plate 3 to form a large divergence angle, granular diffraction spot, thus reducing the requirements for the dynamic response range of the detector. After propagating another 30mm, based on the principle of compressed ultrafast imaging, the beam carrying the complex amplitude information of the object enters a slit 4 with the slit fully open, streak camera 200. The photocathode 5 inside the streak camera converts the light signal into an electronic signal, which passes through microchannel plate 6. During transmission, the electron pulses are deflected by the linear electric field in the deflection plate of scanning electrode 7, causing electrons arriving at different times to be spatially separated. Finally, the deflected electrons bombard the fluorescent screen 8, regenerating the light signal, which is recorded by the built-in camera image acquisition chip 9, resulting in a compressed diffraction holographic pattern, completing the forward detection process. The entire optical path adopts a lensless design, avoiding wavefront deformation caused by optical elements themselves or manufacturing defects.

[0058] See Figure 1 The distance between the object under test 2 and the random phase modulation plate 3 is 70mm, and the distance between the random phase modulation plate 3 and the incident plane of the stripe camera 200 is 30mm.

[0059] See Figure 1 The coherent illumination source 1 is a 532 nm Q-switched laser with a pulse width of 10 nanoseconds (full width at half maximum) and a maximum pulse energy of 5 millijoules. The random phase modulation plate 3 is formed by irregular etching at a depth of 577 nm on a fused silica substrate, with a unit size of 50 μm × 50 μm, and exhibits a binary random distribution of 0 and π phase modulation at a wavelength of 532 nm.

[0060] See Figure 1To capture the complete ultrafast process, the streak camera 200 operates in one-dimensional mode, selecting a time stop close to the event being measured for recording. Depending on requirements, a line-scan ultra-high-speed camera could be considered as an alternative. Typical line-scan ultra-high-speed cameras offer continuous exposure speeds in the hundreds of kHz range and time resolutions in the microsecond range.

[0061] See Figure 1 After the streak camera 200 records the compressed diffraction holographic pattern, the PnP-DIP algorithm, based on a plug-and-play framework and depth image prior, is used to process the recorded compressed diffraction holographic pattern. Through image registration, transformation, cropping, and normalization, the phase and amplitude information of the measured object is reconstructed. The PnP-DIP algorithm is essentially a reconstruction algorithm that restores a spatiotemporal three-dimensional image from acquired two-dimensional spatial data. The CS-CMUI device has high flexibility and tunability, and can achieve three-dimensional imaging of objects through scanning or multi-angle imaging, obtaining more comprehensive and accurate information, thereby improving the phase accuracy and stability of the optical system.

[0062] Example 1

[0063] This embodiment verifies the intensity and phase detection capabilities of the present invention by measuring modulated nanosecond laser pulses. (See also...) Figure 1 and Figure 2 An "E"-shaped modulation pattern is added between the coherent illumination source 1 and the object under test 2. The pattern is formed by a chromium-plated thin film on a hollow E-shaped quartz glass substrate, and the illumination laser pulse is spatially modulated through the pattern to generate a laser pulse with an E-shaped wavefront.

[0064] See Figure 1 The object under test 2 is illuminated by an E-shaped pulsed light. Based on the CMI principle, due to the diffraction of coherent light by the object under test 2, the beam will spread in the propagation direction. After propagating a distance of 70mm, it is modulated by a random phase plate 3 to form a large divergence angle and a granular diffraction spot. After propagating another 30mm, based on the principle of compressed ultrafast imaging, the beam carrying the complex amplitude information of the object enters a stripe camera 200 with the slit 4 fully open. The photocathode 5 inside the stripe camera converts the light signal into an electronic signal and passes it through a microchannel plate 6. During the transmission process, the electronic pulse is deflected by the linear electric field in the deflection plate of the scanning electrode 7, causing the electrons arriving at different times to be separated in space. Finally, the deflected electrons bombard the fluorescent screen 8 and generate a light signal again, which is recorded by the built-in camera image acquisition chip 9 to obtain a compressed diffraction holographic pattern, completing the forward detection process.

[0065] See Figure 1The distance between the object under test 2 and the random phase modulation plate 3 is 70mm, and the distance between the random phase modulation plate 3 and the incident plane of the stripe camera 200 is 30mm.

[0066] See Figure 1 and Figure 2 The coherent illumination source 1 is a 532 nm Q-switched laser with a pulse width of 10 nanoseconds (full width at half maximum) and a maximum pulse energy of 5 millijoules. The center point of the "E"-shaped modulation pattern intersects with the beam, and the plane of the pattern is perpendicular to the direction of light propagation. The random phase modulation plate 3 is formed by irregular etching at a depth of 577 nm on the fused silica substrate. The unit size is 50 μm × 50 μm, and it exhibits a binary random distribution of 0 and π phase modulation at a wavelength of 532 nm.

[0067] See Figure 1 In order to capture the complete ultrafast process, the stripe camera 200 operates in one-dimensional mode, which selects a time setting close to the event being measured for recording.

[0068] See Figure 1 and Figure 3 After the stripe camera 200 records the compressed diffraction holographic pattern, the data processing system 300 registers and transforms the acquired data. After cropping and normalization, the NVIDIA RTX 3090 high-performance graphics card and PnP-DIP algorithm are used to perform parallel calculations on the recorded compressed diffraction holographic pattern to process the data and reconstruct the phase and amplitude information of the measured object 2. Figure 3 This refers to the phase and amplitude data of the E-shaped laser pulse collected by the computer 10 in the data processing system 300.

[0069] See Figure 3 The CS-CMUI device recovered six intensity and six phase images with a frame interval of 2ns, ranging from -5ns to 5ns. Here, the zero point of time was set to the moment corresponding to the maximum laser pulse intensity.

[0070] See Figure 3 To obtain the phase gradient, a first-order difference is performed on all phase images, which can well reflect the changes in curvature inside the laser cavity mirror. The spatial phase distribution exhibits a consistent concave trend, which is consistent with the physical characteristics of the rear cavity mirror in a Q-switched laser.

[0071] Example 2

[0072] This embodiment aims to observe the intensity and phase changes during the ultrafast dynamic process of nanosecond laser ablation of ITO thin films (indium tin oxide semiconductor transparent conductive films). The scientific community has extensive interest in the interaction between nanosecond lasers and materials, both in fundamental research and applications. Ultrafast lasers can alter the state and properties of materials, making them suitable for high-quality, high-precision, and complex three-dimensional structure processing of almost all materials. This invention uses the observation of the ultrafast dynamic process of nanosecond laser ablation of ITO thin films as an example to further explore this field.

[0073] See Figure 1 and Figure 4 The coherent illumination source 1 in the phase modulation optical path 100 of the probe light is Figure 4 The probe laser pulse that passes through after ablation is replaced.

[0074] See Figure 4 A nanosecond pump laser pulse was obliquely incident on the surface of an ITO thin film, and a separate nanosecond probe laser pulse perpendicular to the surface was used to detect the laser-induced ablation dynamics. The ablation process induced by the pump laser pulse was observed by measuring the spatiotemporal intensity and phase evolution of the probe laser pulse transmitted through the ITO thin film. Simultaneously, the spatiotemporal intensity and phase distribution of the laser pulse from the transmission probe without pump laser irradiation were recorded as a reference.

[0075] See Figure 5 The spatiotemporal intensity evolution of the measured laser pulse is normalized to that of the reference laser pulse to obtain the spatiotemporal intensity change of the measured laser pulse. The spatiotemporal phase change of the measured laser pulse can be determined by subtracting the spatiotemporal phase distribution of the reference laser pulse from the spatiotemporal phase distribution of the measured laser pulse.

[0076] See Figure 4 In this embodiment, the sample used is an ITO thin film coated on a soda-lime glass substrate; the pump light is used to induce an ablation dynamic scene on the ITO sample, and it is incident from the glass substrate side to enhance the ablation effect.

[0077] See Figure 4 The thickness of the ITO film is 200±50nm; the pump light and the probe light both come from the same two parts of the 532nm coherent illumination source 1, and the single pulse energy of the incident pump laser pulse is measured to be 1.6mJ.

[0078] See Figure 1 and Figure 4After ablation, the probe laser pulses that pass through replace the coherent illumination source 1 to illuminate the object under test 2. After being diffracted by the object under test 2, the light beam propagates for 70 mm and is modulated by the random phase plate 3 to form a large divergence angle, granular diffraction spot. After propagating for another 30 mm, the beam carrying the complex amplitude information of the object enters a stripe camera 200 with the slit 4 fully open. The photocathode 5 inside the stripe camera converts the light signal into an electronic signal and passes it through the microchannel plate 6. During the transmission process, the electronic pulses are deflected by the linear electric field in the deflection plate of the scanning electrode 7, causing the electrons arriving at different times to be separated in space. Finally, the deflected electrons bombard the fluorescent screen 8 and generate a new light signal, which is recorded by the built-in camera image acquisition chip 9 to obtain a compressed diffraction holographic pattern, completing the forward detection process.

[0079] See Figure 1 , Figure 4 and Figure 5 The measured object 2 is a dynamic event that changes extremely rapidly on the nanosecond scale. After the streak camera 200 records the compressed diffraction holographic pattern, the data processing system 300 registers and transforms the acquired data, and after cropping and normalization, uses the NVIDIA RTX 3090 high-performance graphics card and the PnP-DIP algorithm to perform parallel calculations on the recorded compressed diffraction holographic pattern, thereby reconstructing the phase and amplitude information of the ultrafast dynamic event. Figure 5 This refers to the phase and amplitude data of the nanosecond laser ablation of the ITO thin film acquired by the computer 10 in the data processing system 300.

[0080] See Figure 5 In the selected -9 to 5 ns intensity and phase images, the time zero point corresponds to the maximum pump laser pulse intensity.

[0081] See Figure 5 Before ablation induced by the pump laser pulse occurs, the phase distribution remains almost constant. As the instantaneous intensity within the pump laser pulse increases, the phase value of the ablation region begins to decrease. A sudden phase change is observed at approximately -2 ns, after which the region expands as the phase value decreases.

[0082] See Figure 5 The phase shape of the ablation pit is affected by the laser flux. When the laser flux is low, the phase shape is circular or nearly circular. However, when the laser flux is high, the phase shape becomes very complex, and the maximum phase difference caused by laser ablation can reach 0.516 radians. The detailed observation of ablation dynamics in this invention provides a key solution for analyzing the laser-matter interaction mechanism and optimizing laser parameters, which in turn promotes the further development and application of laser processing in various fields such as manufacturing, medicine, and aerospace.

[0083] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A single-exposure compressed ultrafast coherent modulation imaging device, characterized in that, The device includes a probe light phase modulation optical path (100), a stripe camera (200), and a data processing system (300); The probe light phase modulation optical path (100) includes a coherent illumination source (1), a test object containing phase information (2), and a random phase modulation plate (3) connected in sequence. The probe light phase modulation optical path (100) adopts a lensless design, which reduces the influence of phase distortion of the optical path structure caused by lenses; The stripe camera (200) includes an entrance slit (4), a photocathode (5), a microchannel plate (6), a scanning electrode (7), a fluorescent screen (8), and a built-in camera (9) connected in sequence by optical paths. The data processing system (300) consists of a computer (10), which is equipped with a computing graphics card and a CUDA general-purpose parallel computing architecture; wherein: The lens center of the coherent illumination source (1), the geometric center of the object under test (2), the geometric center of the random phase modulation plate (3), and the lens center of the streak camera (200) are on the same straight line, which is the direction of light propagation. The coherent illumination source (1), the object under test containing phase information (2), the random phase modulation plate (3) and the stripe camera (200) correspond to four planes in the optical path, namely the outgoing plane, the incident plane, the modulation plane and the detection plane, respectively; all four planes are perpendicular to the direction of light propagation, and the intersection of the optical path and the plane is at the geometric center of the plane.

2. The single-exposure compressed ultrafast coherent modulation imaging apparatus according to claim 1, characterized in that, The CUDA general-purpose parallel computing architecture includes an image reconstruction algorithm. This algorithm is a spatial two-dimensional to spatiotemporal three-dimensional image reconstruction algorithm, namely the PnP-DIP algorithm, which is based on a plug-and-play framework and depth image priors. Its specific data processing process is as follows: Let the complex amplitude of the object to be measured be: U0(x,y,t;0)=I(x,y,t;0)exp[iφ(x,y,t;0)] (1) Where i represents the imaginary unit, x and y are the two-dimensional spatial coordinates of the object, t represents time, and the 0 after the semicolon represents the current propagation distance; I(x,y,t;0) represents the amplitude information of the object, and φ(x,y,t;0) represents the phase information of the object; after propagation over a distance d1, the complex amplitude of the diffracted light is expressed as: in The transfer function represents the distance d1, k represents the wave number, and λ represents the wavelength of the coherent illumination source; It is the two-dimensional Fourier transform of U0, f x and f y These are the spatial frequencies of the image in the x and y directions, respectively; the complex amplitude after passing through the random phase modulation plate is expressed as: Where m(x,y) represents the spatial phase distribution of the phase modulation plate; U M After propagating for another d2 distance, the complex amplitude of the diffracted light is expressed as: At this moment, the light intensity information received by the camera is: Inside the streak camera, the time-varying light intensity information is deflected and then integrated and superimposed, resulting in the following compressed result: Where v represents the deflection speed of the streak camera; all the above processes can be represented in matrix form as follows: Where U(x,y,t) represents the complex amplitude of the object under test, and H d1 and H d2 For spatial transfer operators, M represents the phase modulation operator, S represents the spatiotemporal deflection operator of the streak camera, and T represents the integration operator of the streak camera CMOS during a single exposure. A further simplified representation of the forward acquisition process is as follows: y = Ax (8) in This represents the complex amplitude distribution of the object under test. Represents the measurement matrix. This represents the final compressed data result; here, variables W, H, and B represent the number of pixels in the x, y, and t directions of the dynamic scene, respectively, and variable N represents the number of pixels in the y direction in the compressed snapshot coordinate system; the value of N is greater than the value of H, but smaller than the value of H×B; the data recovery process is the process of calculating the test data x from the known signal y; from a mathematical perspective, solving for the original signal x given y and A is a clearly underdetermined problem, because the amount of known data is much smaller than the amount of unknown data, so to accurately recover the signal, other regular constraints must be added to reduce the size of the solution space; at the same time, since the matrix form of A is too large, solving A... -1 The process of finding the inverse is quite computationally and time-consuming. Here, we introduce DIP to replace the inverse process, and rewrite the inverse problem as an optimization problem in the following form: stx = G(θ,z) Where R(x) is a prior term, λ is a regularization parameter used to balance the fidelity term and the prior term; G(θ,z) represents a deep convolutional neural network with updatable parameters, whose input is... It is a standard Gaussian white noise of the same size as the original data x, which remains unchanged during the optimization process. θ is an updatable weight parameter in the network, and G(θ,z) is abbreviated as G(θ). Furthermore, by introducing an updatable auxiliary parameter u and a manually set penalty parameter ρ, an augmented Lagrangian function L is constructed. ρ The constrained optimization problem can be rewritten as an unconstrained optimization problem: Using the ADMM framework, The minimization problem can be broken down into the following three subproblems:

1.

2. 3.u k+1 =u k +(x k+1 -G(θ k+1 )). (13) For the update of x, it is observed that it is the input G(θ) k )-u k about The form of the proximal operator is used here, and a plug-and-play design is used to replace the proximal operator with a denoising algorithm to prevent overfitting of the depth image prior; for the update of θ, a first-order gradient descent step is used in each iteration; It is easy to see from equations (10), (11), (12), and (13) that as the number of iterations k increases, the solution object x... k+1 It will continuously approach the complex amplitude distribution x of the real object under test, and finally converge to x when k is infinitely large. According to equation (8), x is the complex amplitude information of the object under test, U0(x,y,t;0)=I(x,y,t;0)exp[iφ(x,y,t;0)]; by manually selecting the number of iterations k, the phase and amplitude information of the object under test can be reconstructed from the obtained compressed data results.

Images