A single-exposure imaging system and method based on fully-optical link differentiable computation and neural network representation

Abstract

The application discloses a single-exposure imaging system and method based on all-optical link differentiable computing and neural network representation, and belongs to the field of optical computing imaging. In the all-optical link, only a single thick lens is used as an independent wavefront modulation device, and an image sensor is fixed at any reasonable position behind the thick lens. Based on Fourier optics and the space-bandwidth product matching criterion, and in combination with strict constraints on the exit angle of precise ray tracing, the application realizes field calibration without frequency domain aliasing, and breaks the spatial constraint of conjugate image planes. Based on the rotational symmetry of the lens, the application constructs a differentiable thick lens forward diffraction model with local optical path compensation, and combines an implicit neural network with fused hash coding and a variable period activation function to perform optimization. By calculating the loss of network prediction and a single real image and performing reverse propagation, the application can realize high-fidelity and anti-twin image object reconstruction under the condition of no additional customized device and no mechanical scanning, only by using a single intensity image.

Description

Technical Field

[0001] This invention belongs to the field of optical computational imaging, specifically relating to a single-exposure imaging system and method based on all-optical link differentiable computation and neural network representation. The system utilizes a single thick lens as an independent wavefront modulation element. By constructing an all-optical link differentiable computational physical model and a neural network representation architecture, a single light field intensity map can be acquired at any axial position of the image sensor, thereby achieving computational imaging for high-fidelity object reconstruction via the all-optical link. Background Technology

[0002] In the traditional image-forming optics paradigm, system design relies heavily on point-to-point geometric mappings of object and image conjugates. To acquire a clear image in the detector plane that aligns with human visual intuition, traditional optical systems must cascade multiple lenses of different materials and curvatures to correct aberrations. However, this hardware design centered on aberration correction is inherently a highly over-constrained physical engineering problem, inevitably leading to bulky and significantly heavier imaging systems.

[0003] With the development of information optics, computational imaging technology has broken the traditional constraints of WYSIWYG (What You See Is What You Get). This technology no longer focuses on perfect point-to-point light convergence at the physical level, but instead regards optical devices as information encoding and transmission media in the all-optical link. Through the collaborative design of front-end optical modulation and back-end algorithm interpretation, it greatly expands the functional boundaries of the system.

[0004] Introducing lenses with optical positive focal length into the optical path as information modulation and light-collecting elements within the system is a preferred approach to improve the system's information transmission efficiency. However, current computational imaging schemes based on lens modulation still face three major academic and engineering bottlenecks that urgently need to be overcome as the system architecture evolves towards extreme simplification: 1. The ill-conditioned nature of solving the inverse problem leads to a heavy reliance on multi-plane mechanical scanning: When the light wave interference field is recorded only by conventional image sensors (such as CMOS / CCD), only the amplitude intensity information is retained, while the phase information is completely lost. In the inversion of a single interferogram without phase constraints, the algorithm is prone to getting trapped in local minima, resulting in a conjugate background overlapping with the real object on the reconstruction surface, i.e., a twin-image artifact. Existing holographic phase retrieval techniques (such as the Gerchberg-Saxton algorithm) must rely on a high-precision electromechanical translation stage to drive the sensor to perform multiple mechanical displacements and exposure samplings on different axial detection planes behind the lens in order to overcome this ill-conditioned inverse problem. However, this multi-shot mechanism relying on multi-plane mechanical scanning brings significant engineering drawbacks: the large and expensive precision displacement mechanism makes the system bulky and costly, making it difficult to achieve miniaturization and integration of the equipment. Therefore, finding a minimalist imaging mechanism that eliminates mechanical scanning displacement mechanisms and achieves high-fidelity reconstruction with only a single fixed-point exposure has become a core technological bottleneck that urgently needs to be overcome in this field.

[0005] 2. In all-optical information transmission, to accurately match the spatial bandwidth product of the system at the computing end and ensure that the sensor can fully resolve all spatial information of the object without frequency domain aliasing, existing computational imaging solutions often require the design and custom fabrication of optical components with specific parameters. This technical approach, which heavily relies on custom optical devices, raises the manufacturing threshold for imaging equipment and contradicts the goals of computational imaging, which pursue a minimalist hardware architecture and low-cost universality.

[0006] 3. Computational Bottlenecks in Traditional Diffraction Optics Simulation: To reproduce the optical properties of a thick lens in a computational model, traditional physical optics approaches involve densely dividing the three-dimensional lens surface into tiny two-dimensional finite element meshes and performing full-aperture volume propagation using Rayleigh-Sommerfeld diffraction integrals or angular spectral methods. This complex calculus operation based on a two-dimensional mesh is extremely computationally intensive.

[0007] Therefore, there is an urgent need to propose a novel computational imaging system architecture that combines profound optical physics principles with efficient artificial intelligence computing power. This architecture needs to overcome the limitations of mechanical multi-faceted scanning, break down the barriers of relying on expensive custom optical components, and achieve high-fidelity reconstruction of objects in a single, fixed-point shot at any location using only a single thick lens (such as an inexpensive commercial thick lens with inherent aberrations) as the physical encoding medium, under a minimal all-optical link architecture. Summary of the Invention

[0008] The present invention aims to overcome the shortcomings of the prior art and provide a single-exposure imaging system and method based on all-optical link differentiable computing and neural network representation.

[0009] According to a first aspect of the present invention, the present invention provides a single-exposure imaging system based on all-optical link differentiable computation and neural network representation, the system comprising an active illumination module, an independent wavefront modulation device and an image sensor arranged sequentially along the optical path, and a computational processing module communicatively connected to the image sensor; The active illumination module is used to illuminate the object to be imaged, which is located in front of the independent wavefront modulator. Independent wavefront modulators are lenses with a defined effective focal length and physical thickness, used to collect light fields with high-frequency interference information from the object to be imaged, and to realize nonlinear modulation and space-frequency conversion transmission of wavefront phase. The lens is rotationally symmetrical about the central principal optical axis. The image sensor is positioned at any distance along the principal optical axis behind the independent wavefront modulator without beam interruption, and is used to record a single real-shot light field intensity map carrying interference information of the object to be imaged in a single exposure. The computational processing module is equipped with a physical model of forward diffraction transmission through a microthickness lens and an implicit neural network, which are used to reconstruct the light field intensity map into a high-fidelity image of the target object. The physical model for forward diffraction transmission of the microthickness lens is based on the rotationally symmetric geometric prior of the independent wavefront modulator. The light-transmitting region of the lens is discretized radially into multiple concentric rings. The optical path difference between the beam and the vacuum gap through the actual medium is calculated ring by ring and transformed into a complex phase compensation matrix. Finally, the predicted light field intensity distribution at the image sensor plane is calculated by the angular spectrum diffraction propagation operator.

[0010] According to a preferred embodiment of the present invention, the active illumination module includes a light source generator, a spatial filter, and a collimating beam expander assembly; The light source generator is selected from any one of a laser light source, a narrow-linewidth monochromatic LED light source, or a broadband light source with narrow-band filtering; the spatial filter is used to filter out high-frequency stray light; the collimation and beam expansion system is used to collimate and expand the output beam of the light source generator, and the collimated and expanded beam is then used to illuminate the object to be imaged.

[0011] According to a preferred embodiment of the present invention, the independent wavefront modulation device is selected from any one of a spherical lens, an aspherical lens, a freeform lens, a graded refractive index lens, or a composite lens with residual aberrations; the object to be imaged is located within the physical neighborhood of the front focal plane of the independent wavefront modulation device.

[0012] According to a preferred embodiment of the present invention, the photosensitive surface of the image sensor is perpendicular to the main optical axis of the system, and allows it to be fixed at any axial distance along the main optical axis behind the independent wavefront modulator without physical beam interruption, without needing to align with an absolutely conjugate image plane.

[0013] According to a preferred embodiment of the present invention, the computational processing module performs the following steps when reconstructing a high-fidelity image of the target object: 1) In the computation processing module, predefine a continuous spatial domain two-dimensional physical coordinate matrix covering the plane of the object to be reconstructed; initialize an implicit neural network, which takes the continuous spatial domain two-dimensional physical coordinate matrix as input and outputs the predicted distribution of the object to be imaged; 2) Input the predicted distribution of the object to be imaged into the physical model of forward diffraction transmission of the microthickness lens to calculate the predicted light field intensity map at the plane corresponding to the image sensor. 3) Calculate the error between the predicted light field intensity map and the actual light field intensity map, and use it as the loss function; 4) The weight parameters of the implicit neural network are updated by backpropagation based on the automatic differentiation mechanism. The process is iterated until the network converges. After convergence, the implicit neural network directly interprets and outputs a high-fidelity reconstructed image without aberrations.

[0014] According to a preferred embodiment of the present invention, the forward diffraction transmission model of the microthickness-adjustable lens utilizes the rotational symmetry of the lens to equivalently divide the lens into multiple concentric rings. For each concentric ring, when the light wave propagates from the equivalent inner plane to the equivalent outer plane, the difference in optical path length between the real medium and air is converted into a tensorized concentric ring matrix. The predicted distribution of the object to be imaged is obtained by the complex amplitude of the light field modulated by the lens. After modulation, each ring is uniformly coherently superimposed and forward calculated to the specified plane using the angular spectrum diffraction propagation nucleus.

[0015] According to a second aspect of the present invention, a single-exposure imaging method based on all-optical link differentiable computation and neural network representation is provided. This method is applied to the aforementioned single-exposure computational imaging system, and the method includes at least the following steps: 1) Calculate the effective focal length of independent wavefront modulators based on the spatial bandwidth product matching principle. f The feasible range, combined with precise ray tracing calibration, is used to calculate the maximum effective object-side true field of view of the imaging system without frequency domain aliasing under the influence of aberrations in a single exposure; 2) Place the object to be imaged within the front focal plane neighborhood of the maximum effective object-side true field of view and illuminate it. Use an image sensor to capture a single real-shot light field intensity map at a fixed position behind the system's optical path in a single exposure. 3) Construct a computational processing module that includes a forward diffraction transmission model of a microthickness-adjustable lens and an implicit neural network; 4) Using the error between the predicted light field intensity map and the single real-shot light field intensity map as the loss function, the weights of the implicit neural network are updated by backpropagation. After convergence, the implicit neural network directly interprets and outputs a high-fidelity reconstructed image without aberrations.

[0016] The network architecture of the implicit neural network includes at least: The coordinate nonlinear mapping hash layer is used to transform the input continuous spatial domain two-dimensional physical coordinate matrix into high-dimensional discrete features through a hash table mapping mechanism, so as to break the strong spatial correlation between adjacent pixel domains of the image. The multilayer sensing band activation layer is used to receive the high-dimensional discrete features, and its hidden layer adopts a nonlinear activation function containing a variable period sine to capture the low-frequency structure and high-frequency details of the object to be imaged.

[0017] Compared with the prior art, the present invention has the following beneficial effects: 1. Unlike traditional systems that rely on spatial bandwidth product matching to customize lens surface parameters, this system transforms inexpensive commercial single lenses with inherent aberrations into high-bandwidth information encoding devices. This not only achieves the same amount of information collection as traditional bulky imaging systems through an extremely simple system, but also physically removes the spatial constraints on the conjugate image plane, giving the system's detector the ability to receive light field information losslessly at any reasonable position along the optical axis.

[0018] 2. This system relies on the regularized analytical prior ability of implicit neural networks. Under the condition of a single intensity image without direct wavefront phase data constraints, it can robustly converge and reconstruct the measured object, realizing a simplified system architecture that eliminates the need for mechanical displacement scanning components. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 A schematic diagram of the overall optical path architecture and hardware physical layout of a single-exposure computational imaging system based on a single commercial thick lens, provided for embodiments of the present invention; Figure 2 This is a schematic diagram of the system hardware parameter forward optimization and field of view calibration process provided in the embodiment of the present invention; the process sequentially defines the sensor through information theory requirements, defines the lens focal length through bidirectional physical boundaries, and finally verifies the system's true limit performance through software tracing. Figure 3 This is a schematic diagram of the forward diffraction transmission model of a microthick lens constructed according to an embodiment of the present invention. It shows in detail the process of discretizing the thick lens entity into concentric rings along the radial direction and performing local optical path difference (OPD) phase compensation between equivalent virtual planes. Figure 4The following is a single-exposure reconstruction closed-loop data flow graph of an implicit network that integrates hash-encoded implicit neural representation and variable periodic activation function, provided in an embodiment of the present invention. Figure 5 The images provided in this embodiment of the invention are a comparison of the system imaging effects. The left side is the actual light field amplitude distribution of the target surface obtained from a single exposure, and the right side is a high-fidelity, artifact-free object reconstruction image output after joint optimization by the implicit neural network and differentiable physical model of this invention.

[0021] Explanation of reference numerals in the attached figures: 100-Active illumination module; 101-Coherent light source; 102-Spatial filter; 103-Collimating and beam expanding lens group; 200-Object to be imaged; 300-Independent wavefront modulator; 400-Image sensor; 500-Computational processing module. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] In this document, the terms “comprising,” “including,” or any other variations thereof are intended to cover non-exclusive inclusion, which includes not only the elements listed but also other elements not expressly listed.

[0024] In this document, the directional terms such as front, back, top, and bottom are defined based on the location of the components in the accompanying drawings and their relative positions to each other, solely for the purpose of clarity and convenience in expressing the technical solution. It should be understood that the use of these directional terms should not limit the scope of protection claimed in this application.

[0025] As shown in Figure 1, this embodiment of the invention provides a single-exposure imaging system based on all-optical link differentiable computation and neural network representation. In terms of hardware architecture, the system is arranged along the main optical axis as follows: an active illumination module 100 providing monochromatic coherent illumination light waves, an independent wavefront modulator 300, an image sensor 400 for recording the intensity of the interference light field, and a computational processing module 500 connected to the image sensor. When using the system, the object to be imaged 200 is positioned between the active illumination module 100 and the independent wavefront modulator 300, and is located within the physical neighborhood of the front focal plane of the independent wavefront modulator. The specific functional configurations of each component in the system are as follows: The active illumination module 100 is used to emit coherent or partially coherent light beams to uniformly illuminate the object 200 to be imaged; An independent wavefront modulator 300, positioned behind the object to be imaged, collects the light field containing high-frequency interference information from the object and performs nonlinear modulation and space-frequency conversion transmission of the wavefront phase. The independent wavefront modulator is a lens with a defined effective focal length and physical thickness. In this invention, the independent wavefront modulator is chosen as a single thick lens. A thick lens refers to a lens with a thickness close to or exceeding its focal length; this concept distinguishes it from a thin lens. In the industry, a thin lens refers to a lens with a center thickness much smaller than its focal length. Calculations for thin lenses can assume that all refraction occurs only in a virtual plane at the center of the lens, but the thick lens of this invention cannot make such an assumption. In embodiments of this invention, the thickness of the thick lens is greater than 10 wavelengths. The thick lens of this invention is preferably a commercially available thick lens, which is inexpensive and requires no special manufacturing or additional processing.

[0026] The image sensor 400 is positioned at any reasonable distance behind the thick lens along the principal optical axis without beam interruption, and is used to record a single real-shot light field intensity map carrying interference information of the object to be imaged through a single exposure. The computational processing module 500 is communicatively connected to the image sensor and is equipped with a physical model of forward diffraction transmission of a microthickness-capable lens and an implicit neural network. This model is used to reconstruct a high-fidelity image of the target object from the recorded single light field intensity information. The physical model of forward diffraction transmission of the microthickness-capable lens is based on the rotationally symmetric geometric prior of an independent wavefront modulator. It discretizes the light-transmitting region of the lens radially into multiple concentric rings, calculates the optical path difference of the beam through the actual medium and the vacuum gap for each ring, and converts it into a complex phase compensation matrix. Finally, the predicted light field intensity distribution at the plane of the image sensor is calculated by the angular spectrum diffraction propagation operator.

[0027] It should be noted beforehand that, in this invention, the independent wavefront modulation device 300 refers to a single modular component in the optical path used for light collection and wavefront phase modulation. In the preferred embodiments described below, to highlight the low-cost advantage of the minimalist architecture of this invention, the device is specifically implemented and described as a single commercially available thick lens (such as a spherical, aspherical, or freeform single lens with inherent physical thickness). However, this is only one embodiment of the invention, and any lens element with a specific effective focal length, regardless of whether it has residual aberrations (including conventional commercially available compound lenses with residual aberrations), should be equivalently included within the definition of the independent wavefront modulation device of this invention.

[0028] Furthermore, the physical carrier of the computing processing module 500 mentioned in the embodiments is not limited. It can be a local high-performance personal computer or a dedicated processing chip directly connected to the image sensor, an edge computing terminal that is independent of the field equipment, or a cloud server computing platform that transmits the real-shot light field intensity map back via the Internet and performs large-scale parallel computing.

[0029] The complete workflow and core technology implementation scheme of this embodiment can be divided into the following four core steps with close logical connections: 1. Selection of all-optical link parameters based on Fourier space-frequency mapping and spatial bandwidth product matching. This step aims to establish a rigorous set of mathematical and physical boundary conditions, forward deduce and lock in the optimal hardware configuration of the system, and at the same time establish the physical legitimacy of fixing the image sensor at any position along the optical axis.

[0030] First, establish the target field of view (FOV) and target numerical aperture of the system. A commercial image sensor was selected based on the Spatial Bandwidth Product (SBP) matching principle. The lens, acting as a spatial-frequency converter, has an effective focal length... f The choice is constrained by the already selected image sensor: on the one hand, for frequency domain anti-aliasing, the limiting sine of the light emission angle that the image sensor can interpret without distortion is [value missing]. ,in This is the maximum outgoing ray deflection angle (or the limiting ray reception and outgoing angle). For the operating wavelength and the coherent beam wavelength, p Given the physical pixel size of the image sensor, to ensure that light waves at the edge of the field of view do not undergo Nyquist aliasing, the focal length must satisfy the lower bound constraint. ,in On the one hand, it represents the field radius of view of the object-side target; on the other hand, it represents spatial anti-truncation. According to Fourier optics theory, when the object to be imaged is strictly positioned at the front focal plane of the lens on the object side, any spatial coordinates on the object plane... The diverging point source, after wavefront phase modulation by the lens, will be linearly transformed into a parallel plane wave propagating obliquely in a specific direction in the space behind the lens. Therefore, the overall emitted light field of the object can be regarded as a linear superposition of a series of parallel plane wave spectra with definite spatial frequency components. To avoid high-frequency beam overflow from the target surface and resulting information loss, the beam diameter must be less than or equal to the short side of the sensor, that is, the focal length must satisfy the upper bound constraint. , The physical photosensitive short side width of the image sensor is used. The effective focal length is calculated based on the above constraints. f The feasible range is determined, and the required single-piece thick lens is selected within this range (e.g., directly selecting commercially available thick lenses to reduce costs).

[0031] However, in practical all-optical systems engineering, the monolithic thick lenses used often have inherent physical thickness and uncorrected residual aberrations, resulting in strong nonlinearity in the deflection of edge rays. Therefore, this embodiment must import the exact model of the thick lens into the optical software to achieve the limiting exit angle. Using this as a boundary, precise reverse ray tracing is used to determine the true maximum object height of the front focal plane, which is strictly defined as the maximum effective non-aliasing true field of view. As long as the object is placed within this calibrated field of view, the frequency of the emitted light field is strictly limited within the image sensor's interpretation bandwidth, thus allowing the sensor to break free from the constraints of the traditional conjugate image plane and perform fixed-point imaging at any position behind the lens without losing physical information.

[0032] Therefore, this embodiment establishes strict criteria for precise ray tracing calibration: 1) Obtain the physical pixel size of the image sensor p Based on the Nyquist-Shannon sampling theorem, the highest spatial fundamental frequency that an image sensor can interpret without distortion is established. This leads to the deduction of the maximum allowable outgoing light deflection angle of the system. .

[0033] 2) During the design process, input the exact refractive index of the selected single-piece thick lens, the central physical thickness, and the actual surface curvature radius of the front and rear surfaces of the lens to establish a high-fidelity digital physical model.

[0034] 3) Set the target object to be located at the front focal plane of the thick lens, and set the maximum outgoing light deflection angle as described above. As a boundary constraint for tracking rays at the system's edge, it enables precise ray tracing.

[0035] 4) Calculate the maximum radial physical height of the corresponding front focal plane that can satisfy the limiting exit angle constraint, under the condition of including all inherent aberrations of the thick lens. This height is strictly defined as the maximum effective non-aliased object-side true field of view (FOV) of this imaging system.

[0036] 2. Construct a differentiable physical model for the all-optical link (specifically, in this embodiment, a differentiable thick lens forward diffraction transmission model): This step aims to eliminate the intrinsic error of the forward model caused by the thin lens approximation, while avoiding the computational disaster caused by traditional three-dimensional finite element mesh integration.

[0037] Combination Figure 3 As shown, this embodiment fully utilizes the geometric and physical prior that the selected optical lens satisfies strict rotational symmetry about the central principal optical axis.

[0038] 1) In the calculation and processing module, let the maximum effective aperture radius of the thick lens be... The light-transmitting region of the lens is projected from a two-dimensional Cartesian coordinate system to a polar coordinate system. Along a single radial direction, the radius interval [0, ... High-density discrete partitioning is divided into N A concentric ring.

[0039] 2) Regarding the first kA concentric ring (whose radial range is defined as [ ] ) , Extract the maximum axial coordinates of the actual three-dimensional entity of the lens within this radial interval. With minimum axial coordinate Based on this, a pair of equivalent external virtual planes and equivalent internal virtual planes perpendicular to the optical axis are constructed.

[0040] 3) When the incident complex amplitude light field Passing through the first k When the lens is looped, the system uses the height function of the lens's true surface shape. and the refractive index of the medium The optical path difference accumulated during the propagation of light waves through the actual glass medium and the external vacuum gap is rigorously calculated. This optical path difference is mapped to a concentric ring matrix, and the output complex amplitude is calculated. satisfy: 4) All wavefronts passing through the concentric annular matrix are coherently superimposed at a unified virtual cross-section at the lens exit surface. The total complex amplitude at the lens exit surface is then obtained. Then, the diffraction propagation operator based on angular spectrum theory is invoked to propagate the wavefront forward with high fidelity to any specified physical cross-section where the image sensor is located (assuming a transmission distance of ). d To address the issues of non-paraxial and large diffraction angles, an accurate angular spectrum diffraction formula considering the scaling mechanism is employed to calculate the predicted complex amplitude reaching the image sensor plane. Its calculation expression is: In the formula, and These represent the two-dimensional Fourier transform and inverse Fourier transform, respectively. and For frequency domain coordinates, The operating wavelength is used. The final value is the square of the mode of the complex amplitude. This will yield the predicted light field intensity map for subsequent neural network alignment. .

[0041] 3. Object reconstruction based on neural network representation architecture. A single exposure is performed with the image sensor fixed to an arbitrary plane to acquire the actual light field intensity map. Subsequently, this embodiment employs an implicit neural network architecture for object reconstruction, combined with... Figure 4 As shown, the computational processing module initializes an implicit neural network, whose input is only a two-dimensional physical coordinate matrix (x,y) in continuous spatial domain, and whose output is the predicted distribution of objects. The network architecture of the implicit neural network includes at least: 1) A coordinate nonlinear mapping hash layer is used to transform the input object's continuous spatial two-dimensional physical coordinate matrix into high-dimensional discrete features through a hash table mapping mechanism. This rearrangement of the underlying data structure breaks the network's dependence on smooth image distribution and greatly enhances its sensitivity to fitting high-frequency signals.

[0042] 2) A multi-layer sensing frequency band activation layer is used to receive the high-dimensional discrete features. Its hidden layer adopts a non-linear activation function containing a variable period sine to capture the low-frequency structure and high-frequency details of the object to be imaged.

[0043] Through the joint forward inference mapping of the aforementioned network, the decoding layer outputs the predicted distribution of the object under test. This predicted distribution is then input into a pre-set forward diffraction transmission model with a subdimensibly thick lens. After rigorous thick lens phase compensation and scalar diffraction calculations, the predicted light field intensity distribution matrix reaching the target surface of the image sensor is calculated in the forward direction. The system establishes a target loss function. The predicted light field intensity map is compared pixel-by-pixel with the actual light field intensity map. Specifically, this loss function... In addition to including the mean squared error (MSE Loss) between the two, a total variation (TV) regularization term can preferably be introduced to constrain the spatial smoothness of the reconstructed object and suppress background noise. Its specific mathematical expression is as follows: In the formula, The hyperparameter weights are used to adjust the strength of total variation regularization. The network predicts the distribution of objects, where the total variation regularization term... Smoothing constraints are achieved by calculating the gradient magnitudes of the predicted image in the horizontal and vertical directions. The specific discretization calculation formula can be expressed as: .

[0044] Subsequently, the automatic differentiation engine is activated, and the optimizer (such as Adam) calculates the gradient of the loss function with respect to the weights of each node within the implicit neural network, and performs backpropagation to update the hash table and MLP weights. This forward inference and backpropagation process is repeated until the global loss function meets the preset convergence criterion, thus acquiring the reconstructed image and completing the all-optical computational imaging task.

[0045] To more intuitively verify the technical effects of this invention, please refer to the references. Figure 5 . Figure 5 This demonstrates a comparison between the raw acquisition data obtained from a single exposure and the final algorithm reconstruction result under a minimal hardware architecture, as shown in this embodiment of the invention: Figure 5The target surface light field amplitude distribution shown on the left is a single light field intensity map actually recorded by the image sensor 400 in a single exposure (corresponding to the previous text). As can be seen, because the system abandons the expensive aberration correction lens group and only uses a single commercially available thick lens with inherent physical thickness and aberrations, and because the image sensor 400 is freed from the constraints of the traditional conjugate image plane and placed in an arbitrary free position behind the lens, the original image exhibits extremely severe blurring. At this point, the naked eye cannot discern any structural information of the object under test (such as a standard resolution target).

[0046] Figure 5 The high-fidelity object reconstruction image shown on the right is the result of inputting the aforementioned highly degraded single intensity image into this computational processing module 500. After iterative optimization through forward inference and backpropagation using the aforementioned implicit neural network combining fused hash encoding and variable periodic activation function with a microthickness lens forward diffraction transmission model, the final decoded output is the object distribution (i.e., the predicted object after network convergence). As can be clearly seen from the image, the reconstructed image of this invention not only accurately restores high-frequency details (the edges of the test target lines are sharp and clear), but also perfectly overcomes the twin artifacts that are easily generated in traditional single-image interferogram phase retrieval algorithms, resulting in a clean reconstructed background without overlapping interference.

[0047] Furthermore, to expand the universality of this system in practical engineering and enable it to be compatible with broadband light sources or partially coherent light sources (such as broadband LED light sources): in the aforementioned forward diffraction transmission model with a microthickness lens, this system can discretize the input light source into multiple different center wavelengths according to the emission spectrum. Sub-wavelength channels. For each wavelength channel When the incident complex amplitude light field Passing through the k When the lens is looped, the system uses the height function of the lens's true surface shape. and the refractive index of the medium The optical path difference accumulated during the propagation of light waves through the actual glass medium and the external vacuum gap is rigorously calculated. This optical path difference is then mapped to a concentric ring matrix, with the complex amplitude emitted from each wavelength channel. satisfy: .

[0048] Finally, at the plane corresponding to the image sensor, the predicted intensity maps of each sub-wavelength channel are incoherently superimposed, resulting in the total predicted intensity of the target surface. By replacing the single-wavelength model with this incoherent superposition model in the loss function calculation, high-fidelity reconstruction of objects can still be achieved under broadband or partially coherent illumination conditions.

[0049] 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, improvements, etc., 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 imaging system based on all-optical link differentiable computation and neural network representation, characterized in that: The system includes an active illumination module, an independent wavefront modulator and an image sensor arranged sequentially along the optical path, as well as a computing processing module that is communicatively connected to the image sensor. The active illumination module is used to illuminate the object to be imaged, which is located in front of the independent wavefront modulator. Independent wavefront modulators are lenses with a defined effective focal length and physical thickness, used to collect light fields with high-frequency interference information from the object to be imaged, and to realize nonlinear modulation and space-frequency conversion transmission of wavefront phase. The lens is rotationally symmetrical about the central principal optical axis. The image sensor is positioned at any distance along the principal optical axis behind the independent wavefront modulator without beam interruption, and is used to record a single real-shot light field intensity map carrying interference information of the object to be imaged in a single exposure. The computational processing module is equipped with a physical model of forward diffraction transmission through a microthickness lens and an implicit neural network, which are used to reconstruct the light field intensity map into a high-fidelity image of the target object. The physical model for forward diffraction transmission of the microthickness lens is based on the rotationally symmetric geometric prior of the independent wavefront modulator. The light-transmitting region of the lens is discretized radially into multiple concentric rings. The optical path difference between the beam and the vacuum gap through the actual medium is calculated ring by ring and transformed into a complex phase compensation matrix. Finally, the predicted light field intensity distribution at the image sensor plane is calculated by the angular spectrum diffraction propagation operator.

2. The single-exposure imaging system according to claim 1, characterized in that, The active lighting module includes a light source generator, a spatial filter, and a collimation and beam expansion assembly. The light source generator is selected from any one of a laser light source, a narrow-linewidth monochromatic LED light source, or a broadband light source with narrow-band filtering; the spatial filter is used to filter out high-frequency stray light; the collimation and beam expansion system is used to collimate and expand the output beam of the light source generator, and the collimated and expanded beam is then used to illuminate the object to be imaged.

3. The single-exposure imaging system according to claim 1, characterized in that, The independent wavefront modulation device is selected from any one of a spherical lens, an aspherical lens, a freeform lens, a graded refractive index lens, or a composite lens with residual aberrations; the object to be imaged is located within the physical neighborhood of the front focal plane of the independent wavefront modulation device.

4. The single-exposure imaging system according to claim 1, characterized in that, The photosensitive surface of the image sensor is perpendicular to the main optical axis of the system, and it can be fixed at any axial distance along the main optical axis behind the independent wavefront modulator without physical beam interruption, without needing to align with the absolute conjugate image plane.

5. The single-exposure imaging system according to claim 1, characterized in that, The effective focal length of the independent wavefront modulator f The selection follows the spatial bandwidth product matching principle; The physical pixel size of the image sensor p With physical photosensitive short side width Based on the spatial bandwidth product matching principle, and determined by the target object field of view and the target numerical aperture set by the system, the minimum number of discrete information sampling points required to record the field of view without damage is established. The image sensor's anti-aliasing limiting light reception angle As the frequency domain tolerance boundary of the system, The wavelength of the coherent beam, the effective focal length f The lower bound condition must be met. ,in The radius of the object's field of view; The effective focal length f Simultaneously satisfying the upper bound condition ,in The target numerical aperture of the system.

6. The single-exposure imaging system according to claim 1, characterized in that, When performing high-fidelity image reconstruction of the target object, the computational processing module performs the following steps: 1) In the computation processing module, predefine a continuous spatial domain two-dimensional physical coordinate matrix covering the plane of the object to be reconstructed; initialize an implicit neural network, which takes the continuous spatial domain two-dimensional physical coordinate matrix as input and outputs the predicted distribution of the object to be imaged; 2) Input the predicted distribution of the object to be imaged into the physical model of forward diffraction transmission of the microthickness lens to calculate the predicted light field intensity map at the plane corresponding to the image sensor. 3) Calculate the error between the predicted light field intensity map and the actual light field intensity map, and use it as the loss function; 4) The weight parameters of the implicit neural network are updated by backpropagation based on the automatic differentiation mechanism. The process is iterated until the network converges. After convergence, the implicit neural network directly interprets and outputs a high-fidelity reconstructed image without aberrations.

7. The single-exposure imaging system according to claim 6, characterized in that, The microthickness-adjustable lens forward diffraction transmission model utilizes the lens's rotational symmetry to equivalently divide the lens into multiple concentric rings. For each concentric ring, when the light wave propagates from the equivalent inner plane to the equivalent outer plane, the difference in optical path length between the real medium and air is converted into a tensorized concentric ring matrix. The predicted distribution of the object to be imaged is then modulated by the lens to obtain the complex amplitude of the optical field. satisfy: ； In the formula, This is the complex amplitude input at the equivalent inner plane of the ring. and These are the global axial spatial coordinates of the equivalent inner and outer planes, respectively. For the lens surface in this The actual contour height function at coordinates, The refractive index of the lens medium is used; after modulation of each ring zone, they are coherently superimposed and the angular spectrum diffraction propagation nucleus is called to calculate the forward direction to the specified plane.

8. The single-exposure imaging system according to claim 6, characterized in that, The network architecture of the implicit neural network includes at least: The coordinate nonlinear mapping hash layer is used to transform the input continuous spatial domain two-dimensional physical coordinate matrix into high-dimensional discrete features through a hash table mapping mechanism, so as to break the strong spatial correlation between adjacent pixel domains of the image. The multilayer sensing band activation layer is used to receive the high-dimensional discrete features, and its hidden layer adopts a nonlinear activation function containing a variable period sine to capture the low-frequency structure and high-frequency details of the object to be imaged.

9. The single-exposure imaging system according to claim 1, characterized in that, The carrier of the computing processing module is selected from at least one of a local personal computing device, an on-site edge computing terminal, or a cloud server computing platform.

10. A single-exposure imaging method based on all-optical link differentiable computation and neural network representation, characterized in that, This method is applied to the single-exposure imaging system according to any one of claims 1 to 9, and the method includes at least the following steps: 1) Calculate the effective focal length of independent wavefront modulators based on the spatial bandwidth product matching principle. f The feasible range, combined with precise ray tracing calibration, is used to calculate the maximum effective object-side true field of view of the imaging system without frequency domain aliasing under the influence of aberrations in a single exposure; 2) Place the object to be imaged within the front focal plane neighborhood of the maximum effective object-side true field of view and illuminate it. Use an image sensor to capture a single real-shot light field intensity map at a fixed position behind the system's optical path in a single exposure. 3) Construct a computational processing module that includes a forward diffraction transmission model of a microthickness-adjustable lens and an implicit neural network; 4) Using the error between the predicted light field intensity map and the single real-shot light field intensity map as the loss function, the weights of the implicit neural network are updated by backpropagation. After convergence, the implicit neural network directly interprets and outputs a high-fidelity reconstructed image without aberrations.

Images