Deep learning-based three-dimensional reconstruction method and system for holographic turbulence field of turbid water body

Abstract

The invention discloses a deep learning-based turbid water body holographic turbulence field three-dimensional reconstruction method and system, and relates to the technical field of underwater optical imaging, and the method comprises the steps: irradiating a turbid water body through a holographic light path, and driving a camera to capture a transient holographic interferogram sequence; denoising by adopting a noise suppression model, and outputting an optimized hologram; positioning particles through a cascade convolutional neural network to generate an initial plane coordinate set; calculating the axial depth of the particles to obtain a transient space coordinate set; space-time diagram association is executed, and a three-dimensional particle field dynamic sequence is generated; and executing flow field inversion to generate a three-dimensional transient turbulence velocity field. The technical problem that a traditional image-based three-dimensional reconstruction method is difficult to overcome the defects of strong holographic image noise and particle signal distortion in a turbid water body scene is solved, interference caused by turbid water body scattering is effectively resisted, key data needed by turbulent flow field three-dimensional reconstruction are comprehensively and accurately obtained, and the reconstruction efficiency is improved. Therefore, the technical effect of three-dimensional reconstruction of the turbulent flow field of the turbid water body is accurately realized.

Description

Technical Field

[0001] This invention relates to the field of underwater optical imaging technology, and in particular to a method and system for three-dimensional reconstruction of holographic turbulent flow fields in turbid water based on deep learning. Background Technology

[0002] The 3D reconstruction of holographic turbulent fields in turbid water bodies is of great significance for waterway engineering monitoring, and image data processing is a core step. Existing technologies mostly rely on manual or simple algorithms to process holographic images, such as manually suppressing scattering noise and roughly locating particles. However, turbid water bodies have strong scattering noise, and traditional image processing methods cannot accurately denoise and locate particles at the sub-pixel level. They are also ill-suited for processing transient holographic interferogram sequences, resulting in incomplete and inaccurate reconstructed turbulent field data that fails to meet the needs of waterway engineering for precise assessment and effective control of turbulent fields in turbid water bodies. Summary of the Invention

[0003] This application provides a method and system for three-dimensional reconstruction of holographic turbulent flow field in turbid water based on deep learning, which solves the technical problems of strong noise and particle signal distortion in traditional image-based three-dimensional reconstruction methods in turbid water scenarios.

[0004] The first aspect of this application provides a method for three-dimensional reconstruction of holographic turbulent field in turbid water based on deep learning. The method includes: during laser coherent irradiation of turbid water using a dual-pulse off-axis holographic optical path, driving a high-speed CMOS camera to capture a sequence of transient holographic interferograms of the turbid water; using a noise suppression model built based on a deep learning network to suppress turbid water scattering noise in the transient holographic interferogram sequence, outputting an optimized hologram; generating an initial particle plane coordinate set by performing two-dimensional particle projection localization driven by a cascaded convolutional neural network on the optimized hologram; calculating the particle axial depth of the initial particle plane coordinate set, outputting a particle transient spatial coordinate set; performing multi-frame trajectory reconstruction with spatiotemporal graph association on the particle transient spatial coordinate set, generating a three-dimensional particle field dynamic sequence; and generating a three-dimensional transient turbulent velocity field by performing flow field inversion on the three-dimensional particle field dynamic sequence.

[0005] A second aspect of this application provides a deep learning-based 3D reconstruction system for holographic turbulent flow fields in turbid water. The system includes: a transient holographic interferogram sequence acquisition module, used to drive a high-speed CMOS camera to capture a transient holographic interferogram sequence of turbid water during laser coherent irradiation of the turbid water using a dual-pulse off-axis holographic optical path; an optimized hologram acquisition module, used to suppress turbid water scattering noise in the transient holographic interferogram sequence using a noise suppression model built based on a deep learning network, and output an optimized hologram; and an initial particle plane coordinate set acquisition module, used to obtain... The optimized hologram is subjected to cascaded convolutional neural network-driven two-dimensional particle projection localization to generate an initial particle plane coordinate set; a particle transient spatial coordinate set acquisition module is used to calculate the particle axial depth of the initial particle plane coordinate set and output the particle transient spatial coordinate set; a three-dimensional particle field dynamic sequence acquisition module is used to perform multi-frame trajectory reconstruction with spatiotemporal graph association on the particle transient spatial coordinate set to generate a three-dimensional particle field dynamic sequence; a three-dimensional transient turbulent velocity field acquisition module is used to generate a three-dimensional transient turbulent velocity field by performing flow field inversion on the three-dimensional particle field dynamic sequence.

[0006] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0007] This application employs a dual-pulse off-axis holographic optical path to illuminate turbid water and capture a sequence of transient holographic interferograms. Through noise suppression, particle localization, axial depth calculation, multi-frame trajectory reconstruction, and flow field inversion, the transient spatial coordinates of particles and the dynamic sequence of a three-dimensional particle field are obtained. The three-dimensional transient turbulent velocity field is calculated, and optimization adjustments are made based on particle motion physical constraints and fluid characteristics. This allows for accurate reconstruction of the three-dimensional holographic turbulent field information in turbid water, making the three-dimensional reconstruction results more accurate and reliable. It effectively resists interference from scattering by turbid water, comprehensively and accurately obtains the key data required for three-dimensional reconstruction of the turbulent field, and thus precisely achieves the technical effect of three-dimensional reconstruction of the turbulent field in turbid water. Attached Figure Description

[0008] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying 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.

[0009] Figure 1 This is a flowchart illustrating the three-dimensional reconstruction method for holographic turbulent flow field in turbid water based on deep learning, provided in an embodiment of this application.

[0010] Figure 2This is a schematic diagram of the structure of the 3D reconstruction system for holographic turbulent flow field of turbid water based on deep learning provided in the embodiments of this application.

[0011] Figure labeling: Transient holographic interferogram sequence acquisition module 1, optimized hologram acquisition module 2, initial particle plane coordinate set acquisition module 3, particle transient spatial coordinate set acquisition module 4, three-dimensional particle field dynamic sequence acquisition module 5, three-dimensional transient turbulent velocity field acquisition module 6. Detailed Implementation

[0012] This application provides a method and system for three-dimensional reconstruction of holographic turbulent flow field in turbid water based on deep learning, which solves the technical problems of strong noise and particle signal distortion in traditional image-based three-dimensional reconstruction methods in turbid water scenarios.

[0013] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0014] It should be noted that the terms "first," "second," etc., in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or modules not explicitly listed or inherent to such processes, methods, products, or devices.

[0015] Example 1, as Figure 1 As shown, a method for three-dimensional reconstruction of holographic turbulent flow fields in turbid water bodies based on deep learning is described, wherein the method includes:

[0016] During the laser coherent irradiation of turbid water using a dual-pulse off-axis holographic optical path, a high-speed CMOS camera is driven to capture the transient holographic interferogram sequence of the turbid water.

[0017] In this embodiment, the dual-pulse off-axis holographic optical path is an optical system composed of a dual-pulse laser, a reflector, a beam splitter, and other components. It emits two laser beams with an angle between them—a reference beam and an object beam. The object beam, after penetrating the target object, interferes with the reference beam, thereby generating a hologram. This system is commonly used for holographic imaging of dynamic scenes. Turbid water contains impurities such as silt, microorganisms, and organic debris, causing light scattering and reduced transmittance upon penetration. A CMOS camera, based on a complementary metal-oxide-semiconductor image sensor, converts optical signals into electrical signals and generates digital images. It also possesses high frame rate shooting capabilities and can be used to capture dynamic image sequences.

[0018] Specifically, firstly, tracer particles matching the target turbulence scale are added to the turbid water body. Then, a dual-pulse laser is triggered to emit light at a preset laser repetition frequency to coherently irradiate the turbid water body. At the same time, a high-speed CMOS camera is driven at a preset camera exposure frequency to capture a transient holographic interferogram sequence. The ratio of the camera exposure frequency to the laser repetition frequency is 1:(N+1), where N represents the number of frames used for background noise correction within the dual-pulse interval.

[0019] A noise suppression model based on a deep learning network is used to suppress muddy water scattering noise in the transient holographic interferogram sequence, and an optimized hologram is output.

[0020] Optionally, the transient holographic interferogram sequence is first segmented into a transient holographic overlapping block sequence, and the Hanning window is used to suppress boundary effects. Then, the overlapping block sequence is input into a noise suppression model, and the water turbidity parameter is mapped through the physical simulation branch of the model to output noise distribution overlapping blocks. At the same time, multi-scale feature extraction is performed through the data-driven branch to output optimized feature overlapping blocks. Finally, the weighted fusion layer receives and stitches the two types of overlapping blocks, and then performs a weighted average elimination operation to finally output an optimized hologram.

[0021] An initial set of particle plane coordinates is generated by performing cascaded convolutional neural network-driven two-dimensional particle projection localization on the optimized hologram.

[0022] In one embodiment of this application, the optimized hologram is first subjected to dynamic contrast enhancement and multi-directional Gabor filtering to obtain a preprocessed hologram; then the preprocessed hologram is input into a first-level strongly quantized convolutional network to complete global coarse localization and output a candidate region heatmap; subsequently, the candidate region heatmap is input into a second-level residual convolutional network to perform sub-pixel coordinate regression and generate a particle center coordinate distribution; finally, the particle center coordinate distribution is traversed and confidence is filtered to finally output the initial particle plane coordinate set.

[0023] Calculate the particle axial depth of the initial particle plane coordinate set and output the particle transient spatial coordinate set.

[0024] Specifically, before calculating the particle axial depth of the initial particle plane coordinate set, local holographic information corresponding to the particles is first extracted from the optimized hologram. The initial particle plane coordinate set already contains the coordinates (x, y) of each particle in the two-dimensional plane. Based on these coordinates, a fixed-size local holographic sub-image, such as 32×32 pixels, is cropped from the optimized hologram with each (x, y) as the center. This ensures that each sub-image contains only the holographic intensity distribution of a single particle, avoiding interference from the intensity of adjacent particles, and focusing the target particle information for subsequent depth calculations.

[0025] Next, the relationship between local sub-images and axial depth is established using the intensity contrast method. First, each local holographic sub-image is traversed, and the intensity values ​​of all pixels within the sub-image are counted, then the maximum intensity value is selected. and minimum light intensity value The light intensity contrast ratio C is obtained by calculating the ratio of the two. In a dual-pulse off-axis holographic optical path, the axial depth of a particle is fixedly related to the intensity contrast of the hologram. The closer the particle is to the holographic imaging surface, the smaller the axial depth, the more concentrated the intensity distribution, and the larger the contrast C. The farther the particle is from the imaging surface, the greater the axial depth, the more diffuse the intensity distribution, and the smaller the contrast C. This rule provides the basis for depth calculation.

[0026] Pre-calibration was then performed to determine the mapping relationship between contrast and axial depth. Under the same dual-pulse off-axis holographic optical path environment, standard particles with the same material and particle size as the tracer particles were used. These standard particles were placed at multiple fixed positions at known axial depths, for example, from 0 mm to 50 mm, with a position set every 1 mm. A hologram was captured at each position, and the corresponding local sub-image was extracted. The light intensity contrast C corresponding to each known depth z was calculated, resulting in multiple sets of (C, z) calibration data. A linear fitting method was used to process the calibration data, yielding the mapping function z = k*C + b between contrast C and axial depth z, where k is the fitting slope and b is the fitting intercept. This function can be directly used for subsequent calculations of unknown depths.

[0027] Finally, the axial depth of each particle is calculated, and a set of transient spatial coordinates for the particles is generated. The light intensity contrast C corresponding to each particle in the initial particle planar coordinate set is substituted into the mapping function z = k*C + b to obtain the axial depth z of that particle. The two-dimensional coordinates (x, y) are then combined with the axial depth z to form the three-dimensional spatial coordinates (x, y, z) of a single particle. This calculation process is repeated for all particles in the initial particle planar coordinate set, integrating the three-dimensional spatial coordinates of all particles to finally output the set of transient spatial coordinates for the particles.

[0028] Multi-frame trajectory reconstruction with spatiotemporal graph association is performed on the particle transient spatial coordinate set to generate a three-dimensional particle field dynamic sequence.

[0029] Specifically, firstly, a spatiotemporal heterogeneous graph is constructed based on the particle transient spatial coordinate set; then, a graph attention network is used to traverse the graph to calculate the correlation weights between particle nodes and fit the initial trajectory chain; next, the motion continuity of the initial trajectory chain is iteratively optimized based on acceleration smoothing constraints to output a broken trajectory sequence; finally, interpolation is performed on the broken trajectory sequence to complete it, and the final output is a three-dimensional particle field dynamic sequence.

[0030] A three-dimensional transient turbulent velocity field is generated by performing flow field inversion on the dynamic sequence of the three-dimensional particle field.

[0031] Specifically, firstly, the particle spatial neighborhood topological connection is performed frame by frame on the dynamic sequence of the three-dimensional particle field to output the particle-associated topological field of a single frame; then, under the neighborhood constraints of this topological field, the particle position matching of adjacent frames is performed to generate a displacement vector field; subsequently, based on the topological neighborhood scale, the displacement vector field is interpolated to the three-dimensional spatial grid nodes under the fluid incompressibility constraint to generate a gridded three-dimensional velocity field; next, the gridded three-dimensional velocity field is corrected based on the curl distribution of particle motion in adjacent frames to output a vorticity-corrected velocity field; finally, the vorticity-corrected velocity field is time-series superimposed to finally generate a three-dimensional transient turbulent velocity field.

[0032] Furthermore, the method provided in this application embodiment includes:

[0033] Tracer particles matching the target turbulence scale are added to the turbid water body; during the laser coherent irradiation of the turbid water body by triggering a dual-pulse laser to emit dual-pulse off-axis holographic optical path based on a preset laser repetition frequency, the high-speed CMOS camera is driven by a preset camera exposure frequency to capture the transient holographic interferogram sequence, wherein the ratio of the camera exposure frequency to the laser repetition frequency is 1:(N+1), and N is the number of background noise correction frames within the dual-pulse interval.

[0034] In this embodiment, the turbulence scale is a physical quantity that describes the turbulent motion in turbulent water, such as the spatial size, intensity, and range of eddies and disturbances. It is a key basis for selecting suitable tracer particles. A tracer particle is a tiny particle whose particle size and density match the target turbulence scale in the turbulent water, can follow the turbulent motion of the water, and is easy to observe and track using imaging equipment.

[0035] Optionally, in the raw data acquisition stage of the 3D reconstruction of the holographic turbulence field in turbid water, it is first necessary to ensure that the turbulent motion can be effectively tracked by adding tracer particles to the turbid water. A turbulence scale-particle characteristic matching method is adopted: first, key parameters such as the average eddy size and flow velocity of the turbulence in the target turbid water are measured using a portable turbulence analyzer. Then, tracer particles are selected based on these parameters. Generally, particles with a density close to that of the turbid water and a particle size of 1 / 10 to 1 / 5 of the average eddy size are chosen to avoid particles that settle or float too quickly due to excessive density differences, or particles that cannot follow small-scale eddy motion due to inappropriate particle size. After selecting the particles, a micro-injection pump is used to uniformly inject the particle solution into the monitoring area of ​​the turbid water to ensure that the tracer particles are evenly distributed in the water, providing clear motion markers for subsequent capture of turbulent motion trajectories.

[0036] After adding the tracer particles, the repetition frequency of the dual-pulse laser is set using a signal generator. The frequency is adjusted according to the dynamic characteristics of the target turbulence. If the turbulence changes drastically, such as in a near-shore rapid current, the laser repetition frequency can be set to 50-100Hz to ensure that the instantaneous state of the turbulence can be captured. Subsequently, a dual-pulse off-axis holographic optical path is constructed. By adjusting the angles of the reflectors and beam splitters in the optical path, the laser emitted by the laser is split into a reference beam and an object beam. The object beam penetrates the turbid water and interacts with the tracer particles, while the reference beam propagates directly. The angle between the two beams is usually 10-20 degrees to avoid overlap of the light fields, ultimately forming interference fringes on the imaging plane.

[0037] Next, the exposure frequency of the high-speed CMOS camera is set according to the laser repetition frequency, and the ratio between the two is strictly controlled to be 1:(N+1), where N is the number of background noise correction frames within the double pulse interval. For example, when the laser repetition frequency is 50Hz, the camera exposure frequency is set to 50 / (N+1)Hz. N is obtained by observing the static scattering noise intensity of turbid water bodies under different turbidity levels in pre-experiments. For example, the scattering noise of low turbidity water bodies is weak, while that of high turbidity water bodies is strong. Combined with the rate of change of the target turbulence, the more intense the turbulence, the faster the capture is needed to avoid information loss. The number of correction frames N that can effectively cancel the fixed noise is determined. For example, N=2-3 frames can be set for low turbidity water bodies, and N=4-5 frames can be set for high turbidity water bodies. This ensures that the average gray level of the N frames of background images can accurately cover the fixed noise such as static impurities and scattered light.

[0038] During the interval between dual-pulse transmissions, a high-speed CMOS camera first captures N consecutive background images without particle signals, followed by one holographic interferogram containing both particle signals and turbulence information. The grayscale values ​​of the N background images are then averaged, and this average is used to subtract the grayscale value of the concurrently captured holographic interferogram to cancel out interference from static impurities, scattered light, and other fixed noise in the turbid water. This cycle is repeated to continuously capture multiple sets of N background images + one holographic interferogram, ultimately integrating them to form a transient holographic interferogram sequence that reflects the instantaneous state of turbulence in the turbid water.

[0039] By using tracer particles of matched scale to provide observable markers for turbulent motion, and by performing coordinated control of a dual-pulse off-axis holographic optical path and a high-speed CMOS camera, noise suppression is achieved by combining multi-frame background correction. Finally, a sequence of transient holographic interferograms that clearly reflects the motion state of the tracer particles is obtained, providing high-quality raw data support for subsequent deep learning-based processing steps such as noise suppression and particle localization.

[0040] Furthermore, the method provided in this application embodiment includes:

[0041] The noise suppression model includes a parallel physical simulation branch and a data-driven branch. The outputs of the physical simulation branch and the data-driven branch are connected to a weighted fusion layer, and the signal-to-noise ratio of the weighted fusion layer is dynamically adjustable.

[0042] Specifically, in designing the noise suppression model for holographic images of turbid water bodies, a dual-branch parallel model architecture is adopted, taking into account the characteristic that noise in turbid water bodies simultaneously includes scattering noise that can be described by physical laws and randomly distributed image noise. The physical simulation branch focuses on processing scattering noise caused by water turbidity, which has clear optical propagation laws. The data-driven branch handles randomly occurring noise in the holographic image that is difficult to describe by fixed physical formulas, such as sensor thermal noise and ambient light interference. By using two branches to specifically address different types of noise, the problem of insufficient adaptation of a single branch to complex noise is avoided.

[0043] After determining the dual-branch parallel structure, a weighted fusion layer is configured for the noise suppression model, and the signal-to-noise ratio (SNR) is used as the core basis for dynamic weight adjustment. Specifically, the SNR of each frame in the input transient holographic interferogram sequence is analyzed in real time using the gray-scale standard deviation method in image SNR calculation: when the image SNR is low, such as in high-turbidity water bodies where scattering noise is strong, the weight ratio of the physical simulation branch output is increased to enhance the suppression effect on regular scattering noise; when the image SNR is high, such as in low-turbidity water bodies where random noise is more prominent, the weight ratio of the data-driven branch output is increased to better eliminate random interference. The weight adjustment process adopts a linear weighting method, that is, the physical branch weight coefficient α and the data branch weight coefficient β are calculated based on the SNR, where α + β = 1.

[0044] During actual model execution, the transient holographic interferogram sequence to be processed is synchronously input to both the physical simulation branch and the data-driven branch. In parallel processing, each branch suppresses noise based on its own processing logic and outputs pre-processed image data. Subsequently, the weighted fusion layer receives the outputs from both branches and, based on the real-time calculated current image signal-to-noise ratio, automatically calls the preset weight adjustment logic to determine the values ​​of α and β. Then, it fuses the image data output from both branches through linear weighting operations, that is, multiplying the corresponding pixel grayscale values ​​from both branches by α and β respectively and summing the results to obtain the fused pixel values. Finally, all pixels are integrated to generate a single-frame optimized hologram. This process is repeated to complete the noise suppression processing of the entire transient holographic interferogram sequence.

[0045] By designing a dual-branch parallel architecture to specifically handle different types of noise, and combining it with a dynamic weighted fusion layer based on signal-to-noise ratio, the system achieves the effect of adapting to the complex noise environment in turbid water and effectively suppressing various noise interferences, providing higher quality holographic image data for subsequent steps such as particle localization.

[0046] Furthermore, the method provided in this application embodiment includes:

[0047] A differentiable mapping model and an angular spectrum diffraction propagation model are pre-constructed, wherein the differentiable mapping model is used to map the turbidity parameter of the water body to the scattering phase function; the localization configuration of the physical simulation branch is completed by cascading the differentiable mapping model and the angular spectrum diffraction propagation model; a multi-scale encoder and a feature fusion decoder are pre-constructed, and the localization configuration of the data-driven branch is completed by cascading the multi-scale encoder and the feature fusion decoder; after paralleling the physical simulation branch and the data-driven branch, the weight fusion layer is connected at the output end to complete the construction of the noise suppression model.

[0048] Specifically, before constructing the physical simulation branch of the noise suppression model, a differentiable mapping model is first pre-constructed: combining the correlation between water turbidity and scattering characteristics, turbid water samples of different turbidity levels are collected, such as water samples with turbidity values ​​of 50 NTU, 100 NTU, and 200 NTU, respectively. The turbidity parameters of each sample are measured using a turbidimeter, in NTU units. At the same time, the scattering phase function corresponding to each sample is captured by a phase imaging device to describe the angular distribution characteristics of light scattering in the water.

[0049] Subsequently, a multilayer perceptron (MLP) with a single hidden layer was used as the mapping model framework. The measured turbidity parameters were used as the model input, and the scattering angle coefficient and phase shift of the corresponding scattering phase function parameters were used as the output. The mean square error loss function was used to calculate the difference between the model prediction and the actual measurement. The model parameters were iteratively optimized through the gradient descent algorithm until the loss value stabilized and converged, thus completing the construction of the differentiable mapping model. This model can output the corresponding scattering phase function according to the input water turbidity parameters.

[0050] Next, a pre-constructed angular spectrum diffraction propagation model is built. Referring to the angular spectrum simulation method for light field propagation, the basic parameters of the model are first determined: laser wavelength, refractive index of turbid water, and propagation distance of light in the water, which can be set according to the actual detection device. The input of the model is the scattering phase function output by the differentiable mapping model, and the initial light field distribution of the laser, which is set as a plane wave light field with uniform amplitude and 0 phase. According to the angular spectrum propagation formula, a two-dimensional Fourier transform is first performed on the initial light field and the scattering phase function to obtain the angular spectrum distribution of the light field.

[0051] Subsequently, the propagation phase factor is calculated based on the propagation distance. The angular spectrum distribution is then multiplied by the propagation phase factor to obtain the propagated angular spectrum. Finally, an inverse two-dimensional Fourier transform is performed on the propagated angular spectrum to obtain the light field distribution after light propagates in turbid water. This light field distribution is the output of the angular spectrum diffraction propagation model. The output of the differentiable mapping model is connected to the input of the angular spectrum diffraction propagation model to achieve the cascading of the two models and complete the localized configuration of the physical simulation branch. The overall input of this branch is the turbidity parameter of the water body, and the output is the light field distribution after simulated light scattering, which can reflect the law of scattering noise in turbid water.

[0052] Then, a data-driven branch is pre-built, employing the U-Net neural network architecture, which is suitable for image feature extraction and denoising. The multi-scale encoder part of this architecture consists of four sets of convolutional blocks and pooling layers: each convolutional block contains two 3×3 convolutional layers with ReLU activation function for feature extraction; the pooling layers use 2×2 max pooling to achieve feature downsampling to obtain multi-scale information. The feature fusion decoder part also consists of four sets of transposed convolutional blocks and a feature fusion layer: each transposed convolutional block uses 2×2 transposed convolutions with a stride of 2 to achieve feature upsampling to restore resolution; the feature fusion layer uses skip connections to concatenate the feature maps of the encoder at the corresponding scale with the upsampled feature maps of the decoder, supplementing detailed information.

[0053] After constructing the above model, training data was collected, specifically holograms of turbid water bodies with varying degrees of scattering noise, which were used as input data. Clean holograms without noise, which were manually selected and labeled, were used as label data. The two types of data were divided into 256×256 pixel holographic overlapping blocks to suppress boundary effects, and were divided into training and validation sets in an 8:2 ratio.

[0054] During training, noisy holographic overlapping blocks from the training set are input into the U-Net network. A multi-scale encoder extracts noisy features, which are then fused by a feature fusion decoder to output denoised feature blocks. The L1 loss function is used to calculate the difference between the output feature blocks and the clean holographic blocks in the label data. The Adam optimizer iteratively updates the network parameters with a learning rate of 0.001. After each training round, the model performance is evaluated using a validation set. Training stops when the validation set loss value no longer decreases for five consecutive rounds, completing the construction of the multi-scale encoder and feature fusion decoder. These two modules are cascaded to complete the localized configuration of the data-driven branch. The input of this branch is the noisy holographic overlapping blocks, and the output is the denoised optimized feature blocks.

[0055] Finally, the inputs of the physical simulation branch and the data-driven branch are connected in parallel to ensure that both branches can simultaneously receive the same set of transient holographic interferogram data to be processed, namely the water turbidity parameter and the noisy holographic overlap block. A weighted fusion layer is connected to the output of the two branches. The weighted fusion layer receives the light field distribution data output by the physical simulation branch and the optimized feature block output by the data-driven branch. The two types of output data are integrated through linear weighted operation to generate the final optimized holographic data, thus completing the construction of the entire noise suppression model.

[0056] By constructing and cascading the core model of the physical simulation branch in steps, building and training the data-driven branch of the U-Net architecture, and connecting the two branches in parallel and accessing the weight fusion layer, the effect of constructing a noise suppression model that combines physical law adaptability and data learning ability and can effectively suppress the noise of turbid water holograms was achieved.

[0057] Furthermore, the method provided in this application embodiment includes:

[0058] The transient holographic interferogram sequence is segmented into a transient holographic overlapping block sequence, wherein the transient holographic overlapping block sequence employs a Hanning window to suppress boundary effects. The transient holographic overlapping block sequence is input into the noise suppression model, and water turbidity parameters are mapped via the physical simulation branch to output noise distribution overlapping blocks. Multi-scale feature extraction is performed via the data-driven branch to output optimized feature overlapping blocks. The weighted fusion layer receives and overlaps the noise distribution overlapping blocks and optimized feature overlapping blocks, performs weighted average elimination, and outputs the optimized hologram.

[0059] In this embodiment, the Hanning window suppresses boundary effects by multiplying each sub-block with the Hanning window pixel-by-pixel when segmenting the transient holographic interferogram sequence of turbid water into sub-blocks. This smooths the grayscale transition at the edges of the sub-blocks, avoiding false noise caused by abrupt changes in edge grayscale. The transient holographic overlapping block refers to the image sub-block with partially overlapping regions obtained after segmenting the transient holographic interferogram sequence of turbid water according to a preset size and overlap rate, used as input to the noise suppression model for processing.

[0060] Specifically, the image size and boundary effects are addressed first. For large images, a sliding window segmentation method is used. A suitable sub-block size of 256×256 pixels and an overlap rate are set, typically 50%, to avoid information breakage between blocks. Each complete transient holographic interferogram is divided into multiple sub-blocks using a sliding window approach. These sub-blocks together form a transient holographic overlapping block sequence.

[0061] Next, to address the boundary effect caused by abrupt changes in grayscale at the edges of the segmented sub-blocks, the Hanning window method is employed: First, a two-dimensional Hanning window with the same size as the sub-block is generated, with the window function formula w(i,j)=0.5×[1-cos(2πi / (M-1))]×0.5×[1-cos(2πj / (N-1))], where M and N are the width and height of the sub-block, respectively. Then, each holographic overlapping block is multiplied pixel-by-pixel with the corresponding Hanning window to ensure a smooth transition in grayscale at the sub-block edges, avoiding false edge noise during subsequent model processing and ensuring that the edge information of each sub-block is consistent with that of adjacent sub-blocks.

[0062] After completing the overlapping block processing, the transient holographic overlapping block sequence is input into the noise suppression model constructed in the previous steps, and the dual-branch parallel processing is initiated. For the physical simulation branch, the turbidity parameters of the current turbid water body are first collected in real time using a portable turbidimeter, with the unit being NTU. This turbidity parameter is then input into the branch along with the holographic overlapping block. The branch first converts the turbidity parameter into the corresponding scattering phase function through a differentiable mapping model to describe the scattering law of light in the turbid water body. Then, the scattering phase function is input into the angular spectrum diffraction propagation model to simulate the scattering process of laser light passing through the turbid water body. Finally, a noise distribution overlapping block that accurately reflects the current water body scattering noise distribution is output.

[0063] For the data-driven branch, a multi-scale feature extraction logic suitable for image denoising is adopted. The holographic overlapping block is input into the multi-scale encoder, and noise features of different scales are extracted step by step through 3×3 convolution and pooling layers of multiple convolutional layers, that is, from small-scale random noise to large-scale scattered light spots. The extracted multi-scale features are then fed into the feature fusion decoder, and the image resolution is restored by transposed convolution. Skip connections are used to supplement the detailed features extracted by the encoder, and finally the optimized feature overlapping block with most of the noise is output.

[0064] After the two branches output the noise distribution overlap block and the optimized feature overlap block respectively, the weight fusion layer performs subsequent processing. Specifically, for the stitching of the overlap blocks, according to the sliding order during the segmentation steps mentioned above, the overlapping areas of the adjacent noise distribution overlap blocks and optimized feature overlap blocks are averaged pixel by pixel, so as to avoid obvious traces at the stitching point, and restore the complete noise distribution map and the complete optimized feature map with the same resolution as the original transient holographic interferogram.

[0065] Next, the signal-to-noise ratio (SNR) of the current hologram is determined by calculating the ratio of the grayscale standard deviation of the image signal region to that of the noise region. If the SNR is low, the proportion of scattering noise is high, so the weight of the physical simulation branch output is increased; if the SNR is high, the proportion of random noise is high, so the weight of the data-driven branch output is increased, where the sum of the weights of the two branches is 1. Finally, a pixel-wise weighted average operation is performed on the complete noise distribution map and the complete optimized feature map, and the results are integrated into a single-frame optimized hologram. The above process is repeated to process the transient holographic interferograms of all frames, finally obtaining a complete sequence of optimized holograms.

[0066] By using sliding window segmentation and Hanning window to suppress boundary effects, dual-branch parallel processing to extract noise and optimization features, and dynamic weight fusion and splicing based on signal-to-noise ratio, the system effectively suppresses turbid water scattering noise in transient holographic interferogram sequences and outputs high-quality optimized holograms.

[0067] Furthermore, the method provided in this application embodiment includes:

[0068] The optimized hologram is progressively subjected to dynamic contrast enhancement and multi-directional Gabor filtering to output a preprocessed hologram. The preprocessed hologram is loaded into a first-level strongly quantized convolutional network to perform global coarse localization and output a candidate region heatmap. The candidate region heatmap is input into a second-level residual convolutional network for sub-pixel coordinate regression and outputs the particle center coordinate distribution. The particle center coordinate distribution is traversed for confidence filtering and the initial particle plane coordinate set is output.

[0069] In this embodiment, multi-directional Gabor filtering refers to an image processing technique that uses multiple Gabor filters in different directions to perform convolution operations on an image to extract multi-directional texture features, enhance target edges, and suppress noise.

[0070] In one embodiment, the distinction between particles and background is first enhanced through preprocessing. Dynamic contrast enhancement is performed using contrast-limited adaptive histogram equalization (CLAHE): the optimized hologram is divided into 8×8 uniform sub-blocks, and a clipping limit of 0.02 is set to avoid local overexposure. The histogram of each uniform sub-block is calculated and equalized separately to make the grayscale difference in the particle region more significant. Then, multi-directional Gabor filtering is performed. Four Gabor filters with different directions of 0°, 45°, 90°, and 135° are designed, with a wavelength of 5 pixels, a bandwidth of 1.5, and a phase shift of 0. The enhanced image is then convolved with the four Gabor filters in sequence to preserve the texture features of the particle edges and suppress slight noise. Finally, a preprocessed hologram with more prominent particle features is output.

[0071] Next, a first-stage strongly quantized convolutional network is used to achieve global coarse localization of particles. Combining the design principles of lightweight convolutional networks, this network employs a depthwise separable convolutional structure to reduce computational load. Simultaneously, 8-bit weight quantization is introduced to map 32-bit floating-point weights to an integer range of 0-255, reducing storage requirements. The network input is a preprocessed hologram. First, three depthwise separable convolutional layers with 3×3 kernels and a stride of 1 are used to extract multi-scale features. Each convolutional layer is followed by batch normalization and a ReLU activation function. Then, two max-pooling layers with 2×2 kernels and a stride of 2 are used to shrink the feature map size to expand the receptive field, achieving global feature capture. Finally, a 1×1 convolutional layer compresses the number of feature map channels to 1, outputting a candidate region heatmap with the same size as the input image. Higher pixel values ​​in the heatmap indicate a greater probability of particle presence at that location.

[0072] The candidate region heatmap is then input into a second-level residual convolutional network for sub-pixel coordinate regression. Utilizing the high-precision ResNet-18 residual network architecture, residual connections are introduced to avoid gradient vanishing during deep training: the input heatmap first passes through two residual blocks to extract fine features, with each residual block containing two 3×3 convolutional layers. Shallow features are directly passed within the residual block via residual connections. Next, a sub-pixel convolutional layer with a scaling factor of 2 increases the feature map resolution by a factor of 2, achieving sub-pixel-level feature alignment. Finally, a fully connected layer outputs the particle center coordinates (x, y) and corresponding confidence values ​​for each candidate region, forming a particle center coordinate distribution. The coordinate accuracy reaches the 0.1 pixel level, meeting the sub-pixel localization requirements.

[0073] Finally, the distribution of particle center coordinates is filtered by confidence to remove false positioning. First, the average confidence of all coordinate points is calculated, and the threshold is set to 1.2 times the average. Then, each particle center coordinate is traversed, and coordinate points with confidence higher than the threshold are retained, while false coordinates with confidence lower than the threshold, such as false detection points caused by noise, are removed. Finally, the retained coordinate points are sorted to form an initial particle plane coordinate set containing the precise position of the particles in the two-dimensional plane.

[0074] By enhancing particle features through dynamic contrast enhancement and multi-directional Gabor filtering, achieving global coarse localization and sub-pixel precise localization through two-level convolutional networks, and eliminating false coordinates through confidence filtering, the system achieves the effect of accurately generating the initial particle plane coordinate set, providing a reliable two-dimensional positional basis for subsequent particle axial depth calculation.

[0075] Furthermore, the method provided in this application embodiment includes:

[0076] A spatiotemporal heterogeneous graph is constructed based on the particle transient spatial coordinate set; a graph attention network is used to traverse the spatiotemporal heterogeneous graph to calculate the correlation weights between particle nodes and fit an initial trajectory chain; the motion continuity of the initial trajectory chain is iteratively optimized based on acceleration smoothing constraints to output a broken trajectory sequence; interpolation is performed on the broken trajectory sequence to output the three-dimensional particle field dynamic sequence.

[0077] Optionally, the transient spatial coordinate set of particles is first processed to extract the three-dimensional position coordinates (x, y, z) of each particle in each frame and the corresponding time frame information. Each particle in a single frame is taken as an independent node in the spatiotemporal heterogeneous graph. The node data includes the three-dimensional position coordinates of the particle and the time frame to which it belongs. Then, a spatial distance threshold is set, which is set to the actual distance corresponding to 3-5 times the particle size according to the tracer particle size. The time frame interval threshold is set to the particles in two adjacent frames. If two particle nodes in different time frames satisfy that the spatial Euclidean distance is less than the spatial threshold and the time frame interval is less than the time threshold, then an undirected connection edge is established between the two nodes to complete the construction of the spatiotemporal heterogeneous graph.

[0078] Next, the constructed spatiotemporal heterogeneous graph is input into the graph attention network, and a single-head attention mechanism is used to calculate the association weights between particle nodes. First, the features of each node, including its 3D position coordinates and time frame, are linearly transformed to obtain a feature vector. Then, the attention coefficients between the node and all its neighboring nodes are calculated using the Softmax function. The magnitude of the coefficients is inversely proportional to the spatial distance between nodes and directly proportional to the velocity similarity, i.e., the velocity estimated based on the position difference between adjacent frames. The closer the spatial distance and the closer the velocity of the node pair, the higher the attention coefficient, i.e., the association weight. Based on the association weights, the neighboring nodes of each node are sorted, and the two neighboring nodes with the highest weights are selected and connected sequentially to form an initial trajectory chain that spans multiple frames. Each initial trajectory chain corresponds to the preliminary motion path of a single particle.

[0079] Then, based on acceleration smoothing constraints, the initial trajectory chain is iteratively optimized for motion continuity. First, the instantaneous velocity of each adjacent particle in the initial trajectory chain is calculated, specifically by dividing the position difference between two adjacent frames by the time interval between the two frames. Then, the acceleration of each motion segment is calculated based on the instantaneous velocity, specifically by dividing the difference between two adjacent instantaneous velocities by the time interval. An acceleration smoothing threshold is set, such as 0.5 m / s², which is the maximum reasonable acceleration of particle motion in turbid water. If the acceleration of a certain motion segment exceeds the threshold, the position is determined to be a trajectory breakpoint. The positions of particles near the breakpoint are iteratively adjusted, with the adjustment range not exceeding 10% of the spatial distance threshold. The acceleration is recalculated until the acceleration meets the threshold requirement. If the threshold cannot be met after 3 iterations, the breakpoint is retained. Finally, the broken trajectory sequence containing the breakpoint is output.

[0080] Finally, linear interpolation is performed to complete the fracture trajectory sequence. For each fracture trajectory, the effective particle coordinates of the two frames before and after the fracture point are located. For example, if the fracture is located between frame t and frame t+2, and the particle coordinates of frame t are known... and particle coordinates in frame t+2 The particle coordinates of the (t+1)th frame of the broken frame are calculated based on the time interval ratio. The y and z coordinates are calculated using the same method; after completing all the break points, the complete trajectories of all particles are integrated to form a three-dimensional particle field dynamic sequence that can reflect the continuous motion state of particles within multiple frames.

[0081] Furthermore, the method provided in this application embodiment includes:

[0082] In the spatiotemporal heterogeneous graph, each node corresponds to an independent particle in a single frame, and the node data includes the particle's three-dimensional position coordinates and instantaneous velocity. The spatial Euclidean distance between pairs of nodes with undirected connecting edges satisfies a preset distance threshold, and the time frame interval satisfies a preset frame interval.

[0083] In one embodiment, single-frame particle data is first extracted from the particle transient spatial coordinate set to construct nodes of the spatiotemporal heterogeneous graph. For each frame of particles, the three-dimensional position coordinates (x, y, z) of each particle within that frame are directly obtained. These coordinates are the core data in the particle transient spatial coordinate set obtained by calculating the particle axial depth in the previous steps and can be directly used as the basic attributes of the nodes.

[0084] Next, the instantaneous velocity required by the node is calculated. For two adjacent frames, such as frame t and frame t+1, nodes that may belong to the same particle are initially matched by spatial distance. Particles in frame t+1 are selected within a small area around the particle in frame t. The 3D coordinates of the matched particle in frame t+1 are then subtracted from the 3D coordinates of the corresponding particle in frame t to obtain the coordinate difference. This difference is divided by the time interval between the two frames. This time interval is determined by the exposure frequency of the high-speed CMOS camera. For example, when the exposure frequency is 50Hz, the time interval is 0.02s. The instantaneous velocity of the particle in frame t can then be obtained. The 3D position coordinates and instantaneous velocity are integrated as the complete data of a single node. Each independent particle in a single frame corresponds to one such node.

[0085] Subsequently, the conditions for establishing undirected connections in the spatiotemporal heterogeneous graph were determined. The spatial Euclidean distance threshold is referenced to the particle size of the tracer particles, and is usually set to 3-5 times the particle size. For example, when the particle size of the tracer particles is 10μm, the threshold is set to 30-50μm to ensure that only nodes that are spatially close and may have the same particle trajectory are associated. The time frame interval threshold is set to 1, that is, only particle nodes in two adjacent frames are considered to avoid excessive changes in particle motion state and decreased association accuracy due to excessive time intervals.

[0086] Finally, the particle nodes in different frames are traversed to establish undirected connections. For each particle node in frame t, the spatial Euclidean distance between it and all particle nodes in frame t+1 is calculated using the three-dimensional coordinate distance formula: Simultaneously, it is confirmed that the time frame interval between the two nodes is 1. If the spatial Euclidean distance between a pair of nodes is less than a preset spatial threshold, and the time frame interval meets the preset threshold, then an undirected connection edge is established between the two nodes. This process is repeated, traversing all particle nodes in adjacent frames, to complete the construction of all undirected connection edges that meet the conditions. Finally, a spatiotemporal heterogeneous graph is formed, in which nodes contain the three-dimensional position coordinates and instantaneous velocities of particles, and edges are constrained by spatial and temporal thresholds. This provides a structured data foundation for the graph attention network to calculate the particle node association weights.

[0087] Furthermore, the method provided in this application embodiment includes:

[0088] A frame-by-frame particle spatial neighborhood topological connection is performed on the dynamic sequence of the three-dimensional particle field to output a single-frame particle-associated topological field. Under the neighborhood constraints of the single-frame particle-associated topological field, adjacent frame particle position matching is performed to generate a displacement vector field. Based on the topological neighborhood scale of the single-frame particle-associated topological field, under the fluid incompressibility constraint, the displacement vector field is interpolated to the three-dimensional spatial grid nodes to generate a gridded three-dimensional velocity field. The gridded three-dimensional velocity field is corrected based on the curl distribution of particle motion in adjacent frames to output a vorticity-corrected velocity field. The vorticity-corrected velocity field is time-sequentially superimposed to generate the three-dimensional transient turbulent velocity field.

[0089] In one embodiment, spatial neighborhood topological connections are first performed on each frame of particles in the dynamic sequence of the 3D particle field using the K-nearest neighbor method, with K set to 5 to balance association accuracy and computational efficiency: For each particle in a single frame, its distance to all other particles in the frame is calculated using the same 3D coordinate distance formula as described above, and the five closest particles are selected. Topological connections are then established between the current particle and these five particles. This operation is repeated to traverse all particles in the frame, so that each frame of particles forms a network structure with the individual particle as the center and neighboring particles as the associated objects. Finally, a single-frame particle association topological field is output, which can clearly define the spatial neighborhood range of particles in each frame.

[0090] Next, under the neighborhood constraints of the single-frame particle association topological field, adjacent frame particle position matching is performed: taking frame t and frame t+1 as examples, firstly, the neighborhood scale of each particle in frame t is determined, and the maximum distance between the particle and its five associated particles is taken as the neighborhood radius. Then, in frame t+1, candidate particles are searched only within the neighborhood radius of the particle in frame t. For each candidate particle, its three-dimensional position difference with the particle in frame t is calculated, and the candidate particle with the smallest position difference is selected as the matching object. This position difference is taken as the displacement vector, which includes displacement components in the x, y, and z directions. All particle pairs in adjacent frames are traversed, and all displacement vectors are integrated to generate a displacement vector field that reflects the position changes of particles between adjacent frames.

[0091] Subsequently, based on the topological neighborhood scale of the particle-associated topological field in a single frame, the displacement vector field is interpolated to the three-dimensional spatial grid nodes under the fluid incompressibility constraint: First, the three-dimensional spatial grid is divided at 1mm intervals, and the grid node coordinates are preset. The topological neighborhood scale is taken as the average of the neighborhood radii in the particle-associated topological field of each frame, serving as the influence range for interpolation. Using trilinear interpolation, for each grid node, the displacement vectors of all particles within its influence range are selected, and weights are assigned according to the distance from the particle to the grid node, with closer particles receiving larger weights. The velocity components of the grid nodes are calculated by weighted averaging, i.e., the displacement vector divided by the time interval between adjacent frames. Simultaneously, to satisfy the fluid incompressibility constraint, the calculated velocity components are checked for divergence. If the absolute value of the velocity divergence of a certain grid node exceeds a preset small threshold of 0.01m / s, the velocity components of that node are fine-tuned until the divergence meets the requirements, ultimately generating a gridded three-dimensional velocity field covering the entire three-dimensional space.

[0092] Next, the gridded 3D velocity field is corrected based on the curl distribution of particle motion in adjacent frames: first, the curl of particle motion in adjacent frames is calculated using the finite difference method; then, the velocity components of two adjacent grid nodes are taken, and the curl value is approximately calculated using the curl calculation formula, for example, curl in the x-direction = ... This process creates a curl distribution. Referring to the turbulent characteristics of turbulent water, a curl threshold of 0.5 rad / s is set. For nodes in the gridded three-dimensional velocity field whose curl values ​​exceed the threshold, their velocity components are adjusted to the average of the velocity components of the three adjacent nodes with lower curl, in order to reduce local vorticity anomalies. This correction process is repeated until the curl values ​​of all nodes meet the threshold requirement, outputting a vorticity-corrected velocity field that better reflects the actual fluid motion.

[0093] Finally, the vorticity-corrected velocity fields are time-series superimposed. Following the chronological order, the vorticity-corrected velocity fields of all frames are assigned the same weight; for example, when the total number of frames is N, the weight of each frame is 1 / N. The velocity components of the same grid node in different frames are weighted and summed to obtain the average velocity of that node over the entire observation period. The average velocities of all grid nodes are then integrated to generate a three-dimensional transient turbulent velocity field that reflects the overall motion state of the turbid water body during the observation period.

[0094] By performing frame-by-frame particle topology connection, neighborhood constraint particle matching, displacement vector field interpolation correction, and time sequence superposition flow field inversion operations on the dynamic sequence of the three-dimensional particle field, the effect of generating a complete and accurate three-dimensional transient turbulent velocity field is achieved.

[0095] In summary, the deep learning-based holographic turbulent flow field three-dimensional reconstruction method for turbulent water bodies provided in this application has the following technical effects:

[0096] This application uses a dual-pulse off-axis holographic optical path to illuminate turbid water, driving a high-speed CMOS camera to capture a sequence of transient holographic interferograms. After denoising using a deep learning noise suppression model, optimized holograms are obtained. Then, through particle positioning, axial depth calculation, trajectory reconstruction, and finally flow field inversion, a three-dimensional transient turbulent velocity field is accurately generated. This improves the accuracy of three-dimensional reconstruction of the holographic turbulent field of turbid water, effectively resists interference caused by scattering from turbid water, and comprehensively and accurately obtains the key data required for three-dimensional reconstruction of the turbulent field, thereby precisely achieving the technical effect of three-dimensional reconstruction of the turbulent field of turbulent water.

[0097] Example 2, as Figure 2 As shown, based on the same inventive concept as in Embodiment 1 above, this application provides a deep learning-based holographic turbulent flow field three-dimensional reconstruction system for turbid water bodies, the system comprising:

[0098] Transient holographic interferogram sequence acquisition module 1 is used to drive a high-speed CMOS camera to capture the transient holographic interferogram sequence of turbid water during laser coherent irradiation of turbid water using a dual-pulse off-axis holographic optical path.

[0099] The optimized hologram acquisition module 2 is used to perform turbid water scattering noise suppression on the transient holographic interferogram sequence using a noise suppression model built based on a deep learning network, and output an optimized hologram.

[0100] The initial particle plane coordinate set acquisition module 3 is used to generate an initial particle plane coordinate set by performing two-dimensional particle projection positioning driven by a cascaded convolutional neural network on the optimized hologram.

[0101] The particle transient spatial coordinate set acquisition module 4 is used to calculate the particle axial depth of the initial particle plane coordinate set and output the particle transient spatial coordinate set.

[0102] The three-dimensional particle field dynamic sequence acquisition module 5 is used to perform multi-frame trajectory reconstruction with spatiotemporal graph association on the particle transient spatial coordinate set to generate a three-dimensional particle field dynamic sequence.

[0103] The three-dimensional transient turbulent velocity field acquisition module 6 is used to generate a three-dimensional transient turbulent velocity field by performing flow field inversion on the dynamic sequence of the three-dimensional particle field.

[0104] Furthermore, the initial particle plane coordinate set acquisition module 3 is used to perform the following steps:

[0105] The optimized hologram is progressively subjected to dynamic contrast enhancement and multi-directional Gabor filtering to output a preprocessed hologram. The preprocessed hologram is loaded into a first-level strongly quantized convolutional network to perform global coarse localization and output a candidate region heatmap. The candidate region heatmap is input into a second-level residual convolutional network for sub-pixel coordinate regression and outputs the particle center coordinate distribution. The particle center coordinate distribution is traversed for confidence filtering and the initial particle plane coordinate set is output.

[0106] Furthermore, the three-dimensional particle field dynamic sequence acquisition module 5 is used to perform the following steps:

[0107] A spatiotemporal heterogeneous graph is constructed based on the particle transient spatial coordinate set; a graph attention network is used to traverse the spatiotemporal heterogeneous graph to calculate the correlation weights between particle nodes and fit an initial trajectory chain; the motion continuity of the initial trajectory chain is iteratively optimized based on acceleration smoothing constraints to output a broken trajectory sequence; interpolation is performed on the broken trajectory sequence to output the three-dimensional particle field dynamic sequence.

[0108] Furthermore, the three-dimensional transient turbulent velocity field acquisition module 6 is used to perform the following steps:

[0109] A frame-by-frame particle spatial neighborhood topological connection is performed on the dynamic sequence of the three-dimensional particle field to output a single-frame particle-associated topological field. Under the neighborhood constraints of the single-frame particle-associated topological field, adjacent frame particle position matching is performed to generate a displacement vector field. Based on the topological neighborhood scale of the single-frame particle-associated topological field, under the fluid incompressibility constraint, the displacement vector field is interpolated to the three-dimensional spatial grid nodes to generate a gridded three-dimensional velocity field. The gridded three-dimensional velocity field is corrected based on the curl distribution of particle motion in adjacent frames to output a vorticity-corrected velocity field. The vorticity-corrected velocity field is time-sequentially superimposed to generate the three-dimensional transient turbulent velocity field.

[0110] Furthermore, the transient holographic interferogram sequence acquisition module 1 is used to perform the following steps:

[0111] Tracer particles matching the target turbulence scale are added to the turbid water body; during the laser coherent irradiation of the turbid water body by triggering a dual-pulse laser to emit a dual-pulse off-axis holographic optical path based on a preset laser repetition frequency, the high-speed CMOS camera is driven by a preset camera exposure frequency to capture the transient holographic interferogram sequence, wherein the ratio of the camera exposure frequency to the laser repetition frequency is 1:(N+1), and N is the number of background noise correction frames within the dual-pulse interval.

[0112] Furthermore, the optimized hologram acquisition module 2 is used to perform the following steps:

[0113] The noise suppression model includes a parallel physical simulation branch and a data-driven branch. The outputs of the physical simulation branch and the data-driven branch are connected to a weighted fusion layer, and the signal-to-noise ratio of the weighted fusion layer is dynamically adjustable.

[0114] Furthermore, the optimized hologram acquisition module 2 is used to perform the following steps:

[0115] The transient holographic interferogram sequence is segmented into a transient holographic overlapping block sequence, wherein the transient holographic overlapping block sequence employs a Hanning window to suppress boundary effects. The transient holographic overlapping block sequence is input into the noise suppression model, and water turbidity parameters are mapped via the physical simulation branch to output noise distribution overlapping blocks. Multi-scale feature extraction is performed via the data-driven branch to output optimized feature overlapping blocks. The weighted fusion layer receives and overlaps the noise distribution overlapping blocks and optimized feature overlapping blocks, performs weighted average elimination, and outputs the optimized hologram.

[0116] Furthermore, the optimized hologram acquisition module 2 is used to perform the following steps:

[0117] A differentiable mapping model and an angular spectrum diffraction propagation model are pre-constructed, wherein the differentiable mapping model is used to map the turbidity parameter of the water body to the scattering phase function; the localization configuration of the physical simulation branch is completed by cascading the differentiable mapping model and the angular spectrum diffraction propagation model; a multi-scale encoder and a feature fusion decoder are pre-constructed, and the localization configuration of the data-driven branch is completed by cascading the multi-scale encoder and the feature fusion decoder; after paralleling the physical simulation branch and the data-driven branch, the weight fusion layer is connected at the output end to complete the construction of the noise suppression model.

[0118] Furthermore, the three-dimensional particle field dynamic sequence acquisition module 5 is used to perform the following steps:

[0119] In the spatiotemporal heterogeneous graph, each node corresponds to an independent particle in a single frame, and the node data includes the particle's three-dimensional position coordinates and instantaneous velocity. The spatial Euclidean distance between pairs of nodes with undirected connecting edges satisfies a preset distance threshold, and the time frame interval satisfies a preset frame interval.

[0120] The 3D reconstruction system for holographic turbulent flow field of turbid water based on deep learning provided in the embodiments of the present invention can execute the 3D reconstruction method for holographic turbulent flow field of turbid water based on deep learning provided in any embodiment of the present invention, and has the corresponding functional modules and beneficial effects of the execution method.

[0121] Although this application makes various references to certain modules in the system according to the embodiments of this application, any number of different modules can be used and run on user terminals and / or servers. The various units and modules included are only divided according to functional logic, but are not limited to the above division, as long as the corresponding functions can be achieved; in addition, the specific names of each functional unit are only for easy distinction between each other and are not used to limit the scope of protection of this invention.

[0122] The specific embodiments described above do not constitute a limitation on the scope of protection of this application. Those skilled in the art should understand that various modifications, combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the scope of protection of this application. In some cases, the actions or steps described in this application can be performed in a different order than that shown in the embodiments and still achieve the desired results. Furthermore, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

Claims

1. A method for three-dimensional reconstruction of holographic turbulent flow field in turbid water based on deep learning, characterized in that, The method includes: During the laser coherent irradiation of turbid water using a dual-pulse off-axis holographic optical path, a high-speed CMOS camera is driven to capture the transient holographic interferogram sequence of the turbid water. A noise suppression model based on a deep learning network is used to suppress muddy water scattering noise in the transient holographic interferogram sequence, and an optimized hologram is output. An initial set of particle plane coordinates is generated by performing cascaded convolutional neural network-driven two-dimensional particle projection localization on the optimized hologram. Calculate the particle axial depth of the initial particle plane coordinate set and output the particle transient spatial coordinate set; Multi-frame trajectory reconstruction with spatiotemporal graph association is performed on the particle transient spatial coordinate set to generate a three-dimensional particle field dynamic sequence; A three-dimensional transient turbulent velocity field is generated by performing flow field inversion on the dynamic sequence of the three-dimensional particle field.

2. The method for three-dimensional reconstruction of holographic turbulent flow field in turbid water based on deep learning as described in claim 1, characterized in that, The method generates an initial set of particle plane coordinates by performing cascaded convolutional neural network-driven two-dimensional particle projection localization on the optimized hologram. The optimized hologram is progressively subjected to dynamic contrast enhancement and multi-directional Gabor filtering to output a preprocessed hologram; The preprocessed hologram is loaded into a first-level strongly quantized convolutional network to perform global coarse localization and output a candidate region heatmap. The candidate region heatmap is input into the second-level residual convolutional network for sub-pixel coordinate regression, and the particle center coordinate distribution is output. The confidence level is filtered by traversing the distribution of particle center coordinates, and the initial particle plane coordinate set is output.

3. The method for three-dimensional reconstruction of holographic turbulent flow field in turbid water based on deep learning as described in claim 1, characterized in that, Perform multi-frame trajectory reconstruction with spatiotemporal graph association on the particle transient spatial coordinate set to generate a three-dimensional particle field dynamic sequence. The method includes: A spatiotemporal heterogeneous graph is constructed based on the particle transient spatial coordinate set; A graph attention network is used to traverse the spatiotemporal heterogeneous graph to calculate the correlation weights between particle nodes and fit the initial trajectory chain. Based on acceleration smoothing constraints, the motion continuity of the initial trajectory chain is iteratively optimized to output a fracture trajectory sequence; Interpolation is performed on the fracture trajectory sequence to complete the output of the three-dimensional particle field dynamic sequence.

4. The method for three-dimensional reconstruction of holographic turbulent flow field in turbid water based on deep learning as described in claim 3, characterized in that, A three-dimensional transient turbulent velocity field is generated by performing flow field inversion on the dynamic sequence of the three-dimensional particle field. The method includes: Perform frame-by-frame particle space neighborhood topology connection on the dynamic sequence of the three-dimensional particle field and output a single-frame particle-associated topology field; Under the neighborhood constraints of the single-frame particle-associated topological field, adjacent frame particle position matching is performed to generate a displacement vector field; Based on the topological neighborhood scale of the single-frame particle-associated topological field, under the fluid incompressibility constraint, the displacement vector field is interpolated to the three-dimensional spatial grid nodes to generate a gridded three-dimensional velocity field. The gridded three-dimensional velocity field is corrected based on the curl distribution of particle motion in adjacent frames, and the vorticity-corrected velocity field is output. The vorticity-corrected velocity field is time-series superimposed to generate the three-dimensional transient turbulent velocity field.

5. The method for three-dimensional reconstruction of holographic turbulent flow field in turbid water based on deep learning as described in claim 1, characterized in that, In the process of laser coherent irradiation of turbid water using a dual-pulse off-axis holographic optical path, a high-speed CMOS camera is driven to capture a sequence of transient holographic interferograms of the turbid water. The method includes: Tracer particles matching the target turbulence scale are added to the turbid water body; During the laser coherent irradiation of turbid water by triggering a dual-pulse laser to emit dual-pulse off-axis holographic optical path based on a preset laser repetition frequency, the high-speed CMOS camera is driven by a preset camera exposure frequency to capture the transient holographic interferogram sequence. The ratio of the camera exposure frequency to the laser repetition frequency is 1:(N+1), where N is the number of background noise correction frames within the dual-pulse interval.

6. The method for three-dimensional reconstruction of holographic turbulent flow field in turbid water based on deep learning as described in claim 1, characterized in that, The noise suppression model includes a parallel physical simulation branch and a data-driven branch. The outputs of the physical simulation branch and the data-driven branch are connected to a weighted fusion layer, and the signal-to-noise ratio of the weighted fusion layer is dynamically adjustable.

7. The method for three-dimensional reconstruction of holographic turbulent flow field in turbid water based on deep learning as described in claim 6, characterized in that, A noise suppression model based on a deep learning network is used to suppress muddy water scattering noise in the transient holographic interferogram sequence, and an optimized hologram is output. The method includes: The transient holographic interferogram sequence is divided into a transient holographic overlapping block sequence, wherein the transient holographic overlapping block sequence uses a Hanning window to suppress boundary effects; The transient holographic overlapping block sequence is input into the noise suppression model, and the water turbidity parameter is mapped through the physical simulation branch to output the noise distribution overlapping block. Multi-scale feature extraction is performed through the data-driven branch to output the optimized feature overlapping block. The weighted fusion layer receives and stitches together the overlapping blocks of the noise distribution and the overlapping blocks of the optimized features, then performs weighted average elimination and outputs the optimized hologram.

8. The method for three-dimensional reconstruction of holographic turbulent flow field in turbid water based on deep learning as described in claim 6, characterized in that, The method further includes: A differentiable mapping model and an angular spectrum diffraction propagation model are pre-constructed, wherein the differentiable mapping model is used to map the turbidity parameter of the water body to the scattering phase function; The localization configuration of the physical simulation branch is completed by cascading the differentiable mapping model and the angular spectrum diffraction propagation model. A multi-scale encoder and a feature fusion decoder are pre-built, and the localization configuration of the data-driven branch is completed by cascading the multi-scale encoder and the feature fusion decoder. After connecting the physical simulation branch and the data-driven branch in parallel, the weight fusion layer is connected at the output end to complete the construction of the noise suppression model.

9. The method for three-dimensional reconstruction of holographic turbulent flow field in turbid water based on deep learning as described in claim 3, characterized in that, In the spatiotemporal heterogeneous graph, each node corresponds to an independent particle in a single frame, and the node data includes the particle's three-dimensional position coordinates and instantaneous velocity. The spatial Euclidean distance between pairs of nodes with undirected connecting edges satisfies a preset distance threshold, and the time frame interval satisfies a preset frame interval.

10. A three-dimensional reconstruction system for holographic turbulent flow field in turbid water based on deep learning, characterized in that, The system is used to implement the deep learning-based holographic turbulent field three-dimensional reconstruction method for turbulent water bodies according to any one of claims 1-9, the system comprising: The transient holographic interferogram sequence acquisition module is used to drive a high-speed CMOS camera to capture the transient holographic interferogram sequence of turbid water during laser coherent irradiation of turbid water using a dual-pulse off-axis holographic optical path. The hologram acquisition module is optimized to use a noise suppression model based on a deep learning network to suppress muddy water scattering noise in the transient holographic interferogram sequence and output an optimized hologram. The initial particle plane coordinate set acquisition module is used to generate an initial particle plane coordinate set by performing two-dimensional particle projection positioning driven by a cascaded convolutional neural network on the optimized hologram; The particle transient spatial coordinate set acquisition module is used to calculate the particle axial depth of the initial particle plane coordinate set and output the particle transient spatial coordinate set. The three-dimensional particle field dynamic sequence acquisition module is used to perform multi-frame trajectory reconstruction with spatiotemporal graph association on the particle transient spatial coordinate set to generate a three-dimensional particle field dynamic sequence. The three-dimensional transient turbulent velocity field acquisition module is used to generate a three-dimensional transient turbulent velocity field by performing flow field inversion on the dynamic sequence of the three-dimensional particle field.

Images