Particle manifold balance characterization method and device based on multi-view sparseness quantization

Abstract

The invention relates to a particle class manifold balance characterization method and device based on multi-view sparseness quantification. The method comprises the following steps: constructing and training a multi-view sparseness quantification network and a class manifold balance network; in the multi-view sparseness quantization stage, a three-dimensional digital holographic particle field obtained by combining angular spectrum method holographic reproduction with frequency domain high-pass filtering background field removal is mapped along a Z view angle, and coordinates of a target area containing existing particles are obtained through a main network and two auxiliary networks; in the accurate characterization stage, region data are cut out from the three-dimensional digital holographic particle field data according to the target region coordinates obtained in the multi-view sparseness quantification stage, feature extraction is carried out through a manifold-like balance network, and final particle three-dimensional morphological characterization and focusing center coordinates are obtained. Under the condition that holographic particle field imaging data quality is poor and real-time performance is required, detection and segmentation of the particles can be rapidly completed, space coordinates and a focusing layer are determined, and information such as the shape and the size of the particles is obtained.

Description

Technical Field

[0001] The present invention relates to the technical field of particle intelligent detection and characterization based on convolutional neural networks, and in particular to a particle-like manifold balance characterization method and device based on multi-view sparse quantization. Background Art

[0002] Holography utilizes the principles of light interference and diffraction to record and reconstruct three-dimensional images of objects. Depending on the methods and media used in recording and reconstructing holograms, holography can be categorized as traditional optical holography (film-based) and digital holography (digital). Depending on the angle between the object beam and the reference beam during the holographic imaging process, holography can be categorized as on-axis holography and off-axis holography.

[0003] The main imaging process of digital holography is generally divided into two steps: recording the hologram and reproducing it. Digital holography requires a holographic optical experimental system to be established on an optical platform to complete the recording process. Holography can record the amplitude and phase information of an object. Following the basic principle of "interference recording, diffraction reproduction," holography uses reference light to record the complex amplitude of the object light. The coherent superposition of the object light and reference light produces interference fringes, which are the hologram. The hologram is then diffracted to reconstruct the three-dimensional image of the object. In digital holography, when recording holograms, electronic devices such as CCDs and CMOS replace traditional holographic plates, and the captured holograms are then recorded in a computer.

[0004] Particle fields are an important research topic in many fields and can be considered as target clusters. Particles generally refer to geometric entities with specific shapes ranging in size from millimeters to nanometers. Particles include not only solid particles but also liquid particles such as droplets and oil droplets, which are the target entities that constitute particle fields. Accurate measurement of particle fields is of great significance to research in defense, military, aerospace, biomedicine, physics, chemistry, atmospheric science, and other fields. However, particles in particle fields across different application areas often exhibit complex scales, morphologies, distributions, and motion characteristics. Particle fields designed for specific scenarios can be composed of particles that are either static or dynamic, and their distributions can vary widely in concentration, potentially presenting challenges such as cross-scale and arbitrary shapes. Furthermore, particle fields generated in complex environments may be accompanied by severe flow disturbances, resulting in complex background fields, low contrast between faint targets and background, and poor signal-to-noise ratio. On a temporal scale, particle velocities in many high-speed dynamic particle fields reach the order of kilometers per second. The large spatial span, high temporal velocity, and high physical energy place high demands on particle field measurement methods and technologies. Therefore, how to perform transient, large-field-of-view, high-resolution imaging of energetic particle fields in complex scenarios, and thereby achieve dynamic measurement and morphological characterization with high recognition rate, high precision, and high efficiency, is a challenging research topic that the relevant industries focus on, and is also an important bottleneck issue that restricts the development of related fields.

[0005] Convolutional neural networks in deep learning are powerful tools for detecting and segmenting digital holographic particle images. In recent years, many works have been applied to digital holographic particle fields, with significant improvements in positioning accuracy and particle recognition rate. Wu et al. proposed an encoder-decoder network called Dense-U-net to simultaneously obtain particle radius, coordinates, and depth from two-dimensional lensless holograms. Shao et al. proposed an improved UNet structure that takes recorded holograms, reconstructed holographic slices, and longitudinal minimum intensity projections as inputs, and obtains the focal depth and 2D center of mass of each particle. Trujillo et al. used convolutional neural networks (CNNs) to detect and count phase objects in original holograms without the need for holographic reconstruction. Zhao et al. proposed PANet and corresponding training methods to detect three-dimensional holographic particle fields. The detection process is divided into two steps, implemented by PNet and ANet respectively. Experiments show that it has good performance in particle field characterization.

[0006] High-speed particle field analysis and characterization often require a segmented description of the entire dynamic process. The size, velocity, density, and shape of particles within different fragmentation zones can vary significantly. The front end of the holographic recording space records densely distributed metal droplets late in the exposure period; the back end, after the metal droplets are impacted and fragmented, is an earlier exposure period, when the particles are larger and more sparsely distributed. Therefore, different periods correspond to different particle sparsities. Furthermore, within several layers of the same group, the particle distribution is relatively sparse and uneven. Most existing deep learning-based particle detection and characterization work fails to account for differences in particle sparsity and typically directly uses a 3D neural network to traverse the entire 3D digital holographic particle field data. However, the network processes most of the data without background information about the particles, failing to address the issue of redundant background information. Consequently, the proposed method cannot achieve efficient feature extraction in particle fields with varying sparsity. Furthermore, a large number of particles with similar shapes, moderate volumes, and regular morphology are present. In contrast, particles with smaller diameters and blurred features, as well as larger diameters and irregular boundaries, are relatively rare, resulting in a quasi-unbalanced distribution. The holographic particle recognition network mentioned above does not address the class imbalance distribution in the data, and the simulated digital holographic particle data used only changes in the spatial coordinates without distinguishing the shape and size of the particles. Therefore, it cannot effectively simulate and adapt to the actual particle field where the particles have regular and irregular shapes and cross-scale sizes.

[0007] In summary, in practical applications, the particle sparsity of particle fields varies significantly under different exposure times within the holographic recording space; different particles have diverse shapes and a wide range of sizes; the quality of holographic images of different particle fields is low due to hardware and imaging technology; and the imaging process of the imaging system, the complex and changing actual environment, the mutual interference of adjacent particles, and different experimental settings lead to significant differences in the distribution of background interference in the particle field between different experiments. Due to these factors, methods based on traditional image processing algorithms have difficulty in ensuring detection results. Compared with traditional image processing methods, deep learning methods offer significant performance advantages and application prospects when it comes to quickly and efficiently processing complex distributed data. Summary of the Invention

[0008] The purpose of the present invention is to provide a particle-like manifold balance characterization method and device based on multi-perspective sparsity quantization, which can adaptively extract the target area containing particles according to the particle sparsity of the particle field, effectively avoiding background information without targets from being sent to the manifold-like balance network in the precise characterization stage, thereby solving the problem of redundancy of background information.

[0009] The present invention adopts the following technical solutions:

[0010] On the one hand, a particle-like manifold equilibrium characterization method based on multi-view sparse quantization includes:

[0011] S101, constructing a multi-view sparse quantization network including a main network and two auxiliary networks and training the network to obtain a trained multi-view sparse quantization network; constructing a manifold-like balanced network and training the network to obtain a trained manifold-like balanced network;

[0012] S102, in the multi-view sparse quantization stage, the three-dimensional digital holographic particle field obtained by combining angular spectrum holographic reconstruction with frequency domain high-pass filtering to remove the background field is mapped along the Z viewing angle to form reduced-dimensional 2D holographic data, which is input into the trained main network to obtain the coordinates of the regions where particles exist on the X and Y planes; the three-dimensional data divided along the Z viewing angle from the three-dimensional digital holographic particle field data according to the regional coordinates is mapped along the X and Y viewing angles respectively to obtain the respective reduced-dimensional 2D holographic data, which are respectively input into the two trained auxiliary networks to obtain the range of the number of layers of particles existing in the region. The regional extraction results from the three orthogonal viewing angles together constitute the coordinates of the target region containing the existing particles;

[0013] S103, in the precise characterization stage, firstly, the regional data is cut out from the three-dimensional digital holographic particle field data according to the target area coordinates obtained in the multi-view sparse quantization stage, and then the features are extracted through the trained manifold-like balance network to obtain the final three-dimensional particle morphology representation and focusing center coordinates.

[0014] Preferably, the three-dimensional digital holographic particle field obtained by combining angular spectrum holographic reconstruction with frequency domain high-pass filtering to remove the background field is mapped along the Z viewing angle to form 2D holographic data after dimensionality reduction, specifically comprising: taking the mean plus the minimum of pixels of all layers with the same XY plane coordinates to form two-dimensional projection dimensionality reduction data;

[0015] The three-dimensional data divided on the three-dimensional digital holographic particle field data along the Z viewing angle according to the regional coordinates are mapped along the X and Y viewing angles respectively to obtain the respective reduced-dimensional 2D holographic data, specifically including taking the mean plus the minimum value of the pixels of all layers with the same XZ plane and YZ plane coordinates to form the corresponding two-dimensional projection reduced-dimensional data.

[0016] Preferably, the main network and the auxiliary network both include five stages connected in sequence, each stage consisting of a pyramid module with a different number of feature channels and repetitions; specifically, each stage first performs a feature channel change, and then repeats the pyramid module several times; the input of the main network is converted from 1 to 8 after a 3×3 convolution layer, and then repeated feature extraction is performed through the pyramid module in five stages to obtain the region X and Y range coordinates, the number of channels of the main network is [16, 32, 64, 8, 1], and the number of basic unit repetitions is [3, 4, 6, 3, 1]; the input of the auxiliary network is the data divided along the Z perspective on the three-dimensional digital holographic particle field data according to the region coordinates output by the main network, mapped along the X perspective and the Y perspective respectively to obtain 2D holographic data, first passing through a 3×3 convolution layer to convert the number of channels from 1 to 8, and then repeated feature extraction is performed through the pyramid module in five stages to obtain the region Z range coordinates, the number of channels of the auxiliary network is [16, 32, 64, 8, 1], and the number of basic unit repetitions is [1, 1, 2, 1, 1].

[0017] Preferably, the input of the pyramid module undergoes one ordinary 3×3 convolution and three 3×3 convolutions with different diffusivities in parallel, with diffusivities of [1, 3, 8] respectively. Each convolution is followed by Batch Normalization and LeakyReLU activation function, and the output is labeled y={y1, y2, y3, y4}. The operations performed by each group are then expressed as follows:

[0018] y1=f1(x1), y2=f2(x2), y3=f3(x3), y4=f4(x4);

[0019] Among them, f1 represents an ordinary 3×3 convolution layer; f2, f3, and f4 represent three 3×3 diffusion convolution layers with different diffusion rates, respectively. The four groups of outputs are cascaded and labeled as y = {y1, y2, y3, y4}. They are then passed through a 3×3 convolution layer and the LeakyReLU activation function to adjust the number of channels to the number of input channels to obtain the output feature map, and finally a residual connection is made with the input.

[0020] Preferably, the overall structure of the manifold-like balancing network is U-shaped, including an encoder and a decoder; the encoder includes three three-dimensional feature extraction modules and two downsampling modules, which perform feature extraction and scale compression respectively; the decoder includes two three-dimensional feature extraction modules and two upsampling modules, which reorganize the features learned by the encoder; the feature maps output by each basic unit of the encoder are cascaded with the feature maps in the decoder as the input of the three-dimensional feature extraction module in the decoder, so that the decoder obtains shallow feature information; finally, the output is normalized using the Sigmoid function.

[0021] Preferably, the three-dimensional feature extraction module processes the input through two weighted normalization three-dimensional convolutions and a combination of feature map group normalization and LeakyReLU, passes through a channel weighting module, and then passes through a fully connected layer and group normalization to perform feature recombinant fusion to obtain an output feature map, and finally performs a residual connection with the input, as follows:

[0022] F1=WCGL(WCGL(F in ));

[0023] F out =FFN(SEB(F1))+F in ;

[0024] Among them, F in , F1, F out They represent the module input feature map, intermediate feature map and module output feature map respectively; WCGL represents the combination of weight group normalization three-dimensional convolution, feature map group normalization and LeakyReLU activation function; SEB represents channel weighting operation; FFN represents the fully connected layer feature fusion operation.

[0025] Preferably, the training process of the multi-view sparse quantization stage is as follows:

[0026] The main network and the two auxiliary networks are trained separately based on the common cross entropy loss function;

[0027] The main network and two auxiliary networks are trained jointly, and the joint loss function of the three networks is defined as a specific loss function designed based on cross entropy and continues training until convergence.

[0028] Preferably, the joint loss function of the main network is expressed as follows:

[0029]

[0030] Among them, Y (xy) 、Y (yz) 、Y (xz) They are the main network and two auxiliary network outputs, Label (xy) For the region segmentation label of XY plane, GMP Z To perform the minimum projection dimensionality reduction operation on the three-dimensional matrix along the Z perspective, S is the Sigmoid activation function, The weight parameter before the cross entropy positive and negative samples of the main network loss function, loss (xy) The joint loss function of the main network; x ij Represents the output of the network at (i, j) coordinates; y ij Represents the value of the label at the (i, j) coordinate; log represents logarithmic processing of the input; SUM represents the sum of the processing results of each pixel coordinate.

[0031] Preferably, a high-concentration particle field simulation training strategy is adopted in the training process of the manifold-like balance network in the precise characterization stage. According to the particle data characteristics, multiple particle data blocks are taken as the minimum value at the corresponding positions to obtain fused high-concentration particle data blocks to adapt to the high-concentration particle field data; the loss function of the manifold-like balance network is defined as a loss function designed based on cross entropy and trained until convergence.

[0032] In another aspect, a particle-like manifold equilibrium characterization device based on multi-view sparse quantization includes:

[0033] The network construction and training module is configured to construct a multi-view sparse quantization network including a main network and two auxiliary networks and perform training to obtain a trained multi-view sparse quantization network; construct a manifold-like balanced network and perform training to obtain a trained manifold-like balanced network;

[0034] The multi-view sparse quantization module is configured to, during the multi-view sparse quantization stage, map the three-dimensional digital holographic particle field obtained by combining angular spectrum holographic reconstruction with frequency domain high-pass filtering to remove the background field along the Z viewing angle to form reduced-dimensional 2D holographic data, which is input into the trained main network to obtain the coordinates of the regions where particles exist on the X and Y planes; map the three-dimensional data divided along the Z viewing angle from the three-dimensional digital holographic particle field data according to the regional coordinates along the X and Y viewing angles respectively to obtain the respective reduced-dimensional 2D holographic data, which are respectively input into the two trained auxiliary networks to obtain the range of the number of layers of particles existing in the region. The regional extraction results from the three orthogonal viewing angles together constitute the coordinates of the target region containing the particles.

[0035] The precise characterization module is configured to first cut out regional data from the three-dimensional digital holographic particle field data based on the target area coordinates obtained in the multi-view sparse quantization stage, and then perform feature extraction through a trained manifold-like balance network to obtain the final particle three-dimensional morphological representation and focusing center coordinates.

[0036] Compared with the prior art, the present invention has the following beneficial effects:

[0037] (1) The present invention adopts a multi-view sparse quantization and manifold balance method for the intelligent detection and morphological characterization process of digital holographic particle field. Based on the convolutional neural network, the digital holographic particle field dimensionality reduction projection is combined with the two-stage convolutional neural network for the first time according to the data characteristics to adaptively accelerate the three-dimensional digital holographic particle field focusing center detection and morphological characterization for different particle sparsities, thereby improving the particle detection rate and speeding up the network's recognition speed of the sparse particle field; according to the task characteristics of different stages, a two-dimensional multi-view sparse quantization network with a specific structure and a manifold balance network for three-dimensional precise positioning and segmentation are constructed, and specific loss functions for tasks at different stages are designed to complete the holographic reconstruction of the digital holographic particle field and the recognition and segmentation of particle targets, thereby solving the limitations and deficiencies of traditional image processing methods and existing deep learning methods in processing the unbalanced digital holographic particle field data with low efficiency and serious background field interference under different exposure times from dense to sparse, large particle size span and diverse shapes by using dimensionality reduction operation. 3 ) is reduced to O(n 2 ), an efficient and fast multi-view sparse quantization of its three-dimensional distribution information, while having higher robustness to particle fields with unbalanced particle feature classes and uncommon features, thus improving the detection rate and segmentation accuracy;

[0038] (2) In the multi-view sparse quantization stage, the present invention adopts pyramid modules in the five-stage repetitive modules of the main network and the auxiliary network, which can improve the network's ability to represent features of different scales and adaptively filter redundant background field information according to different particle sparsities. At the same time, feature recombination is introduced in the manifold-like balance network in the precise characterization stage, and weight group normalization and feature map group normalization are used, so that the network has high efficiency and robustness in identifying and segmenting particle fields with different characteristics, and has higher detection rate and stability for small category particles than other traditional methods / deep learning methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0040] Figure 1 Schematic diagram of the process of a particle-like manifold balance characterization method based on multi-view sparse quantization according to an embodiment of the present invention;

[0041] Figure 2 This is a model block diagram of a particle-like manifold balance characterization method based on multi-view sparse quantization according to an embodiment of the present invention;

[0042] Figure 3 Schematic diagram of the structure of the main network, auxiliary network, and network module in the multi-view sparse quantization phase of the particle-like manifold balance characterization method based on multi-view sparse quantization according to an embodiment of the present invention; (a) shows the main network and auxiliary network structure; (b) shows the pyramid module structure; (c) shows a schematic diagram of the diffusion convolution kernel;

[0043] Figure 4 Schematic diagrams of the structure of the quasi-manifold balance network and network modules in the precise characterization stage of the particle quasi-manifold balance characterization method based on multi-view sparse quantization according to an embodiment of the present invention; (a) is the quasi-manifold balance network structure; (b) is a schematic diagram of the feature recombination module structure; (c) is a schematic diagram of the three-dimensional pyramid module structure;

[0044] Figure 5 A diagram showing the impact of removing the pyramid module and reducing the number of module repetitions on the multi-view sparse quantization stage through loss curves;

[0045] Figure 6 Schematic diagram showing the impact of removing weight group normalization and feature map group normalization on the accurate representation stage through loss curves;

[0046] Figure 7 Target detection results of different methods on different types of digital holographic particle field dimensionality reduction projection data, including regular shape particle field data, irregular shape particle field data, large size particle field data and array dense particle field data, as well as the corresponding detection area result maps and detection rate calculation;

[0047] Figure 8 To assist the network in detecting the Z coordinate range of the target area of different particle projection data, including the display of dimension reduction data along the X and Y perspectives and the visualization of the corresponding Z coordinate range detection results;

[0048] Figure 9 The particle segmentation results of different methods in the precise characterization stage, including the original particle field map results;

[0049] Figure 10The segmentation detection results of using the high-concentration particle field simulation training strategy and not using the high-concentration particle field simulation training strategy on a high-concentration particle field that the network has never been exposed to, including the original focus layer data map and the segmentation result map;

[0050] Figure 11 The final particle segmentation results using the cross-dimensional dual-stage strategy and not using the cross-dimensional dual-stage strategy, including the original focus layer digital holographic data map, network weight heat map and segmentation result map;

[0051] Figure 12 This is a structural block diagram of a particle-like manifold balance characterization device based on multi-view sparse quantization according to an embodiment of the present invention. DETAILED DESCRIPTION

[0052] To make the objectives, technical solutions, and advantages of the present invention more apparent, the present invention will be further described in detail below with reference to the accompanying drawings. It is apparent that the embodiments described are only some, not all, of the present invention. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without creative effort are intended to fall within the scope of protection of the present invention.

[0053] See also Figure 1 As shown, this embodiment provides a particle-like manifold balance characterization method based on multi-view sparse quantization, which includes the following steps.

[0054] S101, constructing a multi-view sparse quantization network including a main network and two auxiliary networks and training it to obtain a trained multi-view sparse quantization network; constructing a manifold-like balanced network and training it to obtain a trained manifold-like balanced network.

[0055] Specifically, the multi-view sparse quantization network is used in the multi-view sparse quantization phase, taking as input a three-dimensional digital holographic particle field obtained by holographic reconstruction of a single two-dimensional digital hologram using angular spectrum method combined with frequency domain high-pass filtering to remove the background field. The manifold-like balance network is used in the precise characterization phase, taking as input a target region cut out of the original data using the region coordinates output by the multi-view sparse quantization network.

[0056] The original single digital hologram is obtained by recording a particle field composed of metal droplets produced by a high-Mach shock tube bombarding metal with a CCD digital camera. The digital hologram records the diffraction ripples produced by particles at different distances over a large range under laser irradiation with a CCD digital camera, and can be holographically reproduced using the angular spectrum method. During the reproduction process, three-dimensional layered information is observed layer by layer, which is commonly used to perform multi-perspective sparse quantization of the particle field to obtain information such as the particle shape, distribution, and motion state. Therefore, before performing three-dimensional particle position detection and segmentation, it is usually necessary to first perform holographic reconstruction to obtain three-dimensional digital holographic particle field data. However, in the reproduced digital holographic particle field, the particles and background exhibit different pixel intensities. The overall recognition and segmentation task can be divided into two subtasks through projective dimensionality reduction to reduce the complexity of identifying and segmenting the overall particle field. Therefore, based on the characteristics of the task, the embodiments of the present application design a cross-dimensional two-stage network structure that can quickly and one-stop process the positioning and segmentation of the original single digital hologram after reproduction.

[0057] See also Figure 2 As shown, the multi-view sparse quantization and manifold-like balancing method consists of two parts: a multi-view sparse quantization phase and a precise characterization phase. The multi-view sparse quantization phase consists of a main network and two auxiliary networks. The multi-view sparse quantization phase uses particle regions as training labels for 3D target region detection on the dimensionality-reduced projection data of the 3D digital holographic particle field. The precise characterization phase, consisting of a manifold-like balancing network, uses the particle position morphology segmentation map as training labels and aims to learn the 3D particle morphology segmentation process. The precise characterization phase, composed of a manifold-like balancing network, uses the region coordinates output by the precise characterization phase as input, a target region cut out from the original 3D digital holographic particle field data. This eliminates the large amount of insignificant background regions in the sparse particle field, reduces downsampling times, and achieves a smaller network size, accelerating recognition speed. The 3D particle morphology segmentation results estimated by the manifold-like balancing network demonstrate the particle morphology before and after the focusing layer. Furthermore, by introducing a high-density particle field simulation training strategy, the network's robustness to high-density particle field data is further improved.

[0058] In a specific embodiment, the multi-view sparse quantization stage and the precise characterization stage are trained step by step using different data sets respectively, and the manifold-like balance network in the precise characterization stage adopts a high-concentration particle field simulation training strategy. According to the particle data characteristics, multiple particle data blocks are taken at the minimum value at the corresponding position to obtain a fused high-concentration particle data block to adapt to the high-concentration particle field data.

[0059] The multi-view sparse quantization stage and the precise representation stage adopt a step-by-step training method, which includes the following steps.

[0060] S101, training the main network and the auxiliary network separately, and defining a common cross entropy loss function as a loss function;

[0061] S102, jointly train the main network and the two auxiliary networks. The joint loss function of the three networks is defined as a specific loss function designed based on cross entropy and continues training until convergence. The joint loss function is as follows:

[0062]

[0063]

[0064] Among them, Y (xy) 、Y (yz) 、Y (xz) They are the main network and two auxiliary network outputs, Label (xx) For the region segmentation label of XY plane, GMP Z To perform the minimum projection dimensionality reduction operation on the three-dimensional matrix along the Z perspective, S is the Sigmoid activation function, The weight parameter before the cross entropy positive and negative samples of the main network loss function, loss (xy) is the loss function of the main network; x ij Represents the output of the network at (i, j) coordinates; y ij Represents the value of the label at the (i, j) coordinate; log represents logarithmic processing of the input; SUM represents the sum of the processing results of each pixel coordinate.

[0065] It should be noted that the two network loss functions of the auxiliary network are the same as those of the main network, and are similar to loss (xy) The difference is the XZ and YZ perspectives.

[0066] S103, training the manifold-like equilibrium network in the precise characterization stage. During the training process, a high-concentration particle field simulation training strategy is adopted to fuse multiple particle data blocks to increase the particle concentration of the training set. After the fusion is completed, the data set used for training is as follows:

[0067]

[0068] Among them, input1,…,input i is the sub-particle data block to be fused, 2 to 4 particle blocks are randomly fused, P is the data set used for training after the fusion is completed; the loss function is defined as the loss function designed based on cross entropy and trained until convergence.

[0069] S102, in the multi-view sparse quantization stage, the three-dimensional digital holographic particle field obtained by combining angular spectrum holographic reconstruction with frequency domain high-pass filtering to remove the background field is mapped along the Z viewing angle to form 2D holographic data after dimensionality reduction and input into the trained main network to obtain the regional coordinates of the particles on the X and Y planes; the three-dimensional data divided along the Z viewing angle on the three-dimensional digital holographic particle field data according to the regional coordinates are mapped along the X and Y viewing angles respectively to obtain the respective 2D holographic data after dimensionality reduction and input into the two trained auxiliary networks respectively to obtain the layer number range of the particles existing in the region. The regional extraction results of the three orthogonal viewing angles together constitute the target region coordinates containing the existing particles.

[0070] In a specific embodiment, combining frequency domain high-pass filtering to remove background fields in angular spectrum holographic reconstruction includes adding a high-pass filter before inverse Fourier transform, as follows:

[0071]

[0072] Among them, H(f x ,f y ) is a high-pass filter in the frequency domain, f x and f y are the spatial frequencies along the X and Y coordinate directions in the frequency domain, respectively, and threshold is the cutoff frequency of the high-pass filter.

[0073] Furthermore, the method for mapping the three-dimensional hologram reconstructed by the angular spectrum method holography along the Z viewing angle to form the reduced-dimensional 2D holographic data includes taking the mean and adding the minimum value of the pixels of all layers with the same XY plane coordinates to form two-dimensional projection reduced-dimensional data; similarly, the three-dimensional data divided on the three-dimensional digital holographic particle field data along the Z viewing angle according to the regional coordinates are mapped along the X viewing angle and the Y viewing angle respectively to obtain the 2D holographic data, including taking the mean and adding the minimum value of the pixels of all layers with the same XZ plane and YZ plane coordinates to form two-dimensional projection reduced-dimensional data; the main network and the auxiliary network both include five stages connected in sequence, and each stage is composed of pyramid modules with different numbers of feature channels and repetitions.

[0074] Specifically, during the multi-view sparse quantization stage, the original single digital hologram is first reconstructed into a three-dimensional digital holographic particle field using the angular spectrum method. High-pass filtering is then used to remove the background field during the reconstruction process. Feature extraction is performed using a main network consisting of pyramid modules repeated different times across different channels to obtain the x and y coordinates of the target region. The three-dimensional data, divided along the z perspective based on the region coordinates output by the main network, is then mapped along the x and y perspectives to obtain 2D holographic data. The z coordinate range of particles within the target region is then determined by repeating the pyramid modules different times across different channels in the auxiliary network.

[0075] The pixels of all layers with the same XY plane coordinates are averaged and minimum, forming a two-dimensional projection dimensionality reduction data. During the projection dimensionality reduction process, the pixels sharing the same X and Y indexes are calculated along the Z perspective to calculate the mean and minimum, as follows:

[0076]

[0077] Among them, PM 3d To reproduce the three-dimensional digital holographic particle field, PM 2d is the two-dimensional digital holographic particle field after dimensionality reduction, m is the number of voxels along the Z viewing angle with the same XY plane coordinates in the three-dimensional particle field, S i The set of pixels from all layers that have the same XY coordinates for calculating particles or background fields.

[0078] In specific embodiments, see Figure 3 As shown in (a), the number of channels and the number of basic unit repetitions of the main network are [16, 32, 64, 8, 1] and [3, 4, 6, 3, 1], and the number of channels and the number of basic unit repetitions of the auxiliary network are [16, 32, 64, 8, 1] and [1, 1, 2, 1, 1].

[0079] See also Figure 3 As shown in (b), the input of the pyramid module undergoes one regular 3×3 convolution and three 3×3 convolutions with different diffusivity in parallel, with diffusivity of [1, 3, 8]. Each convolution layer is followed by batch normalization and a LeakyReLU activation function. The output is labeled y = {y1, y2, y3, y4}. The operations performed by each group are shown below:

[0080] y1=f1(x1), y2=f2(x2), y3=f3(x3), y4=f4(x4);

[0081] Among them, f1 represents an ordinary 3×3 convolution layer, f2, f3, and f4 represent three 3×3 diffusion convolution layers with different diffusion rates. The four groups of outputs are cascaded and labeled as y = {y1, y2, y3, y4}. They are then passed through a 3×3 convolution layer and the LeakyReLU activation function to adjust the number of channels to the number of input channels to obtain the output feature map, and finally a residual connection is made with the input.

[0082] Figure 3 (c) shows a schematic diagram of diffuse convolution with different diffusion rates.

[0083] S103, in the precise characterization stage, firstly, the regional data is cut out from the three-dimensional digital holographic particle field data according to the target area coordinates obtained in the multi-view sparse quantization stage, and then the features are extracted through the trained manifold-like balance network to obtain the final three-dimensional particle morphology representation and focusing center coordinates.

[0084] In a specific embodiment, the overall structure of the manifold-like balanced network in the precise characterization stage is U-shaped, including an encoder and a decoder; the encoder includes three three-dimensional feature extraction modules and two downsampling modules, which perform feature extraction and scale compression respectively; the decoder includes two three-dimensional feature extraction modules and two upsampling modules, which reorganize the features learned by the encoder; the feature maps output by each basic unit of the encoder are cascaded with the feature maps in the decoder as the input of the three-dimensional feature extraction module in the decoder, so that the decoder obtains shallow feature information; finally, the output segmentation result is normalized using the Sigmoid function.

[0085] See also Figure 4 As shown in (a), the quasi-manifold-balanced network structure consists of an encoder and a decoder. The convolution kernel size in the 3D feature extraction modules in the first layer of the encoder and decoder is 9×9×9, while the convolution kernel size in other 3D feature extraction modules is 3×3×3. Upsampling and downsampling in the 3D network are implemented using 3D convolution. The number of upsampling and downsampling in the network is consistent, with a stride of 2, to ensure that the network input and output scales are the same. LR2B denotes the large kernel feature recombination module, Pyramid Block denotes the 3D pyramid module, WConv3d denotes the weight group normalization module, GN denotes the feature map group normalization operation, aspp denotes the dilated convolution, and numbers such as {1, 8, 16, 32} denote the number of channels.

[0086] See also Figure 4 As shown in (b), the 3D feature extraction module passes the input through two weighted re-normalized 3D convolutions and a combination of feature map group normalization and LeakyReLU, then passes it through a channel weighting module, and then through a fully connected layer and group normalization to re-combine and fuse the features to obtain the output feature map. Finally, a residual connection is made with the input, as shown below:

[0087] F1=WCGL(WCGL(F in ));

[0088] F out =FFN(SEB(F1))+F in ;

[0089] Among them, F in , F1, F outThey represent the module input feature map, intermediate feature map and module output feature map respectively. WCGL represents weighted group normalization three-dimensional convolution, a combination of feature map group normalization and LeakyReLU activation function. SEB represents channel weighting operation. FFN represents fully connected layer feature fusion operation. The weighted group normalization divides the convolution kernel weights into 4 groups in the output channel dimension and standardizes them separately.

[0090] See also Figure 4 As shown in (c), the input of the pyramid module undergoes one normal 3×3×3 convolution and three 3×3×3 convolutions with different diffusivity in parallel, with diffusivity of [1, 3, 8]. Each convolution layer is followed by BatchNormalization and LeakyReLU activation. The output is labeled y = {y1, y2, y3, y4}. The operations performed by each group are shown below:

[0091] y1=f1(x1), y2=f2(x2), y3=f3(x3), y4=f4(x4);

[0092] Among them, f1 represents an ordinary 3×3×3 convolution layer, f2, f3, and f4 represent three 3×3×3 diffusion convolution layers with different diffusion rates. The four groups of outputs are cascaded and labeled as y = {y1, y2, y3, y4}. They are then passed through a 3×3×3 convolution layer and the LeakyReLU activation function to adjust the number of channels to the number of input channels to obtain the output feature map, and finally a residual connection is made with the input.

[0093] The following will explain the network training process and characterization results based on specific experimental data.

[0094] The training data of the embodiment of the present application comes from a single digital holographic particle field shot by 15 groups of CCDs, respectively, through 5 high Mach bombardment metal droplet experiments, the experimental data sources recorded include regular shape particle field, high concentration particle field, large size particle field, array particle field and double exposure irregular particle field, imaging field includes 1300×1300, 1000×1000, 900×900 and 800×800, matrix size is 1300×1300×100, 1000×1000×100, 900×900×100 and 800×800×100. When preparing the label data, two experts were invited to manually mark the particle mask in the field data. The binary mask is the same size as the particle field data, and the particle pixel (positive sample) is set to 1, while the other pixels are 0. These binary images provide labeled data, including particle locations and sizes, for network training. Considering computational cost and increasing the number of training samples, the particle field data and corresponding labeled data are cropped into 64×64×20 three-dimensional tiles during training. This approach preserves the randomness of the particle distribution.

[0095] Pytorch was used as a deep learning framework to implement the multi-view sparse metricization and manifold-like balancing methods. The number of channels in the first layer of the network was 8, and the number of channels doubled with each downsampling. In the loss function, the network adaptively adjusted the weight p_w in the cross-entropy loss based on the recognition results to adjust the training loss's emphasis on the results. The multi-view sparse metricization network training batch size was 32, and the manifold-like balancing network training batch size was 64. ADAM was used as the network optimizer, and the learning rate for steps 1 and 2 was 5×10 -5 After training for 5 rounds, adjust to 10 -5 , step three uses 5×10 -5 After training for 5 rounds, adjust to 10 -5 The computing platform used is an NVIDIA GeForce GTX 3090 GPU. It takes about 24 hours to train the three steps of multi-view sparse quantization and manifold-like balancing method using this platform.

[0096] Figure 5 The main network training loss curve in step 1 shows the effects of removing the pyramid module in the method of the present invention and changing the number of channels and the number of basic unit repetitions to [16, 32, 64, 8, 1] and [2, 2, 4, 2, 1] on the target area detection effect of the two-dimensional main network. The results show that the pyramid module has the most significant improvement on the target area and detection effect. Figure 6 The effects of weight group normalization (GN) and feature map group normalization (BN) on the training loss of step three are removed respectively. As can be seen from the figure, when weight group normalization and feature map group normalization are used in combination, the particle morphology segmentation effect can be effectively improved.

[0097] The main network of the multi-view sparse quantization and manifold-like balance method proposed in the embodiment of the present application is compared with the method of the present invention, which removes the pyramid module and changes the number of channels and the number of repetitions of basic units to [16, 32, 64, 8, 1] and [2, 2, 4, 2, 1] respectively. The ablation experiments are carried out on the regular shape particle field, large size particle field, array particle field and double exposure irregular shape particle field data. The segmentation results of the manifold-like balance network and the convolution-based projection APAUNet, transformer-based UNETR, Bayesian convolution-based BayesNet, ResUNet and the weight group normalization (GN) and feature map group normalization (BN) on the background interference serious particle data, irregular shape particle data, small particle data and dense array particle data are compared and analyzed. The segmentation normalization threshold is 0.99. The main network of the method of the present invention is marked as M-SQNet, and the final manifold-like balance network method is marked as GF-ONet wGN+WGS. The Z coordinate range detection results of the auxiliary network on different types of particles are compared. Figure 8 The results of the high-concentration particle field simulation training strategy are compared before and after the high-concentration particle field data. Figure 10 The particle segmentation results using the two-stage strategy are compared with those using only the manifold-like equilibrium network. Figure 11 exhibit.

[0098] Figure 7 The main network of the multi-view sparse quantization and manifold-like balancing method removes the pyramid module in the method of the present invention and changes the number of channels and the number of basic unit repetitions to [16, 32, 64, 8, 1] and [2, 2, 4, 2, 1] respectively. The ablation experiment results are performed on regular shape particle field, large size particle field, dense array particle field and double exposure irregular shape particle field data. The first two rows of the figure show the target detection results after projection dimensionality reduction of regular shape particle field and dense array particle field, and the last two rows show the target detection results after projection dimensionality reduction of large size particle field and large size particle field. The enlarged area in the figure shows that the recognition rate of this method is improved on different types of particle field data due to the design of the pyramid module and the reasonable number of repetitions. The specific detection accuracy is shown in the figure.

[0099] Figure 8 The Z-axis distribution ranges detected by the auxiliary network for different types of particles during the multi-view sparse quantization phase are shown in the figure. The results clearly show that the Z-axis distribution ranges can be accurately detected for different types of particles or even densely distributed particles, demonstrating the effectiveness of the auxiliary network structure design.

[0100] Figure 9The proposed quasi-manifold balanced network is compared with the convolution-based projection APAUNet, the transformer-based UNETR, the Bayesian convolution-based BayesNet method, the ResUNet method, and the weight group normalization (GN) and feature map group normalization (BN) methods on the segmentation results of particle data with severe background interference, irregularly shaped particles, small particles, and dense array particles. To test the generalization ability of the method, the deep learning-based methods were trained on the test data without fine-tuning. It can be observed from the figure that the quasi-manifold balanced network GF-ONet method fully utilizes the complementarity of feature map group normalization for the sensitivity of small-category particles and weight group normalization for improving the stability of the network, as well as the efficiency of introducing feature channel weighting and reorganization. It has fewer background field misidentifications and more reasonable and accurate segmentation results for different types of particles.

[0101] Table 1 Ablation experiment of the manifold balance network in the particle field data of the high Mach metal droplet experiment;

[0102]

[0103] Type 1: simple background, medium size, medium density; Type 3: complex background, small size, medium density; Type 4: complex background, small size, medium density; Type 5: simple background, large size, medium density. The numbers in brackets indicate the number of particle samples belonging to that category.

[0104] Tables 1 and 2 respectively present ablation experiments on the detection rate of the manifold-like balance network using a cross-dimensional dual-stage strategy, as well as feature map normalization and weight normalization, in a high-Mach metal droplet experimental particle field and a double-exposure particle field. The detection rates of different particle categories using the dual-stage strategy, feature map normalization, and weight normalization methods all lead the other comparison methods. As can be seen from Tables 1 and 2, the results of the present invention are closer to the labeled data.

[0105] Table 2 Ablation experiment of particle field data of manifold-like balance network in high Mach metal droplet double exposure experiment;

[0106]

[0107] Type 1: simple background, medium size, medium density; Type 2: complex background, medium size, medium density; Type 4: complex background, small size, medium density; Type 5: simple background, large size, medium density. The numbers in brackets indicate the number of particle samples belonging to that category.

[0108] Table 3 below shows a comparative ablation experiment comparing the cosine similarity of feature maps and weights across four groups during segmentation using a quasi-manifold balance network using both feature map normalization and weight normalization in a high-Mach metal droplet experimental particle field and a double-exposure particle field. The cosine similarity of feature maps and weights using both feature map normalization and weight normalization was significantly lower than that achieved using neither method or feature map normalization alone. This demonstrates that combining feature map normalization and weight normalization can improve the orthogonality of the network in the segmentation task, making the network more efficient in completing the task.

[0109] Table 3. Comparison of cosine similarity between different groups;

[0110]

[0111] Figure 10 To accurately characterize the phase, the manifold-like equilibrium network was tested on a high-concentration particle field for segmentation results. The test results clearly show that the high-concentration particle field simulation training strategy significantly improves the detection rate and segmentation accuracy of densely distributed, irregularly shaped particle fields, demonstrating the effectiveness of the auxiliary network's high-concentration particle field simulation training strategy.

[0112] Figure 11 This figure compares the particle segmentation results using a two-stage strategy with those using only a manifold-like equilibrium network. The first row shows the original digital holographic particle field data, the second row shows a heat map of the network weight contribution to the results, and the third row shows the segmentation results. The contribution heat map shows that using the cross-dimensional two-stage strategy, the network weights pay more attention to key features such as the edge shape of the particle data. The segmentation results using the two-stage strategy have more accurate edge contours, demonstrating the effectiveness of the two-stage strategy design.

[0113] For details, see Figure 12 As shown, as an implementation of the methods shown in the above figures, the present application provides an embodiment of a particle-like manifold balance characterization device based on multi-view sparse quantization, and the device embodiment is similar to Figure 1 Corresponding to the method embodiment shown, the device can be specifically applied to various electronic devices.

[0114] A particle-like manifold equilibrium characterization device based on multi-view sparse quantization, comprising:

[0115] The network construction and training module 1201 is configured to construct a multi-view sparse quantization network including a main network and two auxiliary networks and perform training to obtain a trained multi-view sparse quantization network; construct a manifold-like balanced network and perform training to obtain a trained manifold-like balanced network;

[0116] The multi-view sparse quantization module 1202 is configured to, during the multi-view sparse quantization stage, map the three-dimensional digital holographic particle field obtained by combining angular spectrum holographic reconstruction with frequency domain high-pass filtering to remove the background field along the Z viewing angle to form reduced-dimensional 2D holographic data, which is input into the trained main network to obtain the coordinates of the regions where particles exist on the X and Y planes; map the three-dimensional data divided along the Z viewing angle from the three-dimensional digital holographic particle field data according to the regional coordinates along the X and Y viewing angles to obtain respective reduced-dimensional 2D holographic data, which are respectively input into the two trained auxiliary networks to obtain the range of the number of layers of particles existing in the region. The regional extraction results from the three orthogonal viewing angles together constitute the coordinates of the target region containing the particles.

[0117] The precise characterization module 1203 is configured to first cut out regional data from the three-dimensional digital holographic particle field data according to the target area coordinates obtained in the multi-view sparse quantization stage, and then perform feature extraction through the trained manifold-like balance network to obtain the final particle three-dimensional morphological representation and focusing center coordinates.

[0118] The above is only a specific implementation of the present invention, but the design concept of the present invention is not limited to this. Any non-substantial changes to the present invention using this concept shall be deemed as an infringement of the protection scope of the present invention.

Claims

1. A particle-like manifold equilibrium characterization method based on multi-view sparse quantization, characterized by: include: S101, constructing a multi-view sparse quantization network including a main network and two auxiliary networks and training the network to obtain a trained multi-view sparse quantization network; constructing a manifold-like balanced network and performing training to obtain a trained manifold-like balanced network; S102, in the multi-view sparse quantization stage, the three-dimensional digital holographic particle field obtained by combining angular spectrum holographic reconstruction with frequency domain high-pass filtering to remove the background field is mapped along the Z viewing angle to form reduced-dimensional 2D holographic data, which is input into the trained main network to obtain the coordinates of the regions where particles exist on the X and Y planes; the three-dimensional data divided along the Z viewing angle from the three-dimensional digital holographic particle field data according to the regional coordinates is mapped along the X and Y viewing angles respectively to obtain the respective reduced-dimensional 2D holographic data, which are respectively input into the two trained auxiliary networks to obtain the range of the number of layers of particles existing in the region. The regional extraction results from the three orthogonal viewing angles together constitute the coordinates of the target region containing the existing particles; S103, in the precise characterization stage, firstly, the regional data is cut out from the three-dimensional digital holographic particle field data according to the target area coordinates obtained in the multi-view sparse quantization stage, and then the features are extracted through the trained manifold-like balance network to obtain the final three-dimensional particle morphology representation and focusing center coordinates.

2. The particle-like manifold equilibrium characterization method based on multi-view sparse quantization according to claim 1 is characterized in that: The three-dimensional digital holographic particle field obtained by combining angular spectrum holographic reconstruction with frequency domain high-pass filtering to remove the background field is mapped along the Z viewing angle to form 2D holographic data after dimensionality reduction. Specifically, the method includes: taking the mean and adding the minimum value of the pixels of all layers with the same XY plane coordinates to form two-dimensional projection dimensionality reduction data; The three-dimensional data divided on the three-dimensional digital holographic particle field data along the Z viewing angle according to the regional coordinates are mapped along the X and Y viewing angles respectively to obtain the respective reduced-dimensional 2D holographic data, specifically including taking the mean plus the minimum value of the pixels of all layers with the same XZ plane and YZ plane coordinates to form the corresponding two-dimensional projection reduced-dimensional data.

3. The particle-like manifold equilibrium characterization method based on multi-view sparse quantization according to claim 1 is characterized in that: The main network and the auxiliary network both include five stages connected in sequence, each stage consisting of a pyramid module with a different number of feature channels and repetitions; specifically, each stage first performs a feature channel change, and then repeats the pyramid module several times; the input of the main network is converted from 1 to 8 after a 3×3 convolution layer, and then repeated feature extraction is performed through the pyramid module of five stages to obtain the region X and Y range coordinates, the number of channels of the main network is [16, 32, 64, 8, 1], and the number of basic unit repetitions is [3, 4, 6, 3, 1]; the input of the auxiliary network is the data divided along the Z perspective on the three-dimensional digital holographic particle field data according to the region coordinates output by the main network, mapped along the X perspective and the Y perspective respectively to obtain 2D holographic data, first passing through a 3×3 convolution layer to convert the number of channels from 1 to 8, and then repeated feature extraction is performed through the pyramid module of five stages to obtain the region Z range coordinates, the number of channels of the auxiliary network is [16, 32, 64, 8, 1], and the number of basic unit repetitions is [1, 1, 2, 1, 1].

4. The particle-like manifold equilibrium characterization method based on multi-view sparse quantization according to claim 3 is characterized in that: The input of the pyramid module is parallelized by one normal 3×3 convolution and three 3×3 convolutions with different diffusivity, respectively [1, 3, 8]. Each convolution is followed by batch normalization and a LeakyReLU activation function. The output is labeled y = {y1, y2, y3, y4}. The operations performed by each group are shown below: y1=f1(x1), y2=f2(x2), y3=f3(x3), y4=f4(x4); Among them, f1 represents an ordinary 3×3 convolution layer; f2, f3, and f4 represent three 3×3 diffusion convolution layers with different diffusion rates, respectively. The four groups of outputs are cascaded and labeled as y = {y1, y2, y3, y4}. They are then passed through a 3×3 convolution layer and the LeakyReLU activation function to adjust the number of channels to the number of input channels to obtain the output feature map, and finally a residual connection is made with the input.

5. The particle-like manifold equilibrium characterization method based on multi-view sparse quantization according to claim 1 is characterized in that: The overall structure of the manifold-like balancing network is U-shaped, consisting of an encoder and a decoder. The encoder includes three 3D feature extraction modules and two downsampling modules, which perform feature extraction and scale compression respectively. The decoder includes two 3D feature extraction modules and two upsampling modules, which reorganize the features learned by the encoder. The feature maps output by each basic unit of the encoder are cascaded with the feature maps in the decoder and used as the input of the 3D feature extraction module in the decoder, so that the decoder can obtain shallow feature information; Finally, the Sigmoid function is used to normalize the output.

6. The particle-like manifold equilibrium characterization method based on multi-view sparse quantization according to claim 5 is characterized in that: The 3D feature extraction module passes the input through two weighted re-normalized 3D convolutions and a combination of feature map group normalization and LeakyReLU, then passes it through a channel weighting module, and then through a fully connected layer and group normalization to perform feature re-combination and fusion to obtain the output feature map. Finally, a residual connection is made with the input as follows: F1=WCGL(WCGL(F in )); F out =FFN(SEB(F1))+F in ; Among them, F in , F1, F out They represent the module input feature map, intermediate feature map and module output feature map respectively; WCGL represents the combination of weight group normalization three-dimensional convolution, feature map group normalization and LeakyReLU activation function; SEB represents channel weighting operation; FFN represents the fully connected layer feature fusion operation.

7. The particle-like manifold equilibrium characterization method based on multi-view sparse quantization according to claim 1 is characterized in that: The training process of the multi-view sparse quantization stage is as follows: The main network and the two auxiliary networks are trained separately based on the common cross entropy loss function; The main network and two auxiliary networks are trained jointly, and the joint loss function of the three networks is defined as a specific loss function designed based on cross entropy and continues training until convergence.

8. The particle-like manifold equilibrium characterization method based on multi-view sparse quantization according to claim 7 is characterized in that: The joint loss function of the main network is expressed as follows: Among them, Y (xy) 、Y (yz) 、Y (xz) They are the main network and two auxiliary network outputs, Label (xy) For the region segmentation label of XY plane, GMP Z To perform the minimum projection dimensionality reduction operation on the three-dimensional matrix along the Z perspective, S is the Sigmoid activation function, The weight parameter before the cross entropy positive and negative samples of the main network loss function, loss (xy) The joint loss function of the main network; x ij Represents the output of the network at (i, j) coordinates; y ij Represents the value of the label at the (i, j) coordinate; log represents logarithmic processing of the input; SUM represents the sum of the processing results of each pixel coordinate.

9. The particle-like manifold equilibrium characterization method based on multi-view sparse quantization according to claim 1 is characterized in that: A high-concentration particle field simulation training strategy is adopted in the training process of the manifold-like balance network in the precise characterization stage. According to the particle data characteristics, multiple particle data blocks are minimum-taken at corresponding positions to obtain fused high-concentration particle data blocks to adapt to the high-concentration particle field data; the loss function of the manifold-like balance network is defined as a loss function designed based on cross entropy and trained until convergence.

10. A particle-like manifold balance characterization device based on multi-view sparse quantization, characterized in that: include: The network construction and training module is configured to construct a multi-view sparse quantization network including a main network and two auxiliary networks and perform training to obtain a trained multi-view sparse quantization network; construct a manifold-like balanced network and perform training to obtain a trained manifold-like balanced network; The multi-view sparse quantization module is configured to, during the multi-view sparse quantization stage, map the three-dimensional digital holographic particle field obtained by combining angular spectrum holographic reconstruction with frequency domain high-pass filtering to remove the background field along the Z viewing angle to form reduced-dimensional 2D holographic data, which is input into the trained main network to obtain the coordinates of the regions where particles exist on the X and Y planes; map the three-dimensional data divided along the Z viewing angle from the three-dimensional digital holographic particle field data according to the regional coordinates along the X and Y viewing angles respectively to obtain the respective reduced-dimensional 2D holographic data, which are respectively input into the two trained auxiliary networks to obtain the range of the number of layers of particles existing in the region. The regional extraction results from the three orthogonal viewing angles together constitute the coordinates of the target region containing the particles. The precise characterization module is configured to first cut out regional data from the three-dimensional digital holographic particle field data based on the target area coordinates obtained in the multi-view sparse quantization stage, and then perform feature extraction through a trained manifold-like balance network to obtain the final particle three-dimensional morphological representation and focusing center coordinates.

Images