CN-121994661-A - Spherical particle size measurement method and device based on generation of antagonism network and deep learning

CN121994661ACN 121994661 ACN121994661 ACN 121994661ACN-121994661-A

Abstract

The invention discloses a spherical particle size measurement method and device based on generation of an antagonism network and deep learning, and belongs to the field of particle measurement and image processing. The method comprises the steps of obtaining an experimental collection spherical particle interference fringe pattern and an ideal fringe pattern based on optical transmission matrix simulation, labeling corresponding spherical particle size labels on each image, converting the simulation pattern into an analog pattern by using a CycleGAN network, constructing an expanded data set, combining the experimental pattern with the analog pattern, training a pre-constructed size prediction model by using the data set, adding a frequency domain attention module and a full-connection layer on a coding and decoding layer and a network output layer respectively on the basis of an Ege-unet network, and carrying out size prediction on an input image by using the trained size prediction model to realize measurement of spherical particles with any size. The invention can finally realize the measurement of spherical particles with any size by generating the data generalization enhancing capability of the countermeasure network.

Inventors

SUN JINLU
WANG XIAOYANG
ZHANG CHENG
LI YUQIANG
WANG DI

Assignees

天津工业大学

Dates

Publication Date: 20260508
Application Date: 20260126

Claims (10)

1. A spherical particle size measurement method based on generation of a countermeasure network and deep learning, comprising the steps of: the preparation method comprises the steps of preparing a data set, namely acquiring an experimental acquired spherical particle interference fringe pattern and an ideal fringe pattern based on optical transmission matrix simulation, and labeling a corresponding spherical particle size label on each image; Converting an ideal fringe pattern simulated by an optical transmission matrix into a simulated pattern with experimental noise characteristics by using CycleGAN networks, and constructing an expanded data set; Combining the spherical particle interference fringe diagram acquired by experiments with the simulation diagram, and dividing the spherical particle interference fringe diagram into a training set, a verification set and a test set according to proportion; Training a pre-constructed size prediction model by using the data set, wherein the size prediction model is based on an Ege-unet network, and a frequency domain attention module and a full-connection layer are respectively added in a coding and decoding layer and a network output layer; and a size prediction step, namely performing size prediction on the input image by using a trained size prediction model, and realizing measurement of spherical particles with any size.
2. The method of claim 1, wherein the CycleGAN network in the image generating step implements real noise migration from the experimental graph to the simulated graph through unpaired training, and generates a simulated graph having both structural features of the simulated graph and noise features of the experimental graph.
3. The method of claim 1, wherein in the dataset preparation step, obtaining an experimentally collected spherical particle fringe pattern comprises: The experimental acquisition is carried out by an interference particle imaging system, the system comprises a laser, a spatial filter, a collimating lens, a cylindrical lens group, a sample cell, an imaging lens and a CCD camera, the diluted polystyrene spherical particles with different sizes are used as experimental particles and are placed in the sample cell to be used as sample carriers, and a spherical particle interference fringe diagram is obtained by the CCD camera through sheet laser beam irradiation.
4. A method according to claim 3, wherein the CCD camera resolution is 2448 x 2048 and the pixel size is 3.45 μm.
5. A method according to claim 3, wherein the experimental particles are polystyrene spherical particles of 30 μm, 45 μm, 60 μm, 90 μm.
6. The method of claim 1, wherein the spherical particle size in the simulated image is 30-90 μm.
7. The method according to claim 1, wherein the pre-constructed size prediction model specifically comprises: adding 1 full connection layer to the decoding final output layer of Ege-unet network to make the network possess the function of numerical value prediction; The 3 frequency domain attention modules are added to three modules after the Ege-unet network is decoded, and the 3 frequency domain attention modules are added to three modules before the Ege-unet network is encoded.
8. The method of claim 1, wherein the step of generating the image comprises a CycleGAN network training process, wherein the optimizer uses Adam, the initial learning rate is set to 2 x 10 -4 , the first moment estimation parameter is set to 0.5, and the training round number is set to 200 epochs, thereby obtaining the optimal network weight.
9. The method of claim 1, wherein the training process in the network construction and training step comprises setting an initial learning rate to 10 -5 , a learning rate momentum to 0.9, a weight decay factor to 10 -8 , and a training round number to 150 epochs, thereby obtaining the optimal weight.
10. A spherical particle size measurement apparatus based on generation of an countermeasure network and deep learning, for realizing the spherical particle size measurement method based on generation of an countermeasure network and deep learning according to any one of claims 1 to 9, the apparatus comprising: the data set preparation module is used for acquiring an experimental acquired spherical particle interference fringe pattern and an ideal fringe pattern based on optical transmission matrix simulation; the image generation module converts an ideal fringe pattern simulated by the optical transmission matrix into a simulated pattern with experimental noise characteristics by using CycleGAN networks, and constructs an expanded data set; The data set dividing module is used for merging the spherical particle interference fringe diagram acquired by experiments with the simulation diagram and dividing the spherical particle interference fringe diagram into a training set, a verification set and a test set according to proportion; The network construction and training module is used for training a pre-constructed size prediction model by using the data set, wherein the size prediction model is based on an Ege-unet network, and a frequency domain attention module and a full-connection layer are respectively added in a coding and decoding layer and a network output layer; And the size prediction module is used for performing size prediction on the input image by using the trained size prediction model, so as to realize measurement of spherical particles with any size.

Description

Spherical particle size measurement method and device based on generation of antagonism network and deep learning Technical Field The invention relates to the technical field of particle measurement and image processing, in particular to a spherical particle size measurement method and device based on generation of an antagonism network and deep learning. Background The measurement of the particle size in any and real time is important to accurately predict the weather and the accuracy of the weather technology affected by the man. The interference particle imaging (Interferometric PARTICLE IMAGING, IPI) technology is a measurement method based on particle scattered light distribution, and the method researches the relationship between the particle diameter and the scattered light distribution thereof through Mie scattering theory, receives the scattered light of the particles through an optical imaging system, and further acquires size information. In the research of realizing particle size measurement by using an interference particle imaging technology, the prior patent CN116559033a extracts the fringe frequency in the interference fringe pattern through a two-dimensional fourier transform algorithm to obtain the particle size, and cannot acquire cloud particle field information in real time. CycleGAN (Cycle-Consistent GENERATIVE ADVERSARIAL Network) is a method based on generating an countermeasure Network, and style conversion between different domains is achieved through unsupervised learning. CycleGAN is characterized in that a loss of 'loop consistency' is introduced, so that the image can be restored back to the original image after being converted by the two generating networks, and the consistency of the structure and the semantics is maintained. Currently, no method for processing an optical transmission matrix simulation graph to be similar to the noise characteristics and styles of an experimental graph by using a CycleGAN network exists in the prior art. The U-net is connected with the cross-layer jump by adopting an encoder-decoder structure, semantic information is extracted through downsampling, spatial details are recovered through upsampling, and high-precision numerical prediction is realized. The U-net network has been applied to the field of particle measurement, and the measurement method of the prior patent document CN119935850A realizes real-time high-precision spherical particle size measurement, but only specific-size particle measurement. Disclosure of Invention The present invention has been made in view of the above problems, and has as its object to provide a spherical particle size measurement method and apparatus based on generating an antagonistic network and deep learning, which overcomes or at least partially solves the above problems. In order to achieve the above purpose, the present invention adopts the following technical scheme: in a first aspect, an embodiment of the present invention provides a spherical particle size measurement method based on generation of an antagonism network and deep learning, including: the preparation method comprises the steps of preparing a data set, namely acquiring an experimental acquired spherical particle interference fringe pattern and an ideal fringe pattern based on optical transmission matrix simulation, and labeling a corresponding spherical particle size label on each image; Converting an ideal fringe pattern simulated by an optical transmission matrix into a simulated pattern with experimental noise characteristics by using CycleGAN networks, and constructing an expanded data set; Combining the spherical particle interference fringe diagram acquired by experiments with the simulation diagram, and dividing the spherical particle interference fringe diagram into a training set, a verification set and a test set according to proportion; Training a pre-constructed size prediction model by using the data set, wherein the size prediction model is based on an Ege-unet network, and a frequency domain attention module and a full-connection layer are respectively added in a coding and decoding layer and a network output layer; and a size prediction step, namely performing size prediction on the input image by using a trained size prediction model, and realizing measurement of spherical particles with any size. In one embodiment, the CycleGAN network in the image generation step realizes real noise migration from the experimental graph to the simulation graph through unpaired training, and generates a simulation graph with structural characteristics of the simulation graph and noise characteristics of the experimental graph. In one embodiment, in the data set preparing step, obtaining an experimentally collected spherical particle interference fringe pattern includes: The experimental acquisition is carried out by an interference particle imaging system, the system comprises a laser, a spatial filter, a collimating lens, a cylindrical l