CN-121767763-B - Micro-plastic component rapid identification method based on multi-mode data driving

CN121767763BCN 121767763 BCN121767763 BCN 121767763BCN-121767763-B

Abstract

The invention relates to the technical field of computer vision and environmental monitoring, and discloses a microplastic component rapid identification method based on multi-mode data driving, which comprises the steps of firstly obtaining a morphological sequence image and a chemical component map of microplastic, and constructing a standardized three-dimensional voxel data set through metadata alignment; the method comprises the steps of constructing a non-coupling training set, executing three-dimensional data enhancement based on space transformation, utilizing an enhanced data training component recognition network, establishing a mapping relation between three-dimensional morphological characteristics of microplastic and chemical components, finally collecting morphological images of a sample to be detected, inputting the morphological images into the network, and outputting a three-dimensional segmentation mask with chemical component codes. The invention realizes the rapid and accurate identification of the chemical components and the three-dimensional distribution of the microplastic by the deep connection of the excavation form and the components only according to the low-cost morphological image, solves the problems of high cost and low speed of the traditional chemical imaging, and is suitable for the efficient monitoring of the pollution of the environmental microplastic.

Inventors

DING YONGCHENG
XIA BING
LI SONGSHUO
WANG QING
WU JIANQIANG
TAN JUAN
RUAN JUNJIE

Assignees

上海市环境科学研究院

Dates

Publication Date: 20260508
Application Date: 20260304

Claims (9)

1. A method for rapidly identifying micro plastic components based on multi-mode data driving is characterized by comprising the following steps: The method comprises the steps of obtaining a morphological sequence image and a chemical component map of a micro-plastic, analyzing and mapping the chemical component map into a three-dimensional voxel level component tag, constructing a standardized three-dimensional voxel data set containing morphological characteristics and component tags through element data alignment of physical space consistency, analyzing the chemical component and generating a corresponding three-dimensional tag file, and forcedly writing the origin coordinates, the direction matrix and the voxel space parameters into the corresponding three-dimensional tag file to realize alignment of the morphological sequence image and the physical coordinate map of the chemical component according to position information recorded by the slice set in metadata, wherein the original two-dimensional slice sequence of the morphological sequence image is grouped by using a sequence instance unique identifier; constructing a training set with a morphological image list and a component label list which are not coupled based on the three-dimensional voxel data set, and executing three-dimensional data enhancement based on space transformation; utilizing a training component recognition network of a training set enhanced by three-dimensional data based on space transformation to establish a mapping relation between three-dimensional morphological characteristics of the microplastic and chemical component distribution; And acquiring morphological sequence images of the sample to be detected, inputting the trained component identification network, and outputting a three-dimensional segmentation mask with chemical component codes by using the mapping relation to realize rapid identification of the micro plastic components.
2. The method for rapidly identifying micro-plastic components based on multi-modal data driving according to claim 1, wherein the process of constructing the training set in which the morphological image list and the component tag list are not coupled comprises: Configuring a keyword list comprising a preset characteristic character string, wherein the preset characteristic character string is used for identifying file attributes containing true value information of chemical components; traversing the data storage path, writing a file path with a file name containing the preset characteristic character string into a component tag list, and writing a file path with a file name not containing the preset characteristic character string into a morphological image list; reading the metadata attribute of a sample, layering based on the type or the form category of the micro-plastic polymer, and ensuring that the proportions of the training set and the verification set are consistent; initializing a random number generator by setting fixed random seed parameters to generate a group of determined index arrangement vectors; And performing synchronous rearrangement on the morphological image list and the component label list by using the index arrangement vector, and cutting off and dividing the training set according to a preset proportion.
3. The method for rapid identification of micro-plastic components based on multi-modal data driving of claim 1, wherein the performing the spatial transformation based three-dimensional data enhancement comprises: Constructing a four-dimensional affine transformation matrix defining a mapping relation from an original voxel coordinate system to a target enhancement coordinate system; randomly generating scaling coefficients along three axial directions, and calculating a new position of the voxel in space by using the four-dimensional affine transformation matrix; For morphological sequence image data in the training set, calculating a gray value at the new position by adopting a tri-linear interpolation algorithm; And selecting an original voxel label value closest to the geometric distance by adopting a nearest neighbor interpolation algorithm as a label value of the new position for label data corresponding to the chemical component map in the training set.
4. The method for rapid identification of micro-plastic components based on multi-modal data driving as set forth in claim 3, wherein the process of performing three-dimensional data enhancement based on spatial transformation further includes: Based on the geometric characteristics of a three-dimensional Cartesian coordinate system, calculating all arrangement sequences of three spatial axial directions and positive and negative reading states of each axial direction; combining the six axial sequences with the eight direction states to generate forty-eight unique space dimension arrangement key values; Randomly extracting one of the space dimension arrangement key values, and executing dimension transposition operation and reverse sequence operation on a preset dimension on samples in the training set; The dimension transposition operation is applied to the morphological input image volume and the chemical component label volume in synchronization with the reverse order operation.
5. The method for rapidly identifying microplastic components based on multi-modal data driving according to claim 1, wherein the process of identifying the network using the enhanced training set training components comprises: constructing a dynamic U-Net network architecture comprising an encoder, a decoder and a jump connection as the component identification network; extracting geometric features and texture features of the microplastic by using a multi-layer convolution module in the encoder, and gradually reducing the spatial resolution of the feature map by setting convolution step length; upsampling a deep feature map by using the decoder through three-dimensional transposed convolution operation, and splicing the corresponding feature map output by the encoder through the jump connection; And mapping the high-dimensional features into corresponding chemical component class probability maps by using convolution operation of an output layer, and carrying out parameter updating by adopting DiceLoss loss functions and combining an Adam optimizer.
6. The method for rapidly identifying micro-plastic components based on multi-modal data driving according to claim 5, wherein the process of using DiceLoss loss function in combination with Adam optimizer for parameter updating comprises: applying a Sigmoid activation function to the feature map of the component identification network output layer to acquire a probability value that the voxel belongs to a preset chemical component category; Calculating an intersection between a predicted volume and a real label volume, and calculating a Dice loss value based on a ratio of the intersection to a sum of the predicted volume and the real label volume; Dynamically adjusting the learning rate by using an Adam optimizer according to the first moment estimation and the second moment estimation of the parameters; And monitoring component identification accuracy on the verification set, and triggering a learning rate attenuation mechanism when the component identification accuracy is not improved within a preset period.
7. The method for rapidly identifying micro-plastic components based on multi-modal data driving according to claim 1, wherein the establishing of the mapping relationship between the three-dimensional morphological characteristics of the micro-plastic and the chemical component distribution comprises the steps of providing a hierarchical label system, wherein the specific construction steps of the label system are as follows: Defining a first level label for distinguishing a micro plastic foreground from an environmental background; defining a second-level label, and dividing the micro-plastic prospect into polyolefin and non-polyolefin according to the chemical structure similarity; Defining a third-level label, and identifying a specific polymer type by utilizing a single-heat coding technology, wherein the polymer type comprises polyethylene, polypropylene and polystyrene; and driving the component identification network to output the prediction probabilities of the first-level label, the second-level label and the third-level label at the output end respectively.
8. The method for rapidly identifying micro plastic components based on multi-modal data driving according to claim 1, wherein the specific steps of outputting the three-dimensional segmentation mask with chemical component coding are as follows: Preprocessing the morphological sequence image of the sample to be detected into a three-dimensional voxel matrix to be detected, wherein the three-dimensional voxel matrix to be detected is consistent with the standardized three-dimensional voxel data set format; inputting the preprocessed three-dimensional voxel matrix to be detected into the component recognition network to obtain a multi-channel probability feature map; Executing maximum value index operation of channel dimension on the probability feature map, and taking the channel index with the maximum probability value as a prediction category label of the voxel; And generating the three-dimensional segmentation mask with the chemical component codes, wherein the background area marks are zero, and the micro plastic area marks are integer codes corresponding to the polymer types.
9. The method for rapidly identifying micro-plastic components based on multi-modal data driving according to claim 8, further comprising generating an environmental monitoring report based on the three-dimensional segmentation mask with chemical component coding, wherein the method comprises the following specific steps: extracting physical space parameters of the three-dimensional voxel matrix to be detected in three orthogonal directions, and calculating the physical volume of a single voxel; counting the total number of voxels under each polymer category, and calculating the absolute volume and the duty ratio of each micro plastic component; Performing connected component analysis on the three-dimensional segmentation mask with the chemical component codes by utilizing twenty-six neighborhood connectivity rules, and identifying independent micro plastic particles; calculating the maximum feret diameter and the minimum feret diameter of the independent micro plastic particles, and classifying the independent micro plastic particles into fiber, chip or pellet based on an aspect ratio index; and summarizing the absolute volume and the duty ratio of each micro plastic component and the classification result of the independent micro plastic particles to generate the environment monitoring report.

Description

Micro-plastic component rapid identification method based on multi-mode data driving Technical Field The invention relates to the technical field of computer vision and environmental monitoring, in particular to a micro-plastic component rapid identification method based on multi-mode data driving. Background At present, microplastic is used as an emerging pollutant and widely exists in water, soil and biological media. The particle size of the particles is tiny, and the chemical components are various. The environmental monitoring work requires accurate identification of the polymer type and physical morphology. Standard detection procedures typically involve environmental sample collection followed by pretreatment steps such as organic digestion and density flotation. The target particles are eventually enriched onto the filter membrane for subsequent detection. The process forms a data base for pollution tracing and ecological risk assessment. Aiming at the enriched particle samples, the prior technical proposal adopts physicochemical analysis means. Microscopic fourier transform infrared spectroscopy and microscopic raman spectroscopy techniques are widely used, both of which determine polymer species by capturing molecular vibration fingerprint spectra of materials. Laser direct infrared imaging technology uses a quantum cascade laser to chemically image and scan a sample. In addition, thermal cracking gas chromatography-mass spectrometry combined techniques analyze the chemical composition of the polymer matrix by measuring the cracked product of the sample after high temperature combustion. However, the above-described technique has limitations in practical applications. The spectrum imaging device mostly adopts a point-by-point or progressive scanning mode, and the acquisition period of the full filter membrane data is longer. This time cost limits the efficiency of high throughput sample monitoring. High-precision spectrum analysis instruments generally rely on precise optical systems, and have specific requirements on spectrogram analysis capability of operators and equipment maintenance environments. Although a general optical microscope acquires an image quickly, only apparent form information such as color and shape can be recorded. Because of the lack of characteristic data of chemical dimensions, it is difficult to distinguish transparent mineral particles from microplastic by simply relying on optical images, and it is also difficult to accurately distinguish components with similar chemical properties such as polyethylene and polypropylene. Therefore, the invention provides a method for rapidly identifying micro plastic components based on multi-mode data driving, which solves the defects in the prior art. Disclosure of Invention Aiming at the defects of the prior art, the invention provides a rapid identification method of micro-plastic components based on multi-mode data driving, which solves the technical problems of long detection time consumption and high cost of spectrum analysis equipment in the existing micro-plastic detection technology, and low identification accuracy caused by the fact that a common optical microscope cannot identify chemical components. The invention provides a micro-plastic component rapid identification method based on multi-mode data driving, which adopts the following technical scheme: A micro-plastic component rapid identification method based on multi-mode data driving comprises the following steps: acquiring a morphological sequence image and a chemical component map of the microplastic, analyzing and mapping the chemical component map into a three-dimensional voxel-level component label, and constructing a standardized three-dimensional voxel data set containing morphological features and the component label through element data alignment of physical space consistency; constructing a training set with a morphological image list and a component label list which are not coupled based on the three-dimensional voxel data set, and executing three-dimensional data enhancement based on space transformation; utilizing a training component recognition network of a training set enhanced by three-dimensional data based on space transformation to establish a mapping relation between three-dimensional morphological characteristics of the microplastic and chemical component distribution; And acquiring morphological sequence images of the sample to be detected, inputting the trained component identification network, and outputting a three-dimensional segmentation mask with chemical component codes by using the mapping relation to realize rapid identification of the micro plastic components. By adopting the technical scheme, as the deep learning network is utilized to excavate and establish the potential nonlinear mapping relation between the external morphological structure and the internal chemical components of the micro-plastic, in the practical application stage, o