CN-122023448-A - Coal rock nano-pore multicomponent image segmentation method

CN122023448ACN 122023448 ACN122023448 ACN 122023448ACN-122023448-A

Abstract

The application discloses a coal rock nanopore multicomponent image segmentation method, and relates to the field of image data processing. The method comprises the steps of obtaining an SEM image of coal rock, carrying out characteristic enhancement pretreatment on adsorbed gas on the SEM image, carrying out coal rock nano-pore multicomponent identification prediction on the treated SEM image by adopting an improved U-Net model to obtain a preliminary prediction result, optimizing the preliminary prediction result by adopting a full-connection conditional random field to correct a region boundary, and removing a noise boundary by adopting morphological post-treatment to obtain segmentation map information. The application can efficiently and accurately realize the image segmentation of coal rock nano pore multicomponent.

Inventors

TANG JIZHOU
MENG YUFAN
ZHANG FENGYUAN
ZHU WEILIN
LI JUNLUN
CHEN WEIHUA
LI YUWEI
SUN ZHE
ZHANG FEIFAN
LIU BOWEN

Assignees

同济大学

Dates

Publication Date: 20260512
Application Date: 20260413

Claims (6)

1. The coal rock nano-pore multicomponent image segmentation method is characterized by comprising the following steps of: Obtaining an SEM image of coal rock, wherein the SEM image is an image with nanoscale resolution; performing characteristic enhancement pretreatment on the adsorption gas on the SEM image to obtain a treated SEM image; The method comprises the steps of performing coal rock nanopore multicomponent identification prediction on a processed SEM image by adopting an improved U-Net model to obtain a preliminary prediction result, wherein the coal rock nanopore multicomponent comprises adsorption gas, a water film, a coal matrix and a pore structure, the improved U-Net model is obtained by training an improved U-Net network structure by adopting an Adam optimizer according to a dataset, the dataset comprises processed SEM images subjected to label screening and expanding, the improved U-Net network structure is obtained by introducing a residual convolution block on the basis of a U-Net framework to ensure that the adsorption gas can be reversely propagated, replacing a ReLU function with a Mish activation function and giving a preset weight to the adsorption gas and the type of the water film by adopting a Focal Loss function; And optimizing the preliminary prediction result by adopting a full-connection conditional random field to correct the region boundary, and removing the noise boundary by adopting morphological post-processing to obtain the segmentation map information.
2. The coal rock nanopore multicomponent image segmentation method according to claim 1, wherein the SEM image is subjected to characteristic enhancement pretreatment on adsorbed gas, and the processed SEM image is obtained, specifically comprising: The SEM image is subjected to Gaussian filtering and median filtering combination noise reduction, and the contrast of the image is enhanced by adopting self-adaptive histogram equalization so as to optimize the gray level difference of a water film and adsorbed gas, so that the enhanced SEM image is obtained; And carrying out blocking treatment on the enhanced SEM image through an overlapped blocking strategy to ensure that the pore structure and the adsorption gas of the wall surface are reserved in a single sub-image, carrying out grid sequence numbering according to a set mode based on a blocking sequence, and recording the original coordinates of each sub-image to splice to obtain the treated SEM image.
3. The coal rock nanopore multicomponent image segmentation method according to claim 1, wherein the improved determination method of the U-Net model specifically comprises: Marking each subgraph in the historical processed SEM image based on gray values by adopting Labelme open source marking tools, and comparing label graphs based on a Dice coefficient to ensure that marked labels correspond to pixels of the historical processed SEM image one by one, and carrying out sample screening and data expansion processing; Constructing an improved U-Net network structure; Dividing the data set into a training set, a verification set and a test set according to a set proportion; Inputting the training set into an improved U-Net network structure, and training model parameters by adopting an Adam optimizer to obtain a trained U-Net network structure, wherein the model parameters comprise convolution kernel weight and BN layer parameters; Inputting the verification set into a trained U-Net network structure, and verifying by adopting a five-fold cross verification method to obtain a verified U-Net network structure; testing the verified U-Net network structure by adopting the test set to obtain a tested U-Net network structure; and determining the tested U-Net network structure as an improved U-Net model.
4. The coal rock nanopore multicomponent image segmentation method according to claim 1, wherein mathematical expressions corresponding to Mish activation functions are: ; Wherein, the Activating a function for Mish; Is the input eigenvalue.
5. The coal rock nanopore multicomponent image segmentation method according to claim 1, wherein the mathematical expression corresponding to the Focal Loss function is: ; Wherein, the Is a Focal Loss function; Is a predictive probability; Is a category weight coefficient; Is a modulation factor.
6. The coal rock nanopore multicomponent image segmentation method according to claim 1, wherein the preliminary prediction result is optimized by using a fully connected conditional random field to correct a region boundary, and noise boundaries are removed by using morphological post-processing to obtain segmentation map information, and the method specifically comprises the following steps: fusing the preliminary prediction result with gray information of the SEM image by adopting a fully connected conditional random field, and carrying out optimization correction on pixel points which do not accord with the distribution rule of the adsorption gas in the preliminary prediction result based on space priori knowledge to obtain an optimized mask map, wherein a Gibbs energy function is adopted for optimization correction, and the expression of the Gibbs energy function is as follows: ; As a gibbs energy function; As a unitary potential function; is the first Class labels corresponding to the pixel points; As a pair potential function; is the first Class labels corresponding to the pixel points; and cutting the optimized mask map to remove the overlapped area, and removing the noise boundary by adopting morphological post-processing to obtain the segmentation map information.

Description

Coal rock nano-pore multicomponent image segmentation method Technical Field The application relates to the field of image data processing, in particular to a coal rock nano-pore multicomponent image segmentation method. Background Along with the advancement of coal bed gas exploitation to deep reservoirs, the accurate quantification of gas-water occurrence states (such as adsorption gas distribution, water film thickness, pore structure characteristics and the like) in coal bed nano pores under in-situ conditions becomes a key scientific problem for revealing the coal bed gas migration law and optimizing the exploitation scheme. The scale effect (usually less than 100 nm) of the coal rock nano-pore and the microscopic occurrence characteristic of the gas-water component make the visualization and quantification research thereof depend on a high-precision microscopic imaging technology. And scanning electron microscopes (Scanning Electron Microscope, SEM) are an important technical support in this field by virtue of the nanoscale resolution imaging capability. However, in the SEM imaging process, because of the coverage characteristics of the molecular layer, the SEM imaging signals are extremely weak, the signal to noise ratio is extremely low, and serious nonlinear overlapping and confusion exist between the gray values, the nanopores and the water film, and noise interference inevitably exists, so that the accurate multi-component distinction faces serious challenges. The traditional coal-rock image segmentation method mainly relies on manual marking or simple threshold segmentation, and has the defects that firstly, the manual segmentation efficiency is extremely low, single image marking usually needs a few hours and is difficult to meet the analysis requirement of a large-scale data set, secondly, subjectivity is strong, different operators have obvious differences in judgment of fuzzy boundaries, so that the consistency of quantification results is poor, thirdly, weak contrast components under the nanoscale cannot be effectively distinguished, the segmentation precision is difficult to guarantee, wherein the gray value of adsorbed gas is especially low and is about 30 to 50, a form attached to the surface of a hole wall appears, the gray difference between the adsorbed gas and a surrounding water film and a matrix is weak, and the boundary is fuzzy, so that the traditional segmentation method is always a problem that the traditional segmentation method is difficult to overcome. In the existing image segmentation technology, the traditional algorithms such as threshold segmentation and region growth are poor in adaptability to complex coal-rock microscopic images, the slight characteristic difference of nanopore multicomponent is difficult to deal with, and the image segmentation method based on deep learning is partially developed in macroscopic coal-rock component extraction, but has limitations on a nanopore scene, on one hand, the traditional convolution network is easy to lose weak boundary information in the characteristic extraction process, so that the segmentation of low-contrast components such as water films and adsorbed gases is incomplete, and on the other hand, the unbalance of sample types (such as low nanopore occupancy rate and water film region fragmentation) can reduce the recognition precision of a model to few types of components, and cannot meet the requirement of multicomponent accurate segmentation. Therefore, aiming at the problem that the multicomponent (coal matrix, nano pore, water film and adsorption gas) in the coal rock nano pore SEM image is difficult to divide efficiently and accurately, an automatic and high-precision image segmentation technology is developed, and the method has important practical significance for pushing quantitative research on the occurrence state of deep coal reservoir gas-water. Disclosure of Invention The application aims to provide a coal rock nanopore multicomponent image segmentation method which can efficiently and accurately realize coal rock nanopore multicomponent image segmentation. In order to achieve the above object, the present application provides the following solutions: the application provides a coal rock nano-pore multicomponent image segmentation method, which comprises the following steps: Obtaining an SEM image of coal rock, wherein the SEM image is an image with nanoscale resolution; performing characteristic enhancement pretreatment on the adsorption gas on the SEM image to obtain a treated SEM image; The method comprises the steps of performing coal rock nanopore multicomponent identification prediction on a processed SEM image by adopting an improved U-Net model to obtain a preliminary prediction result, wherein the coal rock nanopore multicomponent comprises adsorption gas, a water film, a coal matrix and a pore structure, the improved U-Net model is obtained by training an improved U-Net network structure by adopting