CN-121999257-A - Labeling method and system for dense sandstone microscopic image

CN121999257ACN 121999257 ACN121999257 ACN 121999257ACN-121999257-A

Abstract

The invention discloses a labeling method and a labeling system for a dense sandstone microscopic image, wherein the labeling method and the labeling system comprise the steps of carrying out edge enhancement treatment on particles in the dense sandstone microscopic image through an image processing technology to obtain a first image, carrying out particle pre-labeling treatment on the first image through a sandstone microscopic image segmentation model to obtain a second image, wherein the second image comprises at least one editable labeling information formed according to a particle identification result, and modifying the particle identification result of the second image according to the editable labeling information. The method and the device can remarkably improve the labeling efficiency and improve the labeling fineness and accuracy.

Inventors

YU XIAOLU
ZHOU SHENGYOU
JIANG HONG
WANG YUANZHENG
WANG JIYUAN
ZHENG XUYING

Assignees

中国石油化工股份有限公司
中国石油化工股份有限公司石油勘探开发研究院

Dates

Publication Date: 20260508
Application Date: 20241101

Claims (10)

1. A method for labeling a dense sandstone microscopic image, comprising: performing edge enhancement treatment on particles in the compact sandstone microscopic image through an image processing technology to obtain a first image; Performing particle pre-labeling processing on the first image by using a sandstone microscopic image segmentation model to obtain a second image, wherein the second image contains at least one piece of editable information formed according to a particle identification result; And modifying the particle identification result of the second image according to the editable information.
2. The labeling method of claim 1, wherein the labeling method comprises the steps of, in the step of performing edge enhancement processing on particles in a dense sandstone microscopic image by an image processing technique, obtaining a first image includes: Cutting the compact sandstone microscopic image into a plurality of sub-images containing mineral particles in a preset number range according to a preset unit image size; And sequentially carrying out smoothing treatment, image equalization treatment, contrast and brightness collocation adjustment and image sharpening treatment on each sub-image to obtain each sub-image subjected to particle edge enhancement treatment, and marking the sub-image as the first image.
3. The labeling method of claim 2, wherein the labeling method comprises the steps of, And removing Gaussian noise and spiced salt noise in the sub-images by adopting a Gaussian filter and a median filter, thereby finishing the smoothing processing of each sub-image.
4. A method according to claim 2 or 3, wherein the step of cropping the dense sandstone microimage into a plurality of sub-images containing a predetermined number of ranges of mineral particles, according to a predetermined unit image size, comprises: Cutting the compact sandstone microscopic image into n multiplied by n square sub-images, wherein the preset number range is 50-100; Expanding the short side of each sub-image to enable the size of the expanded sub-image to reach a first image size, wherein the first image size is the size of an input image receivable by the sandstone microscopic image segmentation model; and carrying out black area identification on each sub-image subjected to short side expansion, and deleting invalid sub-images which do not meet preset conditions according to a black area identification result.
5. The labeling method of claim 4, wherein the preset condition is that a proportion of the area of the black region of the current sub-image to the area of the sub-image exceeds a preset proportion threshold, and the preset proportion threshold is 50%.
6. The labeling method of any of claims 2-5, further comprising constructing a modified image segmentation generic base model from a sandstone microimage segmentation dataset to obtain the sandstone microimage segmentation model, wherein, The improved image segmentation universal basic model comprises an image decoder, a prompt decoder and a mask decoder with a sandstone microscopic image characteristic self-adaptive identification module, wherein the mask decoder comprises two decoder layers, the self-attention layer in each decoder layer and the rear end of the cross-attention layer from the image mark to the prompt mark are respectively added with the sandstone microscopic image characteristic self-adaptive identification module, The sandstone microscopic image characteristic self-adaptive identification module comprises two full-connection layers and a Gaussian error linear unit positioned between the two full-connection layers.
7. The labeling method of claim 6, wherein the step of constructing the improved image segmentation generic base model comprises: according to the SA-1B data set, diceLoss is adopted as a loss function, and training and verifying are carried out on the original image segmentation universal basic model; constructing the improved image segmentation universal basic model based on the trained image segmentation universal basic model; According to the sandstone microscopic image segmentation dataset, a Log-Cosh DiceLoss is adopted as a loss function, an improved image segmentation universal basic model is trained and verified, and the sandstone microscopic image segmentation model is obtained, so that the sandstone microscopic image segmentation model can generate a corresponding binary mask map with editable grain boundary labeling information for each grain identification area, and the size of the binary mask map is the same as that of the sub-image.
8. The labeling method according to claim 6 or 7, wherein in the step of performing particle pre-labeling on the first image by using a sandstone microscopic image segmentation model to obtain a second image, the method comprises: carrying out particle pre-labeling processing on each first image by using a sandstone microscopic image segmentation model, and generating a corresponding binarization mask map for each particle identification area; All the binarized mask maps are converted to particle contours and aggregated into a second image containing all the particle contour features.
9. The labeling method according to any one of claims 1-8, characterized in that in the step of modifying the particle recognition result of the second image in accordance with the editable labeling information, it comprises: and identifying particles to be adjusted in the image and the corresponding positions and ranges thereof by using the binarization mask map and the particle boundary marking information corresponding to each identified particle region, and modifying the boundary marking result of the particles to be adjusted by polygon marking based on the particles to be adjusted to obtain a refined particle distribution characteristic map.
10. An annotation system for dense sandstone microscopic images, wherein the annotation system is used for implementing the annotation method according to any one of claims 1 to 9, and wherein the annotation system comprises: an image preprocessing module configured to perform edge enhancement processing on particles in the dense sandstone microscopic image by an image processing technique to obtain a first image; The image pre-labeling module is configured to perform particle pre-labeling processing on the first image by using a sandstone microscopic image segmentation model to obtain a second image, wherein the second image contains at least one editable labeling information formed according to a particle identification result; And the refinement modification module is configured to modify the particle identification result of the second image according to the editable labeling information.

Description

Labeling method and system for dense sandstone microscopic image Technical Field The invention relates to the technical field of image processing, in particular to a labeling method and a labeling system for a compact sandstone microscopic image. Background The rock flake microscopic image is an image obtained by a rock appraiser using a professional image acquisition device by grinding a rock sample into flakes, placing the flakes on a polarized microscope stage. In the field of geology, rock slice identification is the classification and naming of rock slice images by observing and analyzing the composition and content of minerals in them. Rock laminate identification is of great importance in the fields of geology and petrology, mainly including determining rock types and compositions, studying the structure and deformation history of the rock, analyzing the physical properties of the rock, assisting mineral and mineral deposit exploration, etc. However, for dense sandstones, microscopic images have the problems of numerous mineral particles, high boundary overlap, and low foreground-background contrast. Therefore, the image identification work is very challenging, and the professional level requirements of the identification workers are high. Meanwhile, the subjectivity of the identification conclusion is strong, the conclusions of different identification experts often have differences, the identification efficiency is low, and a large amount of manpower and material resources are wasted. Under the background of rapid development of artificial intelligence technology, intelligent analysis methods for researching compact sandstone microscopic images by using a computer are gradually emerging. In this case, deep learning is a data-driven method, and when the model is sufficiently trained with sufficient training data, an effect exceeding that of the conventional method is often obtained. Thus, the amount and quality of the training data plays a crucial role in the final effect that the model presents. Labeling is the process of converting original image data into trainable data for a neural network model in deep learning. The traditional data marking method mainly uses marking software to manually mark the data such as texts, pictures and the like, has low marking efficiency, wastes a great deal of time and manpower, and does not meet the requirement of industrial big data on high production efficiency. At present, labeling of compact sandstone microscopic images mainly depends on manual labeling of related identification experts, namely, the types and boundaries of all particles are accurately judged on microscopic images of mineral particles and impurities through labeling software, then the outlines of the particles are drawn through manual dotting, and finally different particles are labeled correspondingly. Taking the segmentation task involved in dense sandstone microscopic image analysis as an example, the labeling form of mineral particles is usually a polygonal frame. When complex contours are involved, the number of points required to delineate the grain boundaries will increase significantly and the effort increases dramatically in order to obtain a label that fits precisely to the grain edges. The software for labeling can be classified into two types according to the degree of automation, wherein the first type is software without automation, namely, the software needs to be completely and manually labeled, and the second type is labeling software with a certain degree of automation. Labeling software typically provides rich forms of labeling, including rectangular boxes, polygonal boxes, keypoint labels, semantic categories, and the like. Since the first type of labeling software does not have automation, labeling personnel need to process objects to be labeled one by one during labeling, and such software includes, but is not limited to LabelIMG, labelme, VIA, and the like. The labeling software has the following defects that mineral particles in the dense sandstone microscopic panoramic image are often large in quantity and complex in contact relationship when being applied to the dense sandstone microscopic panoramic image, the manual labeling of the dense sandstone microscopic panoramic image for finishing a fine labeling consumes great time and energy, and the quality of the labeling contour fitting effect is obviously reduced along with the increase of the fatigue degree of labeling personnel. In addition, the accuracy of labeling results is also limited by the expertise of the labeling personnel. The second type of labeling software aims at improving labeling efficiency, and can automatically label images to a certain extent, and the software comprises, but is not limited to, label studio, EISeg, anylabeling and the like. The labeling forms provided by the method are single, are mostly limited to rectangular frames with regular shapes, or are obvious target images with obvious charac