引用本文:
【打印本页】   【HTML】   【下载PDF全文】   查看/发表评论  【EndNote】   【RefMan】   【BibTex】
←前一篇|后一篇→ 过刊浏览    高级检索
本文已被:浏览次   下载 本文二维码信息
码上扫一扫!
分享到: 微信 更多
基于Unet++网络的数字岩心图像分割泛化能力
赵久玉1,2,蔡建超1,2
(1.油气资源与工程全国重点实验室,中国石油大学(北京),北京 102249;2.中国石油大学(北京)地球科学学院,北京 102249)
摘要:
图像分割是数字岩心技术的重要组成部分,深度学习为数字岩心图像分割提供了新方法。在优选的深度学习模型的基础上确定网络结构、训练数据量来平衡计算效率,进一步在不同类型的岩心数据集上讨论网络的泛化能力及其影响因素。结果表明:Unet、Segnet和Unet++网络中,Unet++网络可以在保证分割精度的同时具有最好的物性参数预测效果;Unet++网络在训练数据量和预测数据量为1∶1,网络结构设计2次采样的条件下,Unet++网络的分割精度可以达到98%;基于多类岩心训练的Unet++网络分割不同岩心图像的平均分割精度达95%,相较于岩心的类型,岩心图像的质量更能影响Unet++网络的识别效果。
关键词:  数字岩心  图像分割  深度学习  Unet++  泛化能力
DOI:10.3969/j.issn.1673-5005.2024.02.013
分类号:
文献标识码:A
基金项目:
ZHAO Jiuyu1,2, CAI Jianchao1,2
(1.National Key Laboratory of Petroleum Resources and Engineering, China University of Petroleum (Beijing), Beijing 102249, China;2.College of Geosciences, China University of Petroleum (Beijing), Beijing 102249, China)
Abstract:
Image segmentation is an important part of the digital rock technology, and development of deep learning provides a new method for digital rock image segmentation. In this study, the network structure and the amount of training data were determined based on optimized deep learning networks to balance the computational efficiency, and the generalization ability of the network and its influencing factors on different types of rock datasets were discussed. The results show that, among the Unet, Segnet and Unet++ networks, the Unet++ network is the best for the prediction of physical parameters while ensuring the segmentation accuracy. The segmentation accuracy of the Unet++ network can reach 98% under the condition that the amount ratio of the training data and the predicted data is 1∶1 and the network has two-time samplings. The average segmentation accuracy of different rock images segmented by the trained Unet++ network based on multi-type rocks can reach 95%. Compared with the rock type, the quality of the rock image is more important on the segmentation results of the Unet++ network.
Key words:  digital rock  image segmentation  deep learning  Unet++  generalization ability
版权所有 中国石油大学学报(自然科学版)编辑部 Copyright©2008 All Rights Reserved
主管单位:中华人民共和国教育部 主办单位:中国石油大学(华东)
地址: 青岛市黄岛区长江西路66号中国石油大学期刊社 邮编:266580 电话:0532-86983553 E-mail: journal@upc.edu.cn
本系统由:北京勤云科技发展有限公司设计