基于多尺度和上下文信息的高分辨网络岩心图像分割  

High-resolution Network Core Image Segmentation Based on Multi-scale and Context Information

作　　者：杨博 潘少伟[1] 李宗强 薛章涛 秦国伟 YANG Bo;PAN Shao-wei;LI Zong-qiang;XUE Zhang-tao;QIN Guo-wei(School of Computer Science,Xi'an Shiyou University,Xi'an 710065,China;The Second Oil Production Plant of Changqing Oilfield Company of PetroChina,Qingyang 745100,China;School of Petroleum Engineering,Xi'an Shiyou University,Xi'an 710065,China)

机构地区：[1]西安石油大学计算机学院,西安710065 [2]中国石油长庆油田分公司第二采油厂,庆阳745100 [3]西安石油大学石油工程学院,西安710065

出　　处：《科学技术与工程》2024年第25期10659-10666,共8页Science Technology and Engineering

基　　金：国家自然科学基金(52174027)。

摘　　要：岩心图像是油气田开发过程中最重要的地质资料之一,它对了解地下的地质状况、确定岩心岩性以及推断沉积环境具有重要意义。针对现有方法在岩心图像分割中存在的计算资源消耗过大、模型容易过拟合、图像细节信息丢失以及像素点冗余等问题,提出基于多尺度和上下文信息的高分辨率网络岩心图像分割方法。由于金字塔池化模块并行叠加使得各分支之间特征信息无法共享,将金字塔池化模块改进为跳跃连接方式,使得不同扩张率的空洞卷积之间产生一定的联系,形成一个更密集的网络结构。然后将改进后的增强空洞空间金字塔池化(enhanced atrous spatial pyramid pooling,EASPP)模块引入高分辨率网络(high-resolution net,HRNet)的每个子网中,解决了HRNet方法中图像像素点冗余等问题,并且可以扩大感受野和捕获多尺度上下文信息。在岩心图像数据集上进行了实验对比,实验结果表明,相比于目前主流的语义分割方法,该方法在平均交并比(mean intersection over union,MIoU)、像素精度(pixel accuracy,PA)评价指标上都有不同程度的提升,而且对过分割或者少分割处理效果更好,最大程度上保留了图像的细节信息。Core image is one of the most important geological data in the process of oil and gas field development,it is of great significance to understand the underground geological condition,determine the core lithology and infer the sedimentary environment.Aiming at the problems such as excessive consumption of computing resources,easy overfitting of models,image details are lost and pixel redundancy,a high resolution network core image segmentation based on multi-scale and context information was proposed.Due to the parallel superposition of the pyramidal module network,the feature information cannot be shared among branches.The pyramidal module was improved into a network model with skip connection structure,which makes certain connections between the cavity convolution with different expansion rates and forms a denser network structure.Then the improved enhanced atrous spatial pyramid pooling(EASPP)module was introduced into each subnet of the high-resolution net(HRNet),which solves the problem of image pixel redundancy in HRNet method,and can enlarge the receptive field and capture multi-scale context information.The experimental results show that compared with the current mainstream semantic segmentation algorithms,this method has different degrees of improvement in mean intersection over union(MIoU)and pixel accuracy(PA)evaluation indicators,and it is better to deal with over-segmentation or under-segmentation,and retains image details to the greatest extent.

关 键 词：语义分割 岩心图像 深度学习 卷积神经网络 

分 类 号：P631[天文地球—地质矿产勘探]