基于岩石薄片图像与改进EfficientNet建模的岩性识别方法  

A Lithology Recognition Method Based on Rock Thin Section Images and Improved EfficientNet Modeling

作　　者：程国建[1] 李宗祥 李秋实[2] 韩江 孙亚招 CHENG Guojian;LI Zongxiang;LI Qiushi;HAN Jiang;SUN Yazhao(School of Computer Science,Xi’an Shiyou University,Xi’an,Shaanxi 710065,China;PetroChina Changqing Oilfield Company,Xi’an,Shaanxi 710018,China;No.8 Oil Production Plant,PetroChina Changqing Oilfield Company,Xi’an,Shaanxi 710016,China)

机构地区：[1]西安石油大学计算机学院,陕西西安710065 [2]中国石油长庆油田公司,陕西西安710018 [3]长庆油田分公司第八采油厂,陕西西安710016

出　　处：《西安石油大学学报(自然科学版)》2025年第2期124-134,共11页Journal of Xi’an Shiyou University(Natural Science Edition)

基　　金：国家自然科学基金项目(40872087);陕西省自然科学基金(2023-JC-YB-601);西安市科技计划高校院所人才服务企业项目(23GXFW0077);陕西省“两链”融合重点专项—两链融合秦创原总窗口产业集群项目(2022QCY-LL-82)。

摘　　要：为了实现岩石薄片岩性高效、准确的分类与识别,提出了一种基于空间注意力与多尺度融合的岩石薄片岩性识别方法。采用多尺度融合的策略,通过多个EfficientNet中的轻量反转瓶颈卷积核(MBConv)对橄榄石、普通辉石、角闪石、黑云母等多种类别的岩石薄片图像进行特征提取,以捕获更多的细节信息。引入空间注意力模块(SGE),融合岩石薄片图像中的空间特征信息。此外,采用Ranger优化器,改善模型的性能及收敛速度。实验表明:提出的MFSRE(Multi-Scale Fusion-SGE-Ranger-EfficientNet)模型在测试集上的召回率、F1分数分别为98.25%、98.29%,具有较高的识别准确率,相较于ShuffleNet、RegNet、MobileNetV2网络具有更好的分类效果。A lithology recognition method based on spatial attention and multi-scale fusion is proposed in order to achieve efficient and accurate classification and lithology recognition of rock thin sections.Employing a multi-scale fusion strategy,the features of various categories of rock thin section images,such as olivine,pyroxene,hornblende,and biotite,are extracted through light weight inverted bottleneck convolution kernels(Mobile Inverted Bottleneck Convolution,MBConv)in multiple EfficientNet models to capture more detailed information.The spatial feature information in rock thin section images is incorporated by introduction of the Spatial Group-wise Enhance(SGE)module.Additionally,the model's performance and convergence speed is improved by using the Ranger optimizer.Experimental results demonstrate that the proposed MFSRE(Multi-Scale Fusion-SGE-Ranger-EfficientNet)model achieves a recall rate of 98.25%and an F1 score of 98.29%on the test set,showing high recognition accuracy of the model.Compared to the networks such as ShuffleNet,RegNet,and MobileNetV2,the MFSRE model exhibits superior classification performance.

关 键 词：岩性识别 岩石薄片图像 EfficientNet 空间注意力机制 

分 类 号：P618.13[天文地球—矿床学] TE122.322[天文地球—地质学]