基于改进ResNet深度残差网络的岩屑岩性识别研究  

Lithology identification of cuttings based on improved ResNet deep residual network

作　　者：杨磊[1] 熊昶 刘文超[1] 王彦秋[1] 侯果[1] YANG Lei;XIONG Chang;LIU Wenchao;WANG Yanqiu;HOU Guo(Tarim Oilfield Company,PetroChina,Korla 841000,Xinjiang)

机构地区：[1]中国石油塔里木油田分公司,新疆库尔勒841000

出　　处：《长江大学学报（自然科学版）》2023年第2期11-19,共9页Journal of Yangtze University(Natural Science Edition)

基　　金：2020年新疆维吾尔自治区创新人才建设专项自然科学计划(自然科学基金)项目“水平井反演优化解释方法的研究与实现”(2020D01A132)。

摘　　要：岩屑岩性识别在油气探测工作中具有重要地位,传统人工识别岩屑受主观因素影响大,操作繁琐。现研究一种采用深度学习方法对高分辨率岩屑图像进行分类,以达到识别岩屑岩性的目的。通过岩屑数据集在主流深度学习网络进行训练后对比研究,提出一种针对沉积岩岩屑的分类网络,该网络采用ResNet的深度残差思想,以残差学习单元进行堆叠,通过恒等映射克服随着网络深度加深导致的网络退化问题,同时加入深度可分离卷积操作,减少参数量以轻量化网络。通过RMSprop优化器最小化损失函数,以Softmax函数作为分类器进行分类。实验证明,网络在拥有更轻量化结构的同时,在岩屑数据集上拥有更优的准确率、精确率、召回率和F 1值。在以5类共计221张高分辨率沉积岩岩屑图像作为实验样本的数据集上,该网络准确率较传统ResNet网络提高1.73个百分点,在与VGG16、DenseNet169、MobileNet、传统ResNet网络的对比中也拥有最高的准确率,达到96.54%。对已知岩屑图像具有良好的分类能力,可为未知岩屑岩性的定义与地层结构描绘提供重要依据。The lithology identification of cuttings plays an important role in oil and gas exploration.The traditional manual identification of cuttings is greatly affected by subjective factors,and the operation is cumbersome.In this paper,a deep learning method was used to classify high resolution cuttings images to achieve the purpose of identifying the lithology of cuttings.A classification network for sedimentary rock debris was proposed by comparing the debris data set after training in the mainstream deep learning network.The network adopts the deep residual idea of ResNet,stacks with the residual learning unit,and overcomes the network degradation problem caused by the deepening of the network depth through the identity mapping.At the same time,the depth separable convolution operation is added to reduce the number of parameters to lightweight the network.The loss function is minimized by RMSprop optimizer,and Softmax function is used as classifier for classification.Experiments show that the network in this paper has better accuracy,precision,recall and F 1 values on the rock debris data set while having a lighter structure.The accuracy of the proposed network is 1.73 percentage points higher than that of the traditional ResNet network on the dataset of 5 types of 221 high resolution sedimentary cuttings images as experimental samples.Compared with VGG16,DenseNet169,MobileNet and traditional ResNet network,the proposed network also has the highest accuracy,reaching 96.54%.It has good classification ability for known debris images,which can provide an important basis for the definition of unknown debris lithology and the description of stratigraphic structure.

关 键 词：深度残差网络 岩性识别 岩屑图像 深度可分离卷积 图像分类 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]