基于地质图像大数据的岩性识别研究  

RECOGNITION OF LITHOLOGY WITH BIG DATA OF GEOLOGICAL IMAGES

作　　者：胡启成 叶为民[1,2] 王琼[1] 陈永贵[1] HU Qicheng;YE Weimin;WANG Qiong;CHEN Yonggui(Key Laboratory of Geotechnical and Underground Engineering of the Ministry of Education,Tongji University,Shanghai 200092,China;United Research Center for Urban Environment and Sustainable Development,the Ministry of Education,Shanghai 200092,China)

机构地区：[1]同济大学岩土及地下工程教育部重点实验室,中国上海200092 [2]教育部城市环境与可持续发展联合研究中心,中国上海200092

出　　处：《工程地质学报》2020年第6期1433-1440,共8页Journal of Engineering Geology

基　　金：国家重点研发计划(资助号:2019YFC1509900).

摘　　要：近些年来,随着大数据、深度学习等技术的飞速发展,大数据的开发与利用为众多行业带来了显著经济与社会效益。借助大数据手段,开展地质文本、图像和序列数据挖掘与应用研究,具有极其重要的理论与社会意义。本文在归纳总结前人工作的基础上,重点针对地质大数据中的图像数据,基于深度学习理论,构建网络学习模型,通过基于网络搜索的数据采集、数据预处理、网络搭建、网络训练及结果/评价等步骤,实现基于地质图像的大数据岩性识别。结果表明,图像识别岩性的测试准确率约为90%;有限的图像数据数,可能是产生识别误差的一个原因;机器对岩石图片所呈现的某些特征相似性,如宏观的形状、颜色等,也会给出正相关评分,从而产生误判。理论上,采用BCNN(Bilinear Convolutional Neural Network)等能够捕捉更精细细节的网络模型,解决计算机视觉中的细粒度识别问题,从而从根本上提升图像识别效率,应该是今后一个研究方向。In recent years,with the rapid development of big data,deep learning and other technologies,the development and utilization of big data have brought significant economic and social benefits to various industries.It is of great theoretical and social significances to carry out researches on data mining and application of geological texts,images and sequence data by means of big data technologies.This paper focuses on the geological image data processing,development of network models using the deep learning theory and lithology recognition through performing network-based data acquisition,data preprocessing,network construction,network training and result/evaluations.Results show that the recognition accuracy of lithology images is about 90%.Limited image data employed can be one of the reasons for deviations of recognitions.Positive correlation scores given by machine for some similar characteristics of the rock images,e.g.,macroscopic shape and color,etc.,also can lead to misjudgments in recognition.Theoretically,adoption of networks like BCNN(Bilinear Convolutional Neural Network)for capturing finer details and solving the problem of fine-grained recognition in computer vision and fundamentally improving the efficiency of image recognition,should be considered in future works.

关 键 词：地质大数据 数据挖掘 深度学习 网络训练 岩性识别 

分 类 号：P588[天文地球—岩石学]