基于深度学习的镜下矿物图像识别方法  

Identification Method of Mineral Image Under Microscope Based on Deep Learning

作　　者：陈博勋 王宏铭 王玲 汤化明 王龙 曹冲 CHEN Boxun;WANG Hongming;WANG Ling;TANG Huaming;WANG Long;CAO Chong(College of Mining Engineering,North China University of Science and Technology,Tangshan,Hebei 063210,China;School of Resources and Safety Engineering,Central South University,Changsha,Hunan 410083,China;School of Computer Science,Beijing University of Posts and Telecommunications,Beijing 100876,China;Hebei Key Laboratory of Mining Development and Safety,North China University of Science and Technology,Tangshan,Hebei 063210,China)

机构地区：[1]华北理工大学矿业工程学院,河北唐山市063210 [2]中南大学资源与安全工程学院,湖南长沙410083 [3]北京邮电大学计算机学院,北京100876 [4]华北理工大学河北省矿业开发与安全技术重点实验室,河北唐山市063210

出　　处：《矿业研究与开发》2022年第11期163-170,共8页Mining Research and Development

基　　金：国家自然科学基金项目(42002098,52004091);河北省自然科学基金项目(D2020209017,E202209119);唐山市科技计划项目(19130216g);华北理工大学大学生创新创业训练计划项目(X2021266).

摘　　要：为探究利用深度学习和图像处理技术实现镜下矿物图像特征智能化识别的可行性,基于卷积神经网络、Adam优化算法等,对采集的辉石、石英、角闪石、橄榄石、斜长石5种矿物图像进行了试验研究,采用OpenCV对有限的数据集进行增广,有效地扩大数据规模、降低样本的不平衡性,基于ResNet-50网络架构优化模型,使用迁移学习的训练策略进行模型训练,以精度和损失作为评价指标。测试结果表明:优化后的网络模型识别精度大幅提高,在模型测试中达到了98.48%的精度,损失控制在0.01~0范围内,并且具备更快的收敛速度、更低的训练耗时,大大提升了网络的训练效率,成功实现了镜下矿物智能化识别及分析。In order to explore the feasibility of deep learning and image processing technology to realize intelligent recognition of mineral image features under microscope,based on the convolution neural network,Adam optimization algorithm and other principles,experimental research was carried out on the collected images of five minerals,including pyroxene,quartz,amphibole,olivine and plagioclase.OpenCV was used to expand the limited data set,effectively expanding the data scale and reducing the imbalance of samples.Based on ResNet-50network architecture optimization model,the training strategy of migration learning was used for model training.Accuracy and loss were used as evaluation indicators.The test results show that the identification accuracy of the optimized network model is greatly improved,reaching 98.48%in the model test,and the loss is controlled within the range of 0.01-0.Moreover,the optimized network model has a faster convergence speed and lower training time consumption,which can greatly improve the training efficiency of the network,and successfully realize the intelligent identification and analysis of minerals under microscope.

关 键 词：矿物鉴定 深度学习 图像处理 计算机视觉 卷积神经网络 

分 类 号：P575[天文地球—矿物学] TP183[天文地球—地质学]