摘要
通过对矿物扫描电镜图像进行分类与鉴定,能够获取矿物的微观信息,确定矿物的组成与类别,对于油气田生、储、盖类型的研究具有重要的意义.由于在一幅图像之中有时不止有一种矿物,且不同矿物之间具有相关性或共生性的特性,而普通的神经网络只提取图像特征或只注意图像局部的特征关系,忽略了矿物之间的相关性.因此如何利用标签之间的关系进行更优秀的多标签图像分类成为扫描电镜图像分类的重要任务.鉴于上述情况,通过构建基于Resnet50的图像特征学习模块与基于图卷积神经网络的分类器模块构成的引入图卷积的卷积神经网络模型可以很好地完成上述任务.使用卷积神经网络模块提取图像特征,并利用GCN模块学习矿物标签之间的相关性,达到提高分类准确率的目的.此模型相比普通的CNN模型准确率提高了5%,相比引入注意力机制的CNN模型,此模型的准确率仍有3%的优势.实验表明,CNN与GCN相结合的分类模型在扫描电镜数据集分类任务中优于其他的分类模型.
Based on the classification and identification of mineral SEM images,the microscopic information of minerals can be obtained,and the composition and classification of minerals can be determined,which is of great significance for the study of source,reservoir and cap types of oil and gas fields.Sometimes there is more than one mineral in an image,and the different minerals have the characteristics of correlation or symbiosis,or-dinary neural networks only extract image features or pay attention to the local feature relationship of the images,which ignore the correlation among the minerals.Therefore,how to use the relationship between tags to better classify multi tag images becomes an important task of SEM image classification.In view of the above situation,the above task can be well completed by building a convolutional neural network model with graph convolution,which is composed of an image feature learning module based on Resnet50 and a classifier module based on graph convolution neural network.The convolution neural network module was used to extract image features,and the GCN module was used to learn the correlation between mineral tags,which can improve the classifica-tion accuracy.Compared with the ordinary CNN model,the accuracy of this model was improved by 5%.Com-pared with the CNN model with attention mechanism,the accuracy of this model still has an advantage of 3%.The experiment showed that the classification model combining CNN and GCN is superior to other classification models in the classification task of SEM dataset.
作者
杜睿山
王栋林
孟令东
张桐
Du Ruishan;Wang Donglin;Meng Lingdong;Zhang Tong(School of Computer and Information Technology,Northeast Petroleum University,Daqing 163318;Key Laboratory of Oil and Gas Reservoir and Underground Gas Storage Integrity Evaluation,Ministry of Heilongjiang Province,Daqing 163318;Research Institute of Exploration and Development of Daqing Oilfield Company Ltd,Daqing 163712)
出处
《海南大学学报（自然科学版）》
CAS
2023年第4期352-358,共7页
Natural Science Journal of Hainan University
基金
国家自然科学基金(41702156)
东北石油大学引导性创新基金(2020YDL-04)。
