融合图像处理与深度学习的亮晶颗粒灰岩岩相学分析应用

Application of sparry grain limestone petrographic analysis combining image processing and deep learning

余晓露 李龙龙 蒋宏 卢龙飞 杜崇娇

YU Xiaolu LI Longlong JIANG Hong LU Longfei DU Chongjiao

中国石化 油气成藏重点实验室，江苏 无锡 214126 中国石化 石油勘探开发研究院 无锡石油地质研究所，江苏 无锡 214126

SINOPEC Key Laboratory of Petroleum Accumulation Mechanisms, Wuxi, Jiangsu 214126, China Wuxi Research Institute of Petroleum Geology, Wuxi, Jiangsu 214126, China

针对传统碳酸盐岩薄片鉴定基于肉眼观察和描述，存在主观性强、定性评价为主、定量困难等问题，以亮晶颗粒灰岩为对象，设计了涵盖流程与技术的智能化岩石薄片图像信息挖掘模型。通过岩相学分析框架构建了岩相特征与薄片图像之间的映射关系。融合图像处理和深度学习设计了全流程的特征提取算法。通过卷积神经网络获得结构组分特征中对颗粒类型的定性识别，即基于改进的ResNet50模型划分颗粒所属类别(内碎屑、生物碎屑、包粒、球粒和团块)。通过数字图像处理技术获得结构组分特征中对颗粒含量、粒径、形状、接触方式的定量识别，即基于阈值分割计算颗粒含量，基于最小外接圆/最小外接矩形，结合面积比/长宽比、交并比(IoU)等算法计算颗粒形态学参数，并通过对染色图像的HSV色彩空间处理获得矿物组分特征中对方解石和其他矿物组分的定性和定量识别。以顺X井亮晶颗粒灰岩薄片样品为例，通过完整的图像识别过程验证了各个特征点提取算法的有效性，并与人工鉴定报告进行对比。岩相学分析框架能够有效地表征亮晶颗粒灰岩中的有意义信息。通过岩相学分析框架结合图像分析算法的模式，实现了对这一类碳酸盐岩的规范化流程化智能鉴定，为岩石薄片图像智能识别研究提供有效的方法支撑。

Traditional carbonate rock slice identification is based on manual observation and description, which is highly subjective, mainly qualitative and difficult to quantify. In this paper, an intelligent rock thin-section image information mining model covering the process and technology is designed. The mapping relationship between lithofacies characteristics and thin slice images is constructed through a petrographic analysis framework. A full-flow feature extraction algorithm is also designed by combing image processing and deep learning. Qualitative recognition of grain type in structural component features is obtained by convolutional neural network, that is, the classification of grains (intraclasts, bioclasts, envelopes, spherulites and agglomerates) based on the improved ResNet50 model. Quantitative recognition of grain content, size, shape and contact mode in structural component features are obtained by digital image processing, that is, grain content is calculated based on threshold segmentation, grain morphology parameters are calculated based on minimum peripheral circle/minimum peripheral rectangle and by combining with area ratio/aspect ratio, intersection ratio (IoU) and other algorithms. And the qualitative and quantitative identification of calcite and other minerals in mineral component features are obtained by HSV colour space processing of the stained images. A thin section example from Shun X well is given as an example, and the validity of each feature extraction algorithm is verified through the complete image recognition process and comparison with the manual identification report. The results show that the petrographic analysis framework is effective in representing meaningful information in sparry grain limestone. Through the model of combing petrographic analysis framework with image analysis algorithm, a standard and intelligent identification of this type of carbonate rocks has been achieved, which can provide effective method support for the study of intelligent identification of rock slice images.

中国石化优秀青年科技创新项目“岩石(矿物)自动化鉴定分析仪” P19028

https://doi.org/10.11781/sysydz2023051026