Abstract: Fluorescent thin section is an important tool to study the properties, distribution characteristics and pore structure of crude oil in reservoirs. However, the data of fluorescent thin section is mainly processed by hand, so that the analysis efficiency is low and easily affected by human factors. This paper proposes an unsupervised automatic segmentation method based on convolution neural network(CNN). Firstly, fluorescent colors generated by different components under the excitation of ultraviolet light source were listed and used to establish fluorescent color chart and standard color system map, thus determining the division standard. Later, after extracting the advanced semantic features of fluorescent images by CNN, feature fusion was achieved through similarity and continuity constraints, and the space distance and angle of fluorescence spectrum was calculated to determine the similarity classification. Finally, the automatic division and quantitative analysis of particles, pores, oily asphalt, colloidal asphalt, and asphaltene in fluorescent images was completed. The experiment of fluorescence thin section images demonstrates that this approach does not rely on a substantial quantity of labeled samples and generally exhibits a low average error, thereby satisfying the practical production demands.

Key words: fluorescent thin section, microscopic remaining oil, convolutional neural network, unsupervised learning, feature clustering

摘要： 荧光薄片是研究储层原油性质、分布特征以及孔隙结构的重要手段。但目前荧光薄片数据处理仍以人工方式为主,分析效率低且受人为因素影响。基于卷积神经网络,提出一种无监督自动分割方法。首先统计出不同组分在紫外光源激发下产生的荧光颜色,建立出荧光颜色图版与标准色系图谱以此确定划分标准,然后使用卷积神经网络提取出荧光图像的高级语义特征,通过相似性和连续性约束进行特征融合,通过计算与荧光色系图谱的空间距离与角度确定相似度划分类别,最终实现荧光图像中颗粒、孔隙、油质沥青、胶质沥青、沥青质沥青等组分的自动划分与定量分析。荧光薄片图像的实验证明,该方法不需要大量标记样本且总体各项平均误差较低,能够满足实际生产需求。

关键词: 荧光薄片, 微观剩余油, 卷积神经网络, 无监督学习, 特征聚类