基于Debseg-Net的岩屑图像语义分割

岩屑的岩性识别是地质工作中的一项重要内容。为解决传统人工鉴别岩性的低效问题和通用机器学习模型在岩屑岩性识别上的不适用性，包括准确率欠佳、网络参数冗杂、网络效率低下，针对岩屑图像的特征设计了一种岩屑图像的语义分割网络Debseg-Net,该网络采用编解码结构，卷积与转置卷积结合实现对岩屑图像特征的提取与像素级分类，采用深度可分离卷积减少参数量从而可进一步加深网络，使用跳级连接避免迭代过程中的信息丢失。同时提出了一种高效的岩屑图像自标记方法。经多次实验，Debseg-Net在10口探井收集的640张共计5类岩屑图像数据集上，识别准确率达到98.43%,平均交并比达到90.01%,领先同类型分割网络2.59%～7.04%,在实现数字化岩屑录井进程中提供了方法。

Semantic segmentation of rock debris image based on Debseg-Net

The lithology identification of rock debris is an important part of geological work. In order to solve the inefficient problem of traditional artificial identification of lithology and the inapplicability of general machine learning model in rock debris lithology identification, including poor accuracy, complex network parameters and low network efficiency, a semantic segmentation network Debseg-Net for rock debris images was designed according to the characteristics of rock debris images. The encoding and decoding structure were adopted. The convolution and transpose convolution were combined to realize the feature extraction and pixel-level classification of rock debris images. The depth separable convolution was used to reduce the number of parameters so as to further deepen the network. The skip connection was used to avoid the loss of information in the iterative process. At the same time, an efficient self-labeling method for rock debris images was proposed. After many experiments, Debseg-Net has a recognition accuracy of 98.43% and a mean intersection over union of 90.01% on 640 rock debris images collected from 10 exploration wells, which is 2.59%~7.04% ahead of the same type of segmentation network. It provides a method for realizing digital rock debris logging.