基于UNet++的煤岩显微图像组分分析  

Component Analysis of Coal Rock Microscopic Image Based on UNet++

作　　者：吕秉略 李忠峰 奚峥皓[1] 姚英茂 季菁菁 LYU Binglve;LI Zhongfeng;XI Zhenghao;YAO Yingmao;JI Jingjing(School of Electric and Electrical Engineering,Shanghai University of Engineering Science,Shanghai 201620;College of Electrical Engineering,Yingkou Institute of Technology,Yingkou 115014)

机构地区：[1]上海工程技术大学电子电气工程学院,上海201620 [2]营口理工学院电气工程学院,营口115014

出　　处：《计算机与数字工程》2022年第2期389-393,404,共6页Computer & Digital Engineering

基　　金：国家自然科学基金项目“面向广义宽基线立体像对的目标三维重建技术研究”(编号:61801286);研究生科研创新项目(编号:20KY0201);辽宁省教育厅2021年度科学研究经费项目(面上项目)“基于神经网络和多目标进化算法的动力配煤模型应用研究”(编号:LJKZ1199)资助。

摘　　要：针对解决煤岩显微图像组分分析过程中,利用图像分割方法遇到的精度较低问题。论文提出了一种基于UNet++模型的图像分割方法。该方法首先将已标记的煤岩显微图像与基于Lovász-Softmax的分割损失相结合,实现对UNet++模型进行训练。再利用训练后的模型对煤岩显微图像按照组分类别进行分割标记。最后,对标记区域进行占比计算,完成煤岩显微图像组分的分析过程。实验结果表明,与K-means算法以及使用交叉熵训练的UNet++模型相比,论文所提算法更关注于各组分的纹理信息差异,且受图像中组分占比不均问题影响较小,对煤岩显微图像组分分割更准确。A method of coal macerals analysis based on UNet++ is proposed in this paper to solve the low accuracy of current coal macerals analysis methods based on image segmentation. The proposed method combines labeled coal images and a loss function based on Lovász-Softmax to train the UNet++ model. Then,the trained model is used to segment and label the coal image by variations of coal maceral. With the labeled coal image,the proportion of different variations of coal maceral is calculated and the analysis process is completed. Compared to the K-means method and the UNet++ with cross entropy,the proposed method focuses on the textural characteristics of different components and less affected by the unbalanced components proportion of image. Thus,the proposed method has more accurate components proportion results of coal maceral images.

关 键 词：图像分割 煤岩分析 Lovász-Softmax损失 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]