基于UNet++的煤岩显微图像组分分析

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。

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

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.