An improved automated testing model for maceral groups in coals based on DeeplabV3+
-
摘要: 煤岩显微组分组的识别对分析煤炭化学性质起到关键作用。人工识别方法费时费力，且专业性要求较高。现有的计算机辅助识别有效方法多以深度学习语义分割模型为手段，但因煤岩显微图像组成复杂，且存在过渡组分，因此无法准确识别煤岩显微组分组。针对此问题，提出改进的DeeplabV3+语义分割模型，在改进模型中引入Swin Transformer骨干网络和SkNet网络。首先，针对煤岩显微图像各个组分组交错杂糅且存在过渡组分，特征提取困难，利用Swin Transformer骨干网络作为基础特征提取网络，提升模型对煤岩显微图像每种组分组的特征提取能力，并使得分割网络获得特征间信息交互的能力；其次，针对在模型中空洞空间卷积池化金字塔模块对特征利用率低的问题，将SkNet网络融入空洞空间卷积池化金字塔模块，强化空洞空间卷积池化金字塔模块对重要特征的提取能力，并抑制非必要特征对最终预测结果的干扰；最后，将改进的DeeplabV3+模型与现有先进算法通过实验进行性能比较，结果表明：改进的DeeplabV3+语义分割模型在煤岩显微图像测试集上的像素准确率为92.06%，与随机森林方法、U-Net语义分割模型和DeeplabV3+语义分割模型相比像素准确率分别提高了9.48%、6.90%和3.40%；改进的DeeplabV3+语义分割模型与人工点测方法测试结果相近。改进的DeeplabV3+语义分割模型较现有煤岩显微组分组自动识别模型性能更优，可作为一种强大的计算机辅助人工识别煤岩显微组分组的手段。Abstract: The identification of maceral groups in coals plays a critical role in analyzing the chemical properties of coals. However, manual identification is laborious and requires high expertise. Existing computer-assisted identification methods, mostly adopting deep learning-based semantic segmentation models, fail to accurately identify maceral groups in coals due to complex compositions of microscopic coal images and the presence of transitional components. Therefore, this study proposed an improved DeeplabV3+ semantic segmentation model integrating the Swin Transformer backbone network and the SkNet. First, to deal with the challenge of feature extraction caused by the intertwined maceral groups and the presence of transitional components in microscopic coal images, the Swin Transformer backbone network was used as the basic feature extraction network to enhance the feature extraction ability of the model for various maceral groups and to enable the information interaction between features of the segmentation network. Second, to improve the feature utilization rate of the Atrous Spatial Pyramid Pooling (ASPP) module in the model, the SkNet network was integrated into the ASPP to enable the ASPP to extract important features and suppress unnecessary features that interfere with the final prediction results. Finally, the improved DeeplabV3+ model was compared with existing advanced algorithms through experiments. As indicated by the comparison results, the improved model yielded pixel accuracy of 92.06% on the test set of microscopic coal images, which was 9.48%, 6.90%, and 3.40% higher than that of the random forest method, the U-Net semantic segmentation model, and the DeeplabV3+ semantic segmentation model, respectively. Furthermore, the improved model showed results similar to the manual point measurement method. Therefore, the improved model, outperforming the existing automatic identification models for coal maceral groups, can serve as a powerful method for the computer-assisted manual identification of maceral groups in coals.
-
Key words:
- microscopic coal image /
- maceral group /
- automated testing /
- semantic segmentation model /
- Swin Transformer /
- SkNet
-
图 2 基于Swin Transformer网络改进的DeeplabV3+编码器
Fig. 2 Improved DeeplabV3+ encoder based on the Swin Transformer network
图 5 基于移动窗口的自注意力计算流程
Fig. 5 Self-attention calculation process based on shifted windows
图 8 煤岩显微图像原图(左)和标注结果示例(右)
Fig. 8 Original microscopic coal images (left) and the example of manual annotation results (right)
图 9 2种模型下损失曲线和模型像素准确率(PA)曲线对比
Fig. 9 Comparison of the loss curves and PA curves for improved DeeplabV3+ and DeeplabV3+
图 10 不同煤岩显微组分识别模型预测结果
Fig. 10 Prediction results of different identification models for coal maceral groups
图 11 使用Swin-Tiny和Swin-Small的分割模型预测结果
Fig. 11 Prediction results of Swin-Tiny and Swin-Small segmentation models
图 12 使用Swin-Small分割模型的损失曲线和模型像素准确率曲线
Fig. 12 Loss curves and PA curves of the Swin-Small segmentation model
表 1 不同煤岩显微组分组识别模型预测结果
Table 1 Pediction results of different identification models for coal maceral groups
单位：% 方法 PA MPA MIoU 随机森林[9] 82.58 63.36 51.12 FCN-16s 80.26 53.29 45.10 U-Net 85.16 67.19 53.25 DeeplabV3 85.45 68.45 52.24 DeeplabV3+ 88.66 70.66 59.00 改进的DeeplabV3+ 92.06 74.66 63.51 
下载: 导出CSV
表 2 改进的DeeplabV3+模型消融实验
Table 2 Ablation experiment results of the improved DeeplabV3+ model
单位：% 方法 PA MPA MIoU DeeplabV3+ 88.66 70.66 59.00 Swin-Deeplab 89.45 71.45 61.24 改进的DeeplabV3+ 92.06 74.66 63.51
下载: 导出CSV
表 3 改进的DeeplabV3+模型识别煤岩显微图像混淆矩阵
Table 3 Confusion matrix of the improved DeeplabV3+ model for identification of microscopic coal images
显微组分组 镜质组 惰质组 壳质组 矿物质 环氧树脂 镜质组 0.95 0.02 0 0 0.03 惰质组 0 0.91 0 0.08 0.01 壳质组 0.02 0 0.79 0 0.19 矿物质 0 0.02 0 0.98 0 环氧树脂 0 0 0.03 0 0.97
下载: 导出CSV
表 4 Swin-Deeplab模型识别煤岩显微图像混淆矩阵
Table 4 Confusion matrix of the Swin-Deeplab model for identification of microscopic coal images
显微组分组 镜质组 惰质组 壳质组 矿物质 环氧树脂 镜质组 0.93 0.03 0 0 0.04 惰质组 0.02 0.90 0 0.08 0 壳质组 0.02 0 0.74 0 0.24 矿物质 0 0.05 0 0.95 0 环氧树脂 0 0 0.05 0 0.95
下载: 导出CSV
表 5 DeeplabV3+模型识别煤岩显微图像混淆矩阵
Table 5 Confusion matrix of the DeeplabV3+ model for identification of microscopic coal images
显微组分组 镜质组 惰质组 壳质组 矿物质 环氧树脂 镜质组 0.90 0.03 0 0.03 0.04 惰质组 0 0.89 0.02 0.08 0.01 壳质组 0.02 0 0.71 0 0.27 矿物质 0.02 0.02 0 0.96 0 环氧树脂 0 0 0.06 0 0.94
下载: 导出CSV
-
[1] GAO Xuerui,ZHAO Yong,LU Shibao,et al. Impact of coal power production on sustainable water resources management in the coal–fired power energy bases of Northern China[J]. Applied Energy,2019,250:821−833.. doi: 10.1016/j.apenergy.2019.05.046 [2] WANG Hongdong,LEI Meng,CHEN Yilin,et al. Intelligent identification of maceral components of coal based on image segmentation and classification[J]. Applied Sciences,2019,9(16):3245.. doi: 10.3390/app9163245 [3] 范章群,宋孝忠. 新疆中生代煤中半镜质组特征及其研究意义[J]. 煤田地质与勘探,2014,42(5):9−12.FAN Zhangqun,SONG Xiaozhong. Characteristics of semi–vitrinite of the Mesozoic coal in Xinjiang area and their research significance[J]. Coal Geology & Exploration,2014,42(5):9−12. [4] 范章群. 我国煤岩学研究现状及前景分析[J]. 中国煤炭地质,2017,29(2):15−19.. doi: 10.3969/j.issn.1674-1803.2017.02.04FAN Zhangqun. Coal petrology study status quo and prospect in China[J]. Coal Geology of China,2017,29(2):15−19.. doi: 10.3969/j.issn.1674-1803.2017.02.04 [5] MLYNARCZUK M ,SKIBA M. The application of artificial intelligence for the identification of the maceral groups and mineral components of coal[J]. Computers & Geosciences,2017,103:133−141. [6] SINGH P K,SINGH M P,SINGH A K,et al. The prediction of the liquefaction behavior of the East Kalimantan coals of Indonesia:An appraisal through petrography of selected coal samples[J]. Energy Sources,2013,35(18):1728−1740.. doi: 10.1080/15567036.2010.529731 [7] 宋孝忠,张群. 煤岩显微组分组图像自动识别系统与关键技术[J]. 煤炭学报,2019,44(10):3085−3097.. doi: 10.13225/j.cnki.jccs.2019.1103SONG Xiaozhong,ZHANG Qun. Automatic image recognition system and key technologies of maceral group[J]. Journal of China Coal Society,2019,44(10):3085−3097.. doi: 10.13225/j.cnki.jccs.2019.1103 [8] 宋孝忠. 煤岩显微图像假边界对显微组分组自动识别的影响[J]. 煤田地质与勘探,2019,47(6):45−50.SONG Xiaozhong. Effect of false boundary of microscopic image on automatic identification of maceral group[J]. Coal Geology & Exploration,2019,47(6):45−50. [9] WANG Hongdong,LEI Meng,LI Ming,et al. Intelligent estimation of vitrinite reflectance of coal from photomicrographs based on machine learning[J]. Energies,2019,12(20):3855.. doi: 10.3390/en12203855 [10] WANG Yue,BAI Xiangfei,WU Linlin,et al. Identification of maceral groups in Chinese bituminous coals based on semantic segmentation models[J]. Fuel,2022,308:121844.. doi: 10.1016/j.fuel.2021.121844 [11] CHEN L C, ZHU Yukun, PAPANDREOU G, et al. Encoder–decoder with atrous separable convolution for semantic image segmentation[C]//Proceedings of the European conference on computer vision (ECCV), 2018: 801–818. [12] HASAN P. Multi–task semantic segmentation of CT images for COVID–19 infections using DeepLabV3+ based on dilated residual network[J]. Physical and Engineering Sciences in Medicine,2022,45(2):443−455.. doi: 10.1007/s13246-022-01110-w [13] LIU Ze, LIN Yutong, CAO Yue, et al. Swin transformer: Hierarchical vision transformer using shifted windows[C]//Proc. of the 2021 International Conference on Computer Vision, 2021: 10012–10022. [14] FU Jun,LIU Jing,JIANG Jie,et al. Scene segmentation with dual relation–aware attention network[J]. IEEE Transactions on Neural Networks and Learning Systems,2021,32(6):2547−2560.. doi: 10.1109/TNNLS.2020.3006524 [15] MOU Lichao,HUA Yuansheng,ZHU Xiaoxiang. Relation matters:Relational context–aware fully convolutional network for semantic segmentation of high–resolution aerial images[J]. IEEE Transactions on Geoscience and Remote Sensing,2020,58(11):7557−7569.. doi: 10.1109/TGRS.2020.2979552 [16] ZHENG Sixiao, LU Jiachen, ZHAO Hengshuang, et al. Rethinking semantic segmentation from a sequence–to–sequence perspective with transformers[C]//Proceedings of the 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Nashville: IEEE, 2021: 6877–6886. [17] DOSOVITSKIY A, BEYER L, KOLESNIKOV A, et al. An image is worth 16x16 words: Transformers for image recognition at scale[C]//International Conference on Learning Representations, 2021. [18] LI Xiang, WANG Wenhai, HU Xiaolin, et al. Selective Kernel networks[C]//2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 2020. [19] 中华人民共和国国家质量监督检验检疫总局, 中国国家标准化管理委员会. 煤的显微组分组和矿物测定方法: GB/T 8899—2013[S]. 北京: 中国标准出版社, 2013. [20] 中华人民共和国国家质量监督检验检疫总局, 中国国家标准化管理委员会. 烟煤显微组分分类: GB/T 15588—2013[S]. 北京: 中国标准出版社, 2014. [21] EVAN S,JONATHAN L,TREVOR D. Fully convolutional networks for semantic segmentation[J]. IEEE Transactions on Pattern Analysis & Machine Intelligence,2017,39(4):640−651. [22] RONNEBERGER O, FISCHER P, BROX T. U–Net: Convolutional networks for biomedical image segmentation[C]//International Conference on Medical Image Computing and Computer–assisted Intervention. Cham: Springer, 2015: 234–241. [23] CHEN L C, PAPANDREOU G, SCHROFF F, et al. Rethinking atrous convolution for semantic image segmentation[J]. Computer Vision and Pattern Recognition, 2017. -
下载:
点击查看大图
计量
- 文章访问数: 86
- HTML全文浏览量: 8
- PDF下载量: 14
- 被引次数: 0
